Development and Validation of AI-Enhanced Auscultation for Valvular Heart Disease Screening through a Multi-Centre Study

Abstract

Valvular heart disease (VHD) is a growing public health concern, yet over half of cases remain undiagnosed due to late symptom onset, limited public awareness, and low sensitivity of traditional stethoscope-based screening. Current AI-enabled tools rely on murmur detection as a proxy for VHD but lack sensitivity for common subtypes like mitral regurgitation and are limited by small datasets. This study presents a novel neural network that directly predicts clinically significant VHD from stethoscope recordings, trained using echocardiographic targets rather than heart murmur labels. A diverse dataset of 1,767 patients across UK primary care and hospital settings was developed, combining stethoscope recordings with echocardiographic labels. The trained recurrent neural network achieved an AUROC of 0.83, outperforming general practitioners and demonstrating exceptional sensitivity for severe aortic stenosis (98%) and severe mitral regurgitation (94%). This algorithm shows promise as a scalable, low-cost screening tool, enabling earlier diagnosis and timely referral for intervention. This research was registered with ClinicalTrials.gov (CAIS: NCT04445012 registered on 2020-06-21, DUO-EF: NCT04601415 registered on 2020-10-19).