Machine learning approach for predicting state transitions via shank acceleration data during freezing of gait in Parkinson’s disease

Abstract

Rehabilitation using external cues has shown positive impact on Freezing-of-Gait (FoG) severity in Parkinson's disease (PD) patients. Research studies on FoG detection and prediction are widely accessible, however little is known about transitions from “FoG” to “normal-gait” phases and vice-versa. Especially, for the patient to recover completely from the freeze event as well as to avoid undesirable distractions due to long-term cue persistence, post-FoG periods must be optimally chosen for managing automatic cue deactivation. Therefore, our study attempts to device a machine-learning based approach for labelling of accelerometer data with state-transitions from “normal-gait” to “FoG” (i.e., FoG-prediction) as well as from “FoG” to “normal-gait” (i.e., unfreeze-prediction). Our prediction approach distinguishes itself by innovating in sensor utilization, feature-extraction, and latency reduction. We introduce a novel four-class labeling scheme based on state transitions, diverging from fixed-time periods utilized in prior studies. Model training involves regression and classification with a streamlined set of 23 hybrid features, minimizing latency with reduced window segments and operating on 4-second and 3-second pre- and post-FoG time-frames, respectively, exhibiting minimal 0.5-second latency. Our approach simply required one shank sensor providing minimum-instrumentation for the user's comfort. Four-class (No-FoG, pre-FoG, FoG, and post-FoG) data have been fed to train random-forest classifier. Our FoG prediction model achieved exceptional metrics, particularly for the “post-FoG” class, with 99.0 % accuracy, 97.4 % precision, 99.1 % sensitivity, and 98.1 % f1-score, along with 99.0 % cohen’s-kappa, 98.3 % jaccard-index, and 99.0 % Matthews-correlation-coefficient. Therefore, an intelligent-strategy for timely activation and deactivation of cues is the major contribution of our work.