A Novel Semi-supervised Model for Pre-impact Fall Detection with Limited Fall Data

Abstract

Due to the high prevalence and severe consequences of falls among older people, pre-impact fall detection algorithms which aim to detect falls before body-ground impacts are of critical importance. The success of existing algorithms based on wearable sensors relies heavily on numerous labeled data from both falls and activities of daily living (ADLs). However, fall data is much more difficult to acquire compared with ADLs, and moreover, labeling falling period apart from ADLs is laborious and time-consuming. To this end, we proposed a novel semi-supervised model for pre-impact fall detection, called Semi-PFD, which maximizes the utilization of easily obtainable ADL data and reduces the dependency on labeled fall data. This model was evaluated on two large-scale public fall datasets (KFall and SisFall) and compared with its supervised baseline under different ratios of fall data. Cross-validation results showed that Semi-PFD outperformed the supervised baseline across all conditions on both datasets. More importantly, Semi-PFD achieved considerable improvement of 2–4% F1-score when the ratios of fall data were low (≤23.1%). Remarkably, with only 76.9% of fall data, Semi-PFD achieved comparable and even higher F1-scores than the state-of-the-art benchmark models which utilized 100% of fall data. We further observed that the incorporation of unsupervised training in Semi-PFD mitigated the typical overconfidence problem associated with supervised training without affecting lead time. These findings underscore Semi-PFD's practical potential, reducing the burden of fall data collection and labeling while maintaining superior performance.