Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Effects of a Personalized Fitness Recommender System Using Gamification and Continuous Player Modeling: System Design and Long-Term Validation Study

Effects of a Personalized Fitness Recommender System Using Gamification and Continuous Player... Background: Gamification and persuasive games are effective tools to motivate behavior change, particularly to promote daily physical activities. On the one hand, studies have suggested that a one-size-fits-all approach does not work well for persuasive game design. On the other hand, player modeling and recommender systems are increasingly used for personalizing content. However, there are few existing studies on how to build comprehensive player models for personalizing gamified systems, recommending daily physical activities, or the long-term effectiveness of such gamified exercise-promoting systems. Objective: This paper aims to introduce a gamified, 24/7 fitness assistant system that provides personalized recommendations and generates gamified content targeted at individual users to bridge the aforementioned gaps. This research aims to investigate how to design gamified physical activity interventions to achieve long-term engagement. Methods: We proposed a comprehensive model for gamified fitness recommender systems that uses detailed and dynamic player modeling and wearable-based tracking to provide personalized game features and activity recommendations. Data were collected from 40 participants (23 men and 17 women) who participated in a long-term investigation on the effectiveness of our recommender system that gradually establishes and updates an individual player model (for each unique user) over a period of 60 days. Results: Our results showed the feasibility and effectiveness of the proposed system, particularly for generating personalized exercise recommendations using player modeling. There was a statistically significant difference among the 3 groups (full, personalized, and gamified) for overall motivation (F =22.49; P<.001), satisfaction (F =22.12; P<.001), and preference 3,36 3,36 (F =15.0; P<.001), suggesting that both gamification and personalization have positive effects on the levels of motivation, 3,36 satisfaction, and preference. Furthermore, qualitative results revealed that a customized storyline was the most requested feature, followed by a multiplayer mode, more quality recommendations, a feature for setting and tracking fitness goals, and more location-based features. Conclusions: On the basis of these results and drawing from the gamer modeling literature, we conclude that personalizing recommendations using player modeling and gamification can improve participants’ engagement and motivation toward fitness activities over time. (JMIR Serious Games 2020;8(4):e19968) doi: 10.2196/19968 KEYWORDS persuasive communication; video games; mobile apps; wearable electronic devices; motivation; mobile phone http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 1 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al activities have the potential to help users achieve their fitness Introduction goals and increase engagement and pleasure by adding game features to physical activity [12,19]. A sedentary lifestyle is defined as a lifestyle in which an individual does not receive regular amounts of physical activity, However, most existing work on gamification and persuasive which is becoming a significant public health issue [1]. Various games in health and wellness are limited because of their use solutions have been considered to encourage a more active of one-size-fits-all approaches, which have been shown to be lifestyle. Among them, combining exercise with gameplay [2] suboptimal [21]. As discussed in the next section, there are some and the use of wearable trackers to motivate and recommend research efforts that have used more in-depth personalization, physical activities [3] have received widespread popularity, but but they are not focused on exergames, and initial attempts at both have user retention issues [4]. This paper addresses the personalization are mostly limited to a small set of static issue of improving long-term engagement with such physical demographic parameters about the user (such as age, gender, activity recommenders and exercise games. and occupation), which makes them more categorized rather than personalized. In addition, their effectiveness in promoting The popularity of computer games and their engaging nature the desired behavior was mostly evaluated based on a single has created a strong trend to use games for nonentertainment point of use and feedback (short term). Few attempts have been purposes [5]. This trend includes overlapping topics and terms made to resolve these issues. For example, MyBehavior [22] such as gamification (the use of game features and mechanics used a tracking-based and dynamically modified user model to in nongame applications [6]), serious games (aimed primarily recommend activities, but the system is not gamified and is at being an educational yet entertaining tool [7]), persuasive based on a limited model of daily activity information. A proper games (games for promoting behavior change [8]), and combination of a detailed model with features such as exergames (a combination of physical exercise with games [9]). personality types, modeling-based personalization, and In particular, gamification has received significant attention recommendation with adaptive gamified elements is still missing because it can be seen as an umbrella topic covering a range of in the area of exergames, as discussed in more detail in the next options from implementing few game elements (such as section. leaderboards) in regular activities to performing serious tasks as a full game [10,11]. To address these research gaps, we propose a comprehensive model for gamified fitness recommender systems that use Games and gamified activities are effective persuasive tools for detailed and dynamic player modeling and wearable-based motivating human behavior [12]. Although recent years have tracking to provide personalized game features and activity seen an increase in persuasive applications designed to promote recommendations. We also present the results of a long-term more active lifestyles [13], studies have suggested that a investigation on the effectiveness of our recommender system one-size-fits-all approach is ineffective for such persuasive that gradually establishes and updates an individual player model applications because different users are motivated by different (for each unique user) over a relatively long period (60 days). persuasive strategies [14] as well as personal reasons such as curiosity and social rewards [15]. There is an increasing demand We hypothesized that (1) player modeling based on continuous for personalization as a means of tailoring an experience to player activity tracking is an effective approach for personalizing individual needs and interests [15,16]. This is particularly the activity recommendations to individual players and (2) case for persuasive and recommender systems such as those in combining player modeling and gamification can promote marketing, education, and health, where retention is as important long-term engagement with the system. as initial action [17]. Among different personalization solutions, On the basis of these hypotheses, we aim to address the player modeling in games that aims to understand players to following specific research questions in this paper: enhance game experience has been an active research topic [18]. It aims to describe a game player’s traits and preferences as How can we generate and use continuous player modeling well as the players’ cognitive, affective, and behavioral patterns to personalize activity recommendations for each user? [19] within well-defined structures that allow designers to tailor Can the combination of player modeling and gamification game contents or goals automatically to suit the needs or techniques improve user engagement and experience toward preferences of individual players. fitness activities over time? Gamification has rapidly emerged over the past years, especially To achieve this, we designed a new player model and a related in the area of exercise and fitness [20], as a tool to promote system architecture for a gamified fitness activity recommender healthy behaviors and maintain an active lifestyle [9]. system. To evaluate the effectiveness of the model-driven Researchers have used various gameplays and game features gamified fitness activity recommender system, we conducted to make exercise and physical activities more engaging and a long-term study on 40 participants and examined the attractive [2,3]. The use of gamification to promote a more effectiveness of our gamification approach in promoting physical active lifestyle can be through either formal exercises performed activity in comparison with a control group. We randomly as games (ie, exergames) or combining games with other assigned our participants into 4 distinct groups corresponding physical activities that are not as rigorous as formal exercise to 4 experimental conditions: (ie, a walk to work). In this paper, we refer to all these cases as The full group received the application with both gamified gamified physical or fitness activity and use the term features and personalized recommendations (based on “exergame” loosely to indicate the same type of activity. These player modeling). http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 2 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al The personalized group received only personalized for building an adaptive gamification system: (1) consider the recommendations but no gamified features. purpose of adaptivity, (2) define the adaptivity criteria, (3) The gamified group received nonpersonalized design the adaptive gamification mechanics and dynamics, and recommendations with gamified features added. (4) design meaningful adaptive interventions. The researchers The control group received generic nonpersonalized applied this framework to the design of a web-based platform recommendations and nongamified features. for knowledge exchange in postgraduate medical training and reported positive user acceptance, feedback, and increased usage. The results of a 60-day-long study showed that the idea of generating personalized exercise recommendations using player Nicholson [24,27] discussed the idea of meaningful gamification modeling is feasible. Moreover, it showed that personalizing (or playification), which focuses on playfulness and activities recommendations using player modeling in combination with that make sense to each player and rely on intrinsic motivations gamification can improve participants’ level of engagement and as opposed to following specific rules to win. Bertran et al [28] motivation toward fitness activities over time. built on this idea and proposed the situational play design, which is a framework for designing context-based and personalized Our work includes the following major contributions to the field games. Orji et al [14-16], among others, as discussed in a later of fitness recommenders and exergames: section, also discussed the idea of player modeling for designing We offer a conceptual model and system architecture for more effective serious and health games. Overall, research in personalized gamified activity recommendations using a the gamification domain suggests the need for more personalized combination of player modeling, gamifications, and activity games that depend on the players’ context and their motivations. tracking. The research presented in this paper expands on these ideas and We build a comprehensive player model for personalization addresses some of the identified needs, such as tracking and that is dynamically updated by continuously tracking player understanding the player using a comprehensive individual-level contexts. model, personalizing both game features and recommended We design and validate a 24/7 recommender system for activities in exergames, and performing long-term studies, within personalized activities combined with various gamified the context of gamified physical and fitness activity elements that adjust to a player’s and environmental recommenders. We discuss the research in these specific areas contexts. in later sections. Finally, to demonstrate the feasibility of our approach, we Gamified Physical Activities conducted a long-term (60-day) field study with 40 Despite their positive effect on promoting an active lifestyle, participants to evaluate the proposed system and compare gamified physical activities face the problem of sustainability the effectiveness of the 4 experimental groups. (also referred to as player retention here and in other literature) Through the design and development of a new recommender [20,25,26]. Although players may feel excited and motivated system and a long-term study, we hypothesized and evaluated to play at first, over time and sometimes quickly, they may lose the effects of gamification and player modeling on physical their willingness to continue. There are studies focusing on the activity. To the best of our knowledge, this research is the first motivation and sustainability of exergames and gamified fitness to link research on player modeling, gamification, and activity activities. For example, Campbell et al [21] discussed the recommendation to propose an approach for personalized concept of everyday fitness games and suggested that for activity recommendation that is continuously and dynamically applications that people frequently use in their everyday lives, updated to reflect users’ changing contexts and states. the design needs to be fun and sustainable as well as adapt to behavioral changes. Macvean et al [29] reported a 7-week study Related Work on users’ physical activity, motivation, and behavioral patterns Gamification using exergames and suggested that longitudinal studies are necessary for evaluating motivational effects as exergames The motivational effects of gamification have been widely ensure that the intensity of a user’s behavior is appropriate and studied by researchers. It has been shown that common game sustained. Previous work [30] also showed that based on existing elements such as badges, rewards, leaderboards, and avatars are technologies and user needs, the idea of employing wearable commonly and successfully used to motivate players [10,21-23]. activity trackers for gamification of exercise and fitness is On the other hand, researchers have argued for various areas of feasible, motivating, and engaging. Adding dynamic features improvement in current gamification research and applications, could have a positive impact on user motivation toward the including diversity in themes and context, study duration, and gamified exercise system, and the gradual release of application sample size [6,23], increasing motivation by relying on more features could increase the user retention rate. It was also found intrinsic factors [22,24], continuous adaptation [25], and that each user was unique and motivated by different types of personalization [6]. In the study by Loria [25], a framework for game features. Therefore, based on these results, it seems improving the player experience and customizing content reasonable to generate customized workout sessions to fit generation is proposed by continuously monitoring how players different user fitness conditions and interests. interact during the game by analyzing information such as players’ in-game behavior and players’ social network. Other Recently, significant research has been devoted to the design researchers have also tackled adaptive gamification. For of active games (also commonly referred to as exertion games example, Böckle et al [26] identified 4 main elements as the or exergames) to match the needs of specific groups such as basis for defining meta-requirements and designing principles http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 3 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al those with disability [28,29] or senior citizens [31] or to use group-cycling class. These preferences have less to do with specific technologies such as virtual reality (VR) [32]. people’s physical characteristics and are affected more by Particularly in the context of VR-based exergames for personalities. Matching activities to personality type has been adolescents, research shows that game elements such as the use shown to have real-world relevance [42]. Research suggests of rewards, increasing challenge levels, frequent updates, and that people who engage in personality-appropriate activities social or multiplayer options are important aspects for continued will stick with the activities longer, enjoy their workout more, engagement in physical activity [33]. Researchers have also and have a better overall fitness experience [43]. Brue [44] investigated design principles for active games [34,35]. created a system based on the principles of the Although impressive and invaluable in their findings, these Myers-Briggs–Type Indicator (MBTI) assessment. She used efforts primarily focus on the design of particular games, game MBTIs and reworked them into an easily maneuverable features, and actual gameplay mechanisms that are best suited color-coded fitness personality model, the 8 Colors of Fitness, for increasing physical activity for the target group or individual. which is also used in our player model. Each color is associated On the other hand, in 24/7 activity recommenders, the focus is with 2 personality types from the 16 possible MBTI types [45]. less on designing a particular game and more on gamifying the For example, blues are loyal, traditional, dependable, and daily experience and recommending activities based on the daily straightforward, whereas greens are nature lovers who seek to routine using dynamic player modeling. quietly merge with the outdoors [42]. Moreover, in recent years, the use of wearable sensors in human Player Models activity recognition has become popular [46], in which most of Busch et al [30] indicated that the one-size-fits-all approach the measured attributes are related to the user’s movement (eg, does not work for persuasive game design. Thus, player-type using accelerometers or GPS), environmental variables (eg, models could be used when tailoring personalized persuasive temperature and humidity), or physiological signals (eg, heart systems. One of the most frequently used player-type models rate or electrocardiogram). These data types are naturally is the one developed by Richard [36], who identified 4 player indexed over the time dimension, consistent, and convenient to types and proposed that each player has some particular access, which could be used in modeling and predicting a user’s preference for one of the types, which makes them mutually daily activity pattern. exclusive. Another model is the BrainHex model [37], which is a relatively new model but has been validated using a large Although there is a significant amount of research on the subject pool of participants [38]. In BrainHex, player types were not of player modeling, none of the existing studies have examined mutually exclusive. Scores under each category are presented how to use a comprehensive player model. In addition, no to determine the player’s primary type and subtypes. It also previous research has simultaneously considered both game connects player types to the game elements. Moreover, the features and recommended activities in exergames design and Hexad model [39], which is of particular interest in our work, investigated whether it is an effective approach over the long is a gamification player-type model created for mapping user term. personality onto gamified design elements. We considered using Personalized Activity Recommendations the Hexad model in our player model because it specifically targeted gamified systems. It proposes 6 player types, and the Personalized recommender systems for physical activity have player types of individuals are correlated with their preferences been studied by many researchers. For example, Guo et al [47] for different game design elements. Design guidelines for proposed a system that recognizes different types of exercises tailoring persuasive gamified systems to each gamification and interprets fitness data (eg, motion strength and speed) to an player type have also been studied [17]. easy-to-understand exercise review score, which aims to provide a workout performance evaluation and recommendation. Furthermore, Wiemeyer et al [40] discussed the concept of Although it achieved 90% accuracy for workout analysis, it player experience (with a focus on individual) versus game focuses only on recognizing fitness activities and not usability (with a focus on technology) and reviewed various personalizing or gamifying them. He et al [48] introduced a theoretical models that can help understand the player system designed to be context aware for physical activity experience. These models are particularly helpful when recommendations. It focuses on selecting suitable exercises for designing full games as opposed to gamifying everyday individualized recommendations. A smartphone app was activities, which is the goal of this research. However, their developed that could generate individualized physical activity insights, such as an integrative multidisciplinary model of player recommendations based on the system’s database of physical experience, can be helpful in future phases of our research when activity. The focus of their work is to recommend different types we focus on the design of game elements. For the work of activities but does not take into account personal details such presented here, our focus was primarily on showing how the as proper time, location, and intensity or any gamified elements. combination of gamification and player modeling could improve engagement. Better player experiences can be achieved through Broekhuizen et al [49] proposed a framework called PRO-fit, more complicated models and game features that are beyond which is another example that employs machine learning and the scope of this work. recommendation algorithms to track and identify users’ activities by collecting accelerometer data, synchronizes with the user’s Personality type also plays an important role in determining calendar, and recommends personalized workout sessions based people’s fitness tastes [41]. Some people may prefer swimming on the user’s and similar users’ past activities, their preferences, laps solo, whereas others may enjoy attending a rowdy and their physical state and availability. The authors highlighted http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 4 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al that many applications nowadays are more focused on tracking real-time activity tracking and recommendation are also user activities but do not provide a recommendation system that frequently missing in existing research on personalization and would help users choose from activities based on their interests exergames. and accomplishment of goals. Therefore, the authors were However, the existing body of research provides invaluable motivated to design a personalized fitness assistant framework insight into recommendations based on real-time tracking, that acts as a motivator and organizer for fitness activities, important parameters to include in a player model, and the making it easier for users to create and follow their workout design of exergames in general. Expanding on existing studies plan and schedule the sessions according to their availability and trying to fill the abovementioned gaps, we propose a and preference. Compared with PRO-fit, our proposed system conceptual model and system architecture that bring together provides recommendations in real time throughout daily life as game elements, dynamic player modeling, and activity tracking opposed to the prefixed recommendations that are not based on to personalize exergames in terms of both game features and any player or exerciser-type model employed in PRO-fit. recommended activities. We built a comprehensive and dynamic Mittal and Sinha [50] used personal information to recommend player model for personalization that is continuously updated general activities such as visiting attractions and shopping. by tracking the player and offers 24/7 personalized activity Although not focused on fitness activities or gamification, their recommendations. Finally, we conducted a long-term user study notion of modeling user data as the base for recommendation in the wild to evaluate the proposed system. is in line with our proposal. Ni et al [51] used a variety of user To the best of our knowledge, although some of the features of data such as daily routine and heart rate to recommend workout our study have been suggested and/or investigated by others, routes. Their method is more focused on physical activity no long-term comprehensive study has been conducted to recommendations but is limited to recommending routes and integrate and evaluate them in real-life exergame apps. does not include gamification elements. Similarly, Rabbi et al [22] proposed MyBehavior, which is a system for tracking users Methods using mobile devices and suggesting food and physical activities. MyBehavior provides personalized and real-time suggestions Conceptual Model but is not gamified and does not include an explicit player Although the existing