Statistical methods for analysis of Parkinson’s disease gait pattern and classification

Abstract

Understanding the human gait and extracting intrinsic feature helps to classify walking patterns of Parkinson disease patients. The measurement of time series gait pattern is required to detect gait disturbances observed in medical gait data. An attempt is taken to compute Normalized Auto Correlation (NAC) along the temporal axis which calculates the degree of gait fluctuation in control subjects (CO) and Parkinson patient’s (PD) gait. In this paper, an underlying statistical analysis is addressed to understand the statistical nature of data. Identifying the proper distribution of these data in advance discards the unwanted information which helps to preserve more informative features. Four different normality testing methods (i.e. W/S, Kolmogorov-Smirnov, Shapiro Wilk and Anderson Darling) are applied to ensure whether the acquired gait data are modelled by a normal distribution. It precludes the costly error during feature analysis to produce the accurate results. A feature selection method, Fisher Discriminant Ratio (FDR) is applied to select most discriminative feature among all the statistical features (i.e. Mean, Median, Mode, Standard Deviation, Variance, Skewness and Kurtosis) derived from both the classes. A probabilistic classifier based on Bayes’ theorem demonstrates its efficiency in the classification of Parkinson gait with illustrating statistical error metrics (i.e. MAE, RMSE, MCE, MSE, SEM, SSE etc.).