An explainable machine learning approach to aid MDS-UPDRS Part III diagnosis using kinematic data of Parkinsonian gait

The Movement Disorder Society-sponsored revision of the Unified Parkinson's Disease Rating Scale (MDS-UPDRS) Part III is widely used to examine the overall motor severity of patients with Parkinson's disease (PD). However, manual implementation of the MDS-UPDRS scheme for PD diagnosis demands substantial time and effort, which imposes physical and cognitive burdens on both patients and neurologists during PD examinations. To tackle this challenge, this study proposes an explainable machine learning approach to automate PD diagnosis by predicting the total MDS-UPDRS Part III score based on a patient's kinematic data during gait. First, MDS-UPDRS Part III examinations for 30 patients with early PD symptoms were performed by a neurologist. Next, gait experiments using a motion capture system were conducted to obtain the ankle, knee, and hip joint angles of the patients at each gait cycle. The collected joint angle data were transformed into joint angle matrices for gait cycles to represent gait pattern images in lower extremities. Then, a two-dimensional convolutional neural network (2D-CNN) model was trained using the gait pattern image to estimate the total MDS-UPDRS Part III score. In addition, the gradient-weighted class activation mapping (Grad-CAM) method was employed to extract plausible gait characteristics for the obtained MDS-UPDRS score. The proposed approach not only suitably predicts the total MDS-UPDRS Part III score but also reveals that significant features for predicting the MDS-UPDRS score are linked to both the single leg support phase and the swing phase of Parkinsonian gait.