Technical Summary: AI-Enabled Digital Auscultation for Adult Valvular Heart Disease

Problem Statement

Timely recognition of clinically significant valvular heart disease (VHD) in adults is currently constrained by the variability of human cardiac auscultation and limited access to echocardiography. While echocardiography remains the diagnostic reference standard, indiscriminate screening is inefficient due to the high prevalence of benign murmurs (20–30%) versus the lower prevalence of moderate-to-severe VHD (2.5–3.3%). Existing artificial intelligence (AI) studies have shown promise in detecting murmurs or specific lesions, but there is a lack of synthesized evidence regarding the clinical utility of AI-enabled digital auscultation as a pre-echocardiographic triage tool. A critical gap exists in distinguishing between the detection of audible murmurs and the identification of echocardiography-confirmed, clinically significant VHD (defined as moderate-or-greater stenosis/regurgitation or disease requiring intervention). Furthermore, the performance of these systems varies significantly based on lesion type, deployment setting, and validation design, necessitating a focused evaluation of their diagnostic accuracy and downstream clinical yield.

Methodology

This study is a systematic review and diagnostic test accuracy meta-analysis conducted in accordance with PRISMA-DTA guidelines and registered in INPLASY.

• Data Sources: A comprehensive search of PubMed, Web of Science, Cochrane CENTRAL, and Embase was performed up to June 30, 2026.
• Eligibility Criteria: The review included studies involving adults (≥18 years) undergoing AI-enabled digital auscultation or phonocardiography (PCG) with an echocardiographic reference standard. Studies were required to provide extractable data to reconstruct 2×2 diagnostic accuracy tables for clinically significant VHD or specific valve lesions.
• Exclusion Criteria: Studies lacking an echocardiographic reference standard, relying solely on public dataset labels, focusing on pediatric populations, or lacking extractable diagnostic accuracy data were excluded.
• Data Extraction: Two reviewers independently extracted data on study characteristics, participant demographics, index test protocols (acquisition sites, duration, preprocessing), AI model architectures, and diagnostic outcomes.
• Statistical Analysis:
  • Primary Synthesis: A bivariate random-effects meta-analysis (Reitsma approach) was used to pool sensitivity and specificity for clinically significant VHD.
  • Lesion-Specific Analysis: Separate bivariate pooling was conducted for Aortic Stenosis (AS), Mitral Regurgitation (MR), Aortic Regurgitation (AR), and Mitral Stenosis (MS) where at least four independent cohorts were available.
  • Heterogeneity & Subgroups: Heterogeneity was assessed via HSROC plots and leave-one-out analyses. Subgroup analyses explored the effects of severity thresholds, auscultation location, clinical setting (screening vs. enriched), commercial status, and validation credibility.
  • Clinical Yield Modeling: Diagnostic accuracy was translated into expected outcomes per 1,000 tested adults across different prevalence scenarios to calculate Positive Predictive Value (PPV) and Number Needed to Echo (NNE).
  • Risk of Bias: Quality assessment was performed using QUADAS-3, and certainty of evidence was graded using GRADE-DTA.

Key Contributions

1. Differentiation of Murmur vs. Disease: The review explicitly addresses the clinical distinction between AI detection of murmurs and the detection of clinically significant VHD, highlighting that improvements in murmur detection do not automatically equate to improved detection of hemodynamically significant valve disease.
2. Lesion-Specific Performance Profiles: The study provides granular diagnostic accuracy estimates for specific valve lesions, moving beyond aggregate VHD metrics to reveal significant performance disparities between lesions (e.g., high accuracy for AS vs. lower, more heterogeneous accuracy for MR).
3. Context-Dependent Yield Modeling: By modeling downstream consequences, the paper quantifies how disease prevalence impacts the clinical utility of AI triage, demonstrating that false-positive referrals dominate in low-prevalence settings while yield improves in enriched populations.
4. Validation Credibility Assessment: The analysis stratifies results by validation design (e.g., external/prospective vs. internal/cross-validation) and commercial status, offering insights into the robustness of current AI claims.

Results

• Study Base: 13 studies comprising 9,790 participants were included. Four independent validation cohorts (2,730 participants; 320 cases) contributed to the primary meta-analysis for clinically significant VHD.
• Overall Diagnostic Accuracy: For clinically significant VHD, the pooled sensitivity was 69.9% (95% CI, 47.9–85.4) and specificity was 75.6% (95% CI, 58.4–87.3).
• Lesion-Specific Performance:
  • Aortic Stenosis (AS): Showed the strongest performance with a pooled sensitivity of 90.4% (95% CI, 86.5–93.3) and specificity of 88.0% (95% CI, 78.0–93.8) across 12 studies. Multi-site recording significantly improved AS sensitivity compared to single-site recording.
  • Mitral Regurgitation (MR): Demonstrated lower and more heterogeneous performance, with pooled sensitivity of 70.2% (95% CI, 61.0–77.9) and specificity of 77.1% (95% CI, 64.0–86.4). Performance was highly sensitive to clinical setting and validation design.
  • Other Lesions: AR and MS showed wider uncertainty due to fewer studies; TR, PR, and PS could not be pooled due to insufficient independent cohorts.
• Clinical Yield: In low-prevalence community settings, AI-positive results were dominated by false positives (e.g., ~238 false referrals per 1,000 tested for VHD). In enriched or referred populations, the Number Needed to Echo (NNE) improved significantly (from ~15 in community settings to ~3 in referred settings).
• Comparator Performance: AI-enabled auscultation generally outperformed standard primary-care auscultation and general practitioner reviews in sensitivity, though specificity varied. Comparisons with cardiologist benchmarks showed mixed results, with AI performing comparably or slightly better in specific lesion detection tasks but not universally superior.
• Quality of Evidence: The certainty of evidence was rated as low for AS and very low for clinically significant VHD and MR, primarily due to risk of bias (enriched cohorts, internal validation), inconsistency, and imprecision.

Significance and Claims

The paper concludes that AI-enabled digital auscultation is a promising but context-dependent tool for pre-echocardiographic triage. It explicitly states that current evidence does not support the use of these systems as a replacement for echocardiography or as a broad, indiscriminate screening tool for all adult VHD in low-prevalence populations.

Instead, the authors claim the technology is best positioned as a targeted triage aid for:

• Patients with a higher pre-test probability of clinically relevant valve disease (e.g., symptomatic patients, older adults with suspected cardiac disease).
• Settings where echocardiography access is constrained.
• Specific lesion detection, particularly Aortic Stenosis, where performance is robust.

The study emphasizes that the clinical value of AI auscultation is inextricably linked to the target population, lesion type, and the implementation pathway. It argues that future deployment must focus on defining clear triage pathways, managing false-positive burdens, and validating performance in real-world, unselected populations rather than relying solely on diagnostic accuracy metrics from selected cohorts. The authors call for future research to shift from diagnostic accuracy estimation toward pathway-level evaluation, including outcomes such as time to diagnosis, missed disease rates, and cost-effectiveness.