Contactless optical phonocardiogram device with deep learning for detection of valvular heart disease

Background

Valvular heart disease (VHD), including stenotic, regurgitant, and multi-valve pathology lesions, frequently remains underdiagnosed due to asymptomatic presentation, limitations of conventional auscultation, and delayed access to echocardiography1. Moderate to Severe VHD is associated with substantial morbidity and mortality, often progressing to heart failure when intervention is delayed2. Earlier detection may enable timely referral for minimally invasive transcatheter valve interventions. Contactless optical devices3 capture cardiac vibration signals which, when combined with deep learning, may improve detection sensitivity and reduce operator dependence, enabling primary care adoption.

Purpose

To evaluate the diagnostic performance of the optical stethoscope with deep learning for detection of clinically significant VHD from Carotid-only and Valve-area-only recordings, benchmarked against transthoracic echocardiography (TTE).

Methods

Data derived from an ongoing diagnostic accuracy study (n=117, phase-1) and a completed cohort (n=25). Participants were categorized into: (i) no/non-significant VHD (healthy to mild-to-moderate) and (ii) clinically significant VHD (mild-to-moderate to severe), simulating real-world screening using TTE as the reference.

Initial model development used a multi-cohort dataset (>25,000 recordings; >50 hours). The prospective cohort was split: 50% (n=58) for self-supervised pre-training and fine-tuning; 50% (n=59) strictly held out for validation. Robustness was assessed via subject-level stratified cross-validation across anatomical recording sites.

Results

High performance was observed across anatomical region groupings. Carotid-only recordings achieved accuracy 92.6 ± 1.5%, sensitivity 89.6 ± 1.9%, and specificity up to 100%. Valve-area recordings demonstrated comparable performance (accuracy 94.3 ± 2.1%, sensitivity 93.8 ± 3.2%, specificity 95.2 ± 1.6%), confirming equivalent diagnostic capability between neck and chest acquisition sites. Permutation testing and bootstrapping confirmed performance exceeded chance (p<0.001).

Conclusion

The contactless optical phonocardiogram device combined with deep learning demonstrates high accuracy for detecting clinically significant VHD against TTE. Recording from the carotid region alone achieved strong diagnostic performance, indicating that neck-based acquisition is sufficient for clinically meaningful screening without multi-site chest recordings. In European and U.S. populations, where moderate-to-severe valve disease prevalence in older adults is approximately 6–10%, the observed sensitivity and specificity translate to a strong negative predictive value (>99%) in primary care settings. This approach may enable earlier identification and triage for confirmatory imaging and timely valve intervention, improving workflow efficiency and healthcare resource allocation.