Classification of Patient’s Heartbeat Obtained by ECG Using Active Learning

Abstract

While physicians can correctly detect various anomalies in heartbeat signals by using electrocardiograms (ECGs) of different individuals, supervised machine learning has mixed results. The diversity of activities performed by patients is often severe class disproportion, and the cost of obtaining and mapping them to label data for individual patients makes the issue more complicated. We generate synthetic biological data using SimGANs and solve these issues by performing patient-adaptive and task-adaptive heartbeat categorization via active learning. Our approach performed much better than the previous methods of heartbeat classification when evaluated on a benchmark database of ECG recordings on the two main classification tasks specified by the medical sciences. Furthermore, our method needed almost 90% less patient-customized training data than the techniques we evaluated.