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

Learn More →

Gait analysis for early Parkinson’s disease detection based on deep learning

Gait analysis for early Parkinson’s disease detection based on deep learning Current Directions in Biomedical Engineering 2019;5(1):9-12 Jyothsna Kondragunta*, Christian Wiede, Gangolf Hirtz Gait analysis for early Parkinson’s disease detection based on deep learning https://doi.org/10.1515/cdbme-2019-0003 Abstract: Better handling of neurological or 1 Introduction neurodegenerative disorders such as Parkinson’s Disease (PD) is only possible with an early identification of relevant Parkinson’s disease (PD) is considered to be the second most symptoms. Although the entire disease can’t be treated but the common age-related neurodegenerative disorder after effects of the disease can be delayed with proper care and Alzheimer’s disease. It is estimated that 7 to 10 million people treatment. Due to this fact, early identification of symptoms worldwide have PD [1]. Scientifically, it is well known that for the PD plays a key role. Recent studies state that gait abnormalities of gait are seen in people with PD [2]. This is abnormalities are clearly evident while performing dual due to the result from varying combinations of hypokinesia, cognitive tasks by people suffering with PD. Researches also rigidity as well as from the defects of posture and equilibrium proved that the early identification of the abnormal gaits leads that includes the characteristics of shuffling gait with small to the identification of PD in advance. Novel technologies steps and poverty of movements in the trunk. Due to this fact, provide many options for the identification and analysis of it was identified that the early identification of this abnormal human gait. These technologies can be broadly classified as gait leads to the identification of PD in advance. This leads our wearable and non-wearable technologies. As PD is more research to identify different possible gait analysis mechanism prominent in elderly people, wearable sensors may hinder the which can subsequently be used for the early identification of natural persons movement and is considered out of scope of PD. this paper. Non-wearable technologies especially Image In medical research, changes in gait reveal key information Processing (IP) approaches captures data of the person’s gait about the person’s quality of life [3]. There is a specific interest through optic sensors Existing IP approaches which perform when searching for reliable information on the evolutions of gait analysis is restricted with the parameters such as angle of different diseases: (a) neurological diseases such as multiple view, background and occlusions due to objects or due to own sclerosis or Parkinson’s; (b) systemic diseases such as body movements. Till date there exists no researcher in terms cardiopathies (in which gait is clearly affected); (c) alterations of analyzing gait through 3D pose estimation. As deep leaning in deambulation dynamic due to sequelae from stroke and (d) has proven efficient in 2D pose estimation, we propose an 3D diseases caused by ageing, which effect a large percentage of pose estimation along with proper dataset. This paper outlines the population. Accurate and reliable knowledge of gait the advantages and disadvantages of the state-of-the-art characteristics, monitoring and evaluating them at a given methods in application of gait analysis for early PD instance of time are key aspects. This enables early diagnosis identification. Furthermore, the importance of extracting the of diseases and their complications to guarantee the best gait parameters from 3D pose estimation using deep learning possible treatment. is outlined. With the advancements in technologies, many techniques are Keywords: Gait analysis, deep learning, Parkinson’s disease available for the identification and analysis of human gait that 3D pose estimation can be broadly classified as wearable and non-wearable technologies. Wearable technologies are considered to be out https://doi.org/10.1515/cdbme-2019-0003 of scope of this paper because they will obstruct the natural movement of the elderly person. Non wearable technolgy such as IP systems captures data of the subject’s gait through optic ______ *Corresponding author: Jyothsna Kondragunta: Faculty of sensors and take objective measurements of the different Electrical Engineering and Information Technology, Technische parameters through digital image processing. Universität Chemnitz, Chemnitz, Germany, In this paper, only IP are considered for reviewing. The recent jyothsna.kondragunta@etit.tu-chemnitz.de advancements in deep learning techniques make it applicable Christian Wiede, Gangolf Hirtz: Faculty of Electrical Engineering for many domains and present implementations who perform and Information Technology, Technische Universität Chemnitz, Chemnitz, Germany Open Access. © 2019 Jyothsna Kondragunta, Christian Wiede, Gangolf Hirtz, published by De Gruyter. This work is licensed under the C reative Commons Attribution 4.0 License. J. Kondranguta et. al., Gait analysis for early Parkinson’s disease detection based on deep learning — 10 pose estimation almost use only deep learning-based 3 Literature review approaches. In order to limit the review, only technqiues which use deep learning for pose