Multi-cohort machine learning identifies predictors of cognitive impairment in Parkinson's disease

Abstract

Cognitive impairment is a frequent complication of Parkinson's disease (PD), affecting up to half of newly diagnosed patients. To improve early detection and risk assessment, we developed machine learning models using clinical data from three independent PD cohorts, which are (LuxPARK, PPMI, ICEBERG). Models were trained to predict mild cognitive impairment (PD-MCI) and subjective cognitive decline (SCD) using Explainable Artificial Intelligence (XAI) for classification and time-to-event analysis. Multi-cohort models showed greater performance stability over single-cohort models, while retaining competitive average performance. Age at diagnosis and visuospatial ability were identified as key predictors. Significant sex differences observed highlight the importance of considering sex-specific factors in cognitive assessment. Men were more likely to report SCD. Our findings highlight the potential of multi-cohort machine learning for early identification and personalized management of cognitive decline in PD.

Fig. 1. SHAP value plot revealing key predictors’ influence on PD-MCI classification in the cross-cohort analysis.

Each row shows a predictor’s impact on mild cognitive impairment (PD-MCI) classification, with SHAP values indicating the direction and magnitude of effect. Points represent individual patients, with colors indicating the predictor’s value (red = high, blue = low). Positive SHAP values (right side) indicate increased likelihood of PD-MCI, while negative values (left side) suggest decreased likelihood. The Benton Judgment of Line Orientation score shows the strongest effect, with lower scores (blue) associated with increased PD-MCI risk. Age at PD diagnosis demonstrates the second strongest impact, with later onset (red) correlating with higher PD-MCI probability. Additional predictors include MDS-UPDRS subscores (Parts II, IV, and I) and weight, each showing varying degrees of influence on cognitive impairment classification.

Fig. 2. SHAP value plot revealing key predictors’ influence on SCD classification in the cross-cohort analysis.

Each row shows a predictor’s impact on subjective cognitive decline (SCD) classification, with SHAP values indicating the direction and magnitude of effect. Points represent individual patients, with colors indicating the predictor’s value (red = high, blue = low). Positive SHAP values (right side) indicate increased likelihood of SCD, while negative values (left side) suggest decreased likelihood. The MDS-UPDRS Part I score shows the strongest effect, with higher scores (red) associated with increased SCD risk. Age at PD diagnosis demonstrates the second strongest impact, with later onset (red) correlating with higher SCD probability.

Fig. 3. Forest plot of median conversion times (years) and hazard ratios for key predictors of time-to-PD-MCI in the cross-cohort analysis.

Forest plot showing the relationship between two key predictors (Benton Judgment of Line Orientation score and age at PD diagnosis) and the development of mild cognitive impairment (PD-MCI). The left panel displays median conversion times with 95% confidence intervals (CIs), stratified by predictor thresholds (≥16 vs <16 for Benton score; ≥53 vs <53 years for age at PD diagnosis). Solid blue lines indicate statistically significant differences between groups (p-value < 0.05), while grey lines indicate non-significant differences. The right panel shows corresponding hazard ratios (HR) with 95% CIs, where HR greater than 1 indicate increased risk of PD-MCI for the higher category compared to the reference group (lower category). For age at PD diagnosis, patients diagnosed at ≥53 years show significantly higher risk of developing PD-MCI compared to those diagnosed earlier.

Fig. 4. Forest plot of time-to-SCD predictors showing median conversion times (years) and hazard ratios in the cross-cohort analysis.

The plot illustrates how multiple clinical predictors affect subjective cognitive decline (SCD). The left panel shows median time to SCD onset with 95% confidence intervals (CIs), comparing subgroups for each predictor. Solid blue lines indicate statistically significant differences between groups (p-value < 0.05), while grey lines indicate non-significant differences. Key predictors include MDS-UPDRS Part I score (≥12 vs <12), postural abnormalities, tremor characteristics, and age at PD diagnosis (≥62 vs <62 years). The right panel displays corresponding hazard ratios (HR) with 95% CIs, where values > 1 indicate increased risk. Notable findings include a significantly higher SCD risk for patients diagnosed after age 62 and those with higher MDS-UPDRS Part I scores. Some features such as sleep problems and REM sleep behavior disorder show wider confidence intervals, suggesting more uncertainty in their predictive value.

Fig. 5. Machine learning pipeline for predicting cognitive impairment in Parkinson’s disease.

Machine learning analysis pipeline for predicting cognitive impairment in Parkinson’s Disease. Schematic representation of the data processing and analysis workflow. Input data from three independent cohorts (LuxPARK, PPMI, and ICEBERG) is pre-processed and then analyzed using both single-cohort and multi-cohort approaches. These analyses are applied to predict both mild cognitive impairment (PD-MCI) and subjective cognitive decline (SCD) outcomes in Parkinson’s disease. The models are evaluated using cross-validation, decision curve and calibration analyses.