Applying pre-trained deep learning models for Multi-Label Classification of Realistic and Noisy Electrocardiogram Images

ECG is an indispensable tool for diagnosing cardiovascular diseases because it can quickly and non-invasive monitor heart rhythms. Recent research has demonstrated that innovative AI approaches have great potential to
advance ECG interpretation. Despite these advances, widespread application of such techniques is hindered by several factors such as digital copies of ECGs remain unavailable in many resource-limited areas. Additionally, ECGs scanned from paper contain noise such as shadows or creases, which are often not accounted for in digital approaches. Furthermore, contemporary research is concentrated on detecting single conditions or diseases, despite the possibility of their co-occurrence. To address these shortcomings, this work investigates the application of pretrained CNNs by fine-tuning popular models such as
AlexNet and ResNet separately for multi-label classification of realistic and noisy ECG images. The results underscore the potential for the improvement of image-based models and suggest that future research should target this
approach to provide a scalable method for interpreting ECG images. Our Team name is ECG UVA and got the classification F-measure of 0.516, achieving a rank of 5/16 on the leaderboard on the hidden dataset in the official phase of the ‘Digitization and Classification of ECG Images: George B. Moody PhysioNet Challenge 2024’.