Feature selection and construction for the discrimination of neurodegenerative diseases based on gait analysis

Abstract

Gait disorder is one symptom of neurodegenerative disease. Using wearable motion sensors to monitor the motor function of patients with neurodegenerative disease has attracted more attention. Research has shown that machine learning techniques can be applied to the classification of neurodegenerative diseases from the gait data recorded by footswitches. In order to identify the most valuable features from 10 raw temporal variables extracted from gait cycles to improve the classification performance, we examine four types of feature selection and construction methods, namely, maximum signal-to-noise ratio based feature selection method, maximum signal-to-noise ratio combined with minimum correlation based feature selection method, maximum prediction power combined with minimum correlation based feature selection method and principal component analysis. Results show that using a set of four features, a relatively high prediction performance has been achieved with classification accuracies ranging from 79.04% to 93.96%. The continual increase of the number of features does not significantly contribute to the improvement of classification performance. This is consistent with clustering-based feature analysis.