Aims: Electrocardiogram (ECG) classification remains a critical research focus, particularly in analyzing one-dimensional ECG signals. Despite the technological advances in digital ECG methodologies, which utilize computational algorithms to enhance accuracy and accessibility of ECG-based diagnoses, traditional paper-based ECGs are still prevalent, especially in lower-income settings. This paper explores a hybrid approach that integrates modern digital techniques with conventional practices to classify physical (paper-based) ECG records as normal or abnormal, as part of the George B. Moody PhysioNet Challenge 2024.
Methods: We employed a synthetic image ECG generator to create datasets from the CinC2021, CODE15, PTB-XL datasets, in addition to a local dataset, InCor-DB, which includes 99,746 ECG exams from 64,192 patients. Our methodology utilized the ConvNext architecture, pre-trained on ECG images from CinC2021, CODE15, and InCor-DB datasets, resized to 224x224 pixels. We fine-tuned the model on the PTB-XL dataset, incorporating demographic data (age, sex, weight, height) for classification of abnormalities. The training strategy involved cross-entropy loss minimization, class-weight adjustments, and AdamW optimizer with a learning rate of 1e-3 over 30 epochs, aided by a cosine scheduler and early stopping to prevent overfitting. We also implemented 10-fold cross-validation, with the final model being an ensemble of the models from each fold.
Results: The model achieved an F1-score of 0.82±0.02 across internal 10-fold cross-validation. On the challenge's hidden validation set, our team, AIMED, achieved an F1-score of 0.71, securing second place in team rankings and fifth overall.
Conclusion: This study demonstrates significant advances in integrating digital technologies into clinical settings, potentially enhancing ECG diagnostic processes globally.