Answers are grounded in this paper and cite their sources.
This study investigates the efficacy of deep recurrent neural networks, specifically Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) architectures, for fall detection utilizing time-series sensor data. The experimental evaluation is conducted on the publicly available SisFall dataset, encompassing a diverse range of activities of daily living (ADL) and fall scenarios. To mitigate the deleterious effects of class imbalance, we comprehensively evaluate multiple data balancing strategies, including a baseline (Normal) approach, Synthetic Minority Over-sampling Technique (SMOTE), Adaptive Synthetic Sampling (ADASYN), and KMeans-SMOTE. Empirical results demonstrate that oversampling can substantially affect classification performance on sequential sensor data, but its effectiveness depends on the specific sampling strategy and model architecture. Among the evaluated configurations, the GRU model trained with SMOTE achieved the best overall performance, showing improved accuracy and sensitivity to fall events
No takes yet. Share an insight, caveat, or question.
Van et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: