A deep-learning ECG model predicted incident AF with AUC 0.82 in FHS and improved prediction over CHARGE score alone with NRI 0.08 to 0.12 in FHS and UK Biobank.
Does an ECG-based deep-learning model predict incident atrial fibrillation comparably to or better than a clinical risk score in community-based populations?
An ECG-based deep-learning model performs comparably to a multivariable clinical risk score for predicting incident atrial fibrillation and provides incremental predictive value when combined.
Absolute Event Rate: 0% vs 0%
Abstract Background Deep-learning models from the electrocardiogram (ECG) can identify individuals with high AF risk, however evaluations in community-based studies are still scarce. Early identification of individuals at higher risk of AF is the first step to implement tailored prevention strategies and may also facilitate diagnosis and prevent complications by enhancing screening of high-risk individuals. Purpose We aimed to refine an ECG-AF model in the Framingham Heart Study (FHS) and compare its performance to predict AF to a clinical risk prediction score in samples from FHS and the UK Biobank. We also evaluated whether ECG-AF could predict other cardiovascular outcomes. Methods We refined the ECG-AF model for AF risk prediction in 60% of the FHS samples free of AF or atrial flutter. We then tested the model’s performance in the remaining FHS samples and in the UK Biobank, comparing it to a clinical risk prediction score (CHARGE). Discrimination was assessed by the area under the receiver operating characteristic curve (AUC) and reclassification by the discrete net reclassification index (NRI). The association of ECG-AF with other cardiovascular outcomes and mortality was assessed using Cox proportional hazards models. Results The study sample included 10,097 FHS participants (mean baseline age 53±12 years; 54.9% women) with 28,151 ECGs, and 49,280 participants from the UK Biobank (mean age 64±8 years, 47.9% women). The ECG-AF model showed moderate discrimination for incident AF (AUC=0.82, 95% CI 0.80-0.84) in the FHS test set, comparable to the CHARGE score (AUC=0.83, 95% CI 0.81-0.85), and incremental when combined (AUC=0.85, 95% CI 0.83-0.87) (Figure 1). In the UK Biobank, the integrative score (CHARGE + ECG-AF) was also incremental to CHARGE alone (AUC=0.81, 95% CI 0.79-0.82 and AUC=0.78, 95% CI 0.76-0.79, respectively) (Figure 2). Compared to CHARGE score alone, the addition of ECG-AF demonstrated improved reclassification for the prediction of AF cases with NRI of 0.08 (95% CI 0.04-0.11) in the FHS and 0.12 (95% CI 0.07-0.14) in the UK Biobank. A higher ECG-AF score was also significantly associated with a higher risk of heart failure, myocardial infarction, and all-cause mortality in both the FHS and UK Biobank. Conclusions In a community-based study, a deep-learning model from ECG had a good performance in predicting AF and other cardiovascular outcomes. This performance was comparable to a multivariable clinical risk score and was incremental when combined. Considering that ECGs are low-cost and widely available, even in areas with limited health access through telemedicine, ECG-AF could be used alone or in combination for AF risk prediction, potentially enabling tailored prevention strategies and more efficient screening.ROC curves for AF risk prediction in FHS ROC curves for AF risk prediction in UKB
Brant et al. (Sat,) reported a other. A deep-learning ECG model predicted incident AF with AUC 0.82 in FHS and improved prediction over CHARGE score alone with NRI 0.08 to 0.12 in FHS and UK Biobank.