The eAS deep learning model using phonocardiogram and single-lead ECG signals achieved an AUROC of 0.9632 with 92.9% sensitivity and 87.9% specificity for detecting moderate-or-greater aortic stenosis in a multicenter external validation cohort, suggesting AI-based PCG/ECG analysis may support decisions to refer patients for echocardiography.
The eAS deep learning model achieved high overall diagnostic accuracy for detecting moderate-or-greater aortic stenosis in the external validation cohort.
Using a pre-determined threshold, the eAS model demonstrated high sensitivity and specificity for moderate-or-greater AS detection.
Model-derived probability scores showed strong associations with echocardiographic measures of aortic stenosis severity.
The external validation cohort exhibited more advanced valvular disease than the training/development cohort.
Exploratory subgroup analyses showed generally consistent model performance, though precision was limited in small subgroups.
The eAS model uses phonocardiogram and single-lead ECG signals acquired via a dedicated PCG device called Cardio-EGG for aortic stenosis screening.
This research suggests that an artificial intelligence model trained to analyze heart sound recordings (phonocardiograms) and simple single-lead heart electrical tracings (ECGs) can accurately identify patients with moderate-to-severe aortic stenosis — a condition where the heart's main outflow valve narrows and restricts blood flow. The study tested the model across 951 patients at three different medical centers, finding that it correctly identified about 93% of patients who had significant aortic stenosis while correctly clearing about 88% of patients who did not have the condition. The device used to collect the recordings, called Cardio-EGG, is simpler and more portable than the ultrasound imaging (echocardiography) currently needed to definitively diagnose the condition. Aortic stenosis is particularly dangerous because it often causes no symptoms until it has progressed to a serious stage. Traditional stethoscope-based heart sound listening by clinicians has limited accuracy for detecting it reliably. This research suggests that a brief recording with a specialized device analyzed by AI could serve as an effective first-line screening tool — helping doctors decide which patients should be referred for a full echocardiogram. The model's probability scores also tracked closely with how severe the disease actually was on echocardiography, suggesting it captures meaningful information about disease progression, not just a yes/no signal. The practical implication of this research is that AI-assisted heart sound analysis could potentially extend reliable aortic stenosis screening beyond specialist cardiology settings to primary care offices, community clinics, or other locations where echocardiography is not immediately available. This could help more patients get timely diagnosis and treatment before the disease reaches dangerous stages. However, the study notes that estimates in smaller patient subgroups were less precise, and further research would be needed to confirm performance across diverse populations.
Saito S, Ishii M, Ogawa S, Seki H, Ikeda K, Kayano Y, et al.. (2026). Diagnostic Performance of a Deep Learning Model Using Phonocardiogram and Single-Lead ECG Signals for Detecting Moderate-or-Greater Aortic Stenosis: Multicenter External Validation.. Clinical cardiology. https://doi.org/10.1002/clc.70498