Explainable classification of Parkinson’s disease with different motor subtypes by analyzing the synthetic MRI quantitative parameters of subcortical nuclei

Objectives

To explore differences in quantitative parameters of subcortical nuclei using synthetic MRI across different motor subtypes of Parkinson’s Disease (PD), and to develop an interpretable model for distinguishing PD subtypes.

Methods

A total of 102 PD patients, including 43 Tremor-Dominant (TD) subtype and 59 Postural Instability and Gait Difficulty (PIGD) subtype and 42 age- and gender-matched healthy controls (HCs) were included. T1, T2, Proton Density (PrD), and Myelin Content (MYC) were extracted from 16 subcortical nuclei. We used Least Absolute Shrinkage and Selection Operator (LASSO) regression to select features for Support Vector Machine (SVM) classification models and construct three classification models using one-vs-one strategy. Shapley analysis (SHAP) was used to explain the model.

Results

The PIGD subtype exhibited more extensive changes in SyMRI parameters in basal ganglia nuclei than the TD subtype, particularly in T1 and T2. AUC values in the training and validation sets were as follows: 0.897/0.818 (PIGD vs. HC), 0.847/0.787 (TD vs. HC), and 0.820/0.769 (PIGD vs. TD). SHAP analysis revealed that in the PIGD vs. HC comparison, the top three features were T2_R_putamen (positive association) and MYC_L_GPi and MYC_R_SN (both negatively associated). In the TD vs. HC comparison, the top three features were T2_R_putamen, T1_R_SN, and T1_L_STN (all positively associated). In the PIGD vs. TD comparison, PrD_R_GPe and T2_R_SN were positively correlated, while MYC_L_GPi and MYC_R_GPe were negatively correlated.

Conclusion

SyMRI effectively detect brain microdamage in PD and distinguish between motor subtypes. Additionally, SHAP analysis identifies key predictive features for distinguishing these subtypes.