estimation, gait analysis and their An early research from [12] and [13] gives an overview of the combination for early identification of PD were discussed. initial methods used for human pose estimation. More recent surveys from [14], [15] covers a more detailed and in-depth survey on vision-based approaches. They provide a good 2 Gait abnormalities and PD overview on the latest techniques related to 3D pose estimation. Especially [15] has covered the important aspects Studies show that changes in gait characteristics leads to gait such as data acquisition, feature representation, data reduction deficiency [4]. The first symptoms of neurological disorders and classification of gait parameters. In recent years, several are poor balance, a significantly slower pace with a stage other technologies are implemented using deep learning showing support on both feet [5]. Some patients also show gait techniques which are out of scope of [15]. But there doesn’t alternations such as shorter steps, lower free speed when exist a literature overview, where the combination of human walking and higher cadence than healthy persons. Several pose estimation for gait analysis is used and a review on most other symptoms can be identified for different types of health recent methodologies. This paper gives a detailed overview on disorders such as osteoporosis [6], multiple sclerosis [7] or PD. the uncovered topics. Gait disorders are commonly observed by people with PD and [16] a dual-source approach for 3D pose estimation from probably occur as a result of progressive loss of dopamine- monocular images was proposed. They used the MoCap producing cells in the substantia nigra compacta of the central dataset with 3D poses and other source of images with nervous system. The absence of dopamine in the basal ganglia annotated 2D poses as data sources. A combination of circuit ultimately results in the loss of gait automaticity. Convolutional Pose Machines (CPM) [17] for 2D pose Clinically, people with PD usually have the hallmark features estimation with 3D pose estimation is performed. 2D joint of slowness (bradykinesia) [8], cessation of movement estimation from a single raw RGB image and 3D pose (akinesia) [9] or freezing of gait. As the disease advances, reconstruction was used for an efficient 3D pose estimation. these gait disorders become more prominent, disabling the [18] used SIMPLify method as the base method and extended patients and severely limiting their quality of life [6]. The it in different directions. They fit a 3D human body model reduction of stride length is considered the most prominent based on 2D features detected in multiple view images. [19] feature of PD gait and is often accompanied by lower walking proposes a two-stage depth ranking based method speed and the tendency towards a longer duration in the double (DRPose3D) to improve the 3D pose estimation. The depth support phase [10]. ranking is used as an additional geometric feature which can Apart from spatio-temporal variables such as stride length, be identified by humans intuitively and contains rich 3D cadence or walking velocity, focusing on kinematic information. parameters as features for PD gait analysis is advantageous. [20] proposes an efficient and effective direct prediction based Kinematic parameters of gait can be pelvis tilt/rotation, hip on ConvNets. An incorporation of a parametric statistical body extension/rotation or knee extension/flexion. There are studies shape model (SMPL) within an end-to-end framework is the which used spatio-temporal [5], kinematic parameters [11] and key part. This allows to generate a very detailed 3D mesh, their combinations to identify the abnormal gait patterns in which results from 2D key points and masks. patients with PD. [2] indicates that the observation and In the above-mentioned review, none of them are involved in analysis is a key parameter in the early detection of PD. gait analysis from the estimated poses. A hierarchical Hence, an accurate and reliable knowledge of gait representation of the human body is proposed by [21] and characteristics for certain movements of a person are more showed that this sort of representation is well suited for human important. The early diagnosis of disease and its complications gait analysis. Hierarchical graphical models were developed enable the medical facilities to handle the situation by splitting the human body into parts. This may lead the appropriately in the near future. algorithms to identify the arms and legs similar as they share the same visual primitives. [22] proposed an incremental Gaussian Mixture Model with Hidden Markov Model (GMM-HMM) for 2D structure-based gait recognition in video. In order to make the pose estimation simple and efficient, the experiments were performed in fixed regions with specific gait rules. This algorithm cannot identify J. Kondragunta et al., Gait analysis for early Parkinson’s disease detection based on deep learning — 11 multiple peoples pose in the single video frame as they assume [6] mentioned that a quantitative analysis of gait parameters is only one person will be in frame. necessary for successful management of an individual patient Joint Gait Pose Manifold (JGPM) based Visual Gait with PD. Generative Model (VGGM) is proposed by [23]. Initially, A bigger study on the effects of quantitative parameters of gait