Artificial intelligence (AI)-based analysis of 12-lead electrocardiograms (ECGs) has similar predictive ability for incident atrial fibrillation (AF) as a clinical risk score, according to a study published online Nov. 8 in Circulation.
Shaan Khurshid, M.D., M.P.H., from Massachusetts General Hospital (MGH) in Boston, and colleagues trained a convolutional neural network (ECG-AI) to infer five-year incident AF risk using 12-lead ECGs in patients receiving care at MGH. Three hazard models were fit and included: ECG-AI five-year AF probability; the Cohorts for Heart and Aging in Genomic Epidemiology AF (CHARGE-AF) clinical risk score; and terms for both ECG-AI and CHARGE-AF (CH-AI). Model performance was assessed in an internal test set and two external test sets (Brigham and Women’s Hospital [BWH] and U.K. Biobank). The training set and test sets included 45,770 and 83,162 individuals, respectively.
The researchers found that the area under the receiver operating characteristic curve (AUROC) was comparable using CHARGE-AF (0.802, 0.752, and 0.732 for MGH, BWH, and U.K. Biobank, respectively) and ECG-AI (0.823, 0.747, and 0.705 for MGH, BWH, and U.K. Biobank, respectively). The highest AUROC was seen with CH-AI (0.838, 0.777, and 0.746 for MGH, BWH, and UK Biobank, respectively). Low calibration error was seen with ECG-AI and CH-AI. The ECG P-wave had the greatest influence on AI model predictions in saliency analyses.
“The application of such algorithms could prompt clinicians to modify important risk factors for atrial fibrillation that may reduce the risk of developing the disease altogether,” Khurshid said in a statement.
Several authors disclosed financial ties to the pharmaceutical industry.