Optimal window lengths, features and subsets thereof for freezing of gait classification

Abstract

Freezing of gait (FoG) is a common gait impairment in Parkinson's disease that puts patients at risk of falls and deteriorates their quality of life. Relief is sought after by evaluating the possibility of wearable systems that detect FoG in real-time and provide gait-reinforcing biofeedback cues. The successful detection relies on the extraction of high quality features, which have to be computed from recent samples of an inertial measurement unit in order to ensure real-time applicability. Unfortunately, the amount of samples considered for a feature's computation, i.e. the data window length, has been subjected to widespread disagreement: With no thorough analysis available, employed window lengths differed by several seconds among implementations. We derive optimal window lengths for a broad number of features used throughout literature by using mutual information as an evaluation metric, and elaborate on a window length's significance in affecting classification performance. With conventional feature selection methods, feature subsets tailored to various machine learning algorithms are established. Relying on these feature subsets for FoG classification, whereby all features are extracted with optimal window lengths, F1-scores increase up to 17.1% for individual classifiers and up to 12.7% on average when compared to previously proposed feature sets that are extracted with sub-optimal window lengths.