Explainable multimodal imaging–based machine learning for cardiovascular risk stratification in early Parkinson’s disease

Aims: This study aimed to characterize convergent biochemical and cardiovascular imaging endophenotypes of atherosclerotic risk in early PD and to implement an interpretable machine-learning architecture integrating multimodal clinical and imaging parameters for precision cardiovascular risk profiling and cardiometabolic classification. Methods: A total of 125 early-stage idiopathic PD patients and age- and sex-matched controls were analyzed. Feature space was reduced using cross-validated Least Absolute Shrinkage and Selection Operator (LASSO) with penalty tuning and stability selection to mitigate multicollinearity. XGBoost, Random Forest, RBF-kernel Support Vector Machine, and Stochastic Gradient Boosting models were trained under a train–hold-out framework with Bayesian hyperparameter optimization. Model performance was assessed via bootstrapping, and interpretability was provided using SHapley Additive exPlanations (SHAP). Results: PD patients showed increased hypertension (54.4%; OR 3.2, p=0.007) and hypercholesterolemia (OR 2.8, p=0.01), with excess left ventricular systolic dysfunction (28.8% vs 0%, p