studies have addressed many aspects of model. As such, it does not take advantage of full these diverse fields, as discussed earlier, they have not been personalization or more engaging features that a game can offer. properly integrated to develop engaging and sustainable In line with this, Ghanvatkar et al [52] conducted a exergames. For example, the effect of various game features comprehensive review of user models used in recommender and continuously adapting the game to player needs and interests systems. They highlight that activity profile, demographic have not been investigated in the context of exergames. information, and contextual data such as location are among the top items to include in user models. In this research, we In this section, we describe our proposed system architecture have defined our player model to include gamer information and related research methods. This proposal is based on our using Hexad and demographic, activity, and exercise submodels, new conceptual model developed after reviewing related work, as suggested by Ghanvatkar et al [52]. consisting of the following principles: Summary of Research Gaps 1. Advances in wearable technologies allow game designers to use commercially available activity tracking sensors and As reviewed in previous sections, research on gamified physical mobile devices as a major element of exergames. activities and related topics has achieved significant results but A game with a static design, no matter how interesting, will requires more work to fill the existing gaps. We identified that lose its attraction after a while. As such, it is important to the main research gap within the context of exergames is the add new features over time to keep players engaged. notion of personalized gamification (a combination of gamified Although different methods exist for adding dynamic physical activities with player model–based personalization), features to games, designers have a limited ability to provide including understanding the players and their environment and new features constantly, and there is no guarantee that they adapting the game features and physical activities dynamically. will be attractive to users. An alternative (or None of the existing studies have successfully investigated the complementary) approach is to dynamically modify the effect of gamification and personalization individually with game by adapting to the player. respect to promoting the efficacy of an intervention, specifically a physical activity intervention within a single application. In Our conceptual model, which builds on our previous work [28] addition, long-term studies outside controlled environments and along with 2 new components, is illustrated in Figure 1. http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 5 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al Figure 1. Conceptual model of the proposed system. Components surrounded by parallelograms were explored in our previous work, whereas components surrounded by rectangles are the aim of this research for personalizing physical activity recommendations. Arrows show the connection between the components. In our previous work [53], we investigated the effect of using player model takes different types of user data and predicts user wearable activity tracking in exergames and the long-term preference for physical activities and finds the proper time and effectiveness of using a dynamic game feature–releasing system location for recommending activity sessions. It consists of in sustaining exergames (marked parallelograms). In this paper, several submodels that cover the user’s general, personality, we aim to further investigate the gamified features for increasing and daily activity data. The recommendation engine used the retention, exploring 2 additional components: (1) player output of the player model and generated customized physical modeling in the personalization of exergames and (2) how to activity session recommendations for individual users (including use such a system to generate personalized physical activity the proper time, location, intensity, and potential type of physical recommendations (marked rectangles). The arrows demonstrate activity). The game generator adds customized game elements how each of the components are related and can directly to the recommendation and generates the final game content influence each other. We believe that tracking activity using that users can interact with. Wearable activity trackers or wearable technology, providing dynamic game features, and smartwatches are used in the system to track the user’s activity detecting a player’s preferences using a player modeling and introduce diverse interactions. The combined use of mobile approach can all contribute to creating a more engaging apps and wearable apps will allow users to interact with the exergame experience, which in turn can generate more system with different modes. The detailed design and personalized physical activity recommendations that players development of the system are introduced in the following will likely find motivating and satisfying. section. System Design Overall, a wearable-based exergame system, with a comprehensive player model for physical activity On the basis of our proposed conceptual model, bringing recommendation and game customization, is proposed as a wearable activity trackers or smartwatches into exergames, solution to the exergame retention problem. dynamically updating game features, and using player modeling for personalization of exergames is being proposed as a solution Application Design and Implementation to the research problem. Therefore, a wearable-based exergame On the basis of the proposed system architecture, a Wear OS with a comprehensive player model for personalization, (formerly Android Wear) application is implemented as the user recommending customized activities, is proposed as a potential interface (UI), which tracks the user’s activities and provides system for further investigation. gamified fitness recommendations. The overall architecture of The proposed system contains 3 main components: a player the application is illustrated in Figure 2. model, a recommendation engine, and a game generator. The http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 6 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al Figure 2. App architecture. How data were collected and transferred to each submodel of the system and how they were used to generate recommendations and game content. Info: information. In this application, and based on the conceptual model in Figure in charge of generating one part of the recommendation, as 1, the player model consists of 4 submodels: (1) an activity shown in Table 1. Choosing game features and physical recognition model that tracks player activities, (2) a general activities are the 2 main personalization options and each has model that holds basic information about the player, (3) an its own submodel. Tracking daily activities is an essential part exerciser-type model that includes information required for of the system, which also has a submodel. The fourth submodel recommending activities, and (4) a gamer-type model that is holds general player information, such as gender, age, weight, used to choose game features. Each of the submodels is mainly and height. Table 1. The roles of each submodel. Submodel Role Activity recognition model Time and location General model Intensity and duration Exerciser-type model Exercise type Gamer-type model Game elements Although each submodel is designed to generate one particular intensity. We use the 8 Colors of Fitness model [44] to suggest part of the recommendation, they are still connected to each different types of activities for the personalized groups. This other to create a more reliable overall recommendation. For model is one of the few that uses personality type as the basis example, the exerciser-type model is built for each individual of activity recommendations and is suggested by other user for recommending different types of activities based on researchers and practitioners [49,50]. their personalities but it also relies on the general model, which The recommendation engine is a decision tree–based module is built based on a user’s fitness and health condition, to exclude that uses all the information generated from the player model those activities that may be suitable for their personality type to create personalized recommendations for each individual but not for their health condition. We refer to the theoretical user. It could either extend an existing activity (eg, by foundations from the Global Recommendations on Physical recommending a longer exercise time, a longer running path, Activity for Health (GRPAH) [54] to determine proper exercise or appropriate intensity), recommend some activities on the recommendations in nonpersonalized cases. The GRPAH is an user’s idle time, or simply recommend a different type of accepted tool approved by the World Health Organization for activity. An example of a decision tree is illustrated in Figure recommending the appropriate exercise type, duration, and 3. As the recommendation system for physical activity itself is http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 7 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al a relatively complicated topic, we do not consider it as the would have an impact on user experience. Therefore, after highest priority for this study. Therefore, we only employed verifying the feasibility of the proposed idea and the roles simple decision tree methods to generate basic recommendations personalization and gamification performed in this type of (we have ensured that all the recommendations followed the system, our research goal will be to investigate the recommender GRPAH guideline for daily physical activities). We are aware system of physical activities. that the rationality and quality of the recommended activities Figure 3. Example of the decision tree used in the recommendation engine (simple version). PA: physical activity. The game generator is responsible for adding game elements [17], we used the game elements recommended by the Hexad to the recommendations to gamify the activity suggestions player–type model, which is more in line with our objective in generated by the recommendation engine. The type of game this work, designing gamified physical activity elements to be added is determined by the Hexad player–type recommendations. Details of the game and activity model [39]. Our work is also partly based on Orji et al [17], as recommendations are provided later in this section. it adopted a similar Hexad player model. However, as opposed to using the persuasive strategies recommended by Orji et al http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 8 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al A Wear OS app was developed for this study. The app is a develop with many possible confounding factors. We also did conversation-based game in which all the interactions happen not want to introduce various esthetic and design variables to in the form of a conversation between the user and the future the study that may interfere with our studied research variables self. The game is based on a story in which a 1-day user receives and influence our results. For the same reason, we did not try a message from the future self in 20 years telling him or her to incorporate our study within an existing game, even though that the world is about to end in that future world but only the adding these features to games that the user prefers may be user can save it by completing a series of tasks. Then, the future another motivating factor in the future. However, it is essential self will guide the user through daily activities, which are to establish their effectiveness first in isolation. The current generated by the recommendation system in a gamified structure. system UI was created using a rapid prototyping approach. A The choice of this game was informed by our need to have a pilot study was also conducted before the formal study to ensure simple design that is capable of incorporating our research that the labels and buttons are clear. The main UI and app icons requirements but at the same time is not too complicated to are shown in Figure 4. Figure 4. Example app interface and icon. (a) A snippet of one conversation between the system and the user. (b) A display of new mission for the user that he or she can choose to accept or decline. UI: user interface. The app tracks the user’s daily activity through Android processing to understand human language and we used its ActivityRecognition [55] and Google Fit Application message API to create a chatbot, which aims to understand a Programming Interface (API) [56], which allows up to 6 user user’s intents and lead participants to designed storylines. activities to be recognized in real time: in vehicle, on foot, Moreover, we have included a weather assistant in the system running, walking, on bicycle, and still. (through the Weather API [58]) to help participants in planning activities around the weather. The Google Fit API provides encapsulated daily activity–related data such as calories burned, daily steps, and heart rate history When designing the game features, we employed the Hexad (if applicable) tracked by both phone and watch sensors. All player types [39] and the game design elements guide [17]. the collected activity data, along with their time stamps and Hexad suggests that game design elements are preferred by each location information, are used as input features to train a daily player type and we implemented 1 element for each type of user activity model for each individual user by which possible in this study for a personalized game experience (in addition to exercise time and location are predicted. As shown in Figure 4, the game storyline). We integrated the following gamification the app is a conversation-based game. We used Wit.ai [57] to elements in our game (Table 2). Figure 5 shows some generate storylines and to build a bot that can talk to participants screenshots of example game elements for different player types. and perform some general greetings, tell the time, and talk about If there was a tie in scores between the 6 types, we randomly the weather. Wit.ai is a tool that uses natural language chose 1 element of the highest score to add. http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 9 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al Figure 5. Example game elements: (a) profile in daily view (including points and challenges), (b) profile in weekly view (including points and challenges), (c) connect to Facebook view, (d) hacker mode view, and (e) theme color customization view. http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 10 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al Table 2. The motivation and corresponding game elements added for each type of player. Hexad type Motivation Game element Socializers Relatedness Link to social network Free spirits Autonomy and self-expression Theme color customization Achievers Mastery Challenge Philanthropists Purpose and meaning Game experience sharing Players Rewards Points Disruptors Change Hacking mode As mentioned, the application is a Wear OS game that requires Link to Social Networks combining the use of both an Android phone and an Android Socializers are motivated by relatedness. They want to interact watch to optimize its recognition accuracy and gamified with others and create social connections [39]. Therefore, we experience. For activity recognition, our app uses watch sensors provided them with an interface for linking the game to their for better accuracy. However, in many situations, participants social network as their unique feature so that they could share may choose not to wear the watch. In those cases, when the their game performance or achievements to their Facebook page, watch was not connected, we used the built-in phone sensors team up with those friends who are already in the game, or invite instead such that the game could run individually on the phone new players to the game. without the watch. A phone clearly offers more screen space and abilities, such as typing messages, compared with a watch. Theme Color Customization User Study Design Free spirits are motivated by autonomy and self-expression. They want to create and explore the game and prefer features Multiphase Research such as unlockable content and customization [17,39]. Thus, The proposed conceptual model and the system were evaluated we added a feature of theme color customization so that they based on a multiphase user study. In our previous work [53], could customize their game UI by unlocking different themes. we introduced phases 1 and 2, which can be summarized as Challenge follows: Achievers are motivated by mastery . They are looking to learn For research phase 1, in-lab user tests of 20 participants were new things and want to overcome challenges [39]. Therefore, conducted to evaluate the effectiveness of the combined use of we added a challenge system for them in our game, in which games and wearable devices in promoting exercise and to tasks were assigned to them as challenges. investigate the usability of the proposed approach and the effects of different factors within the system. Game Experience Sharing Philanthropists are motivated by purpose and meaning. They In research phase 2, a 70-day user study of 36 participants was want to give to other people and enrich the lives of others in designed to verify the hypothesis that adding different game some way with no expectation of reward [39]. For features and gradually releasing them can positively affect user philanthropists, we added a feature for them to share their game engagement and retention. experience with other players. A forum-like interface was added In this paper, we present research phase 3, which is a 60-day to their version of the game in the main screen that allowed long-term study with 40 participants, to demonstrate the them to browse and answer questions of other players. They feasibility and effectiveness of using a player modeling also receive notifications when there is a new question in the technique in the personalization of exergames. forum. Participants and Groups Points A total of 40 participants were recruited locally from the Ottawa Points have been shown to positively affect players [17,39]. area by posters as well as via the web through the Android Wear They will do what is needed for them to collect rewards from Forum [59]. Of the 40 participants, 23 were men and 17 were a system. For players, points in our game can be collected and women. Their average age was 26.93 years, with an SD of 6.07 used as virtual currency to buy extra themes or virtual years. We randomly divided our participants into 4 groups based equipment. on the versions of the app they received: full (gamified and Hacking Mode personalized), gamified only, personalized only, and the control (neither personalization nor gamified, as the control group). Disruptors are motivated by change. In general, they want to Participants were randomly allocated to groups and the disrupt the system [39]. We added a hacking mode for distribution with respect to exercise or player type and physical disruptors, in which they can use the command-line interface activity level did not seem to be particularly biased (Table 3). to access their own game database to make changes to the Participants’ physical activity levels were collected at baseline storyline or delete their records of the game and, eventually, before beginning the study. they may destroy the system. http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 11 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al Table 3. Demographical data for the 4 participant groups (N=40). Characteristic Participant group Full Gamified Personalized Control Age (years), mean (SD) 24.93 (7.27) 26.65 (5.58) 27.85 (6.26) 25.78 (5.93) Gender, n (%) Male 6 (60) 6 (60) 5 (50) 6 (60) Female 4 (40) 4 (40) 5 (50) 4 (40) Hexad user types, n (%) Philanthropist 1 (13) 1 (11) 1 (10) 1 (11) Socializer 1 (13) 2 (22) 1 (10) 2 (22) Free spirit 1 (13) 3 (33) 2 (20) 2 (22) Achiever 2 (25) 2 (22) 3 (30) 2 (22) Disruptor 1 (13) 0 (0) 0 (0) 1 (11) Player 2 (25) 1 (11) 3 (30) 1 (11) 8-color personalities, n (%) Blue 3 (30) 2 (20) 1 (10) 2 (20) Gold 1 (10) 1 (10) 1 (10) 2 (20) White 2 (20) 1 (10) 3 (30) 0 (0) Purple 0 (0) 1 (10) 1 (10) 0 (0) Green 1 (10) 3 (30) 1 (10) 2 (20) Red 1 (10) 0 (0) 1 (10) 2 (20) Saffron 0 (0) 0 (0) 2 (20) 1 (10) Silver 2 (20) 2 (20) 0 (0) 1 (10) 4.04 (2.35) 3.95 (3.21) 4.82 (2.53) 3.83 (2.92) Physical activity level (hours per week), mean (SD) Physical activity levels were self-reported at baseline. To increase the duration in each group and reduce the chance of recommendation, was our hypothesis when other variables of groups affecting each other, all participants remained in the are held constant. Therefore, we did not use existing commercial same group for the entire study duration rather than randomly apps for comparison in this study because we tried to avoid trying all 4 groups. bringing in possible extraneous or confounding variables such as esthetic and gameplay features that were not our focus. The recommendations for the control group and the gamified group were created based on established exercise guidelines Figure 6 shows an example of how recommending the same and were reasonable recommendations for the general 30-min walking activity will look for the 4 study groups. The population. To ensure this, we referred to the theoretical full group received the recommendation through a gamified foundations from the GRPAH [54] to determine proper exercise story (guided by the future self) with the game element of recommendations. Our choice for nonpersonalized groups challenge based on their player type of achiever and a closely follows the one-size-fits-all recommendation method, personalized walking path. The personalized group also received which has been generally used in most physical activity a personalized route but no game story or elements. The recommendation applications, such as Apple Watch gamified group received no personalized route but had the game (recommends a 30-min walk per day) or Fitbit (daily 10,000 story and the game element of points (randomly assigned steps). because no player model was used for the gamified group). The control group received no personalization or gamification as a Furthermore, the main purpose of our study was to demonstrate control group. In the screenshot of the control group, we showed the effectiveness of personalized recommendations. Although an example of how the weather assistant worked. Note that the we tried to offer a reasonable experience for nonpersonalized example conversations were from screenshots and some details groups, the effectiveness of personalization, especially in terms related to the context were not fully displayed. http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 12 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al Figure 6. Example of recommendations for different groups: (a) full, (b) personalized, (c) gamified, and (d) control. To control the gamification level between groups, for the We also limited the number of personalized recommendations gamified group, as there was no player model used, we randomly to 2 times a day to eliminate the variability of engagement assigned a game element from Table 2 to each participant to caused by frequent recommendations. The gamified and control bring them to the same gamification level as the full group. As groups (without personalization) received 2 messages per day the members of personalized groups received different game at 9 AM and 5 PM. We chose these 2 times because 9 AM is elements based on their individual player model, we decided the time of day that most of our participants were active. We that a random selection for nonpersonalized groups would be did not send the notification earlier because we did not want the closest nonpersonalized option. their sleep to be interrupted. We chose 5 PM because most people are off from work at 5 PM. The full and the personalized http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 13 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al group received messages based on when they got up and when a game that can lead to the feelings of relatedness. It is usually they left work, as recorded in their individual player model. The used in multiplayer games, which allow for interactions between results presented in this paper are based on data from a 60-day real players, and was not applicable in our case. Perceived experiment. choice is often used in situations where a person is given a certain task or activity to complete. In our case, we indicated Procedures and Data Collection in the beginning that the users have the full choice of either The study was approved by the research ethics board. We asked using or not using our system as well as how to use it. Therefore, participants to complete a prestudy questionnaire before this subscale was considered not necessary as the participants providing them the app (Multimedia Appendix 1). The were explicitly given full choice. There are different versions questionnaire asked demographic questions including age, of the IMI that have been used in previous studies, which consist gender, height, weight, number of hours they spend per week of different subscales that are only relevant to their unique exercising, type of Android Wear owned, and types and duration context. of playing video games (eg, PC, console, and mobile). Two Data Analysis web-based questionnaires were provided and participants completed them, which provided us with the results to determine For each question in the poststudy questionnaire including their player and exerciser type [60,61]. The app was distributed general perception, IMI subscales, and EMIC subscales, a to participants through the HockeyApp (now Microsoft Visual one-way between-group analysis of variance (ANOVA) and Studio App Center) [62] after receiving participants’ gamer and post hoc Tukey-Kramer Honestly Significant Difference (HSD) exerciser-type results. Application features were selected based test [66] was conducted to analyze the main effects among the on the participant’s player model. 