Diagnostic Accuracy of artificial intelligence for paroxysmal atrial fibrillation detection using ECGs: A Systematic Review and Meta-Analysis

Abstract

Background: PAF can cause stroke, weaken heart muscles, and lead to blood clot formation. The diagnostic method for PAF is electrocardiography (ECG), which requires physician interpretation. However, the increasing use of AI in healthcare has enabled automated diagnosis of cardiac arrhythmias. PAF, often underdiagnosed owing to its transient nature, can now potentially be detected by AI models applied to ECGs. This study evaluated the diagnostic performance of AI models in detecting PAF. Methods: PubMed, Cochrane Library, and Google Scholar were searched, and studies were included if they assessed the diagnostic accuracy of AI models for PAF. Sensitivity and specificity were the primary outcomes for assessing diagnostic accuracy, and subgroup analyses were performed based on the sample size. Results: A total of A total of 956,664 ECGs from eight studies were assessed using AI models. The pooled sensitivity and specificity of the AI models for PAF detection were 0.8585 (95% CI: 0.85, 0.86) and 0.9771 (95% CI: 0.97, 0.98), respectively. Subgroup analysis revealed higher sensitivity in smaller sample size studies (0.91, 95% CI: 90.3, 91.92) compared to larger sample size studies (0.8579, 95% CI: 85.69, 85.89). Conversely, larger sample size studies demonstrated higher specificity (0.9775, 95% CI: 97.73, 97.77) than smaller sample size studies (0.90, 95% CI: 89.42, 90.69). Conclusion: “These findings suggest that sample size may influence the diagnostic accuracy of AI models. AI models exhibit high diagnostic accuracy for detecting PAF, offering an efficient alternative for clinical diagnosis