Embedding Artificial Intelligence into Wearable IoMT Systems

Recent advances in sensors, wearable technologies, and real-time embedded Artificial Intelligence (AI) allows the creation of wearable systems for continuous health monitoring and real-time detection of critical health issues. In the case of heart arrhythmias, this currently requires data to be transmitted to the cloud for processing by computationally heavy models and human interaction. Instead, real-time analysis by AI models on the wearable device would reduce communication transmissions, improve energy usage, and provide immediate feedback to the user and physicians. We have shown that AI models trained using Big Data datasets on larger platforms can be optimized and then deployed as wearable sensors integrated into Internet of Medical Things (IoMT) systems. We used Artificial Intelligence models to detect abnormal heart beats trained using the MIT-BIH dataset. Our objective is to identify heartbeat abnormalities using a model optimized for embedded wearable sensors. The system would warn the user in the case of abnormal heart beats and upload all detected abnormal beats to the medical server for evaluation by a cardiologist. The system was 98.4% accurate in classification of normal/abnormal heartbeats and only missed approximately 0.13% arrhythmias per day with an execution time per heart beat classification of 65.6 ms and uses 0.911 microwatt-hours of energy. The program requires 409 KB of additional memory of ESP32 embedded controller. Utilizing the classification model and only transmitting heartbeats classified as abnormal improves battery usage by 85.5% over streaming heartbeats to cloud services for analysis assuming 10% abnormal classifications per day at an average of 100,000 heartbeats per day.