The Classification of ECG Signal

Problem Description:

This project was one of the two finals of the course Statistics Learning. The students were asked to design an algorithm, which could classify the electrocardiogram (ECG) of distinct patients, with the methods introduced in this course. All the students needed to test their algorithms with the uniform data set, which contains the ECG signals collected from about 100 patients. The accuracy of the classification algorithm was supposed to be over 70%. Several methods were recommended: independence component analysis (ICA), linear discriminant analysis(LDA) and principal component analysis (PCA).

The Solution:

The algorithm framework was shown in Figure 1, which consisted of three parts: preprocessing, dimension reduction and classification. The preprocessing module made sure that all the ECG signals were separated into beat samples and their peaks were aligned, as shown in Figure 2, which was implemented by detecting the R peaks of each beat. Dimension reduction was the most significant part in my algorithm, which directly increased the accuracy from 50% to 90%. I utilized wavelet transformation to represent each signal in Figure 2 as a 16-point signal and then extracted the feature bases of the 16-point signals by ICA algorithm, as shown in Figure 3. In classification module, I compared 6 typical classification algorithms, as shown in Figure 1. The logistic regression and KNN used the wavelet results as input directly and the others used the signal projected in ICA bases shown in Figure 3. The classification results (5-fold cross validation) showed that the KNN has the best performance, even better than the results in references, and the second one was ICA-SVM method. Theoretically, ICA-SVM method seemed to be more robust than KNN method. However, in my experiments, no evidence showed that KNN was inferior to the SVM method. Besides, the data set given by the instructor was too small to give a general conclusion, so the experiments on a bigger dataset might be necessary to confirm the results in this work.

The project report is available at the bottom of this page.