4 groups. ANOVA is commonly used to determine whether there are any statistically significant differences between the For in-game data collection, we used Google Analytics API means of 3 or more independent groups, whereas the Tukey test [63] to track all participants’ comprehensive in-app behavior provides deeper insights into patterns and comparing specific data, including screen views and tapped events with associated groups [66]. Parametric tests were selected for conducting the timestamps. We used Google Fit API to track user daily activity analysis because the samples were drawn independently of each data and a pop-up question asking participants if the other and the shapes of the distributions were normal. The alpha recommendation they received that day was useful. The value was set at .05 for all statistical tests. notification was sent to participants every night at 9 PM. For groups with personalized features (the full group and the For other users’ daily log data, such as the number of active personalized group), we also asked to access the user’s calendar users, the number of conversations, the active calories, and the and location data to be used in recommendations. number of useful recommendations, we visualized them along the timeline to see how the pattern differentiated among the 4 A poststudy questionnaire was conducted at the end of the study groups. to evaluate participants’ experiences during the first 60 days (Multimedia Appendix 2). First, we provided 3 general For qualitative data regarding the possible improvement of the close-ended statements to measure participants’ overall system, because our participants’ answers were mostly short motivation, satisfaction, and preference with the in-game and concise, we simply categorized them and reported the most experience. Participants responded on a 7-point Likert scale commonly mentioned suggestions. ranging from 1 (strongly disagree) to 7 (strongly agree). The statements were as follows: Results I find this kind of application motivating to exercise. General Information I was overall satisfied with this application. The participants’ self-reported average hours of exercise per I prefer using this type of application for exercise over week before the study were 4.16 hours with an SD of 2.96 hours, regular exercises. whereas the average hours per week spent playing video games We used the Intrinsic Motivation Inventory (IMI; Multimedia (including PC, console, and mobile games) were 5.44 hours Appendix 3) [64] to assess participants’ level of intrinsic with an SD of 4.13 hours. The self-reported average active hours motivation related to the game experience. Furthermore, we increased from 4.16 to 4.58 hours after the study. used the European Microsoft Innovation Center (EMIC) Participants could interact with the app through their Android recommender system evaluation measurement (Multimedia phones or watches. Data show that participants read 53.00% Appendix 4) [65] to evaluate the quality of our recommended (10,270/19,377) of messages on their phones and 47.00% activities. We also included open-ended questions to obtain (9107/19,377) of messages on their watches. They tapped participants’ comments and suggestions to improve the system. 35.81% (1785/4985) of prompted choices on their phones and By the end of the 60 days, each participant received a Can $10 63.99% (3190/4985) on their watches. The results suggest that (US $7.7) gift card as an honorarium to thank them for their smartwatches were not only effective and more accurate for participation in the study. tracking activity data but also feasible for some simple Moreover, we customized the IMI scale to fit the current game interactions such as reading messages and tapping a choice from context. We did not use relatedness and perceived choice IMI prompts. Participants tended to interact with watches subscales. Relatedness evaluates the experience of doing independently when completing simple tasks and switched to something with another person, that is, social interactions with http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 14 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al phones when different interactions were necessary (eg, typing On Wednesday, he works from home. He plays basketball every messages). Wednesday night from 8 PM to 9 PM and every Saturday morning from 9 AM to 11 AM. Exerciser type of reds prefer Case Studies exercises such as basketball, tennis, racquetball, in-line skating, Below, we present 2 case studies as examples to show how our frisbee, mountain biking, soccer, and skiing. Our system found system recommends activities to different participants in a that participant B was almost as active as recommended by the typical week. If there is any more information or any change in GRPAH, but the type of activities he performed was limited to the system found during the week, the recommendations adjust basketball. accordingly. Both participants were from the full group receiving System-Generated Activity Recommendations activity recommendations in the form of a gamified story. A 1-hour tennis or racquetball session on Wednesday night Case Study A instead of basketball. A daily 15-min walk after work. Participant Information A 60-min walk (in a nearby park) on Sunday morning. Participant A was a female, 26-year-old student, height 5’8’’, Player Type–Based Game Features weight 61 kg, BMI 20.5 kg/m (normal weight), no serious health issues, and currently taking no medications. Player type: The player type of achiever was assigned the game feature of free spirit; fitness color: white. Our system detected that challenge. Thus, the system provided recommendations to player participant A takes the bus to university every Monday, Tuesday, B in the challenge style. and Thursday and mostly stays at home for the rest of the week. Overall Motivation and Satisfaction She goes to a group-cycling class once a week, on Friday Figure 5 shows the averages and SDs of the scores for the first evenings, for half an hour. According to the GRPAH, adults 3 general questions assessing participant motivation, satisfaction, aged 18 to 64 years were encouraged to perform 300 min of and game preference. The asterisk indicates significant results moderate-intensity aerobic physical activity throughout the found between groups. week for good health benefits [54]. People with exerciser type of white prefer hiking, running, yoga, cardio, and gym strength The results show that there were statistically significant training. When accessing her calendar, the system found she differences between groups as determined by one-way ANOVA had 2 dinner reservations on Thursday and Saturday night, both for overall motivation (F =22.49; P<.001), satisfaction 3,36 at 6 PM for the coming week. (F =22.12; P<.001), and preference (F =15.0; P<.001). 3,36 3,36 System-Generated Activity Recommendations Post hoc comparisons using the Tukey HSD test indicated that for all 3 questions, the mean score for the full, personalized, Extending the walking distance to bus stops on every school and gamified groups was significantly different from that for day (both morning and afternoon, overall 45 min of walking the control group, respectively. This means that, in general, per school day). both gamification and personalization have positive effects on A 30-min walk for non–school days after dinner. participants’ motivation, satisfaction, and preference, as seen A 1-hour home yoga session on Tuesday 7 PM when the in the groups full, personalized, and gamified compared with user is generally not active. the control group. Moreover, for motivation, the mean score A hiking morning on Saturday in a nearby park. for the full group (mean score for full group [MF] 5.8, SD for Player Type–Based Game Features full group [SDF] 0.79) was significantly different from that of The player type of free spirit was assigned the game feature of the personalized group (mean score for personalized group [MP] theme color customization. Thus, the reward of completing 4.7, SD for personalized group [SDP] 1.5). Statistically recommended activities for participant A was to unlock different significant pairwise comparisons are also marked in Figure 7. theme colors. This means that gamification can also add more motivation to a personalized fitness recommendation system, as seen between Case Study B the full group and the personalized group in motivation. It should Participant Information also be noted that the distribution of the dominant player types across the 4 different groups could have influenced these results, Participant B was a male, 35-year-old, software developer, 2 as some player types may have had a stronger preference for height 5’11’’, weight 75 kg, BMI 23.0 kg/m (normal weight), gamification or personalization in general. However, the no serious health issues, and currently taking an over-the-counter distribution with respect to player types did not seem to be pain reliever for his back pain. Player type: achiever; fitness particularly biased (Table 3). color: red. Our system detected that participant B drives to work every Monday, Tuesday, Thursday, and Friday (15-min drive). http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 15 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al Figure 7. Results for poststudy questions 1, 2, and 3. (a) Overall motivation, (b) overall satisfaction, and (c) overall preference over regular exercise. full group (MF 5.7, SDF 0.46) and the personalized group (MP IMI Subscales 4.8, SDP 0.72) as well as between the full group and the control Figure 8 shows the average and SDs of the scores for each IMI group (mean score for control group [MC] 5.0, SD for control subscale question. The results show that there were statistically group [SDC] 0.54). For effort or importance, significant significant differences between groups as determined by a differences were found between the full group (MF 5.8, SDF one-way ANOVA for interest or enjoyment (F =24.24; 0.47) and the gamified group (mean score for gamified group 3,36 P<.001), perceived competence (F =4.60; P=.007), effort or [MG] 4.7, SD for gamified group [SDG] 0.67); between the full 3,36 group and the control group (MC 4.7, SDC 0.70), the importance (F =8.01; P<.001), and value or usefulness 3,36 personalized group (MP 5.6, SDP 0.73), and the gamified group; (F =15.90; P<.001). 3,36 and between the personalized group and the control group. For The Tukey-Kramer HSD test results indicated that for interest value or usefulness, significant differences were also found or enjoyment, the mean score for the full, personalized, and between the full group (MF 5.8, SDF 0.63) and the gamified gamified groups was significantly different from that for the group (MG 4.8, SDG 0.55); between the full group and the control group. Moreover, the pairwise comparison result showed control group (MC 4.6, SDC 0.41), the personalized group (MP that MF (MF 5.9, SDF 0.40) was significantly different from 5.7, SDP 0.63), and the gamified group; and between the the personalized group (MP 5.0, SDP 0.56). For perceived personalized group and the control group. The pairwise competence, significant differences were found between the comparison significance is also marked in Figure 8. http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 16 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al Figure 8. Averages and SDs as evaluated by the Intrinsic Motivation Inventory. From top to bottom: (a) interest or enjoyment, (b) perceived competence, (c) effort or importance, (d) pressure or tension, and (e) value or usefulness. The IMI results indicate that gamifying the exercise increases EMIC Recommender System Evaluation players’ interest in and enjoyment of the personalized Figure 9 shows the averages and SDs of the scores for each recommendation system (significant between the full group and EMIC subscale (under perceived recommendation quality, the personalized group in interest or enjoyment). Personalization perceived system effectiveness, general trust in technology, and contributes more toward promoting effort or importance as well system-specific privacy concerns). The results showed that there as value or usefulness compared with gamification (significant were statistically significant differences between groups as between the personalized group and the gamified group). determined by a one-way ANOVA for perceived recommendation quality (F =108.77; P<.001), perceived 3,36 system effectiveness (F =26.52; P<.001), and system-specific 3,36 privacy concern (F =58.37; P<.001). 3,36 http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 17 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al Figure 9. Average and SD of European Microsoft Innovation Center recommendation. From top to bottom: (a) perceived recommendation quality, (b) perceived system effectiveness, (c) general trust in technology, and (d) system-specific privacy concern. The Tukey-Kramer HSD test results indicated that for both about privacy when the system had a player model and asked perceived recommendation quality and perceived system for more permissions to access their personal data (comparing effectiveness, the mean scores for the full and personalized the full or personalized and gamified or control groups). On the groups were significantly different from the gamified and control other hand, gamification reduced some of the concerns groups that were not personalized. For system-specific privacy (significant difference found between the full and the concerns, the mean scores for the full group and the personalized personalized groups). Note that for the system-specific privacy group were also significantly different from the gamified group question, a higher score indicates less concern. Privacy concerns and the control group because, for nonpersonalized groups, we are important yet beyond the scope of this work. Yet, we believe did not ask to access participants’ personal data (except Google that the noticed effect of gamification can be of value in future Analytics for in-app tracking). Moreover, a significant difference research and design. was also found between the full group (MF 4.1, SDF 0.69) and Daily Statistical Data the personalized group (MP 3.0, SDP 0.73). Statistically As mentioned earlier, we used Google Analytics API to track significant pairwise comparisons are also marked in Figure 9 participants’ comprehensive in-app behavior data and we used using asterisks. Google Fit API to track user daily activity data, including steps Our results suggest that our system is effective in providing and calories burned. Figure 10 shows some daily statistical data: daily fitness recommendations to participants (comparing the the number of active participants of all 4 groups during the 60 full group with the gamified group and the personalized group days of study (a), the daily total number of conversations sent with the control group) with respect to both perceived to the system (b), the daily average active calories burned recommendation quality and perceived system effectiveness. excluding basal metabolism (c), and the daily number of We also found that, as expected, participants were concerned self-reported useful recommendations (d). http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 18 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al Figure 10. Daily statistical data showing patterns of daily active users, conversations, or active calories and percentage of useful recommendations along the timeline. From Figure 8, we can see that for daily active users (a) and exclusive gamification may not. Adding gamified elements to daily conversations (b), there is an overall descent in trends personalized recommendations in the earlier phase (when the appearing as time grows for all 4 groups. Among them, the full player model was not well established yet and the group maintained a relatively higher value compared with the recommendation quality was not steady enough) may negatively other 3 groups and participants in the full, personalized, and affect the amount of physical activity people performed, which gamified groups interacted with the system more than the control requires further research. Note that the active calorie measures group (Figure 10). With respect to the daily active use and daily the calories burned during fitness activities. Basal metabolic conversations (Figure 10), when comparing the personalized parameters were excluded. group and the gamified group, we can see that the value of the For the percentage of useful recommendations, Figure 10 gamified group was higher than that of the personalized group (calculated by the daily number of useful replies divided by in the early phase of the study but was surpassed by the daily active users), the percentage of the full group and the personalized group in the late phase of the experiment (around personalized group increased in the first half of the study and 35-40 days). These results indicate that both personalization then remained flat, with the full group remaining slightly higher and gamification could have a positive impact on promoting than the personalized group. The increase in the full group and participants’ engagement with the system. However, although the personalized group can be attributed to the continuously gamification could bring more interactions in the short term updating player model that will improve recommendations over (within 1 month), personalization could lead to a more sustained time. The gamified group and the control group (without player engagement (over a longer time). Note that Figure 10 shows model) showed descending trends approaching zero. The results that the control group was not active during the last week of the suggest that our system is able to generate useful fitness study. This only indicates that they did not open the app but recommendations by using a player model, and participants they still received recommendations as usual (pop-up considered the recommendation more useful when gamification notifications). Physical activity data were also collected from elements were added. the Google Fit API without opening the app. Player Types For active calories (Figure 10), we can see slight ascent trends In this study, we did not find any significant difference in terms for both the full group and the personalized group and flat trends of different player or exerciser types. Although we had a limited for the gamified group and the control group. The full group sample size for conducting a meaningful statistical analysis, began with a lower average calorie burden compared with the there were still some interesting findings worth mentioning, personalized group and then showed an almost equal value near which may help inspire future research in this area. Table 4 the end. These results indicate that personalization could have shows the distribution of the combinations of player and a positive impact on promoting actual physical activity, whereas exerciser types of our participants. http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 19 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al From Table 4, we can see that certain player types and exerciser type of purples. This indicates that users’ preferences toward types were highly related. For example, we have 5 participants game elements and exercise types may be linked. This idea in total with the exerciser type of silver, with 4 of them could be used to further improve the personalization of the belonging to the player type of free spirit. Similar relations are exercise and game experience but requires further research with shown between the player type of socializer and the exerciser a larger sample size. Table 4. Distribution of the combination of player or exerciser type (N=40). The 8 colors Achiever Player Socializer Philanthropist Disruptor Free spirit Blue 3 3 0 2 0 0 Gold 1 1 1 2 0 0 White 2 1 0 1 0 2 Purple 0 0 2 0 0 0 Green 0 1 3 2 0 1 Red 2 0 0 1 1 0 Saffron 1 0 0 0 1 1 Silver 0 1 0 0 0 4 Figure 11 shows the overall motivation for participants experienced lower overall motivation compared with the other belonging to different player types. Although we were not able 4 player types. This may indicate that the game features and to run a valid statistical analysis based on the small sample sizes, experience we provided to the player type of socializers and we saw that the player type of socializer and disruptor disruptors had more room for improvement. Figure 11. Overall motivation for different player types. Figure 12 shows the average active calories burned for level as the other 6 exerciser types. When looking at the 8 Colors participants belonging to different exerciser types. For the same of Fitness activity suggestions (Multimedia Appendix 5), we reason of small sample sizes, we could not run a valid statistical found that the activity of hiking was the main variable that may analysis. However, we found that the fitness colors of whites lead to the result and it was only recommended for the exerciser and greens were relatively more active during the study. We type of greens and whites. This indicates that hiking might be checked their motivation as well as their self-reported an effective activity that makes people consume more calories, recommendation quality and found that both were at the same which could be further investigated. http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 20 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al Figure 12. Average active calories for different exerciser types. mainly focused on 5 aspects as shown in Figure 13. It shows Qualitative Results that a customized storyline was the most requested feature, One open-ended question was asked of each participant at the followed by multiplayer mode, more quality recommendations, end of the poststudy questionnaire to collect their general a feature for setting and tracking fitness goals, and more feedback (Multimedia Appendix 2). We received many location-based features. These feedback laid the foundation for comments and suggestions on how to improve our system, which planning our future work in this project. Figure 13. Number of main suggestions received from open-ended questions for improving our system. of player modeling and gamification could enhance users’ Discussion engagement with the system as well as promote actual physical activity. Specifically, gamification was found to promote Overall, in this 60-day user study, we verified our hypotheses engagement, but only in the short term, as seen in the gamified that (1) it is feasible to generate personalized exercise group where the members were engaged early on. However, as recommendations with player modeling and (2) the combination http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 21 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al the experiment moved on, the trend changed and the single types (eg, Bartle [68] and the BrainHex model [36]), personalized group became more engaged. This can be attributed more recently, researchers have examined the effectiveness of to the player modeling aspect in that it requires time to get to a trait-oriented models for understanding player choices in games minimum level of precision in reflecting a player’s [69-71]. Trait-oriented models are preferred in recent studies characteristics before it can offer reasonable recommendations. because an individual is rarely motivated by a single factor and Player modeling helped sustain the activity level in the long because of their applicability to game user research in that they term. This suggests that activity recommendation based on aim to characterize players using a set of scores rather than player modeling can be an effective and promising approach categorizing players into a single type. In this study, we decided for creating personalized fitness experiences over longer periods, to use the dominant player type as evaluated by the Hexad model whereas gamification can help attract the users and create the rather than considering all 6 scores because we wanted to control initial interest. the variable by adding only 1 additional element to each user; therefore, we could make sure it is the gamification itself that Our research was motivated by the need to keep players engaged affected the engagement, without interfering with the amount and motivated in exergames. We were inspired by previous of it. A future study can explore the effects of considering the work that suggested a more player-centric and personalized full range of scores. approach to game design and gamification [6,17,28,67] to increase player engagement and the overall effectiveness of the Adaptive and Continuous Modeling intervention. We