Assessment of Parkinson's disease: Classification and complexity analysis

Abstract

In this paper we present a new algorithm to analyze gait patterns of subjects with Parkinson's disease (PD) under two conditions of deep brain stimulation (DBS) off and on. Our goal is to study the complexity and predictability of gait of PD subjects in comparison to the gait of normal subjects. At the same time we used the features extracted for this purpose to classify PD subjects with DBS off or on from healthy control subjects. The angular velocity signals obtained from four uni-axial gyroscopes were used for feature extraction. These four gyroscopes were attached to the right and left shanks and thighs. Approximate Entropy (ApEn), Hurst Exponent (HE) and Higuchi Fractal Dimension (HFD) are the three nonlinear features used for classification. 9 PD subjects and 10 healthy controls were participated in the experiments. The results showed that ApEn and HFD have greater values for PD subjects (with DBS off) indicating more complexity and irregularity in their walking dynamics. Also HE had smaller values for PD subjects (with DBS off) indicating less predictability in their gait time series. The k-nearest neighbor classifier with leave-one-out cross validation method were used for classification and we succeeded to discriminate PD subjects (with DBS off) from the normal ones with the high accuracy of 100%. The accuracy for discrimination of PD subjects with DBS on from the normal ones was 89.47%. Then we analyzed the signals of each single gyroscope, one by one, to find the most informative signal. We achieved the high classification accuracy of 100% (for PD subjects with DBS off) using only the signals of the right shank or the right thigh. It indicates that, among our participants, the signals obtained from the right side of the body are more informative than the left side.