Abstract
<title>Abstract</title> <p>Fallings are a critical public health challenge among aging populations, with approximately 28–35% of individuals aged 65 and older experiencing at least one fall annually. Existing fall detection systems face significant limitations: vision-based approaches raise privacy concerns and require high computational resources, while traditional wearable systems lack robustness and adaptability for real-world environments. Current deep learning methods, despite achieving notable accuracy, suffer from static decision boundaries that cannot adapt to individual movement patterns. This paper introduces a novel hybrid deep reinforcement learning methodology for wearable fall detection that uniquely integrates Temporal Convolutional Networks (TCN), Transformer encoders, and Soft Actor-Critic (SAC) within a unified framework. The proposed methodology captures multi-scale temporal features, models global contextual relationships, and learns adaptive decision policies through a carefully designed reward function. The complete training process requires approximately 9 minutes and 20 seconds on a standard CPU-based workstation without dedicated GPU acceleration, making the methodology accessible to researchers without high-end computational resources. The lightweight architecture comprises approximately 1.2 million trainable parameters, enabling deployment on resource-constrained edge devices. Extensive evaluation on the UP-FALL dataset demonstrates state-of-the-art performance with 99.0% accuracy, perfect precision (100%), 97.0% recall, and an F1-score of 99.0%, achieving a false positive rate of only 1.2% on challenging fall-like Activities of Daily Living (ADLs) compared to 3–8% for baseline methods. Cross-dataset generalization on the SISFALL dataset confirms the methodology's ability to generalize across diverse populations. Ablation studies validate the synergistic contribution of each component, with the complete framework significantly outperforming variants lacking any component. This methodology contributes the first unified integration of temporal convolutions, self-attention, and reinforcement learning for wearable sensor-based classification, providing a valuable blueprint for healthcare monitoring applications beyond fall detection. Despite limitations including evaluation on laboratory-collected data and sensor-specific optimization, the methodology establishes a new state-of-the-art approach for wearable sensor-based fall detection, successfully bridging the gap between high accuracy, operational reliability, and practical deployability.</p>