Abstract Body (Do not enter title and authors here): Background
Evaluation of left ventricular diastolic function is integral to diagnosing heart failure with preserved ejection fraction (HFpEF), but echocardiography is resource-intensive and not always available, leading to delayed detection. We developed an AI-enhanced ECG (AI-ECG) model to detect echocardiography-determined diastolic dysfunction (DD).
Methods
The AI-Risk Estimator for DD (AIRE-DD) is a residual neural network with a discrete-time survival loss function. It was trained on 89,100 ECG-echocardiography (ECG–TTE) pairs from Beth Israel Deaconess Medical Center (BIDMC) and externally validated in four cohorts: King’s College Hospital (KCH; n=1 635), CODE (Brazil; n=882,212), ELSA-Brasil (n=13,739) and UK Biobank (UKB; n = 65 610).
Results
In the BIDMC holdout set (n = 35,760 ECG–TTE pairs), AIRE-DD detected increased LV filling pressures with an AUC of 0.883 (95 % CI 0.874–0.891), sensitivity 0.780, specificity 0.822, PPV 0.564 and NPV 0.927. AIRE-DD also detected grades of diastolic dysfunction with AUCs of 0.792 (≥ grade I), 0.882 (≥ grade II) and 0.906 (grade III). For incident DD prediction in BIDMC participants without baseline DD and LVEF ≥ 50 %, AIRE-DD achieved a C-index of 0.751 (95 % CI 0.719–0.791). In KCH, AIRE-DD identified clinician-confirmed HFpEF with an AUC of 0.850 (0.809–0.887).
In BIDMC, increased filling pressures predicted by AIRE-DD stratified incident outcomes at least as well as echocardiography (age- and sex-adjusted hazard ratios [HRs]: mortality 2.09 vs 1.98; atherosclerotic cardiovascular disease 2.02 vs 1.70; atrial fibrillation 2.12 vs 1.79; heart failure 2.24 vs 2.40; chronic kidney disease 1.93 vs 1.70). Across BIDMC, CODE, ELSA-Brasil and UKB, AIRE-DD-predicted increased filling pressures were associated with age- and sex-adjusted HRs for all-cause mortality of 2.05 (1.87–2.26), 2.83 (2.75–2.92), 4.38 (3.46–5.52) and 1.67 (1.33–2.09), respectively.
Explainability analyses showed that AIRE-DD predictions correlated with broad QRS morphology, T-wave inversion/flattening and poor precordial R-wave progression, and were associated with echocardiographic metrics of impaired relaxation, chamber enlargement and myocardial remodelling.
Conclusion
AIRE-DD provides a non-invasive, scalable method for detection and prediction of diastolic dysfunction and stratification of mortality risk, supporting its potential as a first-line screening tool to prioritise patients for confirmatory imaging and early intervention.