Less is More: Univariate Modelling to Detect Early Parkinson's Disease from Keystroke Dynamics

We analyse keystroke hold times from typing logs to detect early signs of Parkinson’s disease. We develop a feature that captures the dynamic variation between consecutive keystrokes and demonstrate that it can be be used in a univariate model to perform classification with AUC=0.85 from only a few hundred keystrokes. This is a substantial improvement on the current baseline. We argue that previously proposed methods are based on overcomplicated models—our simpler method is not only more elegant and transparent but also more effective.