sEMG-Based Knee Angle Prediction: An Efficient Framework With XGBoost Feature Selection and Multiattention LSTM

Abstract

Accurate prediction of lower limb joint angles is essential for enabling natural human-exoskeleton interaction in rehabilitation robotics. This study proposes a novel framework for knee joint angle prediction using surface electromyography (sEMG) signals, integrating an XGBoost-driven feature selection algorithm and a multiattention hybrid-enhanced long short-term memory (LSTM) network. First, sEMG signals were acquired from healthy participants during dynamic lower limb movements. After preprocessing, temporal and spectral features were extracted, after which the eXtreme Gradient Boosting (XGBoost) algorithm was applied to eliminate redundant features, reducing input dimensionality while maintaining predictive accuracy. Finally, the reduced features were fed into the proposed model, which leverages hybrid attention mechanisms to enhance temporal dependencies and feature relevance. The experimental results validate that the XGBoost-driven feature selection framework significantly minimizes redundancy in sEMG feature extraction. When evaluating the performance of joint angle prediction, the mean absolute error (MAE), root mean square error (RMSE), adjusted ${R}^{{2}}$ , and Pearson correlation coefficient (CC) of the proposed model were 2.47°, 3.55°, 0.95, and 0.98, outperforming traditional machine learning (ML) algorithms and the benchmarks CNN, LSTM, TCN, and CNN-BiLSTM. The framework’s superior computational efficiency and prediction accuracy highlight its potential for real-time implementation in exoskeleton systems, addressing critical limitations in existing control paradigms. This advancement paves the way for adaptive human-robot collaboration in clinical rehabilitation settings.