LSTM-Based Gait Event Prediction Using Surface Electromyography

Abstract:
Gait event prediction plays a pivotal role in medical and rehabilitation domains, where surface electromyography (sEMG) provides more direct measurements of muscular activity compared to conventional kinematic data. This study utilizes the SIAT-LLMD dataset—containing synchronized sEMG, kinematic, and kinetic recordings from 40 healthy adults during normal walking—to develop a long short-term memory (LSTM) neural network for predicting three fundamental gait events: heel-strike (HS), mid-stance (MS), and toe-off (TO). The proposed model architecture employs one LSTM layer complemented by batch normalization and dropout mechanisms to enhance temporal feature extraction robustness. Preprocessing procedures involved gait phase redefinition using hip flexion zero-crossing points and Z-score normalization of sEMG signals. Through comprehensive evaluation employing Hold Validation, 5-fold Cross-Validation, and Leave- One-Subject-Out Cross-Validation protocols, the proposed model demonstrated Notable performance metrics. Quantitative analysis revealed a mean accuracy exceeding 0.99 across all validation paradigms, complemented by remarkably precise temporal detection capabilities (MAE < 0.0006s) for event timing estimation. These robust validation outcomes underscore the model's dual strengths in both classification accuracy and temporal resolution. The achieved performance benchmarks represent significant improvements over existing solutions for movement segmentation (MS), with direct implications for enhancing clinical assessment tools in fall risk prediction and rehabilitation monitoring. Particularly noteworthy is the model's consistent performance across different validation methodologies, suggesting strong generalizability— a critical requirement for real-world clinical implementation.