ECG Delineation with Deep Learning

Electrocardiogram (ECG) delineation identifies the boundaries of P waves, QRS complexes, and T waves, providing structural annotations that can guide cardiovascular research and diagnosis. Traditional ECG delineators, like Pan-Tompkins, can annotate intervals using rules, but often suffer from misplacement or missed waves. In this demo we present R-U-Net, a deep-learning model for ECG delineation using a ResNet-18 encoder and U-Net decoder to produce sample level annotations of P / QRS / T waves. These annotations can then be reviewed and corrected by a human, rather than being manually annotated on an empty strip.

Users are provided with a series of 10 second single-lead ECG and are asked to manually annotate intervals to gain a sense of the annotation cost. Then, both a draft of the Pan-Tompkins rules and a draft of the R-U-Net are loaded onto the same ECG and compared against ground truth. Per-example Mean IoU for both the Rules and R-U-Net are reported for quantitative evaluation of Pan-Tomkins, R-U-Net, and human annotations. Users are able to delete, add, and reset intervals. The ultimate goal is for attendees to understand the difficulty of manual ECG segmentation, and that using a deep learning model allows for both faster and more accurate delineations.

Demo Link: https://ecg-seg-demo.vercel.app/