JGPM was used to represent gait kinematics by coupling the using 218 healthy people and 168 PD patients was performed two nonlinear variables pose and gait in a unified latent space. by [27]. They used force plates for extracting gait parameters, Subsequently, the Gaussian Process Latent Variable Model their study highlights the tight coupling and importance of (GPLVM) for JPGM learning. abnormalities in gait to PD. DeepPose method that estimates human pose is applied in [24] Deep learning techniques were used for pose estimation to extract 2D joint locations of 18 different body parts. A algorithms for vision-based assessment of parkinsonism and cascade convolutional deep neural networks-based pose levodopa-induced dyskinesia (LID) [28]. predictor is used to increase precision of joint localization. Despite promising results, two major drawbacks are the use of two cameras for frontal and side view and the use of markers 4 Proposed approach on human body for pose estimation. This makes this approach only usable in a controlled indoor environment. The literature review shows that there exists research in the The initial review on the parkinsonian gait date backs to 1972 direction of gait analysis through pose estimation and gait [10]. Gait patterns of free speed walking in 21 parkinsonian analysis for early PD identification is necessary. But no patients was analysed by means of intermittent-light research in the direction of using pose estimation on elderly photography and filming. The speed of the forward subjects to extract gait information and identify the progression in these patients was reduced due to diminished possibilities or symptoms related to PD was performed. This stride length and increased cycle duration. Moreover, it is led us to combine these techniques together to develop a observed that the time ratio between swing and stance phases framework with which the patterns related to early detection was either increased or decreased. Their research is a base for PD must be extracted. further research in the direction of identifying the relation In the scope of our project, the data from elderly subjects between PD and human gait. starting from 80+ years is acquired. The data from the subjects [2] proposed a quantitative gait analysis in PD in comparison include cognitive dual tasks and regular gait data through with a healthy control group. This study did a detailed analysis RGB-D cameras. Deep learning algorithms such as on the impact and spatiotemporal, kinematics and kinetics of Convolutional Neural Networks (CNN) will be used for gait parameters in patients with PD. Consequently, this study estimating the 2D poses through obtained RGB images. These contributed in identifying which gait parameters help in 2D poses are projected into an 3D environment based on the therapy phase for the recovery of gait. Finally, they concluded obtained depth data by mapping the depth information with the that ankle plantar flexors are mostly affected in PD gait. estimated 2D poses. Later these estimated 3D poses are used A relationship between clinical features and freezing of gait to extract gait parameters The gait parameters such as stride with respect to the gait abnormalities was outlined in [25]. 30 length, cadence, step width, step angle, step length etc are patients with PD divided into two subgroups: (i) tremor- considered to be in differentiating the gait deviations. Based dominant (TD) group and postural instability and gait on the requirements, additional gait parameters can also be disturbance (PIGD) group (ii) freezing of gait and non- extracted from the estimated 3D pose. The data from the freezing of gait group were taken for these studies using a subjects are acquired and the gait parameters are extracted computerised video motion analysis system. The comparison periodically in 6 months interval. This will continue for 3 years between these subgroups results in the identification of and the extracted gait parameters are used to identify the gait significant reduction in walking velocity and stride length in abnormalities and the symptoms of PD. This research is PIGD group compared to the TD group. compared with the MoCA (Montreal Cognitive Assessment) A similar study from [26] found that dysfunctional kinematics data of the patients for evaluation. and abnormal kinetic parameters play an important role in the characterization of gait in PD patients off therapy, They proposed that these parameters can be used to document treatment effects of parkinsonian gait disorders. J. Kondragunta et al., Gait analysis for early Parkinson’s disease detection based on deep learning — 12 [8] A. Berardelli, “Pathophysiology of bradykinesia in 5 Conclusion and future work Parkinson’s disease,” Brain, vol. 124, no. 11, pp. 2131–2146, [9] V. Kaasinen, J. Joutsa, T. Noponen, and M. Päivärinta, In this paper, we first introduced the importance of PD “Akinetic crisis in parkinson’s disease is associated with a severe loss of striatal dopamine transporter function: A report identification followed by its relation to the abnormalities in of two cases,” Case Rep. Neurol., vol. 6, no. 3, pp. 275–280, gait of the subjects with PD. Subsequently, a discussion on the [10] E. Knutsson, “An Analysis of Parkinsonian Gait,” Brain, vol. relation between the efficient human pose estimation and 95, no. 3, pp. 475–486, 1972. extraction