Improved Arrhythmia Detection from Electrocardiogram

TLDR: The proposed research shows an improved binary classification accuracy using deep autoencoder (DAE) neural network with a support vector machine (SVM) for six types of arrhythmic ECG beat recognition.

Abstract: Computerized detection and monitoring of arrhythmia, or irregular heartbeats using electrocardiogram (ECG) is a typical pattern recognition problem and has been under research from decades. The challenge is mainly with universality of features and their minimization. The proposed research shows an improved binary classification accuracy using deep autoencoder (DAE) neural network with a support vector machine (SVM) for six types of arrhythmic ECG beat recognition. It also compares the efficiency of feature extraction and classification of arrhythmic beats using principal component analysis (PCA), and discrete wavelet transform (DWT) combined with binary K-nearest neighbours (KNN) and SVM classifier. After filtering the raw ECG signal, the R-peaks were detected by a DWT based approach. The window around the QRS zone of each ECG beat was used for feature extraction and a reduced set of 40 features was fed to the binary classifiers. In the present work, six type of abnormalities viz. ‘A’, ‘F’, ‘L’, ‘R’, ‘V’, ‘f’ from MIT-BIH arrhythmia (mitdb) database were used for the evaluating the classifiers’ performance. Among the all six combination of proposed feature extraction and classification techniques, the average sensitivity (SE) and positive predictive value (PPV) achieved by DAE-SVM (DAE- KNN) were 99.99% and 99.98% (99.97% and 99.95%) respectively. The results are competitive with some published works.