Quadratic Programming Data Descriptors for Abnormal Beat Detection in ECG Recordings

This paper analyzes the efficacy of applying one class classifiers (OCCs) to the problem of abnormal beat detection in ECG. It also proposes a novel OCC called Quadratic Programming Dissimilarity representation based Data Descriptor (QPDDD). A comparison of the proposed classification technique with existing classifiers over the MIT-BIH arrhythmia database is presented. Results show that OCCs coupled with wavelet domain features present a practical, robust and scalable solution for handling inter-individual variability in ECG patterns of different types of cardiac beats. An equal error rate of 90-95% was obtained for the MIT-BIH arrhythmia database depending upon the amount of training data used. A major advantage of the proposed scheme is that it requires only normal beats during its training. Another advantage is that it is able to handle inter-individual differences in ECG morphologies as the training takes place separately for each individual.