Abstract
Background: Aortic stenosis (AS) is frequently identified at an advanced stage after clinical symptoms appear. The aim of this systematic review and meta-analysis is to evaluate the diagnostic accuracy of artificial intelligence (AI) algorithms for AS screening.Methods: We conducted a thorough search of six databases. Several evaluation parameters, such as sensitivity, specificity, diagnostic odds ratio (DOR), negative likelihood ratio (NLR), positive likelihood ratio (PLR), and area under the curve (AUC) value were employed in the diagnostic meta-analysis of AI-based algorithms for AS screening. The AI algorithms utilized diverse data sources including electrocardiograms (ECG), chest radiographs, auscultation audio files, electronic stethoscope recordings, and cardio-mechanical signals from non-invasive wearable inertial sensors.Results: Of the 295 articles identified, 10 studies met the inclusion criteria. The pooled estimates for AI-based algorithms in diagnosing AS were as follows: sensitivity 0.83 (95% CI: 0.81-0.85), specificity 0.81 (95% CI: 0.79-0.84), PLR 4.78 (95% CI: 3.12-7.32), NLR 0.20 (95% CI: 0.13-0.28), and DOR 27.11 (95% CI: 14.40-51.05). The AUC value was 0.909 (95% CI: 0.889-0.929), indicating outstanding diagnostic accuracy. Subgroup and meta-regression analyses showed that continent, type of AS, data source, and type of AI-based method constituted sources of heterogeneity. Furthermore, we demonstrated proof of publication bias for DOR values analyzed using Egger's regression test (P = 0.002) and a funnel plot.Conclusion: Deep learning approaches represent highly sensitive, feasible, and scalable strategies to identify patients with moderate or severe AS.