Combining AI-enabled acquisition and interpretation for screening aortic stenosis: a prospective validation and pragmatic workflow
E Lee, J A Naser, C J Kane, J G Malins, J J Thaden, V T Nkomo, T J Poterucha, P A Pellikka, P A Friedman, J K Oh, G C Kane, Z I Attia, S V Pislaru, J G Bird, G TsabanAbstract
Background
Timely identification of moderate or greater aortic stenosis (AS) is critical for appropriate management but is constrained by reliance on comprehensive echocardiography. We developed and validated a deep learning algorithm for AS detection from transthoracic echocardiography (TTE) and evaluated its performance on focused cardiac ultrasound (FoCUS) acquired by novice operators.
Methods
The algorithm was trained on TTE studies from 6,753 patients at a large healthcare network in the United States Midwest, with internal validation (n=852) and testing (n=844). External validation was performed in cohorts from Arizona (n=959) and Florida (n=953). Diagnostic performance was assessed on handheld ultrasound acquired by experienced sonographers (n=602) and in a prospective cohort by true novice operators (n=1,303) performing AI-guided FoCUS within 15 minutes of comprehensive TTE.
Results
Model discrimination was excellent, with an area under the receiver operating characteristic curve (AUROC) of 0.99 (95% CI, 0.98-1.00) internally and 0.99 (95% CI, 0.96/0.97-1.00) in both external cohorts. On handheld ultrasound acquired by experienced sonographers, sensitivity was 95% (95% CI 82%-99%) and specificity was 97% (95% CI 95%-98%). In the prospective novice-operator FoCUS cohort (3.7% >moderate AS prevalence), 96.6% of studies were suitable for AI analysis, with sensitivity of 93% (95% CI, 82%-99%) and specificity of 96% (95% CI, 95-97%). Expert overread of AI-positive and uninterpretable cases (10% of cases) improved positive predictive value from 49.4%to 91.1%.
Conclusions
Deep learning enables accurate detection of at least moderate AS from both TTE and FoCUS. Robust performance on AI-guided FoCUS acquired by novice operators supports scalable, low-resource AS screening in settings with limited echocardiography access.Graphical summary of the studySuggested screening workflow