of gait parameters was carried out. A brief [11] A. P. Rocha, H. Choupina, J. M. Fernandes, M. J. Rosas, R. introduction to gait abnormalities and its relation to PD was Vaz, and J. P. S. Cunha, “Parkinson’s disease assessment based on gait analysis using an innovative RGB-D camera deliberated by mentioning the relevant literature. system,” 2014 36th Annu. Int. Conf. IEEE Eng. Med. Biol. An initial discussion was carried out on the most recent Soc. EMBC 2014, pp. 3126–3129, 2014. [12] A. Agarwal and W. Triggs, “3D Human Pose from Silhouettes technologies related to human pose estimation. Subsequently, by Relevance Vector Regression,” Int. Conf. Comput. Vis. the review on different gait analysis techniques and the effects Pattern Recognit. (CVPR ’04), pp. 882–888, 2004. of human pose to gait parameters assessment was discussed. [13] D. M. Gavrila, “The Visual Analysis of Human Movement: A Survey,” Comput. Vis. Image Underst., vol. 73, no. 1, pp. 82– The relevant literature related to the abnormalities in gait to 98, 1999. the PD was presented. This discussion involves several [14] N. Sarafianos, B. Boteanu, B. Ionescu, and I. A. Kakadiaris, “3D Human pose estimation: A review of the literature and technological implementations as well as clinical proofs. analysis of covariates,” Comput. Vis. Image Underst., vol. From the review, we found that this area of research still needs 152, pp. 1–20, 2016. a lot of improvement in estimation of the accurate 3D poses as [15] K. Soni and A. Singh, “A Survey on Human Gait Recognition Techniques,” IJSTE -International J. Sci. Technol. Eng., vol. well as quantitative gait parameter estimation. In future, we 2, no. 10, pp. 435–438, 2016. aim to estimate the maximum number of gait parameters such [16] U. Iqbal, A. Doering, H. Yasin, B. Krüger, A. Weber, and J. Gall, “A Dual-Source Approach for 3D Human Pose as cadence, step length, step width etc. efficiently and Estimation from a Single Image,” pp. 1–13. accurately from the estimated human pose, which can [17] D. Tome and C. Russell, “Lifting from the Deep : Convolutional 3D Pose Estimation from a Single Image.” consequently be used for the PD detection in advance. [18] Y. Huang et al., “Towards Accurate Marker-less Human Shape and Pose Estimation over Time,” 2017. [19] M. Wang, X. Chen, W. Liu, C. Qian, L. Lin, and L. Ma, “DRPose3D : Depth Ranking in 3D Human Pose Estimation.” References [20] G. Pavlakos, L. Zhu, X. Zhou, and K. Daniilidis, “Learning to Estimate 3D Human Pose and Shape from a Single Color [1] P. N. Today, “Parkinson’s Disease Statistics,” Parkinson’s Image.” News Today, 2017. [Online]. Available: [21] J. Spehr, S. Winkelbach, and F. M. Wahl, “Hierarchical pose https://parkinsonsnewstoday.com/parkinsons-disease- estimation for human gait analysis,” Comput. Methods statistics/. [Accessed: 22-Oct-2018]. Programs Biomed., vol. 106, no. 2, pp. 104–113, 2011. [2] O. Sofuwa, A. Nieuwboer, K. Desloovere, A. M. Willems, F. [22] R. Pu and Y. Wang, “2-D structure-based gait recognition in Chavret, and I. Jonkers, “Quantitative gait analysis in video using incremental GMM-HMM,” Lect. Notes Comput. Parkinson’s disease: Comparison with a healthy control Sci. (including Subser. Lect. Notes Artif. Intell. Lect. Notes group,” Arch. Phys. Med. Rehabil., vol. 86, no. 5, pp. 1007– Bioinformatics), vol. 9008, pp. 58–70, 2015. 1013, 2005. [23] X. Zhang, M. Ding, S. Member, G. Fan, and S. Member, [3] N. Giladi, F. B. Horak, and and J. M. Hausdorff, “Video-based Human Walking Estimation by Using Joint Gait “Classification of gait disturbances: distinguishing between and Pose Manifolds,” no. December, pp. 1–14, 2015. continuous and episodic changes,” PMC, vol. 22, no. 11, [24] H. Ismail, I. Radwan, and H. Suominen, “Gait Estimation and Analysis from Noisy Observations Observations,” 2018. [4] N. Messenger and P. Bowker, “The role of gait analysis in [25] S.-B. Koh, K.-W. Park, D.-H. Lee, J. Kim, and J.-S. Yoon, clinical medicine: a survey of UK centres.,” Eng. Med., vol. “Gait Analysis in Patients With Parkinson’s Disease: 16, no. 4, pp. 221–227, 1987. Relationship to Clinical Features and Freezing,” J. Mov. [5] A. Gaenslen and D. Berg, Early diagnosis of Parkinson’s Disord., vol. 1, no. 2, pp. 59–64, 2008. disease, vol. 90, no. C. Elsevier Inc., 2010. [26] M. Švehlík et al., “Gait Analysis in Patients With Parkinson ’ [6] P. H. Chen, R. L. Wang, D. J. Liou, and J. S. Shaw, “Gait s Disease Off,” YAPMR, vol. 90, no. 11, pp. 1880–1886, disorders in Parkinson’s disease: Assessment and management,” Int. J. Gerontol., vol. 7, no. 4, pp. 189–193, [27] S. Okuda et al., “Gait analysis of patients with Parkinson ’ s disease using a portable triaxial accelerometer,” vol. 4, no. i, [7] A. Muro-de-la-Herran, B. García-Zapirain, and A. Méndez- pp. 93–97, 2016. Zorrilla, “Gait analysis methods: An overview of wearable and [28] M. H. Li, T. A. Mestre, S. H. Fox, and B. Taati, “Vision - Based non-wearable systems, highlighting clinical applications,” Assessment of Parkinsonism and Levodopa - Induced Sensors (Switzerland), vol. 14, no. 2, pp. 3362–3394, 2014. Dyskinesia with Deep Learning Pose Estimation,” pp. 1–8. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Current Directions in Biomedical Engineering de Gruyter

Gait analysis for early Parkinson’s disease detection based on deep learning

Download PDF
4 pages

Loading next page...
