Region-based Active Learning for Reducing Annotation Effort of ECG Waveform Segmentation

Electrocardiogram (ECG) signals consists of various beat morphologies including P wave, QRS complex, and T wave. Waveform segmentation to identify those waves in ECG signals is critical to monitor heart health status and can be leveraged for automatic diagnosis of heart-related diseases. Instead of manually annotating those waves, this work studies an active learning (AL)-based ECG waveform segmentation method to reduce the annotation effort where a learning model is trained to automatically annotate ECG waves. To adapt AL for ECG waveform segmentation, we introduce a region-based AL method where multiple signal regions containing fewer consecutive signal samples within ECG frames are selected iteratively for human annotation instead of querying at the frame level. The performance of the proposed method is validated on the QT database, as it has a faster convergence rate and higher annotation accuracy compared to benchmark schemes which are based on random query and frame-level query strategies. In addition, the proposed method consumes 25% less annotation effort in average compared to the benchmark schemes.