A Multi-Head Attention Transformer Model for Wearable in Situ Fall Detection.

Falls represent a critical health hazard for elderly individuals, often leading to severe injuries and decreased quality of life. While existing fall detection systems predominantly rely on threshold-based algorithms or conventional machine learning approaches like CNNs and RNNs, they frequently exhibit limitations in capturing the complex temporal dependencies and spatial features inherent in fall data. This paper presents a novel multi-head attention transformer architecture specifically designed for wearable in situ fall detection. Our approach leverages the transformer's self-attention mechanism to effectively model long-range dependencies in inertial measurement unit (IMU) sensor data collected from a shinbone-mounted device. The proposed architecture introduces several key innovations: 1) a position-aware embedding layer that preserves the temporal structure of IMU signals, 2) multi-head self-attention blocks that simultaneously attend to different signal features across varying time scales, and 3) a hybrid feature fusion module that combines global context with local patterns. Experimental results on our human subjects dataset demonstrate the critical importance of addressing dataset imbalance. A baseline model trained on imbalanced data achieved only