 
/lp/de-gruyter/gait-analysis-for-early-parkinson-s-disease-detection-based-on-deep-8ppp0Lp9CU
Publisher
de Gruyter
Copyright
© 2019 by Walter de Gruyter Berlin/Boston
eISSN
2364-5504
DOI
10.1515/cdbme-2019-0003
Publisher site
See Article on Publisher Site

Abstract

Current Directions in Biomedical Engineering 2019;5(1):9-12 Jyothsna Kondragunta*, Christian Wiede, Gangolf Hirtz Gait analysis for early Parkinson’s disease detection based on deep learning https://doi.org/10.1515/cdbme-2019-0003 Abstract: Better handling of neurological or 1 Introduction neurodegenerative disorders such as Parkinson’s Disease (PD) is only possible with an early identification of relevant Parkinson’s disease (PD) is considered to be the second most symptoms. Although the entire disease can’t be treated but the common age-related neurodegenerative disorder after effects of the disease can be delayed with proper care and Alzheimer’s disease. It is estimated that 7 to 10 million people treatment. Due to this fact, early identification of symptoms worldwide have PD [1]. Scientifically, it is well known that for the PD plays a key role. Recent studies state that gait abnormalities of gait are seen in people with PD [2]. This is abnormalities are clearly evident while performing dual due to the result from varying combinations of hypokinesia, cognitive tasks by people suffering with PD. Researches also rigidity as well as from the defects of posture and equilibrium proved that the early identification of the abnormal gaits leads that includes the characteristics of shuffling gait with small to the identification of PD in advance. Novel technologies steps and poverty of movements in the trunk. Due to this fact, provide many options for the identification and analysis of it was identified that the early identification of this abnormal human gait. These technologies can be broadly classified as gait leads to the identification of PD in advance. This leads our wearable and non-wearable technologies. As PD is more research to identify different possible gait analysis mechanism prominent in elderly people, wearable sensors may hinder the which can subsequently be used for the early identification of natural persons movement and is considered out of scope of PD. this paper. Non-wearable technologies especially Image In medical research, changes in gait reveal key information Processing (IP) approaches captures data of the person’s gait about the person’s quality of life [3]. There is a specific interest through optic sensors Existing IP approaches which perform when searching for reliable information on the evolutions of gait analysis is restricted with the parameters such as angle of different diseases: (a) neurological diseases such as multiple view, background and occlusions due to objects or due to own sclerosis or Parkinson’s; (b) systemic diseases such as body movements. Till date there exists no researcher in terms cardiopathies (in which gait is clearly affected); (c) alterations of analyzing gait through 3D pose estimation. As deep leaning in deambulation dynamic due to sequelae from stroke and (d) has proven efficient in 2D pose estimation, we propose an 3D diseases caused by ageing, which effect a large percentage of pose estimation along with proper dataset. This paper outlines the population. Accurate and reliable knowledge of gait the advantages and disadvantages of the state-of-the-art characteristics, monitoring and evaluating them at a given methods in application of gait analysis for early PD instance of time are key aspects. This enables early diagnosis identification. Furthermore, the importance of extracting the of diseases and their complications to guarantee the best gait parameters from 3D pose estimation using deep learning possible treatment. is outlined. With the advancements in technologies, many techniques are Keywords: Gait analysis, deep learning, Parkinson’s disease available for the identification and analysis of human gait that 3D pose estimation can be broadly classified as wearable and non-wearable technologies. Wearable technologies are considered to be out https://doi.org/10.1515/cdbme-2019-0003 of scope of this paper because they will obstruct the natural movement of the elderly person. Non wearable technolgy such as IP systems captures data of the subject’s gait through optic ______ *Corresponding author: Jyothsna Kondragunta: Faculty of sensors and take objective measurements of the different Electrical Engineering and Information Technology, Technische parameters through digital image processing. Universität Chemnitz, Chemnitz, Germany, In this paper, only IP are considered for reviewing. The recent jyothsna.kondragunta@etit.tu-chemnitz.de advancements in deep learning techniques make it applicable Christian Wiede, Gangolf Hirtz: Faculty of Electrical Engineering for many domains and present implementations who perform and Information Technology, Technische Universität Chemnitz, Chemnitz, Germany Open Access. © 2019 Jyothsna Kondragunta, Christian Wiede, Gangolf Hirtz, published by De Gruyter. This work is licensed under the C reative Commons Attribution 4.0 License. J. Kondranguta et. al., Gait analysis for early Parkinson’s disease detection based on deep learning — 10 pose estimation almost use only deep learning-based 3 Literature review approaches. In order to limit the review, only technqiues