Application Of Statistics On Disease Discrimination Based On ECG Signals

This paper focuses on the application of a variety of statistical methods in disease identi-fication based on the electrocardiogram (ECG) signals. Recently, the ECG-related researches are mainly about the classification of ECG abnormal beats. In this paper, we extract features from the II lead of ECG signals, and then compare the effects of different classifier. In the data preparation phase, we use WFDB package detecting the boundary of QRS wave, and use SAS to do the feature extraction. There are four types of calculated features, including time span, amplitude, T-wave category and the rate of waveform change. For each feature of each sample, we use statistics mean and quantile forming two feature sets. Besides, we also use PCA due to the high dimension and correlation among the features. In terms of the clas-sifier, we use discriminant analysis based on Mahalanobis distance, support vector machines (SVM) and logistic regression with LASSO. According to the classification results, we can find that discriminant analysis based on Mahalanobis distance is more stable and effective, that quantile statistic has a better extraction of the abnormal information from the ECG sig-nals, and that PCA can improve the performance of these three classifiers. However, due to the small sample size of data and the limitations of feature extraction, the multi-class results are not that good. The discriminant accuracy of multi-class in this paper is about61%, but accuracy of binary classification is up to88%.