Temporal Convolutional Network for Gait Event Detection.

In this study, we propose a novel deep learning-based framework for automatic gait event detection (GED) in diverse and complex walking scenarios, aiming to address the challenge of accurately identifying biomechanical markers linked to movement disorders. Current research methods in GED frequently encounter difficulties in maintaining high accuracy levels, especially when confronted with the intricate challenges of daily-life activities. Our proposed framework employs a Temporal Convolution Network (TCN) to address GED across diverse walking conditions and environments. As a post-processing step, we introduce a peak detection algorithm on the outputs of the TCN to accurately identify the gait events. The effectiveness of this framework is evaluated using a publicly available gait dataset, reporting event detection rates and time agreement through F1 score and Mean Absolute Error (MAE) for Heel Strike (HS) and Toe-Off (TO) events. The proposed methodology achieves a mean F1-score of 0.96 ± 0.07 and 0.92 ± 0.11 for HS and TO, respectively, indicating consistent performance. Similarly, the time agreement results in a mean MAE of 6.25 ms ± 3.67 ms and 16.87 ms ± 11.56 ms for HS and TO across all indoor and outdoor activities. These findings highlight the robustness of the proposed framework in GED under varying conditions, suggesting its potential for various pathological scenarios in daily-life activities.