AN ATTENTION-BASED HYBRID ENSEMBLE RECURRENT NETWORK FOR PRE-IMPACT FALL DETECTION

Falls are one of the leading causes of injury and death among older adults and often result in severe physical, emotional, and psychological consequences. Unlike conventional wearable fall detection systems that react only after a fall has occurred, prior fall detection offers a critical opportunity to intervene before ground impact and reduce the severity of injuries. However, existing approaches largely rely on a single recurrent neural network architecture and often evaluate performance using metrics that do not adequately reflect the challenges of detecting the highly imbalanced pre-fall phase. To address this challenge, this study proposes an attention-based hybrid ensemble model for prior fall detection. The model combines a lightweight convolutional neural network for feature extraction with LSTM, BiLSTM, and GRU networks that capture different temporal characteristics of human motion, while the attention mechanism focuses on the most critical signal segments for accurate prediction. The model is evaluated on the temporally annotated SisFall dataset, in which the pre-fall phase is approximately 108 times rarer than ordinary activity, using an imbalance-aware protocol reporting per-class precision, recall, and F1 alongside macro-F1 and the area under the precision–recall curve. A macro-F1 of 0.744 is obtained, exceeding three baselines retrained under an identical protocol, at 86,249 parameters and 98.5 ms per window. A damped class-weighting scheme reduces pre-fall false alarms by 52%; however, the threshold-independent AUPRC is shown to remain unchanged across all weighting and calibration settings, establishing that such corrections relocate the operating point along a fixed precision–recall curve without improving separability. Reliable window-level pre-impact detection, therefore remains an open problem, and the routine reporting of threshold-independent metrics is recommended.