Classification of Faller and Non-Faller Parkinson's Disease Patients using Wavelet-based Multifractal Spectrum of Center of Pressure Signal

The Center of pressure (COP) signal has an erratic behavior. Quantifying this behavior using the time-scale multi-resolution analysis provides valuable information to the prediction of fall risk in Parkinson's disease patients. In this paper, multifractal spectrum measures, as complexity or smoothness measures of signal, were used to classify faller Parkinson's disease patients from the non-fallers. From data previously gathered, in 11 faller PD participants and 49 non-fallers, the center of pressure signal was recorded on a rigid and foam surface with open and closed eyes. Multifractal spectrum (MFS) measures were calculated using discrete wavelet transform (DWT). Based on MFS measures, several classification algorithms including decision trees, naive Bayes, logistic regression, artificial neural network (ANN), and support vector machine (SVM) were tested searching for the best one. With the aim of better classification performance, stratified cross-validation, as a sample balancing method, and ensemble methods such as Bagging and Adaboost were applied. Results showed the most prevalent smooth behavior in MFS (
$a_{0}$
) on the rigid surface can separate the faller PD participants effectively. Additionally, proprioception (standing on foam or rigid surface), as compared to vision, had a more important role in signal smoothness. A decision tree based on the ID3 algorithm with Bagging showed an acceptable performance (accuracy=0.88, sensitivity=0.57, precision=0.75). Multifractal spectrum analysis of COP disclosed the postural instability characteristics in faller PD patients, that dispose patients to fall, and which is discriminable by fine-tuned classifiers.