extended these ideas to exergames, combining Although many games and other applications rely on a certain them with the notion of real-time activity tracking and user model, in most cases, this is done as a one-time static recommendation as suggested by others [22,47-49] to develop decision assigning the user to a certain group. Our study shows a new theoretical dynamic and individual-level model that brings the value of not only having a more comprehensive personal together various game elements that can help solve the player model but also allowing it to evolve and adapt using ongoing retention problem. The presented results have direct implications data from the user. This constantly tunes the model and makes for the design of fitness assistants and potentially other recommendations more effective. Using such adaptive and recommender systems. dynamic models can enhance the performance of such applications, and we recommend that designers consider it when Gamification Is Good but Not Enough! possible. Previous work by authors and other researchers has shown the potential value of gamification to increase engagement, but they 24/7 Recommendation have also highlighted the issue of retention. Players tend to Fitness and health are not limited to the gyms. Being active is leave the game once it is well experienced. Although adding a lifestyle; therefore, activity recommendations should not be new features can be a reasonable way of keeping participants limited to a particular time. In the absence of a dedicated engaged, it is difficult and costly to implement because of personal trainer, an intelligent fitness assistant equipped with constant designing and upgradation. The ability to understand a detailed player model can offer 24/7 recommendations for participants and their dynamic life and provide gameplay being active that considers various user contexts. Our results features that match the participants’ activities can be a way to show the potential value of this approach, which can be introduce change and novelty when maintaining the development improved with more comprehensive personal data and a better cost under control. database of activities and gameplay features. Although our system provided all-day and continuous modeling and Player Modeling: Personalization Versus recommendation, it is worth noting that the participants did not Categorization wear the activity trackers during sleep and we did not track any The idea of categorizing participants to provide them with sleeping activities. As such, although the system was able to customized service is appealing but ignores individual perform nonstop, in practice, it was paused during sleep times differences, which are often significant. The availability of (night or day). personal data, as a result of various methods of collecting information, suggests that the participants can be understood Limitations as individuals and not members of a category. This true There were certain limitations in the proposed system and the personalization allows a new level of customization that will performed study, some mentioned by the participants, which potentially offer participants a much more appealing and we believe were not critical enough to significantly affect the effective experience. Our results show the potential relevance findings but are still worth noting and improving in future work. of this idea to the field of fitness assistants. The more we We relied on a simple game that we designed ourselves with a understand the user, the more personalized our recommendations simple story or gameplay. This may have negatively affected will be, which will, in turn, result in more effective the players’ attraction and engagement. The game could be recommendations. Developing a comprehensive model that designed through a more rigorous process or we could somehow involves various user characteristics (from personality type to allow customization and choice or potentially use another daily routines) can help understand the user properly. existing game. There was also no multiplayer option, which Furthermore, the idea of personalization versus categorization ignores the social aspects of gaming and active lifestyle and is also related to differences in player types and player traits. could negatively affect the level of user engagement. When Although earlier works have attempted to classify players into designing different gamified features for different types of http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 22 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al players, we assigned only 1 game element to each type of player. to refer to the app instead of the more common term game, or This may not be adequate for targeting individual participants. for EMIC, we used items and activities for the same purpose. Although these terms could have caused some confusion, which The 8 Colors of Fitness system (Multimedia Appendix 5) [44] we will improve in the future, we did not receive any negative was used as a model to suggest activities. This system was used feedback and do not believe that the inconsistencies significantly because the research group did not find any other alternatives affected our findings. and needed to rely on a fairly acceptable method. This system is by no means ideal and has its own limitations. It can be Conclusions replaced with any other method, such as other models, an In this paper, we proposed a system for personalized fitness interactive trainer, or a trained expert system. assistants using gamification and continuous player modeling and reported on a long-term study that investigates the We used the Android Activity Recognition API for activity effectiveness of our proposed system. Our findings show that tracking and prediction in this work. This API is only able to it is possible to provide personalized activity recommendations recognize 6 simple physical activities. For more complex daily by continuously updating a player model based on activity activities, we required manual labeling from participants within tracking. Our study also shows the positive effect of this the conversation. This may bring complexity to the participants. modeling and gamification on user engagement and overall We also only used Android Wear participants and limited each activity. These findings can be used to inform the design of group to 10 members, which may not be adequate. We were personalized and gamified recommender systems in health and also aware that the age range of our participants was relatively fitness and potentially other apps, as they highlight the role of narrow. Most of our participants in this study were young adults; an adaptive model and gamification as long-term and short-term hence, our results may not apply to older adults. Furthermore, factors, respectively. This research opens opportunities for future comparing active calories burned as an absolute value could work, especially in the area of exploring more gameplay have negatively influenced the reliability of the results because features, adding a personalized storyline, multiplayer of potential confounding variables such as gender, weight, and gamification, better activity recognition, suggestion models, height. and evaluation with a larger and more diverse sample. The language of our questions could be improved by being more neutral and consistent. For example, we occasionally used task Conflicts of Interest None declared. Multimedia Appendix 1 Prestudy questionnaire for collecting general information. [DOCX File , 813 KB-Multimedia Appendix 1] Multimedia Appendix 2 Poststudy questionnaire for collecting general feedback. EMIC: European Microsoft Innovation Center; IMI: Intrinsic Motivation Inventory. [DOCX File , 13 KB-Multimedia Appendix 2] Multimedia Appendix 3 Intrinsic Motivation Inventory for evaluating the level of enjoyment related to the game experience. [DOCX File , 14 KB-Multimedia Appendix 3] Multimedia Appendix 4 European Microsoft Innovation Center recommender system evaluation measurement tool. [DOCX File , 14 KB-Multimedia Appendix 4] Multimedia Appendix 5 Eight Colors of Fitness activity suggestions. [DOCX File , 13 KB-Multimedia Appendix 5] References 1. Wilmot EG, Edwardson CL, Achana FA, Davies MJ, Gorely T, Gray LJ, et al. Sedentary time in adults and the association with diabetes, cardiovascular disease and death: systematic review and meta-analysis. Diabetologia 2012 Nov;55(11):2895-2905. [doi: 10.1007/s00125-012-2677-z] [Medline: 22890825] http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 23 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al 2. Hagen K, Chorianopoulos K, Wang AI, Jaccheri L, Weie A. Gameplay as Exercise. In: Proceedings of the 2016 CHI Conference Extended Abstracts on Human Factors in Computing Systems. 2016 Presented at: CHI EA'16; October 21-24, 2016; San Jose, CA. [doi: 10.1145/2851581.2892515] 3. Dharia S, Eirinaki M, Jain V, Patel J, Varlamis I, Vora J, et al. Social recommendations for personalized fitness assistance. Pers Ubiquit Comput 2017 Jun 29;22(2):245-257. [doi: 10.1007/s00779-017-1039-8] 4. Ananthanarayan A, Siek KA. Persuasive Wearable Technology Design for Health and Wellness. In: 6th International Conference on Pervasive Computing Technologies for Healthcare and Workshops. 2012 Presented at: PervasiveHealth'12; May 2-7, 2012; San Diego, CA. [doi: 10.4108/icst.pervasivehealth.2012.248694] 5. McGonigal J. Reality is broken: why games make us better and how they can change the world. Choice Rev Online 2012 Jul 1;49(11):49-6095. [doi: 10.5860/choice.49-6095] 6. Nacke LE, Deterding S. The maturing of gamification research. Comput Hum Behav 2017 Jun;71:450-454. [doi: 10.1016/j.chb.2016.11.062] 7. Prensky M. Digital game-based learning. Comput Entertain 2003 Oct;1(1):21. [doi: 10.1145/950566.950596] 8. Fogg BJ. Persuasive technologies. Commun ACM 1999;2(4):26-29. [doi: 10.1007/978-3-319-07127-5_11] 9. Papastergiou M. Exploring the potential of computer and video games for health and physical education: a literature review. Comput Educ 2009 Nov;53(3):603-622. [doi: 10.1016/j.compedu.2009.04.001] 10. Hamari J. Do badges increase user activity? A field experiment on the effects of gamification. Comput Hum Behav 2017 Jun;71:469-478. [doi: 10.1016/j.chb.2015.03.036] 11. Hamari J, Koivisto J, Sarsa H. Does Gamification Work? -- A Literature Review of Empirical Studies on Gamification. In: 47th Hawaii International Conference on System Sciences. 2014 Presented at: CSS'14; January 6-9, 2014; Waikoloa. [doi: 10.1109/HICSS.2014.377] 12. Baranowski T, Buday R, Thompson DI, Baranowski J. Playing for real: video games and stories for health-related behavior change. Am J Prev Med 2008 Jan;34(1):74-82. [doi: 10.1016/j.amepre.2007.09.027] [Medline: 18083454] 13. Consolvo S, Klasnja P, McDonald DW, Landay JA. Goal-setting Considerations for Persuasive Technologies That Encourage Physical Activity. In: Proceedings of the 4th International Conference on Persuasive Technology. 2009 Presented at: Persuasive'09; April 11-16, 2009; California, Claremont, USA. [doi: 10.1145/1541948.1541960] 14. Orji R, Mandryk RL, Vassileva J, Gerling KM. Tailoring Persuasive Health Games to Gamer Type. In: Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. 2013 Presented at: CHI '13; March 3-8, 2013; Paris, France. [doi: 10.1145/2470654.2481341] 15. Marquet O, Alberico C, Adlakha D, Hipp JA. Examining motivations to play pokémon go and their influence on perceived outcomes and physical activity. JMIR Serious Games 2017 Oct 24;5(4):e21. [doi: 10.2196/games.8048] [Medline: 29066423] 16. Orji R, Mandryk RL, Vassileva J. Improving the efficacy of games for change using personalization models. ACM Trans Comput-Hum Interact 2017 Nov 13;24(5):1-22. [doi: 10.1145/3119929] 17. Orji R, Tondello GF, Nacke LE. Personalizing Persuasive Strategies in Gameful Systems to Gamification User Types. In: Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems. 2018 Presented at: CHI'18; April 3-5, 2018; Montreal, Canada. [doi: 10.1145/3173574] 18. Charles D, McNeill M, McAlister M, Black M, Adrian M, Stringer K, et al. Player-Centred Game Design: Player Modelling and Adaptive Digital Games. In: Proceedings of DiGRA 2005 Conference: Changing Views – Worlds in Play. Digital Games Research Association. 2005 Presented at: DiGRA'05; March 4-8, 2005; Vancouver, Canada URL: https://www. researchgate.net/publication/228636408_Player-centred_game_design_Player_modelling_and_adaptive_digital_games 19. Yannakakis GN, Spronck P, Loiacono D, Andre E. Player Modeling. Dagstuhl Publishing. 2013. URL: http://yannakakis. net/wp-content/uploads/2013/08/pm_submitted_final.pdf [accessed 2020-11-10] 20. Lister C, West JH, Cannon B, Sax T, Brodegard D. Just a fad? Gamification in health and fitness apps. JMIR Serious Games 2014 Aug 4;2(2):e9 [FREE Full text] [doi: 10.2196/games.3413] [Medline: 25654660] 21. Campbell T, Ngo B, Fogarty J. Game Design Principles in Everyday Fitness Applications. In: Proceedings of the 2008 ACM Conference on Computer Supported Cooperative Work. 2008 Presented at: CSCW'08; March 1-6, 2008; San Diego, CA. [doi: 10.1145/1460563.1460603] 22. Rabbi M, Aung MH, Zhang M, Choudhury T. MyBehavior: Automatic Personalized Health Feedback From User Behaviors and Preferences Using Smartphones. In: Proceedings of the 2015 ACM International Joint Conference on Pervasive and Ubiquitous Computing. 2015 Presented at: UbiComp'15; April 1-4, 2015; Osaka, USA. [doi: 10.1145/2750858.2805840] 23. Johnson D, Deterding S, Kuhn K, Staneva A, Stoyanov S, Hides L. Gamification for health and wellbeing: a systematic review of the literature. Internet Interv 2016 Nov;6:89-106 [FREE Full text] [doi: 10.1016/j.invent.2016.10.002] [Medline: 30135818] 24. Mekler ED, Brühlmann F, Tuch AN, Opwis K. Towards understanding the effects of individual gamification elements on intrinsic motivation and performance. Comput Hum Behav 2017 Jun;71:525-534. [doi: 10.1016/j.chb.2015.08.048] 25. Loria E. A Framework to Infer Player Experience and Inform Customized Content Generation in Gameful Systems. In: Extended Abstracts: Extended Abstracts of the Annual Symposium on Computer-Human Interaction in Play Companion Extended Abstracts. 2019 Presented at: CHI PLAY'19; August 21-29, 2019; Barcelona, Spain. [doi: 10.1145/3341215.3356330] http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 24 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al 26. Böckle M, Micheel I, Bick M, Novak J. A Design Framework for Adaptive Gamification Applications. In: Hawaii International Conference on System Sciences. 2018 Presented at: ICSS'18; January 6-9, 2018; Hawaii, USA. [doi: 10.24251/hicss.2018.151] 27. Koivisto J, Hamari J. The rise of motivational information systems: a review of gamification research. Int J Inf Manag 2019 Apr;45:191-210. [doi: 10.1016/j.ijinfomgt.2018.10.013] 28. Bertran FA, Segura EM, Duval J, Isbister K. Chasing Play Potentials: Towards an Increasingly Situated and Emergent Approach to Everyday Play Design. In: Proceedings of the 2019 on Designing Interactive Systems Conference. 2019 Presented at: DIS'19; February 7-10, 2019; San Diego, CA. [doi: 10.1145/3322276.3322325] 29. Macvean A, Robertson J. Understanding Exergame Users' Physical Activity, Motivation and Behavior Over Time. In: Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. 2013 Presented at: CHI'13; March 3-9, 2013; Paris, France. [doi: 10.1145/2470654.2466163] 30. Busch M, Mattheiss E, Hochleitner W, Hochleitner C, Lankes M, Fröhlich P, et al. Using player type models for personalized game design – an empirical investigation. Interact Design 2016;12(44):145-163. [doi: 10.1145/2851581.2892399] 31. Velazquez A, Martínez-García AI, Favela J, Ochoa SF. Adaptive exergames to support active aging: an action research study. Pervasive Mob Comput 2017 Jan;34:60-78. [doi: 10.1016/j.pmcj.2016.09.002] 32. Yoo S, Heywood T, Tang LM, Kummerfeld B, Kay J. Towards a Long Term Model of Virtual Reality Exergame Exertion. In: Proceedings of the 25th Conference on User Modeling, Adaptation and Personalization. 2017 Presented at: UMAP'17; April 6-9, 2017; Bratislava, UK. [doi: 10.1145/3079628.3079679] 33. Farič N, Yorke E, Varnes L, Newby K, Potts HW, Smith L, et al. Younger adolescents' perceptions of physical activity, exergaming, and virtual reality: qualitative intervention development study. JMIR Serious Games 2019 Jun 17;7(2):e11960 [FREE Full text] [doi: 10.2196/11960] [Medline: 31210135] 34. Malone LA, Padalabalanarayanan S, McCroskey J, Thirumalai M. Assessment of active video gaming using adapted controllers by individuals with physical disabilities: a protocol. JMIR Res Protoc 2017 Jun 16;6(6):e116 [FREE Full text] [doi: 10.2196/resprot.7621] [Medline: 28623186] 35. Wiemeyer J, Deutsch J, Malone LA, Rowland JL, Swartz MC, Xiong J, et al. Recommendations for the optimal design of exergame interventions for persons with disabilities: challenges, best practices, and future research. Games Health J 2015 Feb;4(1):58-62 [FREE Full text] [doi: 10.1089/g4h.2014.0078] [Medline: 26181682] 36. Richard B. Hearts, clubs, diamonds, spades: players who suit MUDs. J MUD Res 1996 [FREE Full text] 37. Deterding S, Dixon D, Khaled R, Nacke L. From Game Design Elements to Gamefulness: Defining. In: Proceedings of the 15th International Academic MindTrek Conference: Envisioning Future Media Environments. 2011 Presented at: MindTrek'11; April 9-11, 2011; Tampere, UK. [doi: 10.1145/2181037.2181040] 38. Nacke LE, Bateman C, Mandryk RL. BrainHex: a neurobiological gamer typology survey. Entertain Comput 2014 Jan;5(1):55-62. [doi: 10.1016/j.entcom.2013.06.002] 39. Tondello GF, Wehbe RR, Diamond L, Busch M, Marczewski A, Nacke LE. The Gamification User Types Hexad Scale. In: Proceedings of the 2016 Annual Symposium on Computer-Human Interaction in Play. 2016 Presented at: CHI PLAY'16; March 2-8, 2016; Austin, UK. [doi: 10.1145/2967934.2968082] 40. Wiemeyer J, Nacke L, Moser C, Mueller FF. Player experience. In: Serious Games. New York, USA: Springer; 2016. 41. Cote J, Dreiss A, Clobert J. Social personality trait and fitness. Proc R Soc B 2008 Aug 26;275(1653):2851-2858. [doi: 10.1098/rspb.2008.0783] 42. Courneya KS, Hellsten LM. Personality correlates of exercise behavior, motives, barriers and preferences: an application of the five-factor model. Pers Individ Dif 1998 May;24(5):625-633. [doi: 10.1016/s0191-8869(97)00231-6] 43. Rhodes RE, Smith NE. Personality correlates of physical activity: a review and meta-analysis. Br J Sports Med 2006 Dec;40(12):958-965 [FREE Full text] [doi: 10.1136/bjsm.2006.028860] [Medline: 17124108] 44. Brue S. The 8 Colors of Fitness. Corte Madera, CA: Oakledge Press; 2008. 45. Myers IB, McCaulley MH, Most R. MBTI Manual: A Guide to the Development and Use of the Myers - Briggs Type Indicator. London, UK: Pyschologists Press; 1998. 46. Lara OD, Labrador MA. A survey on human activity recognition using wearable sensors. IEEE Commun Surv Tutorials 2013;15(3):1192-1209. [doi: 10.1109/surv.2012.110112.00192] 47. Guo X, Liu J, Chen Y. FitCoach: virtual fitness coach empowered by wearable mobile devices. In: IEEE Conference on Computer Communications. 2017 Presented at: INFOCOM'17; March 3-10, 2017; Atlanta, USA. [doi: 10.1109/infocom.2017.8057208] 48. He Q, Agu E, Strong D, Tulu B. RecFit: a Context-aware System for Recommending Physical Activities. In: Proceedings of the 1st Workshop on Mobile Medical Applications. 2014 Presented at: MMA'14; November 1-6, 2014; Memphis, USA. [doi: 10.1145/2676431.2676439] 49. Broekhuizen K, van Poppel MN, Koppes LL, Brug J, van Mechelen W. A tailored lifestyle intervention to reduce the cardiovascular disease risk of individuals with Familial Hypercholesterolemia (FH): design of the PRO-FIT randomised controlled trial. BMC Public Health 2010 Feb 15;10(1):69 [FREE Full text] [doi: 10.1186/1471-2458-10-69] [Medline: 20156339] http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 25 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al 50. Mittal R, Sinha V. A Personalized Time-bound Activity Recommendation System. In: 7th Annual Computing and Communication Workshop and Conference. 2017 Presented at: CCWC'17; April 4-9, 2017; Las Vegas, USA. [doi: 10.1109/ccwc.2017.7868485] 51. Ni J, Muhlstein L, McAuley J. Modeling Heart Rate and Activity Data for Personalized Fitness Recommendation. In: The World Wide Web Conference. 2019 Presented at: WWW'19; May 3-8, 2019; San Francisco, CA. [doi: 10.1145/3308558.3313643] 52. Ghanvatkar S, Kankanhalli A, Rajan V. User models for personalized physical activity interventions: scoping review. JMIR Mhealth Uhealth 2019 Jan 16;7(1):e11098. [doi: 10.2196/11098] [Medline: 30664474] 53. Zhao Z, Arya A, Whitehead A, Chan G, Etemad SA. Keeping Users Engaged through Feature Updates: A Long-Term Study of Using Wearable-Based Exergames. In: Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems. 2017 Presented at: CHI'17; April 3-9, 2017; Denver, CO. [doi: 10.1145/3025453.3025982] 54. Global Recommendations on Physical Activity for Health. Geneva, Switzerland: World Health Organization; 2010. 55. Activity Recognition Client. URL: https://developers.google.com/android/reference/com/google/android/gms/location/ ActivityRecognitionClient [accessed 2020-02-11] 56. Google Fit. URL: https://www.google.com/fit/ [accessed 2019-09-11] 57. Wit.ai. URL: https://wit.ai/ [accessed 2019-07-21] 58. Weather API. URL: http://weather.com [accessed 2019-08-01] 59. The Android Wear Forum. URL: https://forums.androidcentral.com/android-wear-wear-os/ [accessed 2018-10-23] 60. Gamified UK User Type Test. URL: https://gamified.uk/UserTypeTest2016/user-type-test.php#.X6XHa2T0lUM [accessed 2019-02-21] 61. The 8 Colors Quiz. URL: http://the8colors.com/quizw.php [accessed 2019-12-02] 62. Visual Studio App Center. Microsoft. URL: https://appcenter.ms/ 63. Google Analytics. URL: https://analytics.google.com/analytics/web/ [accessed 2018-01-13] 64. Deci EL, Ryan RM. Intrinsic Motivation Inventory (IMI). Scale Description. 2003. URL: https://selfdeterminationtheory. org/intrinsic-motivation-inventory/ [accessed 2020-11-10] 65. Knijnenburg BP, Willemsen MC, Gantner Z, Soncu H, Newell C. Explaining the user experience of recommender systems. User Model User-Adap Inter 2012 Mar 10;22(4-5):441-504. [doi: 10.1007/s11257-011-9118-4] 66. Jaccard J, Becker MA, Wood G. Pairwise multiple comparison procedures: a review. Psychol Bull 1984:- epub ahead of print. [doi: 10.4135/9781412985178.n2] 67. Nicholson S. A RECIPE for meaningful gamification. In: Gamification in Education and Business. New York, USA: Springer; 2014. 68. Mueller FF, Edge D, Vetere F, Gibbs MR, Agamanolis S, Bongers B, et al. Designing Sports: a Framework for Exertion Games. In: Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. 2011 Presented at: CHI'11; April 7-9, 2011; Vancouver, Canada. [doi: 10.1145/1978942.1979330] 69. Fortes Tondello G, Valtchanov D, Reetz A, Wehbe RR, Orji R, Nacke LE. Towards a trait model of video game preferences. Int J Hum-Comput Interact 2018 Apr 19;34(8):732-748. [doi: 10.1080/10447318.2018.1461765] 70. Tondello GF, Arrambide K, Ribeiro G, Cen AJ, Nacke LE. 'I don’t fit into a single type': a trait model and scale of game playing preferences. In: Human-Computer Interaction. 2019 Presented at: INTERACT'19; June 2-7, 2019; Paphos, UK. [doi: 10.1007/978-3-030-29384-0_23] 71. Tondello GF, Nacke LE. Player Characteristics and Video Game Preferences. In: Proceedings of the Annual Symposium on Computer-Human Interaction in Play. 2019 Presented at: CHI PLAY'19; July 6-11, 2019; Barcelona, Spain. [doi: 10.1145/3311350.3347185] Abbreviations ANOVA: analysis of variance API: application programming interface EMIC: European Microsoft Innovation Center GRPAH: Global Recommendation on Physical Activity for Health HSD: honestly significant difference IMI: Intrinsic Motivation Inventory MBTI: Myers-Briggs–Type Indicator MC: mean score for control group MF: mean score for full group MG: mean score for gamified group MP: mean score for personalized group SDC: SD for control group SDF: SD for full group SDG: SD for gamified group http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 26 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al SDP: SD for personalized group UI: user interface VR: virtual reality Edited by G Eysenbach; submitted 07.05.20; peer-reviewed by K Blondon, E Loria; comments to author 08.08.20; revised version received 30.09.20; accepted 24.10.20; published 17.11.20 Please cite as: Zhao Z, Arya A, Orji R, Chan G Effects of a Personalized Fitness Recommender System Using Gamification and Continuous Player Modeling: System Design and Long-Term Validation Study JMIR Serious Games 2020;8(4):e19968 URL: http://games.jmir.org/2020/4/e19968/ doi: 10.2196/19968 PMID: 33200994 ©Zhao Zhao, Ali Arya, Rita Orji, Gerry Chan. Originally published in JMIR Serious Games (http://games.jmir.org), 17.11.2020. This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Serious Games, is properly cited. The complete bibliographic information, a link to the original publication on http://games.jmir.org, as well as this copyright and license information must be included. http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 27 (page number not for citation purposes) XSL FO RenderX http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png JMIR Serious Games JMIR Publications

Effects of a Personalized Fitness Recommender System Using Gamification and Continuous Player Modeling: System Design and Long-Term Validation Study

JMIR Serious Games , Volume 8 (4) – Nov 17, 2020

Loading next page...
