Ensemble Deep Learning Algorithm for Structural Heart Disease Screening Using Electrocardiographic Images: PRESENT SHD.

Summary

An ensemble deep learning model operating on ECG images (not raw waveforms) detected multiple left-sided structural heart diseases with AUROCs ~0.85–0.90 across 4 U.S. hospitals and in ELSA-Brasil, with sensitivities ~88–96%. The model generalized to smartphone photographs of printed/monitor ECGs and predicted 2–4-fold higher risk of incident SHD/HF, independent of clinical factors.

Key Findings

• Development cohort: 261,228 ECGs from 93,693 patients; external validation included 11,023 (YNHH test), 44,591 (4 U.S. hospitals), and 3,014 (ELSA-Brasil).
• Composite PRESENT-SHD achieved AUROC 0.886 (YNHH), 0.854–0.900 (external hospitals), and 0.853 (ELSA-Brasil) with sensitivities ~88–96% and specificities ~51–66%.
• Generalized to smartphone photos of ECGs and predicted 2–4× higher risk of incident SHD/HF in clinical cohorts and UK Biobank, independent of traditional risk factors.

Clinical Implications

ECG-image AI screening can triage patients for echocardiography, prioritize valve disease and LV dysfunction detection, and longitudinally flag patients at high risk for SHD/HF, potentially embedding into routine ECG workflows (including smartphone-based capture).

Limitations

• Labeling relies on echocardiography within 30 days; potential misclassification and spectrum bias.
• Retrospective development; prospective impact and cost-effectiveness trials are needed to confirm clinical utility.

Future Directions

Prospective, randomized deployment trials assessing clinical impact, workflow integration, and health economics; calibration for diverse devices and care settings; bias auditing across demographic subgroups.