Embedded Files
A publicly available 12-lead ECG AI model (ECG2HF) predicted 10-year incident Heart Failure (CHF) across 3 health systems, with consistent discrimination (AUC 0.86 at MGH, 0.85 at BWH, 0.84 at BIDMC) in 93,868 adults without HF. In this multi-center validation study, ECG2HF also improved discrimination vs the Pooled Cohorts Equations to Prevent HF (ΔAUC 0.061 at MGH/BWH; 0.038 at BIDMC) and improved net reclassification (NRI 0.16–0.23).
Why It Matters To Your Practice
Most ECG-based HF prediction models are proprietary; ECG2HF is publicly available, which can speed local evaluation, governance review, and reproducibility.
Consistent performance across centers suggests better transferability than single-site models — a key barrier to deploying AI safely in routine ambulatory care.
Using an already-ubiquitous test (12-lead ECG) could enable scalable HF prevention workflows without adding new imaging or lab requirements.
Clinical Implications
Risk stratification: ECG2HF could help identify patients (age 30–79, no known HF) who merit closer follow-up for HF prevention over a 10-year horizon.
Decision support: The model’s favorable discrimination and reclassification vs Pooled Cohorts Equations may support earlier preventive counseling, risk-factor optimization, and targeted monitoring in higher-risk patients.
Workflow design: Consider how a “high ECG2HF risk” flag would route patients (e.g., primary care review, cardiology referral criteria, echocardiography thresholds, or longitudinal biomarker strategies) before implementation.
Insights
Event rates were similar across sites (10-year cumulative HF risk ~4.4%–5.0%), supporting meaningful external validation in broadly comparable ambulatory populations.
Performance was reported with confidence intervals at each site (e.g., AUC 0.86 [0.84–0.87] at MGH), helping clinicians assess uncertainty rather than relying on a single point estimate.
HF outcomes were identified using a validated EHR-based natural language processing approach, reflecting how real-world labels are often constructed for clinical AI development and validation.
The Bottom Line
Multi-center validation shows ECG2HF can predict 10-year incident Heart Failure (CHF) from standard ECG waveforms with stable AUC (~0.84–0.86) and improved risk classification vs a clinical risk equation, supporting its potential as a generalizable preventive-care triage tool.
Page updated
Report abuse