Background: Billions of paper Electrocardiogram (ECG) signals are recorded yearly worldwide, particularly in the global South. Manual review of so many recordings is laborious and time-consuming. The digitalized reconstruction of paper ECGs is critical for efficiently diagnosing cardiac diseases. As part of the George B. Moody PhysioNet Challenge 2024, our team (shenhai_dl) developed deep learning models to reconstruct and classify the signals from simulated paper ECGs.
Method: Leveraging a vast proprietary database of 83,692 12-lead ECGs and simulation software, we generated ECG-image datasets with labels of different lead signals. We then trained a graph convolutional neural network to detect signal regions in simulated ECG images. We proposed two strategies to recover the signal more accurately from detected signal regions. One developed a transformer model to denoise images and remove grids, and the other proposed a modified convolutional layer to learn the rotation angle of an image. Finally, we trained a structured state space model to learn features among long sequences for classifying the digitalized ECG signals.
Results: Our method achieved a cross-validated training performance of 3.84 (SNR for the reconstruction task) and 0.87 (F1 score for the classification task). We received an SNR of -18.12 (reconstruction) and an F1 score of 0.45 (classification) on the hidden test set. The proposed models showed a new clue to reconstructing recordings from paper ECGs and efficiently classifying digital ECG signals.