Edge-optimized machine learning models for real-time personalized health monitoring on wearables

Abstract

Personalized health monitoring, including Human Activity Recognition (HAR) and Fall Detection, is crucial for healthcare. Traditionally, most research in this field has relied on wearable sensors to collect data. The collected data is then typically sent to high-powered devices or servers for processing and analysis. However, there are some challenges with this approach. The reliance on high-powered devices can lead to delays in data processing and might not be suitable for real-time health monitoring. Additionally, the continuous transmission of data can raise privacy concerns and consume significant energy, which is not ideal for wearable devices that are often battery-powered, hence this study. Arduino UNO, based on the ATmega328P with 2 KiB SRAM, and ESP32 AI Thinker, with a dual-core Xtensa LX6 microprocessor, 320 KiB memory, and 3 MiB Flash Memory, are cost-effective and power-efficient, ideal for edge computing. In our research, we utilized the UCI HAPT and UMAFall Detection datasets for Human Activity Recognition and Fall Detection to optimize machine learning models for deployment on Arduino UNO and ESP32. On HAPT dataset, we achieved an impressive accuracy of up to 94% with a precision of 87% while on UMAFall Detection dataset, we achieved an accuracy of 81% with a precision of 77%. Notably, our trained Logistic Regression model for HAPT dataset clocked an average execution time of just 462 microseconds on ESP32 and 17634 micro seconds on Arduino UNO. Similarly, for UMAFall Detection dataset, our trained Decision Tree model clocked an average execution time of just 37 microseconds on ESP32 and 121 micro seconds on Arduino UNO. Furthermore, we significantly optimized resource usage for both HAPT and UMAFall datasets using our trained Decision Tree model, with memory usage minimized to 0.508 KB and 0.509 KB and flash size managed efficiently to a minimum of 6.606 KB and 26.518 KB on Arduino UNO, leaving plenty of resources for developers to add other programs on top of the ML model. It significantly outdid recent studies in terms of highly resource-constrained MCUs and compute resource usage efficiency.