Detecting Early Parkinson's Disease from Keystroke Dynamics using the Tensor-Train Decomposition

Oroojeni M. J. Hooman, James Oldfield, Mihalis A. Nicolaou

Abstract: We present a method for detecting early signs of Parkinson's disease from keystroke hold times that is based on the Tensor-Train (TT) decomposition. While simple uni-variate methods such as logistic regression have shown good performance on the given problem by using appropriate features, the TT format facilitates modelling high-order interactions by representing the exponentially large parameter tensor in a compact multi-linear form. By performing time-series feature extraction based on scalable hypothesis testing, we show that the proposed approach can significantly improve upon state-of-the-art for the given problem, reaching a performance of AUC=0.88, outperforming compared methods such as deep neural networks on the problem of detecting early Parkinson's disease from keystroke dynamics.