Artificial intelligence-based predicting incident atrial fibrillation using sinus rhythm electrocardiogram for post-myocardial infarction patients as an image data structure

Atrial fibrillation (AF) is frequently under-detected but linked to cardiovascular comorbidities, including stroke. Preexisting screening modalities are known to have low performance. We aimed to develop a convolutional neural network (CNN)-based predictive tool for individual risk assessment of incident AF using sinus rhythm electrocardiogram (SRECG) data for post-myocardial infarction (MI) patients.

We developed a predictive tool based on a CNN image detection algorithm, merging it with individual patient characteristics to predict incident AF using 12-lead SRECG. We included 4,656 SRECG from 531 post-MI patients without history of AF and allocated the SRECG to training, internal validation, and testing datasets in a 7:1:2 ratio. We classified patients with incident AF within 24 months after MI.

Our study included 392 patients with 3,262 SRECG for training, 58 patients with 466 SRECG for internal validation, and 81 patients with 928 post-MI SRECG for the testing dataset. The mean age was 68.2 years, with mean left atrial diameter and left ventricular ejection fraction measuring 39.3mm and 62.2%, respectively. Furthermore, 77.8% had a history of hypertension, and 45.6% had diabetes. In the testing dataset, 14 (16.9%) patients were confirmed to have incident AF. The area under the curve (AUC) for predicting incident AF in post-MI patients was 0.88 [0.80-0.95], with a sensitivity of 0.83, specificity of 0.87, and an overall accuracy of 89.0% after training.

Our predictive tool, based on a CNN image detection algorithm merged with individual patient information, can identify post-MI patients at risk of incident AF using SRECG.