The ECG-AI model achieved an AUC of 0.79 for detecting HFpEF up to six months before clinical diagnosis, demonstrating its potential for early identification.
Does an ECG-AI model accurately detect HFpEF years before its clinical diagnosis?
An AI algorithm applied to routine ECGs can identify patients with HFpEF up to three years before their formal clinical diagnosis.
Absolute Event Rate: 0% vs 0%
Abstract Background Heart failure with preserved ejection fraction (HFpEF) accounts for nearly 50% of all heart failure cases but remains significantly underdiagnosed due to its heterogeneity. Compared to heart failure with reduced ejection fraction (HFrEF), treatment options for HFpEF are limited, making prevention and early detection essential to reduce its burden. Using ECGs obtained within one month of diagnosis, we previously developed and externally validated an ECG-AI model capable of detecting HFpEF. Goal This study aims to evaluate Wake Forest’s ECG-AI model using ECGs obtained prior to the first clinical diagnosis of HFpEF, to determine whether early recognition of HFpEF from ECG data alone is feasible and underutilized. Methods The ECG-AI model used in this study is a ResNet-based deep learning algorithm originally developed and validated to detect both left ventricular dysfunction (LVD) and HFpEF. It was trained on over 1 million ECGs from Wake Forest School of Medicine (Winston-Salem, NC) and externally validated on more than 100,000 ECGs from the University of Tennessee Health Science Center (Memphis, TN). The original model was designed for detection using ECGs obtained within 30 days of LVD or HFpEF diagnosis. In this study, we applied the same ECG-AI model to a novel cohort that was not included in the original training set, using historical ECGs obtained prior to HFpEF diagnosis. We evaluated its performance by calculating AUCs across different prediction windows. For HFpEF cases, we used ECGs from up to 6 months before diagnosis as a baseline for detection accuracy (AUC), and compared it with older ECGs (6 months prior) to assess how early the model could detect HFpEF without compromising accuracy. For controls, all available ECGs across all time windows were used. DeLong’s test was applied to compare AUCs from each prediction window to the baseline AUC. Results The analytical cohort comprised 62,146 ECGs from 12,441 patients at Wake Forest Baptist (54% female, 19% Black, mean age 57 ± 16 years). Of these, 922 ECGs (1.5%) were from 305 patients later diagnosed with HFpEF. Using ECGs obtained up to six months prior to diagnosis, the ECG-AI model achieved an AUC of 0.79 (95% CI: 0.76–0.82) for HFpEF detection. AUCs were also calculated for additional prediction windows. As shown in Figure 1, the ECG-AI model was able to detect HFpEF up to three years before clinical diagnosis, without significant loss of accuracy. Conclusion Wake Forest’s ECG-AI model can detect HFpEF years before its clinical diagnosis. Future research is needed to determine whether the model is identifying undiagnosed HFpEF or predicting future risk.
Barr et al. (Thu,) reported a other. The ECG-AI model achieved an AUC of 0.79 for detecting HFpEF up to six months before clinical diagnosis, demonstrating its potential for early identification.