Speed-Variable Gait Phase Estimation During Ambulation via Temporal Convolutional Network

Accurate estimation of gait phases during walking is a crucial prerequisite for both extracting clinically meaningful gait parameters and delivering gait-based feedback control information to rehabilitation devices. In addition, speed variation appears in our daily walking locomotion. However, most existing inertial measurement unit (IMU)-related methods based on heuristic algorithms were reported to be sensitive to walking speed changes. To address this problem, in this study, we presented a temporal convolutional network (TCN)-based approach for automatic and robust recognition of gait phases across multiscene ambulation and different walking speeds. We collected data on both real-world overground walking experiments and public treadmill datasets to validate the performance in terms of the accuracy and robustness of the proposed method. By comparing our method with six machine learning (ML) models and two neural network models, our method achieves 97% accuracy in gait phase estimation for both overground and treadmill walking, outperforming all compared benchmarks. It also excelled in both model generalizability evaluation and velocity robustness comparison tests over the other two neural networks. Notably, TCN can achieve 91% accuracy in velocity robustness tests and outperformed fully convolutional network (FCN) and long short-term memory (LSTM) in mean square error comparison ( ${p} < 0.05$ ). These results show that our method has outstanding estimation performance and high robustness on gait speed variations.