Automated ECG Image Classification with InceptionV3

Despite the rise of digital electrocardiogram (ECG) technology, paper-based ECGs continue to be prevalent, especially in underrepresented and underserved communities.This paper presents the DSAIL team's participation in the George B. Moody PhysioNet Challenge 2024 to develop an open-source algorithm for classifying ECG images.We fine-tuned a pre-trained InceptionV3 model on the PTB-XL dataset, comprising 21,799 12-lead ECG recordings, supplemented with synthetic ECG images from the ECG-Image-Kit.The model was trained using 80% of these images, reserving 20% for validation.Our choice of the InceptionV3 architecture leverages its capability to effectively capture local and global features, which is crucial for the inherent variability in ECG image patterns.The model achieved a validation macro F-measure score of 0.429 on a dataset accessible only to the organizers, securing 6th place on the official classification leaderboard.However, the algorithm struggled with mobile phone images of stained, deteriorated, and cleaned ECGs, yielding a low F-score of 0.08.In contrast, it performed significantly better on color scans of clean and deteriorated paper ECGs, achieving an F-score of 0.5.Although further improvements are necessary, neural network-based algorithms demonstrate promising potential for enhancing access to ECG-based diagnosis and cardiac care.