Abstract
The risk of falling represents a significant barrier preventing many older adults from engaging in mass sports and physical activity. Objective assessment of the fall risk using wearable technologies constitutes an essential support in safe sport participation through early detection of gait instability. This study proposes a hybrid Temporal Convolutional Network (TCN)-based framework for gait-based fall risk identification based on lower back-mounted inertial measurement unit (IMU) sensor data acquired during one-minute laboratory tests and three-day free-living recording. The proposed methodology integrates data-driven temporal modeling of raw inertial signals with clinically interpretable handcrafted gait features. Temporal dependencies are modeled using dilated convolutions, while subject-level predictions are obtained through statistical aggregation of window-level representations. The framework is evaluated using subject-wise cross-validation and demonstrates consistent discrimination between fallers and non-fallers. The proposed methodology: ● Processes waveform-level gait dynamics and captures detailed gait dynamics from raw accelerometer signals from short accelerometer recordings using the TCN. ● Aggregates window-level embeddings at the subject level using statistical descriptors. ● Incorporates clinically interpretable gait features that describe spatiotemporal characteristics, inter-axis coordination, and time- and frequency-domain properties of walking.