Estimation of Mutual Information between Pairs of Biomechanical Signals with Neural Networks

This research showcases the utilization of MINE (Mutual Information Neural Estimator) to assess mutual information between biomechanical signals, specifically focusing on the relationship between toe height and knee angles during the normal walking patterns of healthy individuals. Through simulated data trials, it was determined that the most effective network design consisted of three layers with 50 units each, using ReLU activation functions. Consequently, we developed an approach to examine the mutual information between toe height and knee joint angles, which resulted in two notable outcomes: consistent estimations and agreement with evaluated gait cycles. During the swing phase, there was a higher degree of mutual information between toe height and same-side knee angles as compared to contralateral knee angles, indicating a stronger predictive relationship between toe height and collateral knee angles. Conversely, in the stance phase, mutual information values were nearly zero regardless of knee angle side - an anticipated result since foot height remains constant while knee angles change. Evaluating the entire gait cycle revealed a pattern resembling that of the swing phase, though with differing mutual information values, suggesting an interdependence between the swing and stance phases. In conclusion, MINE proves to be reliable across various hyperparameter configurations. The proposed methodology presents a proficient and precise means for examining complex correlations within biomedical data, holding potential for customized rehabilitation plans and varied biomedical engineering applications.