Predicting dementia in people with Parkinson's disease

Abstract

Parkinson's disease (PD) exhibits a variety of symptoms, with approximately 25% of patients experiencing mild cognitive impairment and 45% developing dementia within ten years of diagnosis. Predicting this progression and identifying its causes remains challenging. Our study utilizes machine learning and multimodal data from the UK Biobank to explore the predictability of Parkinson's dementia (PDD) post-diagnosis, further validated by data from the Parkinson's Progression Markers Initiative (PPMI) cohort. Using Shapley Additive Explanation (SHAP) and Bayesian Network structure learning, we analyzed interactions among genetic predisposition, comorbidities, lifestyle, and environmental factors. We concluded that genetic predisposition is the dominant factor, with significant influence from comorbidities. Additionally, we employed Mendelian randomization (MR) to establish potential causal links between hypertension, type 2 diabetes, and PDD, suggesting that managing blood pressure and glucose levels in Parkinson's patients may serve as a preventive strategy. This study identifies risk factors for PDD and proposes avenues for prevention.

Fig. 1. Boxplot of ROC scores across models.

Boxplot showing the models’ AUC via repeated nested cross-validation. The boxplots are displayed with a median line, the interquantile range (IQR) represented by the box borders, whiskers extending to 1.5 the IQR and outliers represented by hollow circles. Each black dot represents a repetition of the nested cross-validation of each model. ROC receiver operating characteristic, AUC area under the curve.

Fig. 2. Mean ROC curve with variability.

Mean ROC curve of the Random Forest model, derived from five-fold cross-validation. Individual fold ROC curves are shown as thin colored lines, and the bold blue line represents the mean ROC across folds. The shaded area denotes ±1 standard deviation, illustrating variability in performance. The diagonal dashed line indicates chance-level performance (AUC = 0.5). ROC receiver operating characteristic, AUC area under the curve.

Fig. 3. Boxplot of ROC scores on reduced subset of features.

Comparison of the Random Forest model trained on the reduced subset of features in PPMI and UK Biobank datasets. The boxplots are displayed with a median line, the interquantile range (IQR) represented by the box borders, whiskers extending to 1.5 the IQR and outliers represented by hollow circles. Each black dot represents a repetition of the nested cross-validation of each model. ROC receiver operating characteristic, AUC area under the curve.

Fig. 4. SHAP summary plot.

Beeswarm plot showing the contribution of the top features to the model’s output. Each point represents a single observation, with the horizontal axis indicating the SHAP value (impact on model output). Features are ranked by mean absolute SHAP value, from most to least important. Color represents the original feature value (red = high, blue = low), illustrating how feature magnitude influences model predictions. The bottom entry aggregates the combined impact of 44 additional features not shown individually.

Fig. 5. SHAP dependence plots.

SHAP dependence plots for the top contributing features in the model. Each subplot displays the SHAP value (impact on model output) on the vertical axis against the corresponding feature value on the horizontal axis. These plots illustrate both the direction and magnitude of each feature’s contribution to individual predictions. Non-linear relationships and feature interactions are observable in continuous variables such as PGS4281, Age at Recruitment, and Body Mass Index.

Fig. 6. SHAP contribution by feature group.

Cumulative SHAP value percentages grouped by data modality. The bar plot summarizes the relative contribution of each feature group to the model’s output, with SHAP values aggregated within each category and normalized to sum to 100%. This grouping facilitates comparison of the influence of broader data domains.