Patient-Specific Heartbeat Classification Based on I-Vector Adapted Deep Neural Networks

Abstract

Automatic heartbeat classification from electrocardiogram (ECG) signals is important for diagnosing heart arrhythmias. A main challenge in ECG classification is the variability of ECG signals across patients. This paper proposes a patient-specific heartbeat classifier to address the inter-patient variations in ECG signals. Inspired by the success of identity vectors (i-vectors) in speech and speaker recognition, we extracted one i-vector from five minutes of ECG data for each patient and applied it to adapt a patient-independent deep neural network (DNN) to a patient-specific DNN, namely i-vector adapted patient-specific DNN ($iAP$-DNN). Evaluations on the MIT-BIH arrhythmia database show that the $iAP$-DNN is able to classify raw ECG signals of the corresponding patient into normal heartbeats and different types of arrhythmias and that it outperforms existing patient-specific classifiers in terms of sensitivity-vs-specificity and Mathews correlation coefficients.