Long short-term attention memory (LSTAM): a global-feature-integrated model for joint moment prediction in human rehabilitation

Joint moments are critical parameters for evaluating human movement, and time-series models are widely used to predict them from biosignals. However, biosignals collected via accelerometers, gyroscopes, and electromyography (EMG) sensors are often susceptible to local features such as short-term fluctuations and noise, which hinders the models' ability to effectively capture global features and weakens their capability to predict long-term trends. To address this issue, this paper proposes a long short-term attention memory (LSTAM) model that integrates global features. Our main contributions include the use of fast Fourier transform for spectral decomposition, multilayer perceptrons for nonlinear transformation, and convolutional modules to suppress the impact of local features in the sensor data. Additionally, an LSTM network enhanced with attention mechanisms is incorporated to dynamically focus on key temporal and frequency-domain patterns. We evaluated the proposed model on a publicly available dataset and compared its performance with existing methods, including LSTM, TCN, Conv2D, TimeMixer, xPatch, FFN, and TranSEMG. Experimental results show that the LSTAM model achieved a variance accounted for (VAF) of 0.907 ± 0.022 for hip flexion-extension (FE) and 0.927 ± 0.026 for hip abduction-adduction (AA); a root mean square error (RMSE) of 8.04 ± 2.27 (FE) and 5.56 ± 2.01 (AA); and a coefficient of determination (R²) of 0.908 ± 0.029 (FE) and 0.922 ± 0.030 (AA). These results demonstrate that LSTAM significantly outperforms existing models, offering a robust and efficient solution for joint moment prediction and human rehabilitation evaluation.