Knowledge Augmentation via Synthetic Data: A Framework for Real-World ECG Image Classification

TL;DR

<3-5 sentence high-level summary> The paper addresses the gap between ECG photographs and digital signals by introducing a Knowledge Augmentation framework that combines diverse synthetic data with a robust preprocessing pipeline. It employs a two stage Morphology Learning and Task-Specific Adaptation strategy on ConvNeXt and demonstrates strong cross source generalization, achieving a macro-AUROC of $0.9677$ on the hidden test of the 2024 BHF Challenge. The approach mitigates the single source limitation of synthetic data by learning general morphology from heterogeneous sources and adapting to photo like data, yielding state of the art performance. This work has practical significance for real world ECG image interpretation and suggests a robust pathway for deploying AI in photo based ECG workflows.

Abstract

In real-world clinical practice, electrocardiograms (ECGs) are often captured and shared as photographs. However, publicly available ECG data, and thus most related research, relies on digital signals. This has led to a disconnect in which computer assisted interpretation of ECG cannot easily be applied to ECG images. The emergence of high-fidelity synthetic data generators has introduced practical alternatives by producing realistic, photo-like, ECG images derived from the digital signal that could help narrow this divide. To address this, we propose a novel knowledge augmentation framework that uses synthetic data generated from multiple sources to provide generalisable and accurate interpretation of ECG photographs. Our framework features two key contributions. First, we introduce a robust pre-processing pipeline designed to remove background artifacts and reduces visual differences between images. Second, we implement a two-stage training strategy: a Morphology Learning Stage, where the model captures broad morphological features from visually different, scan-like synthetic data, followed by a Task-Specific Adaptation Stage, where the model is fine-tuned on the photo-like target data. We tested the model on the British Heart Foundation Challenge dataset, to classify five common ECG findings: myocardial infarction (MI), atrial fibrillation, hypertrophy, conduction disturbance, and ST/T changes. Our approach, built upon the ConvNeXt backbone, outperforms a single-source training baseline and achieved \textbf{1st} place in the challenge with an macro-AUROC of \textbf{0.9677}. These results suggest that incorporating morphology learning from heterogeneous sources offers a more robust and generalizable paradigm than conventional single-source training.

Figures (6)

• Figure 1: Examples of synthetic ECG images generated by different tools.
• Figure 2: Overview of the proposed Knowledge Augmentation (KA) framework. The process consists of a pre-processing pipeline followed by a two-stage training strategy. Stage I (Morphology Learning) trains on diverse ECG scans (including real ECG scans and generated ECG scans) to learn general ECG morphology. Stage II (Task-Specific Adaptation) fine-tunes the model on photo-realistic images (the BHF Challenge dataset - training data) for final classification.
• Figure 3: Overview of the data pre-processing workflow steps. Step 1: Raw image input. Step 2: ROI extraction using YOLO to remove surrounding background. Step 3: Segmentation using SAM to isolate the paper sheet. Step 4: Perspective rectification to correct the viewing angle. Step 5: CLAHE colour enhancement. Three strategies are evaluated based on these steps: (1) Raw Input (Step 1 only); (2) ROI Crop + CLAHE (Step 2 $\rightarrow$ Step 5); and (3) Full Pipeline (Seg. + Rect.) (Steps 1--5)..
• Figure 4: ConvNeXt architecture: (a) overall structure; (b) stem layer; (c) ConvNeXt block; (d) downsampling layer.
• Figure 5: Grad-CAM visualizations for different pre-processing strategies on two samples: Atrial Fibrillation (AF, Top Row, a-c) and Hypertrophy (HYP, Bottom Row, d-f). (a, d) The Raw Input setting focuses on background noise. (b, e) The ROI Crop + CLAHE model shows a balanced focus that includes rhythm, which is the correct strategy for AF (b). (c, f) The Full Pipeline (Seg. + Rect.) model strongly prioritizes single-beat morphology. While this morphological focus is accurate for HYP (f), this bias toward morphology is less effective for the rhythm-based disorder AF (c).