Continuous Wearable-based Fall Detection using Tiny Machine Learning

Falls among the elderly pose a significant threat to their health, causing physical trauma, psychological issues, and increasing healthcare costs. This paper presents a wearablebased fall detection system (WFDS) that utilizes Tiny Machine Learning (TinyML) techniques. The system employs wearable sensors that offer greater mobility, easier installation, and broader coverage than ambient sensors. The proposed system was designed to provide a practical and cost-effective fall detection solution. The system’s core is a fall detection algorithm employing TinyML to deploy machine learning models on resource-limited devices. The WFDS utilizes a 50 Hz 3 -axis IMU and a TinyML model to process and detect falls. Developed with the SisFall dataset and evaluated using KFall and self-collected datasets, the quantized model measures 9,623 bytes and detects falls with over $\mathbf{9 7. 8 \%}$ sensitivity for SisFall and $97.4 \%$ for selfcollected data. The usage of a TinyML fall detection system on an ESP32-S3 MCU in conjunction with an MPU6050 sensor allows for real-time, continuous monitoring with an inference time of 14 milliseconds per 2000 milliseconds of data. This demonstrates the significant potential of TinyML in wearable fall detection systems to mitigate the severe consequences of geriatric falls by enabling timely intervention and reducing fallrelated morbidity and mortality.