TraxVBF: A hybrid transformer-xLSTM framework for EMG signal processing and assistive technology development in rehabilitation

Abstract

The fact that people with mobility impairments often have great difficulties in performing essential Activities of Daily Living (ADL) shows the importance of developing effective rehabilitation strategies. To address this need, we propose TraxVBF, a multimodal visual biofeedback submodel using surface Electromyography (sEMG) signals and kinematic movement data to exploit muscle synergy patterns. TraxVBF offers innovative real time visual feedback that can be used to enhance neurorehabilitation systems. Pre-processing and extracting muscle synergy patterns is performed by the Hierarchical Fast Alternating Least Squares (Fast-HALS) algorithm, and key movement points are identified with the Modified MediaPipe algorithm to capture temporal and spatial dynamics with precision using TraxVBF, which is driven by Extended Long Short-Term Memory (xLSTM) and Transformer architectures. This allows the model to predict movement trajectories accurately, enabling motor learning and functional recovery of patients through real time feedback without the expensive hardware. The model is shown to significantly improve performance metrics such as Mean Square Error (MSE), Root Mean Square Error (RMSE), and Coefficient of Determination (R²). For healthy participants, TraxVBF-type Base outperforms state of the art models (LSTM and GRU) with an MSE of 0.06 and R² of 0.89. Practical evaluations with an average R² of 0.880 for healthy participants and 0.327 for patients demonstrate the model generalizability. These results indicate that TraxVBF may be a useful tool to improve motor learning and rehabilitation, and longer term clinical trials and multi-sensory biofeedback are needed.