Abstract
OBJECTIVE: Parkinson's disease (PD) diagnosis relies on the evaluation of motor and non-motor symptoms, with gait abnormalities serving as a key marker for early detection. Traditional clinical assessment often relies on visual gait analysis, which is a subjective process prone to bias. This study introduces a PD severity classification method that leverages gait features. METHODS: A Spatial-temporal Convolutional neural network-Transformer (ST-CNN-Transformer) model for PD severity classification was established. Multimodal gait data, including foot acceleration, angular velocity, and Vertical Ground Reaction Force (VGRF), were collected in collaboration with Xiangyang First People's Hospital, Hubei Province. Zero-velocity points (ZVPs) were detected using the Generalized Likelihood Ratio Test (GLRT), and gait cycle features were extracted from inertial measurement unit data for precise segmentation. The ST-CNN-Transformer model captures spatial-temporal features and periodic correlations. RESULTS: Evaluation on a dataset comprising 10 healthy controls and 30 PD patients yielded a classification accuracy of 98.81%, surpassing existing gait-based methods for PD severity classification. CONCLUSION: This study introduces a deep learning (DL) approach to automating PD severity classification by integrating ZVP and gait segmentation derived from multimodal data. The proposed model significantly enhances diagnostic accuracy. SIGNIFICANCE: By combining DL with GLRT-based gait segmentation and multimodal gait analysis, this study proposes a robust and interpretable PD severity assessment framework that contributes to more accurate and objective clinical decision-making.