ECG-Based AI Model Predicts Post-Heart Attack Outcomes
A team of researchers has introduced a pretrained AI model aimed at forecasting the progression of cardiovascular disease following a myocardial infarction (MI). This innovative model employs contrastive learning on unlabelled ECGs, incorporating patient-specific temporal data alongside supervised multitask heads, and is subsequently fine-tuned for predicting post-MI outcomes. It shows superior performance compared to a model developed from the ground up, achieving an AUC of 0.794 versus 0.608. This research highlights the potential of clinically structured ECG modeling to enhance classification, particularly in scenarios with limited data. The findings are available on arXiv (2605.13568) in the Computer Science > Machine Learning category.
Key facts
- Myocardial infarction is a leading cause of death.
- ECG-based prognostic models underperform due to scarce labelled data.
- Foundation models can learn from unlabelled ECGs via self-supervision.
- The proposed model combines contrastive learning with supervised multitask heads.
- The model is fine-tuned on post-MI outcome prediction.
- It achieved 0.794 AUC vs 0.608 AUC for a model trained from scratch.
- Clinically structured ECG modelling improves classification in limited data regimes.
- The paper is available on arXiv.
Entities
Institutions
- arXiv