Feature Learning for Detection and Prediction of Freezing of Gait in Parkinson’s Disease

Freezing of gait (FoG) is a common gait impairment among patients with advanced Parkinson’s disease. FoG is associated with falls and negatively impact the patient’s quality of life. Wearable systems that detect FoG have been developed to help patients resume walking by means of auditory cueing. However, current methods for automated detection are not yet ideal. In this paper, we first compare feature learning approaches based on time-domain and statistical features to unsupervised ones based on principal components analysis. The latter systematically outperforms the former and also the standard in the field - Freezing Index by up to 8.1% in terms of F1-measure for FoG detection.

We go a step further by analyzing FoG prediction, i.e., identification of patterns (pre-FoG) occurring before FoG episodes, based only on motion data. Until now this was only attempted using electroencephalography. With respect to the three-class problem (FoG vs. pre-FoG vs. normal locomotion), we show that FoG prediction performance is highly patient-dependent, reaching an F1-measure of 56% in the pre-FoG class for patients who exhibit enough gait degradation before FoG.