Multicenter study develops an XGBoost-based multimodal machine-learning model combining QSM radiomic features and MRS-derived neurochemical metabolites to classify early Parkinson's disease with high AUC and SHAP-based interpretability.

Offers promising diagnostic biomarkers and a transparent classification tool that could improve early detection and patient stratification for clinical studies, but provides limited direct mechanistic or therapeutic target information for drug discovery.

Parkinson's disease (PD) is a complex, progressive neurodegenerative disorder characterized by high heterogeneity and diagnostic challenges in its early stages. This study aimed to develop and validate a multimodal machine learning model for early PD detection by integrating neurochemical metabolites and QSM-based radiomic features from multicenter PD cohorts. Several model architectures, including Random Forest, Support Vector Machine, eXtreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine, were evaluated for comparing appropriate methods. The XGBoost model demonstrated superior predictive performance, achieving AUC values of 0.984 and 0.973 in the training and test cohorts, respectively. Feature importance analysis identified key diagnostic biomarkers using SHapley Additive exPlanations (SHAP) and enhanced model interpretability. This study can provide new insights into the neurobiological mechanisms underlying early PD.