Patient-adaptive ectopic beat classification using active learning

A major challenge in applying machine learning techniques to the problem of heartbeat classification is dealing effectively with inter-patient differences in electrocardiograms (ECGs). Inter-patient differences create a need for patient-specific classifiers, since there is no a priori reason to assume that a classifier trained on data from one patient will yield useful results when applied to a different patient. Unfortunately, patient-specific classifiers come at a high cost, since they require a labeled training set. Using active learning, we show that one can drastically reduce the amount of patient-specific labeled training data required to build a highly accurate patient-specific binary heartbeat classifier for identifying ventricular ectopic beats. Tested on all 48 half-hour ECG recordings from the MIT-BIH Arrhythmia Database, our approach achieves an average sensitivity of 96.2% and specificity of 99.9%. The average number of beats needed to train each patient-specific classifier was less than 37 beats, approximately 30 seconds of data.