A multitask deep learning model utilizing electrocardiograms for major cardiovascular adverse events prediction.

Summary

Using 2.82 million ECGs and external validation, ECG‑MACE accurately predicted 1‑year first‑ever HF, MI, ischemic stroke, and mortality and outperformed Framingham risk scores for longer‑term outcomes. The model also stratified 10‑year incidence, highlighting its potential for scalable preventive risk screening from routine ECGs.

Key Findings

• Trained on 2,821,889 12‑lead ECGs with external validation (n=113,224), the model achieved AUROCs: HF 0.90, MI 0.85, ischemic stroke 0.76, mortality 0.89.
• Outperformed Framingham risk scores for 5‑year MACEs and 10‑year mortality prediction.
• Over 10 years, model‑positive patients had substantially higher incidence ratios (HF 15.28; MI 7.87; IS 4.74; mortality 13.18) than model‑negative patients.

Clinical Implications

Health systems could integrate ECG‑MACE into EHRs to flag high‑risk patients for early preventive interventions, targeted diagnostics, or closer follow‑up, even before traditional risk factors accumulate.

Limitations

• Retrospective design without prospective clinical deployment
• Generalizability across devices/health systems and model interpretability need further assessment

Future Directions

Prospective, multi-center deployment trials to assess clinical utility, workflow integration, outcomes impact, and health equity; calibration across vendors and regions; interpretable AI methods.