Access the full text.
Sign up today, get DeepDyve free for 14 days.
Current Directions in Biomedical Engineering 2019;5(1):29-32 Pierre Gembaczka*, Burkhard Heidemann, Bernhard Bennertz, Wolfgang Groeting, Thomas Norgall and Karsten Seidl Combination of sensor-embedded and secure server-distributed artificial intelligence for healthcare applications Abstract: The application of artificial intelligence (AI) in the 1 Introduction areas of health, care and social participation offers great opportunities but also involves great challenges. Extensive In modern technology, the use of artificial intelligence (AI) is regulatory, ethical and data-security related requirements becoming increasingly important. In modern smartphones, exist for data recording, storage and processing of respective facial recognition for authentication purposes is almost part personalized and patient-related data. “Artificial Intelligence of everyday life. Other state-of-the-art examples are voice as a Service” (AIaaS) is pushed for consumer applications by control for smart home applications and automated driving. global players, which implies data storage on external Many applications are also conceivable in medicine, ranging database server. However, the available solutions do not meet from AI-based disease diagnosis, through faster drug the requirements. Moreover, small and medium-sized development, to improved gene processing . The enterprises (SMEs) in the field of healthcare fear the loss of integration of AI in medical devices may enable a better early data sovereignty and information outflow. In this paper, we detection of diseases and anticipation of life-threatening propose a secure and resource-efficient approach by events such as a heart attack. embedding AI directly close to the sensor in combination An example is Apple's current smartwatch, which has an with secure and distributed data processing on local server or electrocardiogram (ECG) sensor and can detect atrial certified “Trusted Data Center”. For this purpose, we have fibrillation using an app and cloud-based AI. Apple and developed the Artificial Intelligence for Embedded Systems Stanford University carried out a study involving over (AIfES) platform-independent machine learning library in C 400,000 people. According to Apple, it was the largest study programming language. It contains a fully configurable deep of its kind in which only US users could participate [2–6]. artificial neural network with feedforward structure. The The integration of the AI model into the product is a library can be run directly on a microcontroller and even major hurdle. Most AI software solutions require a PC or at allows to train the neural network. Possible healthcare least edge hardware, such as a Raspberry PI or a applications include direct (pre-) processing of sensor data, smartphone . For embedded systems, such as sensor calibration, pattern recognition and classification. microcontrollers, integration is particularly difficult, as the Keywords: AI, embedded AI, neural network, artificial manufacturers usually develop individual solutions to integrate the AI models. In addition, the performance of intelligence, embedded systems, health avatar microcontrollers is usually not sufficient to calculate very https://doi.org/10.1515/cdbme-2019-0008 deep and large neural networks. A cloud-based solution is often used here, where powerful servers take over the calculation. The disadvantages are that a permanent internet connection is required and the data must be transferred to the ______ server. Hereby, data privacy is an important aspect that is *Corresponding author: Pierre Gembaczka Fraunhofer Institute already problematic in consumer electronics and even for Microelectronic Circuits and Systems (IMS), Duisburg, exacerbated in the field of healthcare as patient data must be Germany, e-mail: email@example.com Burkhard Heidemann, Bernhard Bennertz, Wolfgang transmitted. A system that processes the data offline or at Groeting: Fraunhofer Institute for Microelectronic Circuits and least pre-processes it in such a way that no raw data has to be Systems (IMS), Duisburg, Germany, transferred would be ideal. Thomas Norgall: Fraunhofer Institute for Integrated Circuits (IIS), Our approach is the use of neural networks directly on Erlangen, Germany, the embedded system close to the sensors which has the Karsten Seidl: Department of Electronic Components and advantage that no sensitive data can leave the device and Circuits, University of Duisburg-Essen, Duisburg, Germany Open Access. © 2019 Pierre Gembaczka et al., published by De Gruyter. This work is licensed under the Creative Commons Attribution 4.0 License. P. Gembaczka et al., Combination of sensor-embedded and secure server-distributed artificial intelligence for healthcare applications — 30 privacy can be fully guaranteed. In addition, real-time capability and reliability are relevant for many applications. When connected to a data cloud, transmission times can be critical and connection losses are possible. The Fraunhofer IMS has addressed the mentioned challenges and has developed AIfES (Artificial Intelligence for Embedded Systems), a platform-independent and ever- growing machine learning library in the C programming language, which exclusively uses standard libraries based on the GNU Compiler