Abstract 11917: Structural Heart Disease AI-ECG Signatures Forecast New-Onset Heart Failure Across Two External Health Systems
Abstract
Background: AI-ECG models developed to detect structural heart disease (SHD) may capture subclinical myocardial abnormalities. If these models also identify future HF risk when applied off-the-shelf in external health systems, they could support scalable, opportunistic HF risk stratification without new model development. We evaluated whether validated AI-ECG SHD models predict new-onset HF and improve discrimination beyond basic clinical risk factors across two external health systems. Aim: To test whether validated SHD AI-ECG models predict new-onset HF, add incremental value over clinical risk factors, and behave consistently across two external health systems. Methods: Patients without any prior HF diagnosis were identified at Massachusetts General Hospital, MA (MGH) and Emory University Hospital, GA (EUH) using ≥ 1-year look-back ( Fig 1 ). Six validated AI-ECG classifiers for individual SHDs (LV systolic dysfunction, aortic and mitral stenosis/regurgitation, and severe LV hypertrophy) and a model for identifying an ensemble for these SHDs (all developed at a large, diverse health system) were applied without fine-tuning or recalibration. We used Cox models to assess the association of binary AI-ECG screen results at index ECG with future HF risk, and sequentially adjusted for demographics (age, sex) and comorbidities (hypertension, diabetes, obesity). Harrell's C-statistic was used to assess HF risk discrimination. Results: Among 178,537 at-risk patients at MGH and 55,010 in EUH, followed for median 3.6 and 5.9 years, respectively, 9,573 (5.4%) and 9,840 (18.0%) developed HF. All 7 AI-ECG models developed to detect various SHDs predicted new-onset HF in both external health systems without fine-tuning or recalibration. ( Fig 2 ). Acrosss disease-specific models, positive AI-ECG screens conferred 1.7- to 3.6-fold HF risk at MGH and 2.0- to 4.1-fold at EUH. The ensemble SHD model conferred 3.4-fold risk (aHR 3.40 [3.22-3.59]) at MGH and 4.3-fold at EUH (aHR 4.34 [4.12-4.57]). While the LVSD classifier and the ensemble models provided similar HF risk enrichment, the PRESENT-SHD ensemble identified a larger at-risk population, flagging 38-49% patients and capturing 72-75% of future HF events, compared with 10-19% and 31-46% for LVSD. All disease-specific models improved discrimination over demographics and comorbidities (ΔC +0.023 to +0.116; all p<0.001; Fig 3 ). Conclusion: AI-ECG models built to detect structural heart disease forecast new-onset HF years in advance, with large and consistent incremental value over clinical risk factors across two health systems. Convergent signals across independent disease-specific models point to a shared subclinical ECG signature of HF risk and show that these detection models can be repurposed off-the-shelf for opportunistic HF risk stratification.