Background: Individuals with obstructive sleep apnoea (OSA) experience a higher burden of atrial fibrillation (AF) than the general population, and many cases of AF remain undetected. We tested the feasibility of an artificial intelligence (AI) approach to opportunistic detection of AF from single-lead electrocardiograms (ECGs) which are routinely recorded during in-laboratory polysomnographic sleep studies.

Methods: Using transfer learning, an existing ECG AI model was applied to 1839 single-lead ECG traces recorded during in-laboratory sleep studies without any training of the algorithm. Manual review of all traces was performed by two trained clinicians who were blinded to each other's review. Discrepancies between the two investigators were resolved by two cardiologists who were also unaware of each other's scoring. The diagnostic accuracy of the AI algorithm was calculated against the results of the manual ECG review which were considered gold standard.

Results: Manual review identified AF in 144 of the 1839 single-lead ECGs (7.8%). The AI detected all cases of manually confirmed AF (sensitivity = 100%, 95% CI: 97.5-100.0). The AI model misclassified many ECGs with artefacts as AF, resulting in a specificity of 76.0 (95% CI: 73.9-78.0), and an overall diagnostic accuracy of 77.9% (95% CI: 75.9%-97.8%).

Conclusion: Transfer learning AI, without additional training, can be successfully applied to disparate ECG signals, with excellent negative predictive values, and can exclude AF among patients undergoing evaluation for suspected OSA. Further signal-specific training is likely to improve the AI's specificity and decrease the need for manual verification.


atrial fibrillation, machine learning, artificial intelligence, transfer learning, obstructive sleep apnoea

Link to Publisher Version (URL)


Available for download on Sunday, July 17, 2022

Find in your library