Algorithm for predicting valvular heart disease from heart sounds in an unselected cohort

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Valvular heart disease (VHD) affects a significant portion of the population.
• Early detection of VHD is crucial for effective management and improved patient outcomes.
• Current diagnostic methods can be invasive or require specialized equipment.

Purpose of the Study:

• To evaluate the efficacy of advanced machine learning algorithms in identifying VHD using digital heart sound recordings.
• To assess the performance of these algorithms across different stages of VHD, including asymptomatic individuals.
• To compare algorithm performance with traditional diagnostic standards.

Main Methods:

• A recurrent neural network was developed and trained using heart sound recordings from 2,124 participants in the Tromsø7 study.
• Recordings were collected from four auscultation points using digital stethoscopes.
• The model predicted murmurs, which were then used to diagnose VHD confirmed by echocardiography.

Main Results:

• The algorithm achieved high sensitivity (90.9%) and specificity (94.5%) for detecting aortic stenosis (AS), with an AUC of 0.979.
• Performance for detecting moderate or greater aortic regurgitation (AR) and mitral regurgitation (MR) was moderate (AUCs 0.634 and 0.549, respectively).
• Incorporating clinical variables significantly improved AR and MR detection (AUCs 0.766 and 0.677), with higher accuracy for symptomatic cases.

Conclusions:

• Machine learning models demonstrate excellent potential for AS detection in general populations.
• Detection of AR and MR using heart sounds alone is challenging but improves with clinical data integration and focus on symptomatic patients.
• This AI-driven approach may enhance non-invasive VHD screening and diagnosis.