Diagnostic classification of mild cognitive impairment in Parkinson's disease using subject-level stratified machine-learning analysis

Background The timely identification of mild cognitive impairment (MCI) in Parkinson's disease (PD) is essential for early intervention and clinical management, yet it remains a challenge in practice. Methods We conducted an analysis of 3,154 clinical visits from 896 participants in the Parkinson's Progression Markers Initiative (PPMI) cohort. Participants were divided into two groups: cognitively normal (PD-NC, MoCA ≥ 26) and MCI (PD-MCI, 21 ≤ MoCA ≤ 25). To ensure no visit-level information leakage, subject-level stratified sampling was employed to split the data into training (70%) and hold-out test (30%) sets. From an initial set of 12 routinely assessed clinical features, seven were selected using least absolute shrinkage and selection operator (LASSO) logistic regression: age, sex, years of education, disease duration, UPDRS-I, UPDRS-III, and Geriatric Depression Scale (GDS). Four machine learning models—logistic regression (LR), support vector machine (SVM), random forest (RF), and extreme gradient boosting (XGBoost)—were trained using subject-level stratified 10-fold cross-validation with Bayesian optimization. Probabilistic outputs were dichotomized using three thresholding strategies: default 0.5, F1-score maximization, and Youden index maximization. Results On the independent test set, SVM achieved the highest overall performance with AUC-ROC of 0.7252 and AUC-PR of 0.5008. LR also performed competitively despite its simplicity. RF achieved the top performance in sensitivity, reaching 0.8150. Feature importance analysis consistently highlighted age, years of education, and disease duration as the most informative predictors for distinguishing PD-MCI. Additionally, more stringent site-level split validation yielded slightly decreased overall performance, with LR showing improved AUC-PR. Importantly, the core feature importance ranking remained largely consistent across validation strategies. Conclusion This study developed and validated robust machine learning models for PD-MCI classification using standard clinical assessments alone. Through subject-level or site-level stratified cross-validation combined with Bayesian optimization, we achieved rigorous model evaluation while minimizing overfitting risk. These findings demonstrate the potential for implementing data-driven, interpretable diagnostic tools to enhance early cognitive impairment screening in routine PD care.