ARTIFICIAL intelligence (AI) appears to be able to predict atrial fibrillation (AF) within 1 month amongst patients with sinus rhythm, according to research carried out by Neal Yuan, University of California, San Francisco, USA, and colleagues. The algorithm used in this study was one of few developed to identify hidden AF in normal ECGs in diverse patients. “The AI algorithm worked well across a range of ages, genders, and ethnicities,” added the authors.
The team acquired 907,858 ECGs with normal sinus rhythm (mean age: 62 years; 94% males; 62% White; 11% Black) from the 21 regional Veterans Affairs (VA) sites, spanning from 1987–2022. ECGs collected from the VA sites in San Francisco and Palo Alto, California, USA, were used for model training, validation, and testing. External testing was conducted on 72,483 ECGs from non-VA sites, from 2005–2018 (mean age: 60 years; 53% females; 75% White; 9% Black).
Results showed that the AI model developed by Yuan and colleagues predicted AF within 31 days of sinus rhythm, with an area under the receiver operating characteristic curve of 0.86 (95% confidence interval [CI]: 0.85–0.86), accuracy of 0.78 (95% CI: 0.77–0.78), and an F1 score (which evaluates precision and recall in machine learning) of 0.30 (95% CI: 0.30–0.31), in the ECGs obtained from VA sites. In the group of ECGs collected externally, the AI model predicted AF with an area under the receiver operating characteristic curve of 0.93 (95% CI: 0.93–0.94), accuracy of 0.87 (95% CI: 0.86–0.88) and an F1 score of 0.46 (95% CI: 0.44–0.48). The AI was calibrated to have a Brier score of 0.02 across all sites.
For the individuals deemed high risk by the deep learning AI model, the number needed to screen to detect a positive case of AF was 2.47 individuals for 25% testing sensitivity, and 11.48 for 75%. The AI model performed similarly in patients who were Black, female, or younger than 65 years old. The authors concluded that such AI models have the potential to be implemented in clinical practice soon, provided they are tested on diverse populations beforehand, ensuring the algorithm is tested in populations that have less access to healthcare. These algorithms “should be tested beyond just large academic medical centres,” added the authors.
