Deep Learning Algorithms to Detect Murmurs Associated With Structural Heart Disease.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Cardiac auscultation accuracy varies significantly among healthcare professionals, potentially delaying treatment for structural heart disease.
• Few machine learning (ML) algorithms for cardiac auscultation have reached clinical practice despite growing interest.
• A novel suite of deep learning-trained, FDA-cleared algorithms was evaluated for heart sound analysis.

Purpose of the Study:

• To evaluate the performance of a novel suite of FDA-cleared deep learning algorithms for detecting structural heart murmurs.
• To compare the algorithm's diagnostic accuracy against human clinicians in a simulated clinical setting.

Main Methods:

• Algorithms were trained on over 15,000 heart sound recordings.
• Validation involved 2,375 recordings from 615 subjects, collected using digital stethoscopes in clinical settings.
• Echocardiograms served as the gold standard for diagnosis, and algorithm performance was compared to 10 clinicians.

Main Results:

• The algorithm demonstrated a sensitivity of 85.6% and specificity of 84.4% for detecting structural murmurs.
• Performance improved to 97.9% sensitivity and 90.6% specificity for clearly audible adult murmurs.
• The algorithm's average accuracy (84.7%) surpassed that of clinicians (77.9%), with consistent performance.

Conclusions:

• The evaluated ML algorithms accurately identify murmurs linked to structural heart disease.
• The study highlights the algorithm's consistency compared to significant interobserver variability among clinicians.
• Implementing ML algorithms in clinical practice could enhance structural heart disease detection and patient management.