 
/lp/jmir-publications/effects-of-a-personalized-fitness-recommender-system-using-bh2VpF1EVp

References (86)

Publisher
JMIR Publications
Copyright
Copyright © The Author(s). Licensed under Creative Commons Attribution cc-by 4.0
ISSN
2291-9279
DOI
10.2196/19968
Publisher site
See Article on Publisher Site

Abstract

Background: Gamification and persuasive games are effective tools to motivate behavior change, particularly to promote daily physical activities. On the one hand, studies have suggested that a one-size-fits-all approach does not work well for persuasive game design. On the other hand, player modeling and recommender systems are increasingly used for personalizing content. However, there are few existing studies on how to build comprehensive player models for personalizing gamified systems, recommending daily physical activities, or the long-term effectiveness of such gamified exercise-promoting systems. Objective: This paper aims to introduce a gamified, 24/7 fitness assistant system that provides personalized recommendations and generates gamified content targeted at individual users to bridge the aforementioned gaps. This research aims to investigate how to design gamified physical activity interventions to achieve long-term engagement. Methods: We proposed a comprehensive model for gamified fitness recommender systems that uses detailed and dynamic player modeling and wearable-based tracking to provide personalized game features and activity recommendations. Data were collected from 40 participants (23 men and 17 women) who participated in a long-term investigation on the effectiveness of our recommender system that gradually establishes and updates an individual player model (for each unique user) over a period of 60 days. Results: Our results showed the feasibility and effectiveness of the proposed system, particularly for generating personalized exercise recommendations using player modeling. There was a statistically significant difference among the 3 groups (full, personalized, and gamified) for overall motivation (F =22.49; P<.001), satisfaction (F =22.12; P<.001), and preference 3,36 3,36 (F =15.0; P<.001), suggesting that both gamification and personalization have positive effects on the levels of motivation, 3,36 satisfaction, and preference. Furthermore, qualitative results revealed that a customized storyline was the most requested feature, followed by a multiplayer mode, more quality recommendations, a feature for setting and tracking fitness goals, and more location-based features. Conclusions: On the basis of these results and drawing from the gamer modeling literature, we conclude that personalizing recommendations using player modeling and gamification can improve participants’ engagement and motivation toward fitness activities over time. (JMIR Serious Games 2020;8(4):e19968) doi: 10.2196/19968 KEYWORDS persuasive communication; video games; mobile apps; wearable electronic devices; motivation; mobile phone http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 1 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al activities have the potential to help users achieve their fitness Introduction goals and increase engagement and pleasure by adding game features to physical activity [12,19]. A sedentary lifestyle is defined as a lifestyle in which an individual does not receive regular amounts of physical activity, However, most existing work on gamification and persuasive which is becoming a significant public health issue [1]. Various games in health and wellness are limited because of their use solutions have been considered to encourage a more active of one-size-fits-all approaches, which have been shown to be lifestyle. Among them, combining exercise with gameplay [2] suboptimal [21]. As discussed in the next section, there are some and the use of wearable trackers to motivate and recommend research efforts that have used more in-depth personalization, physical activities [3] have received widespread popularity, but but they are not focused on exergames, and initial attempts at both have user retention issues [4]. This paper addresses the personalization are mostly limited to a small set of static issue of improving long-term engagement with such physical demographic parameters about the user (such as age, gender, activity recommenders and exercise games. and occupation), which makes them more categorized rather than personalized. In addition, their effectiveness in promoting The popularity of computer games and their engaging nature the desired behavior was mostly evaluated based on a single has created a strong trend to use games for nonentertainment point of use and feedback (short term). Few attempts have been purposes [5]. This trend includes overlapping topics and terms made to resolve these issues. For example, MyBehavior [22] such as gamification (the use of game features and mechanics used a tracking-based and dynamically modified user model to in nongame applications [6]), serious games (aimed primarily recommend activities, but the system is not gamified and is at being an educational yet entertaining tool [7]), persuasive based on a limited model of daily activity information. A proper games (games for promoting behavior change [8]), and combination of a detailed model with features such as exergames (a combination of physical exercise with games [9]). personality types, modeling-based personalization, and In particular, gamification has received significant attention recommendation with adaptive gamified elements is still missing because it can be seen as an umbrella topic covering a range of in the area of exergames, as discussed in more detail in the next options from implementing few game elements (such as section. leaderboards) in regular activities to performing serious tasks as a full game [10,11]. To address these research gaps, we propose a comprehensive model for gamified fitness recommender systems that use Games and gamified activities are effective persuasive tools for detailed and dynamic player modeling and wearable-based motivating human behavior [12]. Although recent years have tracking to provide personalized game features and activity seen an increase in persuasive applications designed to promote recommendations. We also present the results of a long-term more active lifestyles [13], studies have suggested that a investigation on the effectiveness of our recommender system one-size-fits-all approach is ineffective for such persuasive that gradually establishes and updates an individual player model applications because different users are motivated by different (for each unique user) over a relatively long period (60 days). persuasive strategies [14] as well as personal reasons such as curiosity and social rewards [15]. There is an increasing demand We hypothesized that (1) player modeling based on continuous for personalization as a means of tailoring an experience to player activity tracking is an effective approach for personalizing individual needs and interests [15,16]. This is particularly the activity recommendations to individual players and (2) case for persuasive and recommender systems such as those in combining player modeling and gamification can promote marketing, education, and health, where retention is as important long-term engagement with the system. as initial action [17]. Among different personalization solutions, On the basis of these hypotheses, we aim to address the player modeling in games that aims to understand players to following specific research questions in this paper: enhance game experience has been an active research topic [18]. It aims to describe a game player’s traits and preferences as How can we generate and use continuous player modeling well as the players’ cognitive, affective, and behavioral patterns to personalize activity recommendations for each user? [19] within well-defined structures that allow designers to tailor Can the combination of player modeling and gamification game contents or goals automatically to suit the needs or techniques improve user engagement and experience toward preferences of individual players. fitness activities over time? Gamification has rapidly emerged over the past years, especially To achieve this, we designed a new player model and a related in the area of exercise and fitness [20], as a tool to promote system architecture for a gamified fitness activity recommender healthy behaviors and maintain an active lifestyle [9]. system. To evaluate the effectiveness of the model-driven Researchers have used various gameplays and game features gamified fitness activity recommender system, we conducted to make exercise and physical activities more engaging and a long-term study on 40 participants and examined the attractive [2,3]. The use of gamification to promote a more effectiveness of our gamification approach in promoting physical active lifestyle can be through either formal exercises performed activity in comparison with a control group. We randomly as games (ie, exergames) or combining games with other assigned our participants into 4 distinct groups corresponding physical activities that are not as rigorous as formal exercise to 4 experimental conditions: (ie, a walk to work). In this paper, we refer to all these cases as The full group received the application with both gamified gamified physical or fitness activity and use the term features and personalized recommendations (based on “exergame” loosely to indicate the same type of activity. These player modeling). http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 2 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al The personalized group received only personalized for building an adaptive gamification system: (1) consider the recommendations but no gamified features. purpose of adaptivity, (2) define the adaptivity criteria, (3) The gamified group received nonpersonalized design the adaptive gamification mechanics and dynamics, and recommendations with gamified features added. (4) design meaningful adaptive interventions. The researchers The control group received generic nonpersonalized applied this framework to the design of a web-based platform recommendations and nongamified features. for knowledge exchange in postgraduate medical training and reported positive user acceptance, feedback, and increased usage. The results of a 60-day-long study showed that the idea of generating personalized exercise recommendations using player Nicholson [24,27] discussed the idea of meaningful gamification modeling is feasible. Moreover, it showed that personalizing (or playification), which focuses on playfulness and activities recommendations using player modeling in combination with that make sense to each player and rely on intrinsic motivations gamification can improve participants’ level of engagement and as opposed to following specific rules to win. Bertran et al [28] motivation toward fitness activities over time. built on this idea and proposed the situational play design, which is a framework for designing context-based and personalized Our work includes the following major contributions to the field games. Orji et al [14-16], among others, as discussed in a later of fitness recommenders and exergames: section, also discussed the idea of player modeling for designing We offer a conceptual model and system architecture for more effective serious and health games. Overall, research in personalized gamified activity recommendations using a the gamification domain suggests the need for more personalized combination of player modeling, gamifications, and activity games that depend on the players’ context and their motivations. tracking. The research presented in this paper expands on these ideas and We build a comprehensive player model for personalization addresses some of the identified needs, such as tracking and that is dynamically updated by continuously tracking player understanding the player using a comprehensive individual-level contexts. model, personalizing both game features and recommended We design and validate a 24/7 recommender system for activities in exergames, and performing long-term studies, within personalized activities combined with various gamified the context of gamified physical and fitness activity elements that adjust to a player’s and environmental recommenders. We discuss the research in these specific areas contexts. in later sections. Finally, to demonstrate the feasibility of our approach, we Gamified Physical Activities conducted a long-term (60-day) field study with 40 Despite their positive effect on promoting an active lifestyle, participants to evaluate the proposed system and compare gamified physical activities face the problem of sustainability the effectiveness of the 4 experimental groups. (also referred to as player retention here and in other literature) Through the design and development of a new recommender [20,25,26]. Although players may feel excited and motivated system and a long-term study, we hypothesized and evaluated to play at first, over time and sometimes quickly, they may lose the effects of gamification and player modeling on physical their willingness to continue. There are studies focusing on the activity. To the best of our knowledge, this research is the first motivation and sustainability of exergames and gamified fitness to link research on player modeling, gamification, and activity activities. For example, Campbell et al [21] discussed the recommendation to propose an approach for personalized concept of everyday fitness games and suggested that for activity recommendation that is continuously and dynamically applications that people frequently use in their everyday lives, updated to reflect users’ changing contexts and states. the design needs to be fun and sustainable as well as adapt to behavioral changes. Macvean et al [29] reported a 7-week study Related Work on users’ physical activity, motivation, and behavioral patterns Gamification using exergames and suggested that longitudinal studies are necessary for evaluating motivational effects as exergames The motivational effects of gamification have been widely ensure that the intensity of a user’s behavior is appropriate and studied by researchers. It has been shown that common game sustained. Previous work [30] also showed that based on existing elements such as badges, rewards, leaderboards, and avatars are technologies and user needs, the idea of employing wearable commonly and successfully used to motivate players [10,21-23]. activity trackers for gamification of exercise and fitness is On the other hand, researchers have argued for various areas of feasible, motivating, and engaging. Adding dynamic features improvement in current gamification research and applications, could have a positive impact on user motivation toward the including diversity in themes and context, study duration, and gamified exercise system, and the gradual release of application sample size [6,23], increasing motivation by relying on more features could increase the user retention rate. It was also found intrinsic factors [22,24], continuous adaptation [25], and that each user was unique and motivated by different types of personalization [6]. In the study by Loria [25], a framework for game features. Therefore, based on these results, it seems improving the player experience and customizing content reasonable to generate customized workout sessions to fit generation is proposed by continuously monitoring how players different user fitness conditions and interests. interact during the game by analyzing information such as players’ in-game behavior and players’ social network. Other Recently, significant research has been devoted to the design researchers have also tackled adaptive gamification. For of active games (also commonly referred to as exertion games example, Böckle et al [26] identified 4 main elements as the or exergames) to match the needs of specific groups such as basis for defining meta-requirements and designing principles http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 3 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al those with disability [28,29] or senior citizens [31] or to use group-cycling class. These preferences have less to do with specific technologies such as virtual reality (VR) [32]. people’s physical characteristics and are affected more by Particularly in the context of VR-based exergames for personalities. Matching activities to personality type has been adolescents, research shows that game elements such as the use shown to have real-world relevance [42]. Research suggests of rewards, increasing challenge levels, frequent updates, and that people who engage in personality-appropriate activities social or multiplayer options are important aspects for continued will stick with the activities longer, enjoy their workout more, engagement in physical activity [33]. Researchers have also and have a better overall fitness experience [43]. Brue [44] investigated design principles for active games [34,35]. created a system based on the principles of the Although impressive and invaluable in their findings, these Myers-Briggs–Type Indicator (MBTI) assessment. She used efforts primarily focus on the design of particular games, game MBTIs and reworked them into an easily maneuverable features, and actual gameplay mechanisms that are best suited color-coded fitness personality model, the 8 Colors of Fitness, for increasing physical activity for the target group or individual. which is also used in our player model. Each color is associated On the other hand, in 24/7 activity recommenders, the focus is with 2 personality types from the 16 possible MBTI types [45]. less on designing a particular game and more on gamifying the For example, blues are loyal, traditional, dependable, and daily experience and recommending activities based on the daily straightforward, whereas greens are nature lovers who seek to routine using dynamic player modeling. quietly merge with the outdoors [42]. Moreover, in recent years, the use of wearable sensors in human Player Models activity recognition has become popular [46], in which most of Busch et al [30] indicated that the one-size-fits-all approach the measured attributes are related to the user’s movement (eg, does not work for persuasive game design. Thus, player-type using accelerometers or GPS), environmental variables (eg, models could be used when tailoring personalized persuasive temperature and humidity), or physiological signals (eg, heart systems. One of the most frequently used player-type models rate or electrocardiogram). These data types are naturally is the one developed by Richard [36], who identified 4 player indexed over the time dimension, consistent, and convenient to types and proposed that each player has some particular access, which could be used in modeling and predicting a user’s preference for one of the types, which makes them mutually daily activity pattern. exclusive. Another model is the BrainHex model [37], which is a relatively new model but has been validated using a large Although there is a significant amount of research on the subject pool of participants [38]. In BrainHex, player types were not of player modeling, none of the existing studies have examined mutually exclusive. Scores under each category are presented how to use a comprehensive player model. In addition, no to determine the player’s primary type and subtypes. It also previous research has simultaneously considered both game connects player types to the game elements. Moreover, the features and recommended activities in exergames design and Hexad model [39], which is of particular interest in our work, investigated whether it is an effective approach over the long is a gamification player-type model created for mapping user term. personality onto gamified design elements. We considered using Personalized Activity Recommendations the Hexad model in our player model because it specifically targeted gamified systems. It proposes 6 player types, and the Personalized recommender systems for physical activity have player types of individuals are correlated with their preferences been studied by many researchers. For example, Guo et al [47] for different game design elements. Design guidelines for proposed a system that recognizes different types of exercises tailoring persuasive gamified systems to each gamification and interprets fitness data (eg, motion strength and speed) to an player type have also been studied [17]. easy-to-understand exercise review score, which aims to provide a workout performance evaluation and recommendation. Furthermore, Wiemeyer et al [40] discussed the concept of Although it achieved 90% accuracy for workout analysis, it player experience (with a focus on individual) versus game focuses only on recognizing fitness activities and not usability (with a focus on technology) and reviewed various personalizing or gamifying them. He et al [48] introduced a theoretical models that can help understand the player system designed to be context aware for physical activity experience. These models are particularly helpful when recommendations. It focuses on selecting suitable exercises for designing full games as opposed to gamifying everyday individualized recommendations. A smartphone app was activities, which is the goal of this research. However, their developed that could generate individualized physical activity insights, such as an integrative multidisciplinary model of player recommendations based on the system’s database of physical experience, can be helpful in future phases of our research when activity. The focus of their work is to recommend different types we focus on the design of game elements. For the work of activities but does not take into account personal details such presented here, our focus was primarily on showing how the as proper time, location, and intensity or any gamified elements. combination of gamification and player modeling could improve engagement. Better player experiences can be achieved through Broekhuizen et al [49] proposed a framework called PRO-fit, more complicated models and game features that are beyond which is another example that employs machine learning and the scope of this work. recommendation algorithms to track and identify users’ activities by collecting accelerometer data, synchronizes with the user’s Personality type also plays an important role in determining calendar, and recommends personalized workout sessions based people’s fitness tastes [41]. Some people may prefer swimming on the user’s and similar users’ past activities, their preferences, laps solo, whereas others may enjoy attending a rowdy and their physical state and availability. The authors highlighted http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 4 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al that many applications nowadays are more focused on tracking real-time activity tracking and recommendation are also user activities but do not provide a recommendation system that frequently missing in existing research on personalization and would help users choose from activities based on their interests exergames. and accomplishment of goals. Therefore, the authors were However, the existing body of research provides invaluable motivated to design a personalized fitness assistant framework insight into recommendations based on real-time tracking, that acts as a motivator and organizer for fitness activities, important parameters to include in a player model, and the making it easier for users to create and follow their workout design of exergames in general. Expanding on existing studies plan and schedule the sessions according to their availability and trying to fill the abovementioned gaps, we propose a and preference. Compared with PRO-fit, our proposed system conceptual model and system architecture that bring together provides recommendations in real time throughout daily life as game elements, dynamic player modeling, and activity tracking opposed to the prefixed recommendations that are not based on to personalize exergames in terms of both game features and any player or exerciser-type model employed in PRO-fit. recommended activities. We built a comprehensive and dynamic Mittal and Sinha [50] used personal information to recommend player model for personalization that is continuously updated general activities such as visiting attractions and shopping. by tracking the player and offers 24/7 personalized activity Although not focused on fitness activities or gamification, their recommendations. Finally, we conducted a long-term user study notion of modeling user data as the base for recommendation in the wild to evaluate the proposed system. is in line with our proposal. Ni et al [51] used a variety of user To the best of our knowledge, although some of the features of data such as daily routine and heart rate to recommend workout our study have been suggested and/or investigated by others, routes. Their method is more focused on physical activity no long-term comprehensive study has been conducted to recommendations but is limited to recommending routes and integrate and evaluate them in real-life exergame apps. does not include gamification elements. Similarly, Rabbi et al [22] proposed MyBehavior, which is a system for tracking users Methods using mobile devices and suggesting food and physical activities. MyBehavior provides personalized and real-time suggestions Conceptual Model but is not gamified and does not include an explicit player