Artificial Intelligence-Enhanced Auscultation to Identify Left Ventricular Dysfunction - Development and External Validation in Prospective Multicenter Studies.
AI auscultation accurately detected left ventricular dysfunction (LVEF ≤40%) among patients referred for clinically indicated echocardiography, supporting its potential as a scalable, accessible pre-echocardiographic risk-enrichment tool, though further validation is warranted before broader screening applications.
Key Findings
Results
A four-site AI auscultation ensemble model detected LVD with high discrimination in the multicenter training cohort and outperformed ECG-only and clinical logistic models.
Training cohort included 2359 patients (mean age 69.7±13.5 years; 40.4% female; 8.9% with reduced LVEF).
AUROC was 0.926 (95% CI, 0.911 to 0.939) in the training cohort.
The ensemble combined a transformer-based network and a one-dimensional convolutional neural network.
Comparators included ECG-only models and clinical logistic regression models incorporating clinical variables, single-lead ECG-derived features, and BNP levels.
Results
In an independent external validation cohort, the four-site AI auscultation model maintained high performance for detecting LVD.
External validation cohort: 964 patients (mean age 68.6±15.4 years; 38.9% female; 7.6% with reduced LVEF) from another independent multicenter registry.
AUROC was 0.904 (95% CI, 0.876 to 0.930).
Sensitivity, specificity, and accuracy were 86.3%, 74.2%, and 75.1%, respectively.
LVD was defined as LVEF ≤40% on concurrent echocardiography.
Results
AI auscultation outperformed cardiologist consensus in accuracy in a 100-patient reader substudy.
Accuracy was 82.0% for AI versus 62.0% for cardiologist consensus (P=0.004).
The reader substudy involved seven cardiologists.
Results
Model performance was consistent across prespecified subgroups.
The abstract states performance was consistent across prespecified subgroups; specific subgroup values are not given in the abstract.
Results
A simplified two-site auscultation model retained high accuracy in the external validation cohort.
Two-site model AUROC was 0.903 (95% CI, 0.878 to 0.927) in the external validation cohort.
This was nearly identical to the four-site model AUROC of 0.904.
Methods
The study was a prospective, multicenter design using digital auscultation at four standard sites with simultaneous single-lead ECG in patients referred for echocardiography.
Adult patients referred for clinically indicated echocardiography were enrolled.
Model was trained on multicenter data and externally validated on a separate registry.
Funded by AMI Inc. and the Japan Society for the Promotion of Science KAKENHI program (grant 25K19371).
What This Means
This study tested whether an artificial intelligence program that analyzes heart sounds recorded with a digital stethoscope could identify people whose heart's main pumping chamber is weak (left ventricular dysfunction, defined as an ejection fraction of 40% or lower). Heart sounds and a single-lead ECG were recorded from adults already scheduled for an echocardiogram. The model was built using 2359 patients and then tested on a separate group of 964 patients from different centers.
In the separate test group, the AI distinguished weak from normal pumping function well (AUROC 0.904). It correctly flagged about 86% of those with weak pumping and correctly cleared about 74% of those without. In a smaller comparison of 100 patients, the AI was more accurate than a consensus of seven cardiologists (82% vs. 62%). It also did better than models using ECG alone or clinical information and blood test results. A simpler version using only two listening positions performed about as well.
This research suggests AI-assisted stethoscope exams could eventually help decide who most needs an echocardiogram, especially where access to heart imaging is limited. However, the study only included people already referred for echocardiography, and the authors note that further validation is needed before the tool is used for wider screening of the general population.
Sato E, Izumi Y, Saito Y, Saijo Y, Kitai T, Hayashida A, et al.. (2026). Artificial Intelligence-Enhanced Auscultation to Identify Left Ventricular Dysfunction - Development and External Validation in Prospective Multicenter Studies.. NEJM AI. https://doi.org/10.1056/AIoa2600258