Collection (GCC). The application of AIfES in the field of healthcare is part of the German Federal Ministry for Economic Affairs and Energy (BMWi) funded project “Care[Ful]KI – Responsible AI Platform for Health, Care and Social Participation”. It Figure 1: Overview of the Artificial Intelligence for Embedded aims an integrated, on open standards based, legally certain Systems (AIfES) library and highly available AI data platform with competence and data pool . It shall empower small and medium-sized It is required that all algorithms are optimized for enterprises (SMEs) in the field of healthcare to implement hardware with “limited resources”. The AIfES functions innovative products and services applying anonymized or work exclusively on the space provided when called and pseudo-anonymized healthcare data. declare only the most necessary variables within a function. This paper gives an overview about the AIfES library for This means that the programmer can create the memory area healthcare applications and outlines a vision of how future for the training data and the weights themselves and only has development of sensor-embedded and secure server- to pass this to the AIfES functions. distributed AI-based medical devices can look like. By optimizing the algorithms even the training of small networks is possible on the microcontroller. This allows to implement e.g., self-calibrating sensors . 2 Artificial Intelligence for Efficient feature extraction is required in order to implement networks with a number of inputs as small as Embedded Systems (AIfES) possible. We have developed small and efficient neural networks using individual data pre-processing strategies. In In order to bring the AI in the form of Artificial Neural addition, distribution of computing tasks is possible, e.g., by Networks (ANN) close to the sensor, it has to be integrated small smart embedded systems taking over data pre- into embedded systems such as microcontrollers . The processing and making the results available to a higher-level AIfES Library offers a feedforward network configurable in system, e.g., on local server or certified “Trusted Data almost all parameters and also deep network structures [10 – Center”. This significantly reduces the amount of data to be 12]. Furthermore, AIfES can be compiled for almost any transferred. platform. For Windows, for example, a DLL (Dynamic Link We have developed a learning process especially for Library) is available that can be integrated into various tools small ANNs based on an evolutionary/genetic algorithm that (such as LabVIEW, MATLAB) or software development can autonomously find its way out of a local minimum. environments (Visual Studio, Python). Figure 1 shows an Practical experiments with sensor data showed that this overview of the AIfES Library. algorithm can have a clear advantage over the classical The optimal and smallest possible network structure has backpropagation method for small networks. It results in a to be determined. There are different approaches for this , lower total error in many experiments and moreover, the but testing different structures is always advantageous. Large parameterization effort was significantly lower. Whereas the network structures do not always show better results  and backpropagation method was sensitive to changes in the therefore optimization with regard to the use on an embedded learning rate, the evolutionary method could be used with a system is meaningful. Another possibility is the use of a constant configuration. This is important, for example, in genetic algorithm to determine a suitable network applications such as independent learning in the field, where structure . no manual adjustments can be made. P. Gembaczka et al., Combination of sensor-embedded and s ecure server-distributed artificial intelligence for healthcare applications 31 In addition to the pure software solution, individual capability can be achieved through data processing directly customer-specific ASICs (application-specific integrated on the system. circuit) can also be developed, where not only the sensors but also a ANN is directly integrated. For example, a RISC-V Author Statement microcontroller with hardware accelerator especially for Research funding: The work is part of the German Federal ANNs is under development, which can be directly integrated Ministry for Economic Affairs and Energy (BMWi) funded with the sensor circuitry. project “Care[Ful]KI – Responsible AI Platform for Health, Care and Social Participation”. 3 Implementation of the AI References model and Medical Approval  Data Revenue GmbH, “Künstliche Intelligenz in der Medizin - KI für Diagnostik, Medikamentenentwicklung, Previously approved medical devices with integrated AI Personalisierung von Behandlungen und Genome- Editierung,” Data Revenue GmbH, 18 07 2018. [Online]. are based on “locked algorithms”. Thus, a pre-trained model Available: is preferred for the intended application, which can be trained https://www.datarevenue.com/de/usecases/kuenstliche- on a powerful system and then brought to the target platform intelligenz-in-der-medizin. [Accessed 10 06 2019]. as a “locked algorithm”. For training purposes, the AIfES  L. Becker, “Früherkennung per Apple Watch: Vorhofflimmern DLL, e.g., can be integrated into MATLAB to perform the "erst