Machine Learning Approaches for Fall Detection Using Integrated Data from Multi-Brand Sensors

Abstract Falls are a major health concern across all age groups, leading to severe injuries and even death. Wearable sensor-based fall detection systems using accelerometers, gyroscopes, and magnetometers (inertial measurement units, IMUs) have emerged as a promising solution. Existing research primarily utilizes data from a single brand of IMU. This study addresses this limitation by proposing a multi-sensor data fusion approach for enhanced fall detection accuracy with Machine Learning. We present a novel approach that combines data from two different commercially available IMUs: Motion Trackers Wireless (MTW) and a custom-designed Activity Tracking Device (ATD). A hybrid dataset encompassing data from 44 volunteers was created, capturing both fall and daily activity information from sensors positioned on the waist. The data was organized in a time-series format to capture the sequential nature of fall events. Ten machine learning (ML) classifiers were trained and evaluated on unseen data using a data splitting method. The Extra Trees algorithm achieved the best performance on the hybrid dataset, with an accuracy of 99.54%, precision of 99.18%, recall of 99.79%, and F-score of 99.49%. This demonstrates the effectiveness of multi-sensor data fusion in creating a highly accurate fall detection system with minimal false alarms, utilizing data from various IMU brands. This study highlights the potential of combining data from different sensors to improve fall detection accuracy, paving the way for more robust and brand-agnostic fall detection systems with time series and ML based approach.