Although the existing studies have addressed many aspects of model. As such, it does not take advantage of full these diverse fields, as discussed earlier, they have not been personalization or more engaging features that a game can offer. properly integrated to develop engaging and sustainable In line with this, Ghanvatkar et al [52] conducted a exergames. For example, the effect of various game features comprehensive review of user models used in recommender and continuously adapting the game to player needs and interests systems. They highlight that activity profile, demographic have not been investigated in the context of exergames. information, and contextual data such as location are among the top items to include in user models. In this research, we In this section, we describe our proposed system architecture have defined our player model to include gamer information and related research methods. This proposal is based on our using Hexad and demographic, activity, and exercise submodels, new conceptual model developed after reviewing related work, as suggested by Ghanvatkar et al [52]. consisting of the following principles: Summary of Research Gaps 1. Advances in wearable technologies allow game designers to use commercially available activity tracking sensors and As reviewed in previous sections, research on gamified physical mobile devices as a major element of exergames. activities and related topics has achieved significant results but A game with a static design, no matter how interesting, will requires more work to fill the existing gaps. We identified that lose its attraction after a while. As such, it is important to the main research gap within the context of exergames is the add new features over time to keep players engaged. notion of personalized gamification (a combination of gamified Although different methods exist for adding dynamic physical activities with player model–based personalization), features to games, designers have a limited ability to provide including understanding the players and their environment and new features constantly, and there is no guarantee that they adapting the game features and physical activities dynamically. will be attractive to users. An alternative (or None of the existing studies have successfully investigated the complementary) approach is to dynamically modify the effect of gamification and personalization individually with game by adapting to the player. respect to promoting the efficacy of an intervention, specifically a physical activity intervention within a single application. In Our conceptual model, which builds on our previous work [28] addition, long-term studies outside controlled environments and along with 2 new components, is illustrated in Figure 1. http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 5 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al Figure 1. Conceptual model of the proposed system. Components surrounded by parallelograms were explored in our previous work, whereas components surrounded by rectangles are the aim of this research for personalizing physical activity recommendations. Arrows show the connection between the components. In our previous work [53], we investigated the effect of using player model takes different types of user data and predicts user wearable activity tracking in exergames and the long-term preference for physical activities and finds the proper time and effectiveness of using a dynamic game feature–releasing system location for recommending activity sessions. It consists of in sustaining exergames (marked parallelograms). In this paper, several submodels that cover the user’s general, personality, we aim to further investigate the gamified features for increasing and daily activity data. The recommendation engine used the retention, exploring 2 additional components: (1) player output of the player model and generated customized physical modeling in the personalization of exergames and (2) how to activity session recommendations for individual users (including use such a system to generate personalized physical activity the proper time, location, intensity, and potential type of physical recommendations (marked rectangles). The arrows demonstrate activity). The game generator adds customized game elements how each of the components are related and can directly to the recommendation and generates the final game content influence each other. We believe that tracking activity using that users can interact with. Wearable activity trackers or wearable technology, providing dynamic game features, and smartwatches are used in the system to track the user’s activity detecting a player’s preferences using a player modeling and introduce diverse interactions. The combined use of mobile approach can all contribute to creating a more engaging apps and wearable apps will allow users to interact with the exergame experience, which in turn can generate more system with different modes. The detailed design and personalized physical activity recommendations that players development of the system are introduced in the following will likely find motivating and satisfying. section. System Design Overall, a wearable-based exergame system, with a comprehensive player model for physical activity On the basis of our proposed conceptual model, bringing recommendation and game customization, is proposed as a wearable activity trackers or smartwatches into exergames, solution to the exergame retention problem. dynamically updating game features, and using player modeling for personalization of exergames is being proposed as a solution Application Design and Implementation to the research problem. Therefore, a wearable-based exergame On the basis of the proposed system architecture, a Wear OS with a comprehensive player model for personalization, (formerly Android Wear) application is implemented as the user recommending customized activities, is proposed as a potential interface (UI), which tracks the user’s activities and provides system for further investigation. gamified fitness recommendations. The overall architecture of The proposed system contains 3 main components: a player the application is illustrated in Figure 2. model, a recommendation engine, and a game generator. The http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 6 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al Figure 2. App architecture. How data were collected and transferred to each submodel of the system and how they were used to generate recommendations and game content. Info: information. In this application, and based on the conceptual model in Figure in charge of generating one part of the recommendation, as 1, the player model consists of 4 submodels: (1) an activity shown in Table 1. Choosing game features and physical recognition model that tracks player activities, (2) a general activities are the 2 main personalization options and each has model that holds basic information about the player, (3) an its own submodel. Tracking daily activities is an essential part exerciser-type model that includes information required for of the system, which also has a submodel. The fourth submodel recommending activities, and (4) a gamer-type model that is holds general player information, such as gender, age, weight, used to choose game features. Each of the submodels is mainly and height. Table 1. The roles of each submodel. Submodel Role Activity recognition model Time and location General model Intensity and duration Exerciser-type model Exercise type Gamer-type model Game elements Although each submodel is designed to generate one particular intensity. We use the 8 Colors of Fitness model [44] to suggest part of the recommendation, they are still connected to each different types of activities for the personalized groups. This other to create a more reliable overall recommendation. For model is one of the few that uses personality type as the basis example, the exerciser-type model is built for each individual of activity recommendations and is suggested by other user for recommending different types of activities based on researchers and practitioners [49,50]. their personalities but it also relies on the general model, which The recommendation engine is a decision tree–based module is built based on a user’s fitness and health condition, to exclude that uses all the information generated from the player model those activities that may be suitable for their personality type to create personalized recommendations for each individual but not for their health condition. We refer to the theoretical user. It could either extend an existing activity (eg, by foundations from the Global Recommendations on Physical recommending a longer exercise time, a longer running path, Activity for Health (GRPAH) [54] to determine proper exercise or appropriate intensity), recommend some activities on the recommendations in nonpersonalized cases. The GRPAH is an user’s idle time, or simply recommend a different type of accepted tool approved by the World Health Organization for activity. An example of a decision tree is illustrated in Figure recommending the appropriate exercise type, duration, and 3. As the recommendation system for physical activity itself is http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 7 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al a relatively complicated topic, we do not consider it as the would have an impact on user experience. Therefore, after highest priority for this study. Therefore, we only employed verifying the feasibility of the proposed idea and the roles simple decision tree methods to generate basic recommendations personalization and gamification performed in this type of (we have ensured that all the recommendations followed the system, our research goal will be to investigate the recommender GRPAH guideline for daily physical activities). We are aware system of physical activities. that the rationality and quality of the recommended activities Figure 3. Example of the decision tree used in the recommendation engine (simple version). PA: physical activity. The game generator is responsible for adding game elements [17], we used the game elements recommended by the Hexad to the recommendations to gamify the activity suggestions player–type model, which is more in line with our objective in generated by the recommendation engine. The type of game this work, designing gamified physical activity elements to be added is determined by the Hexad player–type recommendations. Details of the game and activity model [39]. Our work is also partly based on Orji et al [17], as recommendations are provided later in this section. it adopted a similar Hexad player model. However, as opposed to using the persuasive strategies recommended by Orji et al http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 8 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al A Wear OS app was developed for this study. The app is a develop with many possible confounding factors. We also did conversation-based game in which all the interactions happen not want to introduce various esthetic and design variables to in the form of a conversation between the user and the future the study that may interfere with our studied research variables self. The game is based on a story in which a 1-day user receives and influence our results. For the same reason, we did not try a message from the future self in 20 years telling him or her to incorporate our study within an existing game, even though that the world is about to end in that future world but only the adding these features to games that the user prefers may be user can save it by completing a series of tasks. Then, the future another motivating factor in the future. However, it is essential self will guide the user through daily activities, which are to establish their effectiveness first in isolation. The current generated by the recommendation system in a gamified structure. system UI was created using a rapid prototyping approach. A The choice of this game was informed by our need to have a pilot study was also conducted before the formal study to ensure simple design that is capable of incorporating our research that the labels and buttons are clear. The main UI and app icons requirements but at the same time is not too complicated to are shown in Figure 4. Figure 4. Example app interface and icon. (a) A snippet of one conversation between the system and the user. (b) A display of new mission for the user that he or she can choose to accept or decline. UI: user interface. The app tracks the user’s daily activity through Android processing to understand human language and we used its ActivityRecognition [55] and Google Fit Application message API to create a chatbot, which aims to understand a Programming Interface (API) [56], which allows up to 6 user user’s intents and lead participants to designed storylines. activities to be recognized in real time: in vehicle, on foot, Moreover, we have included a weather assistant in the system running, walking, on bicycle, and still. (through the Weather API [58]) to help participants in planning activities around the weather. The Google Fit API provides encapsulated daily activity–related data such as calories burned, daily steps, and heart rate history When designing the game features, we employed the Hexad (if applicable) tracked by both phone and watch sensors. All player types [39] and the game design elements guide [17]. the collected activity data, along with their time stamps and Hexad suggests that game design elements are preferred by each location information, are used as input features to train a daily player type and we implemented 1 element for each type of user activity model for each individual user by which possible in this study for a personalized game experience (in addition to exercise time and location are predicted. As shown in Figure 4, the game storyline). We integrated the following gamification the app is a conversation-based game. We used Wit.ai [57] to elements in our game (Table 2). Figure 5 shows some generate storylines and to build a bot that can talk to participants screenshots of example game elements for different player types. and perform some general greetings, tell the time, and talk about If there was a tie in scores between the 6 types, we randomly the weather. Wit.ai is a tool that uses natural language chose 1 element of the highest score to add. http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 9 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al Figure 5. Example game elements: (a) profile in daily view (including points and challenges), (b) profile in weekly view (including points and challenges), (c) connect to Facebook view, (d) hacker mode view, and (e) theme color customization view. http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 10 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al Table 2. The motivation and corresponding game elements added for each type of player. Hexad type Motivation Game element Socializers Relatedness Link to social network Free spirits Autonomy and self-expression Theme color customization Achievers Mastery Challenge Philanthropists Purpose and meaning Game experience sharing Players Rewards Points Disruptors Change Hacking mode As mentioned, the application is a Wear OS game that requires Link to Social Networks combining the use of both an Android phone and an Android Socializers are motivated by relatedness. They want to interact watch to optimize its recognition accuracy and gamified with others and create social connections [39]. Therefore, we experience. For activity recognition, our app uses watch sensors provided them with an interface for linking the game to their for better accuracy. However, in many situations, participants social network as their unique feature so that they could share may choose not to wear the watch. In those cases, when the their game performance or achievements to their Facebook page, watch was not connected, we used the built-in phone sensors team up with those friends who are already in the game, or invite instead such that the game could run individually on the phone new players to the game. without the watch. A phone clearly offers more screen space and abilities, such as typing messages, compared with a watch. Theme Color Customization User Study Design Free spirits are motivated by autonomy and self-expression. They want to create and explore the game and prefer features Multiphase Research such as unlockable content and customization [17,39]. Thus, The proposed conceptual model and the system were evaluated we added a feature of theme color customization so that they based on a multiphase user study. In our previous work [53], could customize their game UI by unlocking different themes. we introduced phases 1 and 2, which can be summarized as Challenge follows: Achievers are motivated by mastery . They are looking to learn For research phase 1, in-lab user tests of 20 participants were new things and want to overcome challenges [39]. Therefore, conducted to evaluate the effectiveness of the combined use of we added a challenge system for them in our game, in which games and wearable devices in promoting exercise and to tasks were assigned to them as challenges. investigate the usability of the proposed approach and the effects of different factors within the system. Game Experience Sharing Philanthropists are motivated by purpose and meaning. They In research phase 2, a 70-day user study of 36 participants was want to give to other people and enrich the lives of others in designed to verify the hypothesis that adding different game some way with no expectation of reward [39]. For features and gradually releasing them can positively affect user philanthropists, we added a feature for them to share their game engagement and retention. experience with other players. A forum-like interface was added In this paper, we present research phase 3, which is a 60-day to their version of the game in the main screen that allowed long-term study with 40 participants, to demonstrate the them to browse and answer questions of other players. They feasibility and effectiveness of using a player modeling also receive notifications when there is a new question in the technique in the personalization of exergames. forum. Participants and Groups Points A total of 40 participants were recruited locally from the Ottawa Points have been shown to positively affect players [17,39]. area by posters as well as via the web through the Android Wear They will do what is needed for them to collect rewards from Forum [59]. Of the 40 participants, 23 were men and 17 were a system. For players, points in our game can be collected and women. Their average age was 26.93 years, with an SD of 6.07 used as virtual currency to buy extra themes or virtual years. We randomly divided our participants into 4 groups based equipment. on the versions of the app they received: full (gamified and Hacking Mode personalized), gamified only, personalized only, and the control (neither personalization nor gamified, as the control group). Disruptors are motivated by change. In general, they want to Participants were randomly allocated to groups and the disrupt the system [39]. We added a hacking mode for distribution with respect to exercise or player type and physical disruptors, in which they can use the command-line interface activity level did not seem to be particularly biased (Table 3). to access their own game database to make changes to the Participants’ physical activity levels were collected at baseline storyline or delete their records of the game and, eventually, before beginning the study. they may destroy the system. http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 11 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al Table 3. Demographical data for the 4 participant groups (N=40). Characteristic Participant group Full Gamified Personalized Control Age (years), mean (SD) 24.93 (7.27) 26.65 (5.58) 27.85 (6.26) 25.78 (5.93) Gender, n (%) Male 6 (60) 6 (60) 5 (50) 6 (60) Female 4 (40) 4 (40) 5 (50) 4 (40) Hexad user types, n (%) Philanthropist 1 (13) 1 (11) 1 (10) 1 (11) Socializer 1 (13) 2 (22) 1 (10) 2 (22) Free spirit 1 (13) 3 (33) 2 (20) 2 (22) Achiever 2 (25) 2 (22) 3 (30) 2 (22) Disruptor 1 (13) 0 (0) 0 (0) 1 (11) Player 2 (25) 1 (11) 3 (30) 1 (11) 8-color personalities, n (%) Blue 3 (30) 2 (20) 1 (10) 2 (20) Gold 1 (10) 1 (10) 1 (10) 2 (20) White 2 (20) 1 (10) 3 (30) 0 (0) Purple 0 (0) 1 (10) 1 (10) 0 (0) Green 1 (10) 3 (30) 1 (10) 2 (20) Red 1 (10) 0 (0) 1 (10) 2 (20) Saffron 0 (0) 0 (0) 2 (20) 1 (10) Silver 2 (20) 2 (20) 0 (0) 1 (10) 4.04 (2.35) 3.95 (3.21) 4.82 (2.53) 3.83 (2.92) Physical activity level (hours per week), mean (SD) Physical activity levels were self-reported at baseline. To increase the duration in each group and reduce the chance of recommendation, was our hypothesis when other variables of groups affecting each other, all participants remained in the are held constant. Therefore, we did not use existing commercial same group for the entire study duration rather than randomly apps for comparison in this study because we tried to avoid trying all 4 groups. bringing in possible extraneous or confounding variables such as esthetic and gameplay features that were not our focus. The recommendations for the control group and the gamified group were created based on established exercise guidelines Figure 6 shows an example of how recommending the same and were reasonable recommendations for the general 30-min walking activity will look for the 4 study groups. The population. To ensure this, we referred to the theoretical full group received the recommendation through a gamified foundations from the GRPAH [54] to determine proper exercise story (guided by the future self) with the game element of recommendations. Our choice for nonpersonalized groups challenge based on their player type of achiever and a closely follows the one-size-fits-all recommendation method, personalized walking path. The personalized group also received which has been generally used in most physical activity a personalized route but no game story or elements. The recommendation applications, such as Apple Watch gamified group received no personalized route but had the game (recommends a 30-min walk per day) or Fitbit (daily 10,000 story and the game element of points (randomly assigned steps). because no player model was used for the gamified group). The control group received no personalization or gamification as a Furthermore, the main purpose of our study was to demonstrate control group. In the screenshot of the control group, we showed the effectiveness of personalized recommendations. Although an example of how the weather assistant worked. Note that the we tried to offer a reasonable experience for nonpersonalized example conversations were from screenshots and some details groups, the effectiveness of personalization, especially in terms related to the context were not fully displayed. http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 12 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al Figure 6. Example of recommendations for different groups: (a) full, (b) personalized, (c) gamified, and (d) control. To control the gamification level between groups, for the We also limited the number of personalized recommendations gamified group, as there was no player model used, we randomly to 2 times a day to eliminate the variability of engagement assigned a game element from Table 2 to each participant to caused by frequent recommendations. The gamified and control bring them to the same gamification level as the full group. As groups (without personalization) received 2 messages per day the members of personalized groups received different game at 9 AM and 5 PM. We chose these 2 times because 9 AM is elements based on their individual player model, we decided the time of day that most of our participants were active. We that a random selection for nonpersonalized groups would be did not send the notification earlier because we did not want the closest nonpersonalized option. their sleep to be interrupted. We chose 5 PM because most people are off from work at 5 PM. The full and the personalized http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 13 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al group received messages based on when they got up and when a game that can lead to the feelings of relatedness. It is usually they left work, as recorded in their individual player model. The used in multiplayer games, which allow for interactions between results presented in this paper are based on data from a 60-day real players, and was not applicable in our case. Perceived experiment. choice is often used in situations where a person is given a certain task or activity to complete. In our case, we indicated Procedures and Data Collection in the beginning that the users have the full choice of either The study was approved by the research ethics board. We asked using or not using our system as well as how to use it. Therefore, participants to complete a prestudy questionnaire before this subscale was considered not necessary as the participants providing them the app (Multimedia Appendix 1). The were explicitly given full choice. There are different versions questionnaire asked demographic questions including age, of the IMI that have been used in previous studies, which consist gender, height, weight, number of hours they spend per week of different subscales that are only relevant to their unique exercising, type of Android Wear owned, and types and duration context. of playing video games (eg, PC, console, and mobile). Two Data Analysis web-based questionnaires were provided and participants completed them, which provided us with the results to determine For each question in the poststudy questionnaire including their player and exerciser type [60,61]. The app was distributed general perception, IMI subscales, and EMIC subscales, a to participants through the HockeyApp (now Microsoft Visual one-way between-group analysis of variance (ANOVA) and Studio App Center) [62] after receiving participants’ gamer and post hoc Tukey-Kramer Honestly Significant Difference (HSD) exerciser-type results. Application features were selected based test [66] was conducted to analyze the main effects among the on the participant’s player model. 