der Anfang",” heise online, 18 03 2019. [Online]. Available: https://www.heise.de/mac-and- model development. i/meldung/Frueherkennung-per-Apple-Watch- After the successful training the ANN is ported to the Vorhofflimmern-erst-der-Anfang-4339343.html. [Accessed 10 embedded system. Therefore, AIfES is compiled as a library 06 2019]. for the target platform and integrated into the chosen  Stanford Medicine, “Apple Heart Study demonstrates ability of wearable technology to detect atrial fibrillation,” Stanford development environment. Parameters such as network Medicine, 16 03 2019. [Online]. Available: structure or selected activation functions are taken from the http://med.stanford.edu/news/all-news/2019/03/apple-heart- training results. The optimized weights from the training study-demonstrates-ability-of-wearable-technology.html. phase can be easily extracted and transferred to the target [Accessed 10 06 2019].  J. Clover, “Study Confirms Apple Watch Can Detect system. After porting the model runs on the target system. Abnormal Heart Rhythm With 97% Accuracy,” In the healthcare environment, many products with AI have MacRumors.com, LLC., 21 03 2018. [Online]. Available: already been approved (e.g. by the FDA), but many https://www.macrumors.com/2018/03/21/apple-watch- regulatory questions are still unresolved . Regarding abnormal-heart-rhythm/. [Accessed 10 06 2019]. these regulations, the FDA published a draft in April 2019  Stanford Medicine, “Apple Heart Study,” Stanford Medicine, 01 08 2018. [Online]. Available: . It describes that continuously learning systems pose a http://med.stanford.edu/appleheartstudy.html. [Accessed 10 particular challenge because the algorithm is constantly 06 2019]. changing.  Cardiogram, Inc., “Peer-Reviewed Research,” Cardiogram, Inc., [Online]. Available: https://cardiogram.com/research/. [Accessed 10 06 2019].  J. Schneider, “Edge AI Computing-KI wandert von der Cloud 4 Conclusion an die Edge,” Industry of Things, 10 08 2018. [Online]. Available: https://www.industry-of-things.de/ki-wandert-von- der-cloud-an-die-edge-a-741059/. [Accessed 10 06 2019]. Embedded AI is a promising approach to meet the  M. Unseld, “VDE entwickelt KI-Plattform für Gesundheit, special requirements of medical applications. With AIfES, Pflege und soziale Teilhabe,” Technologieverband VDE, 16 the Fraunhofer IMS has already achieved initial successes 05 2019. [Online]. Available: https://www.vde.com/de/presse/vde-entwickelt-ki-plattform- and the continuous development opens up new areas of careful-ki. [Accessed 10 06 2019]. application for the AI. Clear advantages can be achieved by  A. Hennig, P. Gembaczka, L. Cousin and A. Grabmaier, avoiding cloud-based data processing. For example, data “Smart Self-Sufficient Wireless Current Sensor,” in Smart protection is guaranteed because no data leaves the device as SysTech 2018; European Conference on Smart Objects, Systems and Technologies, Munich, 2018. it processes data offline. It is also possible to guarantee a  S. Singh, S. K. Pandey, U. Pawar and R. R. Janghel, significantly higher level of system stability, since such a “Classification of ECG Arrhythmia using Recurrent Neural system does not require an internet connection and thus Networks,” Procedia Computer Science, vol. 132, no. ISSN failures can be significantly reduced. In addition, real-time 1877-0509, pp. 1290-1297, 2018. P. Gmebaczka et al.,Combination of sensor-embedded and s ecure server-distributed artificial intelligence for healthcare applications — 32  S. Jiménez-Serrano, J. Yagüe-Mayans, E. Simarro- Computer Science and Engineering - Research in Computing Mondéjar, C. J. Calvo, F. Castells and J. Millet, “Atrial Science, vol. 27, pp. 15-24, 2007. Fibrillation Detection Using Feedforward Neural Networks  C. Johner, “Künstliche Intelligenz in der Medizin,” Johner and Automatically Extracted Signal Features,” Computing in Institut GmbH, 11 06 2019. [Online]. Available: Cardiology, no. 44, pp. 1-4, 2017. https://www.johner-institut.de/blog/regulatory-  S. Savalia and V. Emamian, “Cardiac Arrhythmia affairs/kuenstliche-intelligenz-in-der-medizin/. [Accessed 11 Classification by Multi-Layer Perceptron and Convolution 06 2019]. Neural Networks,” Bioengineering, vol. 5, no. 35, pp. 1-12,  F. U. F. &. Drug, “Proposed Reulatory Framework for 2018. Modifications to Artificial Intelligence/Machine Learning  D. Stathakis, “How many hidden layer and nodes?,” (AI/ML) - Based Software as a Medical Device (SaMD),” 02 International Journal of Remote Sensing, vol. 30, no. 8, pp. 04 2019. [Online]. Available: 2133-2147, 2009. https://www.regulations.gov/contentStreamer?documentId=F  T. Kavzoglu, “Determining Optimum Structure for Artificial DA-2019-N-1185- Neural Networks,” Proceeding of the stth Annual Technical 0001&attachmentNumber=1&contentType=pdf. [Accessed Conference and Exhibition of Remote Sensing Society, vol. 11 06 2019]. 25, pp. 675-682, 1999.  A. Fiszelew, P. Britos, A. Ochoa, H. Merlino, E. Fernández and R. García-Martínez, “Finding Optimal Neural Network Architecture Using Genetic Algorithms,” Advances in
Current Directions in Biomedical Engineering – de Gruyter
Published: Sep 1, 2019
Keywords: AI; embedded AI; neural network; artificial intelligence; embedded systems; health avatar
Access the full text.
Sign up today, get DeepDyve free for 14 days.