Explainable Parkinson’s Disease Prediction from Keystroke Dynamics Using AdaBoost Ensemble Learning

Parkinson’s disease (PD) is a common, neurodegenerative disorder with a long asymptomatic period before the manifestation of its well-known motor symptoms and is also notoriously difficult to diagnose in its early stages. The present study is to use machine learning to achieve the diagnosis of Parkinson’s disease based on keystroke dynamics obtained while the users are typing on the keyboard. The Tappy Keystroke dataset in the PhysioNet dataset is utilized. 33 features from the hold time, flight time and latency were extracted from the samples through statistical analysis. The base classifier is the AdaBoost algorithm after combining with Bayesian optimization. The results compared with the Decision Tree, Naive Bayes, K-Nearest Neighbor, Artificial Neural Network, RUSBoost and Bagging all show that the AdaBoost has the highest accuracy of 87.16% and the AUC of 0.9129. It was observed that ensemble-based methods generally showed more balanced and higher performance compared to single classifiers. The decision process of the most successful model was interpreted using the SHAP method; it was determined that the HoldTime and FlightTime variation measures played a decisive role in classification. The results obtained indicate that keystroke dynamics can be used as a non-invasive, scalable PD biomarker that does not require a clinical setting.