An Artificial Intelligence Model for ECG-Based Prediction of Heart Failure with Preserved Ejection Fraction Diagnosis

Background: Heart failure with preserved ejection fraction (HFpEF) accounts for over half of all heart failure (HF) cases and remains diagnostically challenging due to nonspecific symptoms and a lack of accessible noninvasive screening tools, leading to widespread underdiagnosis and delayed treatment. The electrocardiogram (ECG) is a low-cost, widely available tool that reflects myocardial electrical remodeling. We previously developed and externally validated an ECG-based artificial intelligence (ECG-AI) model capable of classifying ECGs into four categories: reduced ejection fraction (rEF), mid-range ejection fraction (mEF), HFpEF, and controls. In this study, we evaluate the ability of this ECG-AI model to predict future HFpEF diagnosis using real-world data from a large integrated health system. Methods: We applied the validated ECG-AI model to an independent cohort of 7713 patients from Wake Forest Baptist Health (WFBH), using one ECG per patient recorded before the clinical diagnosis of HFpEF. Model discrimination was evaluated across multiple prediction windows from six months up to ten years before diagnosis using the area under the receiver operating characteristic curve (AUC), with comparisons performed using DeLong’s test. Performance was also compared head-to-head with the H2FPEF score, a validated clinical score for HFpEF, in the subset of patients with complete data for score calculation. ECG-AI outputs were stratified into quartiles and evaluated using Kaplan–Meier survival analysis and multivariable Cox proportional hazards regression adjusted for demographics and major comorbidities. Results: Of the 7713 patients, 283 (3.7%) were subsequently diagnosed with HFpEF. The ECG-AI model achieved an AUC of 0.79 (95% CI, 0.72–0.86) using ECGs recorded within six months before diagnosis. Discrimination remained stable across prediction windows extending up to ten years before diagnosis (AUC range, 0.78–0.80; all DeLong p > 0.01 vs. the six-month window). Compared with the H2FPEF score, ECG-AI demonstrated superior discrimination, improving the AUC by 0.06–0.07 across the evaluated prediction windows. Patients in the highest ECG-AI risk quartile had an unadjusted hazard ratio (HR) of 11.18 (95% CI, 7.27–17.21; C-index, 0.75) and an adjusted HR of 6.60 (95% CI, 4.24–10.28; C-index, 0.82). Adding clinical covariates to ECG-AI improved the C-index by 0.07 and 0.09 for the five- and ten-year prediction windows, respectively. Conclusions: An independently validated ECG-AI model predicted incident HFpEF up to ten years before clinical diagnosis and outperformed the H2FPEF score using ECG alone, indicating that ECG-derived signatures precede clinical recognition of the syndrome. ECG-AI shows promise as a prognostic, risk-stratification tool to prioritize further evaluation; prospective validation and single-lead assessment are needed before screening or wearable deployment.