Prediction of freezing of gait from Parkinson's Disease movement time series using conditional random fields

Abstract

Freezing of Gait (FOG) in Parkinson's Disease (PD) is a brief episodic impedance of movement that is mostly manifested at the late stages of the PD. Accelerometer sensors are widely utilized to collect dysfunctional movement time series data stemming from patients with PD. In this work, we propose a robust FOG predictive model that employs a combination of wavelets and Conditional Random Fields (CRF) to predict FOG episodes from low level FOG accelerometer time series interleaved with normal movement time series of PD patients. Specifically, in order to derive and extract unique signature features of FOG time series, we utilize wavelets to perform in-depth analysis of PD movement spectral at multiple resolutions. We design a CRF that leverages the extracted signature feature vectors to diligently learn the underlying characteristics of FOG time series and to effectively predict FOG episodes at their onsets. Our empirical evaluations on a real PD dataset demonstrate that our technique delivers enhanced prediction accuracies.