4 groups. ANOVA is commonly used to determine whether there are any statistically significant differences between the For in-game data collection, we used Google Analytics API means of 3 or more independent groups, whereas the Tukey test [63] to track all participants’ comprehensive in-app behavior provides deeper insights into patterns and comparing specific data, including screen views and tapped events with associated groups [66]. Parametric tests were selected for conducting the timestamps. We used Google Fit API to track user daily activity analysis because the samples were drawn independently of each data and a pop-up question asking participants if the other and the shapes of the distributions were normal. The alpha recommendation they received that day was useful. The value was set at .05 for all statistical tests. notification was sent to participants every night at 9 PM. For groups with personalized features (the full group and the For other users’ daily log data, such as the number of active personalized group), we also asked to access the user’s calendar users, the number of conversations, the active calories, and the and location data to be used in recommendations. number of useful recommendations, we visualized them along the timeline to see how the pattern differentiated among the 4 A poststudy questionnaire was conducted at the end of the study groups. to evaluate participants’ experiences during the first 60 days (Multimedia Appendix 2). First, we provided 3 general For qualitative data regarding the possible improvement of the close-ended statements to measure participants’ overall system, because our participants’ answers were mostly short motivation, satisfaction, and preference with the in-game and concise, we simply categorized them and reported the most experience. Participants responded on a 7-point Likert scale commonly mentioned suggestions. ranging from 1 (strongly disagree) to 7 (strongly agree). The statements were as follows: Results I find this kind of application motivating to exercise. General Information I was overall satisfied with this application. The participants’ self-reported average hours of exercise per I prefer using this type of application for exercise over week before the study were 4.16 hours with an SD of 2.96 hours, regular exercises. whereas the average hours per week spent playing video games We used the Intrinsic Motivation Inventory (IMI; Multimedia (including PC, console, and mobile games) were 5.44 hours Appendix 3) [64] to assess participants’ level of intrinsic with an SD of 4.13 hours. The self-reported average active hours motivation related to the game experience. Furthermore, we increased from 4.16 to 4.58 hours after the study. used the European Microsoft Innovation Center (EMIC) Participants could interact with the app through their Android recommender system evaluation measurement (Multimedia phones or watches. Data show that participants read 53.00% Appendix 4) [65] to evaluate the quality of our recommended (10,270/19,377) of messages on their phones and 47.00% activities. We also included open-ended questions to obtain (9107/19,377) of messages on their watches. They tapped participants’ comments and suggestions to improve the system. 35.81% (1785/4985) of prompted choices on their phones and By the end of the 60 days, each participant received a Can $10 63.99% (3190/4985) on their watches. The results suggest that (US $7.7) gift card as an honorarium to thank them for their smartwatches were not only effective and more accurate for participation in the study. tracking activity data but also feasible for some simple Moreover, we customized the IMI scale to fit the current game interactions such as reading messages and tapping a choice from context. We did not use relatedness and perceived choice IMI prompts. Participants tended to interact with watches subscales. Relatedness evaluates the experience of doing independently when completing simple tasks and switched to something with another person, that is, social interactions with http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 14 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al phones when different interactions were necessary (eg, typing On Wednesday, he works from home. He plays basketball every messages). Wednesday night from 8 PM to 9 PM and every Saturday morning from 9 AM to 11 AM. Exerciser type of reds prefer Case Studies exercises such as basketball, tennis, racquetball, in-line skating, Below, we present 2 case studies as examples to show how our frisbee, mountain biking, soccer, and skiing. Our system found system recommends activities to different participants in a that participant B was almost as active as recommended by the typical week. If there is any more information or any change in GRPAH, but the type of activities he performed was limited to the system found during the week, the recommendations adjust basketball. accordingly. Both participants were from the full group receiving System-Generated Activity Recommendations activity recommendations in the form of a gamified story. A 1-hour tennis or racquetball session on Wednesday night Case Study A instead of basketball. A daily 15-min walk after work. Participant Information A 60-min walk (in a nearby park) on Sunday morning. Participant A was a female, 26-year-old student, height 5’8’’, Player Type–Based Game Features weight 61 kg, BMI 20.5 kg/m (normal weight), no serious health issues, and currently taking no medications. Player type: The player type of achiever was assigned the game feature of free spirit; fitness color: white. Our system detected that challenge. Thus, the system provided recommendations to player participant A takes the bus to university every Monday, Tuesday, B in the challenge style. and Thursday and mostly stays at home for the rest of the week. Overall Motivation and Satisfaction She goes to a group-cycling class once a week, on Friday Figure 5 shows the averages and SDs of the scores for the first evenings, for half an hour. According to the GRPAH, adults 3 general questions assessing participant motivation, satisfaction, aged 18 to 64 years were encouraged to perform 300 min of and game preference. The asterisk indicates significant results moderate-intensity aerobic physical activity throughout the found between groups. week for good health benefits [54]. People with exerciser type of white prefer hiking, running, yoga, cardio, and gym strength The results show that there were statistically significant training. When accessing her calendar, the system found she differences between groups as determined by one-way ANOVA had 2 dinner reservations on Thursday and Saturday night, both for overall motivation (F =22.49; P<.001), satisfaction 3,36 at 6 PM for the coming week. (F =22.12; P<.001), and preference (F =15.0; P<.001). 3,36 3,36 System-Generated Activity Recommendations Post hoc comparisons using the Tukey HSD test indicated that for all 3 questions, the mean score for the full, personalized, Extending the walking distance to bus stops on every school and gamified groups was significantly different from that for day (both morning and afternoon, overall 45 min of walking the control group, respectively. This means that, in general, per school day). both gamification and personalization have positive effects on A 30-min walk for non–school days after dinner. participants’ motivation, satisfaction, and preference, as seen A 1-hour home yoga session on Tuesday 7 PM when the in the groups full, personalized, and gamified compared with user is generally not active. the control group. Moreover, for motivation, the mean score A hiking morning on Saturday in a nearby park. for the full group (mean score for full group [MF] 5.8, SD for Player Type–Based Game Features full group [SDF] 0.79) was significantly different from that of The player type of free spirit was assigned the game feature of the personalized group (mean score for personalized group [MP] theme color customization. Thus, the reward of completing 4.7, SD for personalized group [SDP] 1.5). Statistically recommended activities for participant A was to unlock different significant pairwise comparisons are also marked in Figure 7. theme colors. This means that gamification can also add more motivation to a personalized fitness recommendation system, as seen between Case Study B the full group and the personalized group in motivation. It should Participant Information also be noted that the distribution of the dominant player types across the 4 different groups could have influenced these results, Participant B was a male, 35-year-old, software developer, 2 as some player types may have had a stronger preference for height 5’11’’, weight 75 kg, BMI 23.0 kg/m (normal weight), gamification or personalization in general. However, the no serious health issues, and currently taking an over-the-counter distribution with respect to player types did not seem to be pain reliever for his back pain. Player type: achiever; fitness particularly biased (Table 3). color: red. Our system detected that participant B drives to work every Monday, Tuesday, Thursday, and Friday (15-min drive). http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 15 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al Figure 7. Results for poststudy questions 1, 2, and 3. (a) Overall motivation, (b) overall satisfaction, and (c) overall preference over regular exercise. full group (MF 5.7, SDF 0.46) and the personalized group (MP IMI Subscales 4.8, SDP 0.72) as well as between the full group and the control Figure 8 shows the average and SDs of the scores for each IMI group (mean score for control group [MC] 5.0, SD for control subscale question. The results show that there were statistically group [SDC] 0.54). For effort or importance, significant significant differences between groups as determined by a differences were found between the full group (MF 5.8, SDF one-way ANOVA for interest or enjoyment (F =24.24; 0.47) and the gamified group (mean score for gamified group 3,36 P<.001), perceived competence (F =4.60; P=.007), effort or [MG] 4.7, SD for gamified group [SDG] 0.67); between the full 3,36 group and the control group (MC 4.7, SDC 0.70), the importance (F =8.01; P<.001), and value or usefulness 3,36 personalized group (MP 5.6, SDP 0.73), and the gamified group; (F =15.90; P<.001). 3,36 and between the personalized group and the control group. For The Tukey-Kramer HSD test results indicated that for interest value or usefulness, significant differences were also found or enjoyment, the mean score for the full, personalized, and between the full group (MF 5.8, SDF 0.63) and the gamified gamified groups was significantly different from that for the group (MG 4.8, SDG 0.55); between the full group and the control group. Moreover, the pairwise comparison result showed control group (MC 4.6, SDC 0.41), the personalized group (MP that MF (MF 5.9, SDF 0.40) was significantly different from 5.7, SDP 0.63), and the gamified group; and between the the personalized group (MP 5.0, SDP 0.56). For perceived personalized group and the control group. The pairwise competence, significant differences were found between the comparison significance is also marked in Figure 8. http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 16 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al Figure 8. Averages and SDs as evaluated by the Intrinsic Motivation Inventory. From top to bottom: (a) interest or enjoyment, (b) perceived competence, (c) effort or importance, (d) pressure or tension, and (e) value or usefulness. The IMI results indicate that gamifying the exercise increases EMIC Recommender System Evaluation players’ interest in and enjoyment of the personalized Figure 9 shows the averages and SDs of the scores for each recommendation system (significant between the full group and EMIC subscale (under perceived recommendation quality, the personalized group in interest or enjoyment). Personalization perceived system effectiveness, general trust in technology, and contributes more toward promoting effort or importance as well system-specific privacy concerns). The results showed that there as value or usefulness compared with gamification (significant were statistically significant differences between groups as between the personalized group and the gamified group). determined by a one-way ANOVA for perceived recommendation quality (F =108.77; P<.001), perceived 3,36 system effectiveness (F =26.52; P<.001), and system-specific 3,36 privacy concern (F =58.37; P<.001). 3,36 http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 17 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al Figure 9. Average and SD of European Microsoft Innovation Center recommendation. From top to bottom: (a) perceived recommendation quality, (b) perceived system effectiveness, (c) general trust in technology, and (d) system-specific privacy concern. The Tukey-Kramer HSD test results indicated that for both about privacy when the system had a player model and asked perceived recommendation quality and perceived system for more permissions to access their personal data (comparing effectiveness, the mean scores for the full and personalized the full or personalized and gamified or control groups). On the groups were significantly different from the gamified and control other hand, gamification reduced some of the concerns groups that were not personalized. For system-specific privacy (significant difference found between the full and the concerns, the mean scores for the full group and the personalized personalized groups). Note that for the system-specific privacy group were also significantly different from the gamified group question, a higher score indicates less concern. Privacy concerns and the control group because, for nonpersonalized groups, we are important yet beyond the scope of this work. Yet, we believe did not ask to access participants’ personal data (except Google that the noticed effect of gamification can be of value in future Analytics for in-app tracking). Moreover, a significant difference research and design. was also found between the full group (MF 4.1, SDF 0.69) and Daily Statistical Data the personalized group (MP 3.0, SDP 0.73). Statistically As mentioned earlier, we used Google Analytics API to track significant pairwise comparisons are also marked in Figure 9 participants’ comprehensive in-app behavior data and we used using asterisks. Google Fit API to track user daily activity data, including steps Our results suggest that our system is effective in providing and calories burned. Figure 10 shows some daily statistical data: daily fitness recommendations to participants (comparing the the number of active participants of all 4 groups during the 60 full group with the gamified group and the personalized group days of study (a), the daily total number of conversations sent with the control group) with respect to both perceived to the system (b), the daily average active calories burned recommendation quality and perceived system effectiveness. excluding basal metabolism (c), and the daily number of We also found that, as expected, participants were concerned self-reported useful recommendations (d). http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 18 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al Figure 10. Daily statistical data showing patterns of daily active users, conversations, or active calories and percentage of useful recommendations along the timeline. From Figure 8, we can see that for daily active users (a) and exclusive gamification may not. Adding gamified elements to daily conversations (b), there is an overall descent in trends personalized recommendations in the earlier phase (when the appearing as time grows for all 4 groups. Among them, the full player model was not well established yet and the group maintained a relatively higher value compared with the recommendation quality was not steady enough) may negatively other 3 groups and participants in the full, personalized, and affect the amount of physical activity people performed, which gamified groups interacted with the system more than the control requires further research. Note that the active calorie measures group (Figure 10). With respect to the daily active use and daily the calories burned during fitness activities. Basal metabolic conversations (Figure 10), when comparing the personalized parameters were excluded. group and the gamified group, we can see that the value of the For the percentage of useful recommendations, Figure 10 gamified group was higher than that of the personalized group (calculated by the daily number of useful replies divided by in the early phase of the study but was surpassed by the daily active users), the percentage of the full group and the personalized group in the late phase of the experiment (around personalized group increased in the first half of the study and 35-40 days). These results indicate that both personalization then remained flat, with the full group remaining slightly higher and gamification could have a positive impact on promoting than the personalized group. The increase in the full group and participants’ engagement with the system. However, although the personalized group can be attributed to the continuously gamification could bring more interactions in the short term updating player model that will improve recommendations over (within 1 month), personalization could lead to a more sustained time. The gamified group and the control group (without player engagement (over a longer time). Note that Figure 10 shows model) showed descending trends approaching zero. The results that the control group was not active during the last week of the suggest that our system is able to generate useful fitness study. This only indicates that they did not open the app but recommendations by using a player model, and participants they still received recommendations as usual (pop-up considered the recommendation more useful when gamification notifications). Physical activity data were also collected from elements were added. the Google Fit API without opening the app. Player Types For active calories (Figure 10), we can see slight ascent trends In this study, we did not find any significant difference in terms for both the full group and the personalized group and flat trends of different player or exerciser types. Although we had a limited for the gamified group and the control group. The full group sample size for conducting a meaningful statistical analysis, began with a lower average calorie burden compared with the there were still some interesting findings worth mentioning, personalized group and then showed an almost equal value near which may help inspire future research in this area. Table 4 the end. These results indicate that personalization could have shows the distribution of the combinations of player and a positive impact on promoting actual physical activity, whereas exerciser types of our participants. http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 19 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al From Table 4, we can see that certain player types and exerciser type of purples. This indicates that users’ preferences toward types were highly related. For example, we have 5 participants game elements and exercise types may be linked. This idea in total with the exerciser type of silver, with 4 of them could be used to further improve the personalization of the belonging to the player type of free spirit. Similar relations are exercise and game experience but requires further research with shown between the player type of socializer and the exerciser a larger sample size. Table 4. Distribution of the combination of player or exerciser type (N=40). The 8 colors Achiever Player Socializer Philanthropist Disruptor Free spirit Blue 3 3 0 2 0 0 Gold 1 1 1 2 0 0 White 2 1 0 1 0 2 Purple 0 0 2 0 0 0 Green 0 1 3 2 0 1 Red 2 0 0 1 1 0 Saffron 1 0 0 0 1 1 Silver 0 1 0 0 0 4 Figure 11 shows the overall motivation for participants experienced lower overall motivation compared with the other belonging to different player types. Although we were not able 4 player types. This may indicate that the game features and to run a valid statistical analysis based on the small sample sizes, experience we provided to the player type of socializers and we saw that the player type of socializer and disruptor disruptors had more room for improvement. Figure 11. Overall motivation for different player types. Figure 12 shows the average active calories burned for level as the other 6 exerciser types. When looking at the 8 Colors participants belonging to different exerciser types. For the same of Fitness activity suggestions (Multimedia Appendix 5), we reason of small sample sizes, we could not run a valid statistical found that the activity of hiking was the main variable that may analysis. However, we found that the fitness colors of whites lead to the result and it was only recommended for the exerciser and greens were relatively more active during the study. We type of greens and whites. This indicates that hiking might be checked their motivation as well as their self-reported an effective activity that makes people consume more calories, recommendation quality and found that both were at the same which could be further investigated. http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 20 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al Figure 12. Average active calories for different exerciser types. mainly focused on 5 aspects as shown in Figure 13. It shows Qualitative Results that a customized storyline was the most requested feature, One open-ended question was asked of each participant at the followed by multiplayer mode, more quality recommendations, end of the poststudy questionnaire to collect their general a feature for setting and tracking fitness goals, and more feedback (Multimedia Appendix 2). We received many location-based features. These feedback laid the foundation for comments and suggestions on how to improve our system, which planning our future work in this project. Figure 13. Number of main suggestions received from open-ended questions for improving our system. of player modeling and gamification could enhance users’ Discussion engagement with the system as well as promote actual physical activity. Specifically, gamification was found to promote Overall, in this 60-day user study, we verified our hypotheses engagement, but only in the short term, as seen in the gamified that (1) it is feasible to generate personalized exercise group where the members were engaged early on. However, as recommendations with player modeling and (2) the combination http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 21 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al the experiment moved on, the trend changed and the single types (eg, Bartle [68] and the BrainHex model [36]), personalized group became more engaged. This can be attributed more recently, researchers have examined the effectiveness of to the player modeling aspect in that it requires time to get to a trait-oriented models for understanding player choices in games minimum level of precision in reflecting a player’s [69-71]. Trait-oriented models are preferred in recent studies characteristics before it can offer reasonable recommendations. because an individual is rarely motivated by a single factor and Player modeling helped sustain the activity level in the long because of their applicability to game user research in that they term. This suggests that activity recommendation based on aim to characterize players using a set of scores rather than player modeling can be an effective and promising approach categorizing players into a single type. In this study, we decided for creating personalized fitness experiences over longer periods, to use the dominant player type as evaluated by the Hexad model whereas gamification can help attract the users and create the rather than considering all 6 scores because we wanted to control initial interest. the variable by adding only 1 additional element to each user; therefore, we could make sure it is the gamification itself that Our research was motivated by the need to keep players engaged affected the engagement, without interfering with the amount and motivated in exergames. We were inspired by previous of it. A future study can explore the effects of considering the work that suggested a more player-centric and personalized full range of scores. approach to game design and gamification [6,17,28,67] to increase player engagement and the overall effectiveness of the Adaptive and