Diagnostic accuracy of deep learning models for detecting valvular heart diseases using echocardiography: a systematic review and meta-analysis

Valvular heart disease is a major global cause of cardiovascular morbidity and mortality, and timely diagnosis remains challenging despite echocardiography being the diagnostic gold standard. Recent advances in deep learning have shown promising accuracy in echocardiographic assessment of valvular disease. The purpose of this systematic review and meta-analysis was to assess deep learning models’ diagnostic accuracy for valvular heart disease diagnosis using echocardiography. A systematic search of the literature was performed in MEDLINE, Scopus, and Web of Science to identify relevant studies published up to December 1, 2025, using keywords related to deep learning, valvular heart disease, and echocardiography. Study selection was conducted through sequential screening of titles, abstracts, and full-text articles in accordance with predefined eligibility criteria. A bivariate random-effects meta-analysis was applied to pool diagnostic accuracy measures, and subgroup analyses were conducted to explore potential sources of heterogeneity. A total of 749 records were identified, of which 16 studies met the inclusion criteria and were included in the meta-analysis. Pooled diagnostic performance indicated a sensitivity of 0.91 (95% CI: 0.88–0.93) and a specificity of 0.89 (95% CI: 0.83–0.92). Both parameters exhibited considerable between-study variability, with heterogeneity indices reaching I² = 99% ( p < 0.001). Furthermore, analysis of the summary receiver operating characteristic curve revealed an overall AUC of 0.96 (95% CI: 0.93–0.97), while the combined diagnostic odds ratio was estimated at 77.2 (95% CI: 44.57–133.73). Sensitivity analyses yielded broadly comparable pooled estimates across the assessed analyses, and assessment of publication bias using Deeks’ asymmetry test did not indicate a statistically significant effect ( p = 0.30). The findings of this systematic review and meta-analysis suggest that deep learning–driven approaches applied to echocardiography show promising diagnostic performance for the detection of valvular heart disease. However, the certainty of evidence was rated as very low because of substantial between-study heterogeneity and methodological limitations. Therefore, the pooled sensitivity, specificity, and AUC should be interpreted as exploratory summary measures rather than stable or generalizable estimates of diagnostic performance. Multicenter validation, standardized imaging protocols, external validation, and improved interpretability are essential before routine clinical implementation.