Predicting Severity of Parkinson's Disease with Typing Behavior: A Machine Learning Approach

Abstract

Parkinson’s disease (PD) is one of the most common neurological diseases in adults over the age of 65. Current monitoring of disease stage and progression consists of physician visits, which is inefficient and unreliable since symptom severity varies throughout the day. A more practical solution could include monitoring behavior throughout the day. Previous work has found that typing behavior is an accurate way to differentiate people with PD from those without PD. This finding leads to the idea that it may also be useful to use keyboard characteristics to detect PD severity. The current study examined whether this was possible. Additionally, it compared three different machine learning methods: logistic regression, k¬-nearest neighbors, and random forests. Finally, it examined how accuracy of the classification of PD severity differed when including increasing amounts of keystrokes (1000, 2000, and 5000 keystrokes). It was found that the random forests classifier could predict PD stage moderately well in the 5000-keystroke dataset. However, there was no further difference in model type or clear pattern to show that models become more accurate with increased amounts of keystrokes. This study is a first step in examining how computer behaviors might be able to be used for predicting and potentially monitoring PD patients’ disease stage.