which use deep learning for pose estimation, gait analysis and their An early research from [12] and [13] gives an overview of the combination for early identification of PD were discussed. initial methods used for human pose estimation. More recent surveys from [14], [15] covers a more detailed and in-depth survey on vision-based approaches. They provide a good 2 Gait abnormalities and PD overview on the latest techniques related to 3D pose estimation. Especially [15] has covered the important aspects Studies show that changes in gait characteristics leads to gait such as data acquisition, feature representation, data reduction deficiency [4]. The first symptoms of neurological disorders and classification of gait parameters. In recent years, several are poor balance, a significantly slower pace with a stage other technologies are implemented using deep learning showing support on both feet [5]. Some patients also show gait techniques which are out of scope of [15]. But there doesn’t alternations such as shorter steps, lower free speed when exist a literature overview, where the combination of human walking and higher cadence than healthy persons. Several pose estimation for gait analysis is used and a review on most other symptoms can be identified for different types of health recent methodologies. This paper gives a detailed overview on disorders such as osteoporosis [6], multiple sclerosis [7] or PD. the uncovered topics. Gait disorders are commonly observed by people with PD and [16] a dual-source approach for 3D pose estimation from probably occur as a result of progressive loss of dopamine- monocular images was proposed. They used the MoCap producing cells in the substantia nigra compacta of the central dataset with 3D poses and other source of images with nervous system. The absence of dopamine in the basal ganglia annotated 2D poses as data sources. A combination of circuit ultimately results in the loss of gait automaticity. Convolutional Pose Machines (CPM) [17] for 2D pose Clinically, people with PD usually have the hallmark features estimation with 3D pose estimation is performed. 2D joint of slowness (bradykinesia) [8], cessation of movement estimation from a single raw RGB image and 3D pose (akinesia) [9] or freezing of gait. As the disease advances, reconstruction was used for an efficient 3D pose estimation. these gait disorders become more prominent, disabling the [18] used SIMPLify method as the base method and extended patients and severely limiting their quality of life [6]. The it in different directions. They fit a 3D human body model reduction of stride length is considered the most prominent based on 2D features detected in multiple view images. [19] feature of PD gait and is often accompanied by lower walking proposes a two-stage depth ranking based method speed and the tendency towards a longer duration in the double (DRPose3D) to improve the 3D pose estimation. The depth support phase [10]. ranking is used as an additional geometric feature which can Apart from spatio-temporal variables such as stride length, be identified by humans intuitively and contains rich 3D cadence or walking velocity, focusing on kinematic information. parameters as features for PD gait analysis is advantageous. [20] proposes an efficient and effective direct prediction based Kinematic parameters of gait can be pelvis tilt/rotation, hip on ConvNets. An incorporation of a parametric statistical body extension/rotation or knee extension/flexion. There are studies shape model (SMPL) within an end-to-end framework is the which used spatio-temporal [5], kinematic parameters [11] and key part. This allows to generate a very detailed 3D mesh, their combinations to identify the abnormal gait patterns in which results from 2D key points and masks. patients with PD. [2] indicates that the observation and In the above-mentioned review, none of them are involved in analysis is a key parameter in the early detection of PD. gait analysis from the estimated poses. A hierarchical Hence, an accurate and reliable knowledge of gait representation of the human body is proposed by [21] and characteristics for certain movements of a person are more showed that this sort of representation is well suited for human important. The early diagnosis of disease and its complications gait analysis. Hierarchical graphical models were developed enable the medical facilities to handle the situation by splitting the human body into parts. This may lead the appropriately in the near future. algorithms to identify the arms and legs similar as they share the same visual primitives. [22] proposed an incremental Gaussian Mixture Model with Hidden Markov Model (GMM-HMM) for 2D structure-based gait recognition in video. In order to make the pose estimation simple and efficient, the experiments were performed in fixed regions with specific gait rules. This algorithm cannot identify J. Kondragunta et al., Gait analysis for early Parkinson’s disease detection based on deep learning — 11 multiple peoples pose in the single video frame as they assume [6] mentioned that a quantitative analysis of gait parameters is only one person will be in frame. necessary for successful management of an individual patient Joint Gait Pose Manifold (JGPM) based Visual Gait with PD. Generative Model (VGGM) is proposed by [23]. Initially, A bigger study on the effects of