Evaluation of LSTM, GRU, and ANFIS Models for Ankle Angle and Ankle Moment Prediction Using Biomechanical Data

Accurate prediction of joint angles and moments is crucial for understanding human gait and developing assistive technologies, such as exoskeletons and prosthetics. This study compares the performance of Long Short-Term Memory (LSTM) networks, Gated Recurrent Units (GRU), and Adaptive Neuro-Fuzzy Inference Systems (ANFIS) for predicting ankle angle and ankle moment at walking speeds of 1 km/h, 2 km/h, and 3 km/h. A multimodal sensor fusion approach was employed, incorporating electromyography (EMG), Ground Reaction Forces (GRF), and knee angle data as model inputs to enhance predictive accuracy. The performance of the models were assessed using Root Mean Square Error (RMSE) and Coefficient of Determination (R2) metrics, under both subject-specific and Leave-One-Subject-Out (LOSO) cross-subject validation frameworks. The results demonstrate that, under subject-specific evaluation, ANFIS achieved superior accuracy for ankle angle prediction at higher walking speeds, outperforming LSTM and GRU by up to 10.79% and 7.08% in RMSE at 3km/h, respectively. Conversely, GRU provided better performance for ankle moment prediction under subject-specific evaluation, achieving up to 29.04% and 18.27% lower RMSE than ANFIS and LSTM at 1km/h, respectively, highlighting its robustness under higher gait variability. Under the LOSO design, GRU consistently demonstrated superior generalization capability across subjects and walking speeds, reducing RMSE by 13.5% over ANFIS at 1km/h and 2.84% at 3km/h for ankle angle prediction, and achieving up to 17.5% and 4.5% lower RMSE compared to ANFIS and LSTM for ankle moment prediction. The comparative results were further validated using paired t-tests on RMSE values, confirming statistically significant performance differences between models. These findings highlight the importance of selecting appropriate modeling architectures for biomechanical prediction tasks and identify GRU as a promising candidate for robust gait estimation applicable to assistive and rehabilitative technologies.