A Novel Approach for Enhancing Joint Moment Prediction in Complex Gait Patterns Using Pretrained Weights in Time-Series Models

Accurately predicting joint moments is crucial for assessing human motor function; however, obtaining such data directly from individuals remains challenging. While gait data for common activities like walking and running can be easily acquired through accelerometers, gyroscopes, and electromyography (EMG) sensors, sensor data for less common or irregular gait patterns are relatively scarce. This scarcity limits the generalization ability of existing time-series models in predicting joint moments during complex gait patterns. To address this issue, this paper proposes a novel method to improve the performance of time series models in predicting joint moments during complex gait scenarios by utilizing pre-trained weights. These pre-trained weights are derived from gait data collected from a large sample of healthy individuals, capturing key information that represents various muscle activity characteristics. The pre-trained weights are then applied to time series models, significantly enhancing their prediction accuracy in complex gait conditions. Experimental results show notable improvements in metrics such as variance accounted for (VAF), root mean square error (RMSE), and the coefficient of determination (R²), confirming that the pre-trained weights improve the generalization capability of joint moment prediction models.