Machine Learning-based Pre-impact Fall Detection and Injury Prevention for the Elderly with Wearable Inertial Sensors

Abstract

Falls are the leading cause of death and non-fatal injuries among older people, thus pre-impact fall detection that detects a fall before body-ground impact is of crucial significance. 32 young subjects performed different types of falls and daily activities, and their motion data was recorded by a wearable inertial sensor to establish a large-scale motion dataset. Five commonly used machine learning algorithms were applied and compared thoroughly in terms of accuracy and practicality for pre-impact fall detection. Results showed that in terms of sensitivity, specificity and lead time, both support vector machine (SVM: 99.77%, 93.10%, 362 ± 150 ms) and random forest (RF: 100%, 92.90%, 357 ± 145 ms) achieved better results than other 3 models. SVM showed a much shorter latency (66 ms) than RF (1047 ms) running in a microcontroller. Those findings suggest that SVM has the highest potential to be embedded into a wearable sensor based system to provide real-time fall protection for the elderly. © 2021, The Author(s), under exclusive license to Springer Nature Switzerland AG.