Towards Transparent AI-Aided Neurology: Detection and Lateralization of Parkinson's Disease

Research overview

The text below is the source article’s original abstract under a verified reuse license. cc by

Abstract Background: Parkinson's Disease (PD) diagnosis remains challenging due to subjective assessments and delayed detection of asymmetric motor symptoms. While digital biomarkers like keystroke dynamics show promise, most AI approaches lack clinical interpretability and fail to lateralize motor onset. Methods: We propose a transparent machine learning framework named KATE using keystroke data from the Tappy keyboard app. Hand-specific features (hold-time variance, inter-key intervals, asymmetry indices) were engineered to quantify motor laterality. Four models—Logistic Regression, SVM, Random Forest, and XGBoost—were evaluated for binary PD detection and multiclass asymmetry classification (left/right-dominant, symmetric). SHAP and LIME provided global/local explanations. Results: Binary detection: XGBoost achieved near-perfect performance (AUC=1.00, recall=99.6\%, FP rate=4.3\%).\\ Asymmetry classification: XGBoost led with F1=85\% and precision=83\%, though right-dominant PD recall lagged (78\%). Explainability: SHAP identified pathophysiological biomarkers—Latency time\_max (motor initiation delay), Flight\_min\_RS (inter-key coordination deficits), and Hold\_std\_S (motion smoothness degradation). Ensembles reduced false positives by 42\% versus Logistic Regression. Conclusion: Our framework enables accurate, asymmetry-aware PD screening while providing clinically interpretable insights. Integration of hand-specific digital phenotyping with explainable AI pioneers a template for transparent neurodegenerative disease diagnostics.

Related topics