quantitative parameters of gait JGPM was used to represent gait kinematics by coupling the using 218 healthy people and 168 PD patients was performed two nonlinear variables pose and gait in a unified latent space. by [27]. They used force plates for extracting gait parameters, Subsequently, the Gaussian Process Latent Variable Model their study highlights the tight coupling and importance of (GPLVM) for JPGM learning. abnormalities in gait to PD. DeepPose method that estimates human pose is applied in [24] Deep learning techniques were used for pose estimation to extract 2D joint locations of 18 different body parts. A algorithms for vision-based assessment of parkinsonism and cascade convolutional deep neural networks-based pose levodopa-induced dyskinesia (LID) [28]. predictor is used to increase precision of joint localization. Despite promising results, two major drawbacks are the use of two cameras for frontal and side view and the use of markers 4 Proposed approach on human body for pose estimation. This makes this approach only usable in a controlled indoor environment. The literature review shows that there exists research in the The initial review on the parkinsonian gait date backs to 1972 direction of gait analysis through pose estimation and gait [10]. Gait patterns of free speed walking in 21 parkinsonian analysis for early PD identification is necessary. But no patients was analysed by means of intermittent-light research in the direction of using pose estimation on elderly photography and filming. The speed of the forward subjects to extract gait information and identify the progression in these patients was reduced due to diminished possibilities or symptoms related to PD was performed. This stride length and increased cycle duration. Moreover, it is led us to combine these techniques together to develop a observed that the time ratio between swing and stance phases framework with which the patterns related to early detection was either increased or decreased. Their research is a base for PD must be extracted. further research in the direction of identifying the relation In the scope of our project, the data from elderly subjects between PD and human gait. starting from 80+ years is acquired. The data from the subjects [2] proposed a quantitative gait analysis in PD in comparison include cognitive dual tasks and regular gait data through with a healthy control group. This study did a detailed analysis RGB-D cameras. Deep learning algorithms such as on the impact and spatiotemporal, kinematics and kinetics of Convolutional Neural Networks (CNN) will be used for gait parameters in patients with PD. Consequently, this study estimating the 2D poses through obtained RGB images. These contributed in identifying which gait parameters help in 2D poses are projected into an 3D environment based on the therapy phase for the recovery of gait. Finally, they concluded obtained depth data by mapping the depth information with the that ankle plantar flexors are mostly affected in PD gait. estimated 2D poses. Later these estimated 3D poses are used A relationship between clinical features and freezing of gait to extract gait parameters The gait parameters such as stride with respect to the gait abnormalities was outlined in [25]. 30 length, cadence, step width, step angle, step length etc are patients with PD divided into two subgroups: (i) tremor- considered to be in differentiating the gait deviations. Based dominant (TD) group and postural instability and gait on the requirements, additional gait parameters can also be disturbance (PIGD) group (ii) freezing of gait and non- extracted from the estimated 3D pose. The data from the freezing of gait group were taken for these studies using a subjects are acquired and the gait parameters are extracted computerised video motion analysis system. The comparison periodically in 6 months interval. This will continue for 3 years between these subgroups results in the identification of and the extracted gait parameters are used to identify the gait significant reduction in walking velocity and stride length in abnormalities and the symptoms of PD. This research is PIGD group compared to the TD group. compared with the MoCA (Montreal Cognitive Assessment) A similar study from [26] found that dysfunctional kinematics data of the patients for evaluation. and abnormal kinetic parameters play an important role in the characterization of gait in PD patients off therapy, They proposed that these parameters can be used to document treatment effects of parkinsonian gait disorders. J. Kondragunta et al., Gait analysis for early Parkinson’s disease detection based on deep learning — 12 [8] A. Berardelli, “Pathophysiology of bradykinesia in 5 Conclusion and future work Parkinson’s disease,” Brain, vol. 124, no. 11, pp. 2131–2146, [9] V. Kaasinen, J. Joutsa, T. Noponen, and M. Päivärinta, In this paper, we first introduced the importance of PD “Akinetic crisis in parkinson’s disease is associated with a severe loss of striatal dopamine transporter function: A report identification followed by its relation to the abnormalities in of two cases,” Case Rep. Neurol., vol. 6, no. 3, pp. 275–280, gait of the subjects with PD. Subsequently, a discussion on the [10] E. Knutsson, “An Analysis of Parkinsonian Gait,” Brain, vol. relation between the efficient human pose estimation and 95, no. 3, pp. 