Comparison of deep and conventional machine learning methods in predicting joint moments in patients with cerebral palsy

The hip, knee, and ankle joint moments during gait are critical for clinical decision-making in patients with cerebral palsy (CP). These moments are typically calculated using inverse dynamics and human body models based on ground reaction forces (GRF). However, obtaining GRF data from CP patients can be challenging. Recent studies suggest joint moments in CP patients can be predicted using joint angles alone, bypassing the need for GRF, through machine learning (ML). However, the optimal type and scope of input data for such models remain unclear. This study aimed to identify the most feasible ML approach based on prediction accuracy for predicting joint moments from gait kinematics in CP patients. Retrospective gait data from 917 CP patients were analysed; after applying inclusion-exclusion criteria, data from 622 CP patients were used. We evaluated four conventional ML algorithms, ridge regression, k-nearest neighbors, random forest, and multilayer neural network, using feature-based input, and two deep learning algorithms, one-dimensional convolutional neural network and long short-term memory network, using raw data input, with each model’s hyper-parameters optimized in a problem-specific manner. Models were assessed using normalized root mean square error (nRMSE) and Pearson correlation coefficient (PCC). Deep learning models achieved an average nRMSE of 14.75 ± 7.10 and PCC of 0.95, while conventional ML models yielded 16.03 ± 6.55 and 0.94, indicating both conventional and deep learning approaches showed promise for predicting joint moments in patients with CP. By considering factors such as data availability and computational cost, an appropriate ML method can be selected to effectively address gait kinetics prediction in individuals with CP.