Figure 1.

A, Flowchart showing data collection and processing. A total of 134 PD patients with comprehensive neuropsychological evaluation were included in this study. Structural MRI was performed on each subject, and all subjects were randomly assigned to Group A (n = 101) or Group B (n = 33). Association analysis between morphological changes (including cortical thickness, subcortical structure, and white matter volume) and hypokinetic dysarthria (HD) was conducted using the general linear model (GLM) in Group A. B, Flowchart showing prediction of HD by machine learning. In Group A (training set), cortical thickness (in terms of vertex‐wise analysis or atlas) and volumes of white matter and subcortical structures were considered to be features and included in the feature‐based regions of interest (ROIs) and atlas models. All features in each method were ranked based on normalized absolute values of β obtained using the GLM. The mean square error (MSE) of the model established by the top i feature was calculated, and the feature set with minimum MSE was selected and used to establish the final model. The model was then used to predict the severity of HD in the test set (Group B). The performance of the machine learning was evaluated followed by permutation test