Unified Gait Event Detection using Temporal Convolutional Network and Bayesian Optimization

Accurate gait event detection (GED), which involves identifying key events such as heel strikes (HSs) and toe-offs (TOs) from inertial measurement unit (IMU) signals, is critical for quantifying gait abnormalities. However, most existing approaches either rely on heuristic methods tailored to specific activities or require separate deep learning models for different environments, limiting their generalizability and scalability. To address this research gap, we propose a unified deep learning framework based on a temporal convolutional network (TCN) for GED that operates across multiple activities and environments without the need for multiple models. Our approach incorporates: (1) Bayesian optimization for TCN hyperparameter tuning; (2) Gaussian kernel-based ground truth generation; and (3) a weighted loss function to emphasize challenging gait events. Specifically, our unified framework eliminates the need to train separate models for each activity, thereby reducing the computational cost. Evaluation using leave-one-out cross-validation across 20 subjects demonstrates that the proposed framework achieves overall F1 scores of 0.99 ± 0.00 for HS and 0.97 ± 0.06 for TO as well as competitive temporal precision as measured by mean absolute error (MAE). These results outperform existing state-of-the-art methods and underscore the potential of the unified framework for robust GED in diverse real-world scenarios, thereby advancing its application in clinical gait analysis, rehabilitation, and the design of assistive technologies.