Analysis of the effect of the regulating parameters in recurrent deep neural networks on the regression of finger joint angles

Accurate estimation of joint angles during limb movements plays a crucial role in the rehabilitation and diagnosis of neuromuscular and rheumatic disorders. This study aims to predict finger joint kinematics from surface electromyography (sEMG) signals using Long Short-Term Memory (LSTM) networks, which are well-suited for modeling temporal dependencies in physiological data. To enhance the model's generalization and reduce overfitting, four regularization strategies—LASSO, ridge, elastic net, and dropout—were systematically evaluated. Among these, LASSO and ridge regularization showed optimal performance when their coefficients were set to 0.0005, effectively balancing model complexity and prediction accuracy. While dropout was also beneficial, its performance declined at higher rates, with 0.2 identified as the most effective setting. The inclusion of appropriate regularization techniques led to a significant improvement in model accuracy, up to 20%, demonstrating their critical role in refining EMG-based kinematic estimation. The proposed LSTM model achieved a maximum prediction accuracy of 98% and an average of 96%, evaluated using the Pearson correlation coefficient. The results highlight the importance of selecting the appropriate regularization parameters to optimize both prediction accuracy and training speed in deep learning tasks designed to estimate joint angles.