AI Auscultation for Left Ventricular Dysfunction
In an independent cohort of 964 patients referred for echocardiography, AI auscultation detected LVEF ≤40% with 86.3% sensitivity and 74.2% specificity.
Researchers at Juntendo University and collaborating institutions developed AI that combines digital heart sounds with single-lead ECG recordings to identify reduced left ventricular function. They tested the model in an independent multicenter cohort and reported the findings in NEJM AI. The intended role is to help identify patients who may benefit from echocardiography.
Assessment before echocardiography
Left ventricular ejection fraction (LVEF) describes the proportion of blood the ventricle ejects with each contraction. The team defined dysfunction as LVEF of 40% or less on concurrent echocardiography.
In this prospective multicenter study, adults referred for clinically indicated echocardiography underwent digital auscultation at four chest sites and simultaneous single-lead ECG recording. The study did not recruit a general screening population.
Independent validation
The team trained the model using data from 2,359 patients. For external validation, it used 964 patients from a separate multicenter registry, where 7.6% had reduced LVEF.
Sensitivity was 86.3% and specificity was 74.2%. Sensitivity measures how many patients with dysfunction the model detected; specificity measures how many patients without it the model correctly classified as negative. Detection and false-positive referrals both matter when considering this use.
External-validation AUC was 0.904 (95% CI 0.876 to 0.930). AUC measures discrimination, with 0.5 representing chance and 1.0 perfect separation. A simplified two-site model achieved an AUC of 0.903. Discrimination remained high with fewer recording sites.
Comparison with cardiologists
In an additional reader substudy involving 100 patients, AI accuracy was 82.0%, versus 62.0% for cardiologist consensus (P=0.004). The researchers also compared the model with ECG-only models and statistical models incorporating clinical information.
Those accuracy figures belong to the 100-patient reader substudy. External-validation accuracy was 75.1% in a different cohort of 964 patients. The results should not be presented as a physician comparison conducted across the full external-validation cohort.
Clinical scope
All participants already had a clinical indication for echocardiography. The authors call for further validation before broader screening. These findings concern a tool for identifying patients for echocardiographic evaluation; the study does not establish that AI can replace echocardiography.
Research note: AMI Inc. and the Japan Society for the Promotion of Science KAKENHI program funded the study. The public abstract does not provide detailed competing-interest disclosures or Japanese regulatory approval information for this AI method.
Sources
- [1]
- [2]
Images: generated with GPT Image 2.5 (illustrative).