475–486, 1972. extraction of gait parameters was carried out. A brief [11] A. P. Rocha, H. Choupina, J. M. Fernandes, M. J. Rosas, R. introduction to gait abnormalities and its relation to PD was Vaz, and J. P. S. Cunha, “Parkinson’s disease assessment based on gait analysis using an innovative RGB-D camera deliberated by mentioning the relevant literature. system,” 2014 36th Annu. Int. Conf. IEEE Eng. Med. Biol. An initial discussion was carried out on the most recent Soc. EMBC 2014, pp. 3126–3129, 2014. [12] A. Agarwal and W. Triggs, “3D Human Pose from Silhouettes technologies related to human pose estimation. Subsequently, by Relevance Vector Regression,” Int. Conf. Comput. Vis. the review on different gait analysis techniques and the effects Pattern Recognit. (CVPR ’04), pp. 882–888, 2004. of human pose to gait parameters assessment was discussed. [13] D. M. Gavrila, “The Visual Analysis of Human Movement: A Survey,” Comput. Vis. Image Underst., vol. 73, no. 1, pp. 82– The relevant literature related to the abnormalities in gait to 98, 1999. the PD was presented. This discussion involves several [14] N. Sarafianos, B. Boteanu, B. Ionescu, and I. A. Kakadiaris, “3D Human pose estimation: A review of the literature and technological implementations as well as clinical proofs. analysis of covariates,” Comput. Vis. Image Underst., vol. From the review, we found that this area of research still needs 152, pp. 1–20, 2016. a lot of improvement in estimation of the accurate 3D poses as [15] K. Soni and A. Singh, “A Survey on Human Gait Recognition Techniques,” IJSTE -International J. Sci. Technol. Eng., vol. well as quantitative gait parameter estimation. In future, we 2, no. 10, pp. 435–438, 2016. aim to estimate the maximum number of gait parameters such [16] U. Iqbal, A. Doering, H. Yasin, B. Krüger, A. Weber, and J. Gall, “A Dual-Source Approach for 3D Human Pose as cadence, step length, step width etc. efficiently and Estimation from a Single Image,” pp. 1–13. accurately from the estimated human pose, which can [17] D. Tome and C. Russell, “Lifting from the Deep : Convolutional 3D Pose Estimation from a Single Image.” consequently be used for the PD detection in advance. [18] Y. Huang et al., “Towards Accurate Marker-less Human Shape and Pose Estimation over Time,” 2017. [19] M. Wang, X. Chen, W. Liu, C. Qian, L. Lin, and L. Ma, “DRPose3D : Depth Ranking in 3D Human Pose Estimation.” References [20] G. Pavlakos, L. Zhu, X. Zhou, and K. Daniilidis, “Learning to Estimate 3D Human Pose and Shape from a Single Color [1] P. N. Today, “Parkinson’s Disease Statistics,” Parkinson’s Image.” News Today, 2017. [Online]. Available: [21] J. Spehr, S. Winkelbach, and F. M. Wahl, “Hierarchical pose https://parkinsonsnewstoday.com/parkinsons-disease- estimation for human gait analysis,” Comput. Methods statistics/. [Accessed: 22-Oct-2018]. Programs Biomed., vol. 106, no. 2, pp. 104–113, 2011. [2] O. Sofuwa, A. Nieuwboer, K. Desloovere, A. M. Willems, F. [22] R. Pu and Y. Wang, “2-D structure-based gait recognition in Chavret, and I. Jonkers, “Quantitative gait analysis in video using incremental GMM-HMM,” Lect. Notes Comput. Parkinson’s disease: Comparison with a healthy control Sci. (including Subser. Lect. Notes Artif. Intell. Lect. Notes group,” Arch. Phys. Med. Rehabil., vol. 86, no. 5, pp. 1007– Bioinformatics), vol. 9008, pp. 58–70, 2015. 1013, 2005. [23] X. Zhang, M. Ding, S. Member, G. Fan, and S. Member, [3] N. Giladi, F. B. Horak, and and J. M. Hausdorff, “Video-based Human Walking Estimation by Using Joint Gait “Classification of gait disturbances: distinguishing between and Pose Manifolds,” no. December, pp. 1–14, 2015. continuous and episodic changes,” PMC, vol. 22, no. 11, [24] H. Ismail, I. Radwan, and H. Suominen, “Gait Estimation and Analysis from Noisy Observations Observations,” 2018. [4] N. Messenger and P. Bowker, “The role of gait analysis in [25] S.-B. Koh, K.-W. Park, D.-H. Lee, J. Kim, and J.-S. Yoon, clinical medicine: a survey of UK centres.,” Eng. Med., vol. “Gait Analysis in Patients With Parkinson’s Disease: 16, no. 4, pp. 221–227, 1987. Relationship to Clinical Features and Freezing,” J. Mov. [5] A. Gaenslen and D. Berg, Early diagnosis of Parkinson’s Disord., vol. 1, no. 2, pp. 59–64, 2008. disease, vol. 90, no. C. Elsevier Inc., 2010. [26] M. Švehlík et al., “Gait Analysis in Patients With Parkinson ’ [6] P. H. Chen, R. L. Wang, D. J. Liou, and J. S. Shaw, “Gait s Disease Off,” YAPMR, vol. 90, no. 11, pp. 1880–1886, disorders in Parkinson’s disease: Assessment and management,” Int. J. Gerontol., vol. 7, no. 4, pp. 189–193, [27] S. Okuda et al., “Gait analysis of patients with Parkinson ’ s disease using a portable triaxial accelerometer,” vol. 4, no. i, [7] A. Muro-de-la-Herran, B. García-Zapirain, and A. Méndez- pp. 93–97, 2016. Zorrilla, “Gait analysis methods: An overview of wearable and [28] M. H. Li, T. A. Mestre, S. H. Fox, and B. Taati, “Vision - Based non-wearable systems, highlighting clinical applications,” Assessment of Parkinsonism and Levodopa - Induced Sensors (Switzerland), vol. 14, no. 2, pp. 3362–3394, 2014. Dyskinesia with Deep Learning Pose Estimation,” pp. 1–8.

Journal

Current Directions in Biomedical Engineering – de Gruyter

Published: Sep 1, 2019

Keywords: Gait analysis; deep learning; Parkinson’s disease 3D pose estimation

There are no references for this article.