Optimal Quantile Level Selection Method And Its Application To ECG Data For Disease Classification

Classification with a large number of predictors become increasingly important in biological and medical research in the big data era. Motivated by cardiovascular disease classification and the observation that the distributions of various measurement variables on the electrocardiogram (ECG) signals of the diseased group are often skewed, heavy-tailed or multi-modal, we characterize the distributions by an optimal quantile level which maximizes the standardized difference of the sample quantiles between two distributions. And we expand this method to distinguish the difference between multiple distributions. In this paper, numeric studies and an intensive study of a real dataset which is from the II lead of the PTB Diagnostic ECG Database are performed to illustrate the performance of the proposed method. According to the result of classification, we find that the optimal quantile levels has a better extraction of the characteristics of abnormal heartbeat from ECGs, and the classification results by using quantile statistic is almost always better than using the mean statistic. The discriminant accuracy to distinguish healthy people from patient is up to 84.96%. However, because of the small sample size of data for some diseases, the accuracy of the multi-class is only about 67%. Besides, in the study of real data, three classifiers of stepwise discriminant analysis based on Mahalanobis distance, support vector machine (SVM) and classification and regression tree (CART) are used, and then the performance of these classifiers are evaluated by Leave-One-Out cross validation and re-substitution method.