Machine learning-based stratification of mild cognitive impairment in Parkinson’s disease: a multicenter cross-sectional analysis

TL;DR

This study creates a machine learning tool to identify Parkinson’s patients at risk of mild cognitive impairment using routine clinical data, helping doctors prioritize those needing detailed cognitive tests.

Contribution

A clinic-ready, externally validated machine learning model for PD-MCI risk stratification using MoCA-based labels and routinely collected variables.

Findings

Logistic regression showed balanced performance with AUCs of 0.789 (training), 0.778 (internal), and 0.772 (external).

Education and motor severity were the strongest predictors of PD-MCI risk, followed by sex and age at disease onset.

The tool prioritizes MoCA-normal patients for further neuropsychological evaluation and closer monitoring.

Abstract

Cognitive impairment is a prominent non-motor manifestation of Parkinson’s disease (PD) and is associated with reduced quality of life, increased mortality, and higher healthcare utilization. We aimed to develop and externally validate a machine-learning model, trained on Montreal Cognitive Assessment (MoCA)—based Movement Disorder Society (MDS) Level I labels, that estimates the contemporaneous probability of mild cognitive impairment in PD (PD-MCI) from routinely collected clinical variables, enabling clinicians to prioritize MoCA-normal patients with higher model-estimated probability for MDS Level II neuropsychological evaluation and closer follow-up.

We analyzed 799 participants with PD from the Parkinson’s Progression Markers Initiative (PPMI), randomly assigning them to training (n = 559) and internal validation (n = 240) cohorts. An independent external cohort comprised 70 consecutive patients recruited at The Affiliated Hospital of Guilin Medical University between February 2024 and March 2025. The reference outcome was MoCA-based PD-MCI (21–25) versus cognitively normal PD (26–30). Candidate predictors were screened by LASSO (1-SE criterion). To handle class imbalance, SMOTE was applied only during model fitting; both validation cohorts retained native class distributions. Five machine-learning models (logistic regression [LR], support vector machine, XGBoost, neural network, LightGBM) were evaluated on non-resampled data for discrimination (area under the receiver operating characteristic curve, AUC), calibration, and clinical utility (decision-curve analysis, DCA). Interpretability combined a nomogram with Shapley additive explanations (SHAP); a bilingual web calculator was also implemented.

Of 799 PPMI participants, 169 (21.2%) met the MoCA-based PD-MCI definition. Seven routinely collected predictors were retained (sex, age, education, age at disease onset, MDS-UPDRS Part III, GDS, UPSIT). LR showed the most balanced performance: AUC 0.789 (training), 0.778 (internal), and 0.772 (external). At a fixed threshold of 0.50 in the external cohort, LR’s sensitivity was 89.7%, specificity 43.9%, and F1-score 66.7%. Calibration and DCA favored LR. SHAP indicated education and motor severity as dominant contributors, followed by sex and age at onset; depressive burden (GDS) and hyposmia (UPSIT) increased risk, whereas chronological age had a smaller marginal effect.

We developed and externally validated a probability-based, clinic-ready risk-stratification tool for PD-MCI using routinely available variables and MoCA-based MDS Level I labels. Implemented as a nomogram and bilingual calculator, it supports sensitivity-oriented triage—especially among MoCA-normal patients—by prioritizing timely MDS Level II evaluation and closer follow-up. The tool complements, rather than replaces, formal diagnostic assessment and does not predict long-term conversion.

Not applicable. The PPMI study is registered with ClinicalTrials.gov (NCT01141023) and the registration date is June 8, 2010.

The online version contains supplementary material available at 10.1186/s12911-025-03215-0.

Figures