Continuous Modeling intervention. We extended these ideas to exergames, combining Although many games and other applications rely on a certain them with the notion of real-time activity tracking and user model, in most cases, this is done as a one-time static recommendation as suggested by others [22,47-49] to develop decision assigning the user to a certain group. Our study shows a new theoretical dynamic and individual-level model that brings the value of not only having a more comprehensive personal together various game elements that can help solve the player model but also allowing it to evolve and adapt using ongoing retention problem. The presented results have direct implications data from the user. This constantly tunes the model and makes for the design of fitness assistants and potentially other recommendations more effective. Using such adaptive and recommender systems. dynamic models can enhance the performance of such applications, and we recommend that designers consider it when Gamification Is Good but Not Enough! possible. Previous work by authors and other researchers has shown the potential value of gamification to increase engagement, but they 24/7 Recommendation have also highlighted the issue of retention. Players tend to Fitness and health are not limited to the gyms. Being active is leave the game once it is well experienced. Although adding a lifestyle; therefore, activity recommendations should not be new features can be a reasonable way of keeping participants limited to a particular time. In the absence of a dedicated engaged, it is difficult and costly to implement because of personal trainer, an intelligent fitness assistant equipped with constant designing and upgradation. The ability to understand a detailed player model can offer 24/7 recommendations for participants and their dynamic life and provide gameplay being active that considers various user contexts. Our results features that match the participants’ activities can be a way to show the potential value of this approach, which can be introduce change and novelty when maintaining the development improved with more comprehensive personal data and a better cost under control. database of activities and gameplay features. Although our system provided all-day and continuous modeling and Player Modeling: Personalization Versus recommendation, it is worth noting that the participants did not Categorization wear the activity trackers during sleep and we did not track any The idea of categorizing participants to provide them with sleeping activities. As such, although the system was able to customized service is appealing but ignores individual perform nonstop, in practice, it was paused during sleep times differences, which are often significant. The availability of (night or day). personal data, as a result of various methods of collecting information, suggests that the participants can be understood Limitations as individuals and not members of a category. This true There were certain limitations in the proposed system and the personalization allows a new level of customization that will performed study, some mentioned by the participants, which potentially offer participants a much more appealing and we believe were not critical enough to significantly affect the effective experience. Our results show the potential relevance findings but are still worth noting and improving in future work. of this idea to the field of fitness assistants. The more we We relied on a simple game that we designed ourselves with a understand the user, the more personalized our recommendations simple story or gameplay. This may have negatively affected will be, which will, in turn, result in more effective the players’ attraction and engagement. The game could be recommendations. Developing a comprehensive model that designed through a more rigorous process or we could somehow involves various user characteristics (from personality type to allow customization and choice or potentially use another daily routines) can help understand the user properly. existing game. There was also no multiplayer option, which Furthermore, the idea of personalization versus categorization ignores the social aspects of gaming and active lifestyle and is also related to differences in player types and player traits. could negatively affect the level of user engagement. When Although earlier works have attempted to classify players into designing different gamified features for different types of http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 22 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al players, we assigned only 1 game element to each type of player. to refer to the app instead of the more common term game, or This may not be adequate for targeting individual participants. for EMIC, we used items and activities for the same purpose. Although these terms could have caused some confusion, which The 8 Colors of Fitness system (Multimedia Appendix 5) [44] we will improve in the future, we did not receive any negative was used as a model to suggest activities. This system was used feedback and do not believe that the inconsistencies significantly because the research group did not find any other alternatives affected our findings. and needed to rely on a fairly acceptable method. This system is by no means ideal and has its own limitations. It can be Conclusions replaced with any other method, such as other models, an In this paper, we proposed a system for personalized fitness interactive trainer, or a trained expert system. assistants using gamification and continuous player modeling and reported on a long-term study that investigates the We used the Android Activity Recognition API for activity effectiveness of our proposed system. Our findings show that tracking and prediction in this work. This API is only able to it is possible to provide personalized activity recommendations recognize 6 simple physical activities. For more complex daily by continuously updating a player model based on activity activities, we required manual labeling from participants within tracking. Our study also shows the positive effect of this the conversation. This may bring complexity to the participants. modeling and gamification on user engagement and overall We also only used Android Wear participants and limited each activity. These findings can be used to inform the design of group to 10 members, which may not be adequate. We were personalized and gamified recommender systems in health and also aware that the age range of our participants was relatively fitness and potentially other apps, as they highlight the role of narrow. Most of our participants in this study were young adults; an adaptive model and gamification as long-term and short-term hence, our results may not apply to older adults. Furthermore, factors, respectively. This research opens opportunities for future comparing active calories burned as an absolute value could work, especially in the area of exploring more gameplay have negatively influenced the reliability of the results because features, adding a personalized storyline, multiplayer of potential confounding variables such as gender, weight, and gamification, better activity recognition, suggestion models, height. and evaluation with a larger and more diverse sample. The language of our questions could be improved by being more neutral and consistent. For example, we occasionally used task Conflicts of Interest None declared. Multimedia Appendix 1 Prestudy questionnaire for collecting general information. [DOCX File , 813 KB-Multimedia Appendix 1] Multimedia Appendix 2 Poststudy questionnaire for collecting general feedback. EMIC: European Microsoft Innovation Center; IMI: Intrinsic Motivation Inventory. [DOCX File , 13 KB-Multimedia Appendix 2] Multimedia Appendix 3 Intrinsic Motivation Inventory for evaluating the level of enjoyment related to the game experience. [DOCX File , 14 KB-Multimedia Appendix 3] Multimedia Appendix 4 European Microsoft Innovation Center recommender system evaluation measurement tool. [DOCX File , 14 KB-Multimedia Appendix 4] Multimedia Appendix 5 Eight Colors of Fitness activity suggestions. [DOCX File , 13 KB-Multimedia Appendix 5] References 1. Wilmot EG, Edwardson CL, Achana FA, Davies MJ, Gorely T, Gray LJ, et al. Sedentary time in adults and the association with diabetes, cardiovascular disease and death: systematic review and meta-analysis. Diabetologia 2012 Nov;55(11):2895-2905. [doi: 10.1007/s00125-012-2677-z] [Medline: 22890825] http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 23 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al 2. Hagen K, Chorianopoulos K, Wang AI, Jaccheri L, Weie A. Gameplay as Exercise. In: Proceedings of the 2016 CHI Conference Extended Abstracts on Human Factors in Computing Systems. 2016 Presented at: CHI EA'16; October 21-24, 2016; San Jose, CA. [doi: 10.1145/2851581.2892515] 3. Dharia S, Eirinaki M, Jain V, Patel J, Varlamis I, Vora J, et al. Social recommendations for personalized fitness assistance. Pers Ubiquit Comput 2017 Jun 29;22(2):245-257. [doi: 10.1007/s00779-017-1039-8] 4. Ananthanarayan A, Siek KA. Persuasive Wearable Technology Design for Health and Wellness. In: 6th International Conference on Pervasive Computing Technologies for Healthcare and Workshops. 2012 Presented at: PervasiveHealth'12; May 2-7, 2012; San Diego, CA. [doi: 10.4108/icst.pervasivehealth.2012.248694] 5. McGonigal J. Reality is broken: why games make us better and how they can change the world. Choice Rev Online 2012 Jul 1;49(11):49-6095. [doi: 10.5860/choice.49-6095] 6. Nacke LE, Deterding S. The maturing of gamification research. Comput Hum Behav 2017 Jun;71:450-454. [doi: 10.1016/j.chb.2016.11.062] 7. Prensky M. Digital game-based learning. Comput Entertain 2003 Oct;1(1):21. [doi: 10.1145/950566.950596] 8. Fogg BJ. Persuasive technologies. Commun ACM 1999;2(4):26-29. [doi: 10.1007/978-3-319-07127-5_11] 9. Papastergiou M. Exploring the potential of computer and video games for health and physical education: a literature review. Comput Educ 2009 Nov;53(3):603-622. [doi: 10.1016/j.compedu.2009.04.001] 10. Hamari J. Do badges increase user activity? A field experiment on the effects of gamification. Comput Hum Behav 2017 Jun;71:469-478. [doi: 10.1016/j.chb.2015.03.036] 11. Hamari J, Koivisto J, Sarsa H. Does Gamification Work? -- A Literature Review of Empirical Studies on Gamification. In: 47th Hawaii International Conference on System Sciences. 2014 Presented at: CSS'14; January 6-9, 2014; Waikoloa. [doi: 10.1109/HICSS.2014.377] 12. Baranowski T, Buday R, Thompson DI, Baranowski J. Playing for real: video games and stories for health-related behavior change. Am J Prev Med 2008 Jan;34(1):74-82. [doi: 10.1016/j.amepre.2007.09.027] [Medline: 18083454] 13. Consolvo S, Klasnja P, McDonald DW, Landay JA. Goal-setting Considerations for Persuasive Technologies That Encourage Physical Activity. In: Proceedings of the 4th International Conference on Persuasive Technology. 2009 Presented at: Persuasive'09; April 11-16, 2009; California, Claremont, USA. [doi: 10.1145/1541948.1541960] 14. Orji R, Mandryk RL, Vassileva J, Gerling KM. Tailoring Persuasive Health Games to Gamer Type. In: Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. 2013 Presented at: CHI '13; March 3-8, 2013; Paris, France. [doi: 10.1145/2470654.2481341] 15. Marquet O, Alberico C, Adlakha D, Hipp JA. Examining motivations to play pokémon go and their influence on perceived outcomes and physical activity. JMIR Serious Games 2017 Oct 24;5(4):e21. [doi: 10.2196/games.8048] [Medline: 29066423] 16. Orji R, Mandryk RL, Vassileva J. Improving the efficacy of games for change using personalization models. ACM Trans Comput-Hum Interact 2017 Nov 13;24(5):1-22. [doi: 10.1145/3119929] 17. Orji R, Tondello GF, Nacke LE. Personalizing Persuasive Strategies in Gameful Systems to Gamification User Types. In: Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems. 2018 Presented at: CHI'18; April 3-5, 2018; Montreal, Canada. [doi: 10.1145/3173574] 18. Charles D, McNeill M, McAlister M, Black M, Adrian M, Stringer K, et al. Player-Centred Game Design: Player Modelling and Adaptive Digital Games. In: Proceedings of DiGRA 2005 Conference: Changing Views – Worlds in Play. Digital Games Research Association. 2005 Presented at: DiGRA'05; March 4-8, 2005; Vancouver, Canada URL: https://www. researchgate.net/publication/228636408_Player-centred_game_design_Player_modelling_and_adaptive_digital_games 19. Yannakakis GN, Spronck P, Loiacono D, Andre E. Player Modeling. Dagstuhl Publishing. 2013. URL: http://yannakakis. net/wp-content/uploads/2013/08/pm_submitted_final.pdf [accessed 2020-11-10] 20. Lister C, West JH, Cannon B, Sax T, Brodegard D. Just a fad? Gamification in health and fitness apps. JMIR Serious Games 2014 Aug 4;2(2):e9 [FREE Full text] [doi: 10.2196/games.3413] [Medline: 25654660] 21. Campbell T, Ngo B, Fogarty J. Game Design Principles in Everyday Fitness Applications. In: Proceedings of the 2008 ACM Conference on Computer Supported Cooperative Work. 2008 Presented at: CSCW'08; March 1-6, 2008; San Diego, CA. [doi: 10.1145/1460563.1460603] 22. Rabbi M, Aung MH, Zhang M, Choudhury T. MyBehavior: Automatic Personalized Health Feedback From User Behaviors and Preferences Using Smartphones. In: Proceedings of the 2015 ACM International Joint Conference on Pervasive and Ubiquitous Computing. 2015 Presented at: UbiComp'15; April 1-4, 2015; Osaka, USA. [doi: 10.1145/2750858.2805840] 23. Johnson D, Deterding S, Kuhn K, Staneva A, Stoyanov S, Hides L. Gamification for health and wellbeing: a systematic review of the literature. Internet Interv 2016 Nov;6:89-106 [FREE Full text] [doi: 10.1016/j.invent.2016.10.002] [Medline: 30135818] 24. Mekler ED, Brühlmann F, Tuch AN, Opwis K. Towards understanding the effects of individual gamification elements on intrinsic motivation and performance. Comput Hum Behav 2017 Jun;71:525-534. [doi: 10.1016/j.chb.2015.08.048] 25. Loria E. A Framework to Infer Player Experience and Inform Customized Content Generation in Gameful Systems. In: Extended Abstracts: Extended Abstracts of the Annual Symposium on Computer-Human Interaction in Play Companion Extended Abstracts. 2019 Presented at: CHI PLAY'19; August 21-29, 2019; Barcelona, Spain. [doi: 10.1145/3341215.3356330] http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 24 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al 26. Böckle M, Micheel I, Bick M, Novak J. A Design Framework for Adaptive Gamification Applications. In: Hawaii International Conference on System Sciences. 2018 Presented at: ICSS'18; January 6-9, 2018; Hawaii, USA. [doi: 10.24251/hicss.2018.151] 27. Koivisto J, Hamari J. The rise of motivational information systems: a review of gamification research. Int J Inf Manag 2019 Apr;45:191-210. [doi: 10.1016/j.ijinfomgt.2018.10.013] 28. Bertran FA, Segura EM, Duval J, Isbister K. Chasing Play Potentials: Towards an Increasingly Situated and Emergent Approach to Everyday Play Design. In: Proceedings of the 2019 on Designing Interactive Systems Conference. 2019 Presented at: DIS'19; February 7-10, 2019; San Diego, CA. [doi: 10.1145/3322276.3322325] 29. Macvean A, Robertson J. Understanding Exergame Users' Physical Activity, Motivation and Behavior Over Time. In: Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. 2013 Presented at: CHI'13; March 3-9, 2013; Paris, France. [doi: 10.1145/2470654.2466163] 30. Busch M, Mattheiss E, Hochleitner W, Hochleitner C, Lankes M, Fröhlich P, et al. Using player type models for personalized game design – an empirical investigation. Interact Design 2016;12(44):145-163. [doi: 10.1145/2851581.2892399] 31. Velazquez A, Martínez-García AI, Favela J, Ochoa SF. Adaptive exergames to support active aging: an action research study. Pervasive Mob Comput 2017 Jan;34:60-78. [doi: 10.1016/j.pmcj.2016.09.002] 32. Yoo S, Heywood T, Tang LM, Kummerfeld B, Kay J. Towards a Long Term Model of Virtual Reality Exergame Exertion. In: Proceedings of the 25th Conference on User Modeling, Adaptation and Personalization. 2017 Presented at: UMAP'17; April 6-9, 2017; Bratislava, UK. [doi: 10.1145/3079628.3079679] 33. Farič N, Yorke E, Varnes L, Newby K, Potts HW, Smith L, et al. Younger adolescents' perceptions of physical activity, exergaming, and virtual reality: qualitative intervention development study. JMIR Serious Games 2019 Jun 17;7(2):e11960 [FREE Full text] [doi: 10.2196/11960] [Medline: 31210135] 34. Malone LA, Padalabalanarayanan S, McCroskey J, Thirumalai M. Assessment of active video gaming using adapted controllers by individuals with physical disabilities: a protocol. JMIR Res Protoc 2017 Jun 16;6(6):e116 [FREE Full text] [doi: 10.2196/resprot.7621] [Medline: 28623186] 35. Wiemeyer J, Deutsch J, Malone LA, Rowland JL, Swartz MC, Xiong J, et al. Recommendations for the optimal design of exergame interventions for persons with disabilities: challenges, best practices, and future research. Games Health J 2015 Feb;4(1):58-62 [FREE Full text] [doi: 10.1089/g4h.2014.0078] [Medline: 26181682] 36. Richard B. Hearts, clubs, diamonds, spades: players who suit MUDs. J MUD Res 1996 [FREE Full text] 37. Deterding S, Dixon D, Khaled R, Nacke L. From Game Design Elements to Gamefulness: Defining. In: Proceedings of the 15th International Academic MindTrek Conference: Envisioning Future Media Environments. 2011 Presented at: MindTrek'11; April 9-11, 2011; Tampere, UK. [doi: 10.1145/2181037.2181040] 38. Nacke LE, Bateman C, Mandryk RL. BrainHex: a neurobiological gamer typology survey. Entertain Comput 2014 Jan;5(1):55-62. [doi: 10.1016/j.entcom.2013.06.002] 39. Tondello GF, Wehbe RR, Diamond L, Busch M, Marczewski A, Nacke LE. The Gamification User Types Hexad Scale. In: Proceedings of the 2016 Annual Symposium on Computer-Human Interaction in Play. 2016 Presented at: CHI PLAY'16; March 2-8, 2016; Austin, UK. [doi: 10.1145/2967934.2968082] 40. Wiemeyer J, Nacke L, Moser C, Mueller FF. Player experience. In: Serious Games. New York, USA: Springer; 2016. 41. Cote J, Dreiss A, Clobert J. Social personality trait and fitness. Proc R Soc B 2008 Aug 26;275(1653):2851-2858. [doi: 10.1098/rspb.2008.0783] 42. Courneya KS, Hellsten LM. Personality correlates of exercise behavior, motives, barriers and preferences: an application of the five-factor model. Pers Individ Dif 1998 May;24(5):625-633. [doi: 10.1016/s0191-8869(97)00231-6] 43. Rhodes RE, Smith NE. Personality correlates of physical activity: a review and meta-analysis. Br J Sports Med 2006 Dec;40(12):958-965 [FREE Full text] [doi: 10.1136/bjsm.2006.028860] [Medline: 17124108] 44. Brue S. The 8 Colors of Fitness. Corte Madera, CA: Oakledge Press; 2008. 45. Myers IB, McCaulley MH, Most R. MBTI Manual: A Guide to the Development and Use of the Myers - Briggs Type Indicator. London, UK: Pyschologists Press; 1998. 46. Lara OD, Labrador MA. A survey on human activity recognition using wearable sensors. IEEE Commun Surv Tutorials 2013;15(3):1192-1209. [doi: 10.1109/surv.2012.110112.00192] 47. Guo X, Liu J, Chen Y. FitCoach: virtual fitness coach empowered by wearable mobile devices. In: IEEE Conference on Computer Communications. 2017 Presented at: INFOCOM'17; March 3-10, 2017; Atlanta, USA. [doi: 10.1109/infocom.2017.8057208] 48. He Q, Agu E, Strong D, Tulu B. RecFit: a Context-aware System for Recommending Physical Activities. In: Proceedings of the 1st Workshop on Mobile Medical Applications. 2014 Presented at: MMA'14; November 1-6, 2014; Memphis, USA. [doi: 10.1145/2676431.2676439] 49. Broekhuizen K, van Poppel MN, Koppes LL, Brug J, van Mechelen W. A tailored lifestyle intervention to reduce the cardiovascular disease risk of individuals with Familial Hypercholesterolemia (FH): design of the PRO-FIT randomised controlled trial. BMC Public Health 2010 Feb 15;10(1):69 [FREE Full text] [doi: 10.1186/1471-2458-10-69] [Medline: 20156339] http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 25 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al 50. Mittal R, Sinha V. A Personalized Time-bound Activity Recommendation System. In: 7th Annual Computing and Communication Workshop and Conference. 2017 Presented at: CCWC'17; April 4-9, 2017; Las Vegas, USA. [doi: 10.1109/ccwc.2017.7868485] 51. Ni J, Muhlstein L, McAuley J. Modeling Heart Rate and Activity Data for Personalized Fitness Recommendation. In: The World Wide Web Conference. 2019 Presented at: WWW'19; May 3-8, 2019; San Francisco, CA. [doi: 10.1145/3308558.3313643] 52. Ghanvatkar S, Kankanhalli A, Rajan V. User models for personalized physical activity interventions: scoping review. JMIR Mhealth Uhealth 2019 Jan 16;7(1):e11098. [doi: 10.2196/11098] [Medline: 30664474] 53. Zhao Z, Arya A, Whitehead A, Chan G, Etemad SA. Keeping Users Engaged through Feature Updates: A Long-Term Study of Using Wearable-Based Exergames. In: Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems. 2017 Presented at: CHI'17; April 3-9, 2017; Denver, CO. [doi: 10.1145/3025453.3025982] 54. Global Recommendations on Physical Activity for Health. Geneva, Switzerland: World Health Organization; 2010. 55. Activity Recognition Client. URL: https://developers.google.com/android/reference/com/google/android/gms/location/ ActivityRecognitionClient [accessed 2020-02-11] 56. Google Fit. URL: https://www.google.com/fit/ [accessed 2019-09-11] 57. Wit.ai. URL: https://wit.ai/ [accessed 2019-07-21] 58. Weather API. URL: http://weather.com [accessed 2019-08-01] 59. The Android Wear Forum. URL: https://forums.androidcentral.com/android-wear-wear-os/ [accessed 2018-10-23] 60. Gamified UK User Type Test. URL: https://gamified.uk/UserTypeTest2016/user-type-test.php#.X6XHa2T0lUM [accessed 2019-02-21] 61. The 8 Colors Quiz. URL: http://the8colors.com/quizw.php [accessed 2019-12-02] 62. Visual Studio App Center. Microsoft. URL: https://appcenter.ms/ 63. Google Analytics. URL: https://analytics.google.com/analytics/web/ [accessed 2018-01-13] 64. Deci EL, Ryan RM. Intrinsic Motivation Inventory (IMI). Scale Description. 2003. URL: https://selfdeterminationtheory. org/intrinsic-motivation-inventory/ [accessed 2020-11-10] 65. Knijnenburg BP, Willemsen MC, Gantner Z, Soncu H, Newell C. Explaining the user experience of recommender systems. User Model User-Adap Inter 2012 Mar 10;22(4-5):441-504. [doi: 10.1007/s11257-011-9118-4] 66. Jaccard J, Becker MA, Wood G. Pairwise multiple comparison procedures: a review. Psychol Bull 1984:- epub ahead of print. [doi: 10.4135/9781412985178.n2] 67. Nicholson S. A RECIPE for meaningful gamification. In: Gamification in Education and Business. New York, USA: Springer; 2014. 68. Mueller FF, Edge D, Vetere F, Gibbs MR, Agamanolis S, Bongers B, et al. Designing Sports: a Framework for Exertion Games. In: Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. 2011 Presented at: CHI'11; April 7-9, 2011; Vancouver, Canada. [doi: 10.1145/1978942.1979330] 69. Fortes Tondello G, Valtchanov D, Reetz A, Wehbe RR, Orji R, Nacke LE. Towards a trait model of video game preferences. Int J Hum-Comput Interact 2018 Apr 19;34(8):732-748. [doi: 10.1080/10447318.2018.1461765] 70. Tondello GF, Arrambide K, Ribeiro G, Cen AJ, Nacke LE. 'I don’t fit into a single type': a trait model and scale of game playing preferences. In: Human-Computer Interaction. 2019 Presented at: INTERACT'19; June 2-7, 2019; Paphos, UK. [doi: 10.1007/978-3-030-29384-0_23] 71. Tondello GF, Nacke LE. Player Characteristics and Video Game Preferences. In: Proceedings of the Annual Symposium on Computer-Human Interaction in Play. 2019 Presented at: CHI PLAY'19; July 6-11, 2019; Barcelona, Spain. [doi: 10.1145/3311350.3347185] Abbreviations ANOVA: analysis of variance API: application programming interface EMIC: European Microsoft Innovation Center GRPAH: Global Recommendation on Physical Activity for Health HSD: honestly significant difference IMI: Intrinsic Motivation Inventory MBTI: Myers-Briggs–Type Indicator MC: mean score for control group MF: mean score for full group MG: mean score for gamified group MP: mean score for personalized group SDC: SD for control group SDF: SD for full group SDG: SD for gamified group http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 26 (page number not for citation purposes) XSL FO RenderX JMIR SERIOUS GAMES Zhao et al SDP: SD for personalized group UI: user interface VR: virtual reality Edited by G Eysenbach; submitted 07.05.20; peer-reviewed by K Blondon, E Loria; comments to author 08.08.20; revised version received 30.09.20; accepted 24.10.20; published 17.11.20 Please cite as: Zhao Z, Arya A, Orji R, Chan G Effects of a Personalized Fitness Recommender System Using Gamification and Continuous Player Modeling: System Design and Long-Term Validation Study JMIR Serious Games 2020;8(4):e19968 URL: http://games.jmir.org/2020/4/e19968/ doi: 10.2196/19968 PMID: 33200994 ©Zhao Zhao, Ali Arya, Rita Orji, Gerry Chan. Originally published in JMIR Serious Games (http://games.jmir.org), 17.11.2020. This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Serious Games, is properly cited. The complete bibliographic information, a link to the original publication on http://games.jmir.org, as well as this copyright and license information must be included. http://games.jmir.org/2020/4/e19968/ JMIR Serious Games 2020 | vol. 8 | iss. 4 | e19968 | p. 27 (page number not for citation purposes) XSL FO RenderX

Journal

JMIR Serious GamesJMIR Publications

Published: Nov 17, 2020

Keywords: persuasive communication; video games; mobile apps; wearable electronic devices; motivation; mobile phone

There are no references for this article.