Digitization and Classification of ECG Images: The George B. Moody PhysioNet Challenge 2024

Abstract

The George B. Moody PhysioNet Challenge 2024 invited teams to develop algorithmic approaches for digitizing and classifying electrocardiograms (ECGs) from photographed or scanned images of paper ECGs. Paper ECGs have existed for decades, capturing the variability and evolution of cardiovascular diseases (CVDs) across demographics, geography, and time. Physical and digital ECG images remain common in cardiac care. However, ECG-based interpretation algorithms typically require digital time-series representations of ECG data, so existing algorithms cannot interpret them, and new algorithms cannot learn from them. Therefore, digitizing ECG images is important for improving the accessibility and quality of cardiac care. To support this goal, the Challenge introduced ECG-Image-Kit, a synthetic ECG image generator with various realistic distortions, such as wrinkles, creases, shadows, rotations, and handwriting, to allow teams to create arbitrary large and diverse datasets for training generaliz-able models. The Challenge also introduced ECG-Image-Database, a dataset of 35,595 real ECG papers from 1,977 distinct ECG records, to assess and support the generaliz-abilty of the Challenge approaches. A total of 62 teams participated in the Challenge, representing diverse approaches from both academia and industry worldwide.