Eye-movement-based machine learning for identifying Parkinson's disease dementia in a retrospective cohort: an
Jing Wu1,2, Qing Liu2, Song Yang1,3
1Department of Neurology, The Affiliated Hospital of Guizhou Medical University, Guiyang, China.
Background:
Parkinson's disease dementia (PDD) is a major non-motor complication of Parkinson's disease (PD), but conventional cognitive screening is influenced by education, language, motor impairment, fatigue, and testing conditions. We developed and interpreted machine-learning models using eye-movement and demographic features to identify PDD among patients with PD.
Methods:
In this retrospective single-center study, 486 patients were screened between December 2024 and February 2025. After exclusions, 297 patients with idiopathic PD were analyzed, including 84 without dementia and 213 with PDD. Thirteen machine-learning algorithms were trained using demographic and task-derived eye-movement variables. Performance was assessed using ROC AUC, PR AUC, classification metrics, calibration, the Brier score, and decision-curve analysis. SHapley Additive exPlanations (SHAP) were used to interpret the final model.
Results:
Patients with PDD showed poorer cognitive performance and abnormalities in antisaccade inhibition, saccade execution, and fixation stability. Seven predictors were retained: age, sex, education level, antisaccade accuracy, mean overlap-saccade completion duration, mean gap-saccade velocity, and lateral-gaze deviation duration >4°. Logistic regression and elastic net achieved the highest test-set ROC AUCs (both 0.962), while logistic regression had the lowest Brier score (0.077). A radial basis function support vector machine (SVM) was retained as the primary classifier for detailed evaluation because it provided balanced performance across threshold-dependent metrics and could capture potentially nonlinear relationships and interactions. In the test set, the SVM achieved a ROC AUC of 0.947 and PR AUC of 0.979. Accuracy, sensitivity, specificity, and F1 score were 0.867, 0.870, 0.857, and 0.904, respectively. SHAP identified education, antisaccade accuracy, mean overlap-saccade completion duration, mean gap-saccade velocity, and lateral-gaze deviation duration >4° as the leading contributors.
Conclusion:
A seven-feature model combining demographic and eye-movement measures showed promising internal discrimination of PDD from PD without dementia. Eye-movement-based machine learning may complement clinical cognitive assessment, but external validation in larger and more representative cohorts is required before clinical use.