Improving Valvular Pathologies and Ventricular Dysfunction Diagnostic Efficiency Using Combined Auscultation and...

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence in Medicine

Background:

• Early diagnosis of valvular heart disease and ventricular dysfunction is crucial for timely intervention.
• Traditional diagnostic methods like echocardiography can be resource-intensive.
• Sensor-based methods like auscultation and ECG offer accessible preliminary screening.

Purpose of the Study:

• To evaluate the efficacy of integrating machine learning with sensor-based data (auscultation and ECG) for diagnosing valvular abnormalities and left ventricular dysfunction.
• To assess the diagnostic performance of a composite sensor model compared to individual sensor contributions.
• To explore the potential of multimodal artificial intelligence in cardiovascular diagnostics.

Main Methods:

• Collected simultaneous auscultation and 12-lead ECG data from 1052 patients.
• Utilized a fourfold cross-validation approach to train neural networks on heart sounds and ECG leads.
• Employed a stacking technique to combine outputs from individual neural networks for a composite model.
• Validated the model on an independent cohort of 103 patients screened for severe aortic stenosis, severe mitral regurgitation, and left ventricular dysfunction (ejection fraction ≤ 40%).

Main Results:

• The composite sensor model demonstrated high diagnostic efficiency: AUC of 0.93 for aortic stenosis, 0.80 for mitral regurgitation, and 0.75 for left ventricular dysfunction.
• Individual sensor contributions varied depending on the specific condition being diagnosed, highlighting synergistic potential.
• The machine learning model successfully identified severe aortic stenosis, severe mitral regurgitation, and left ventricular dysfunction.

Conclusions:

• Machine learning models integrating auscultation and ECG data can effectively detect cardiovascular conditions typically diagnosed by imaging.
• The sensor fusion approach leveraging multimodal AI shows significant promise for efficient and accessible cardiovascular diagnostics.
• This study underscores the potential of combining non-invasive sensor data with AI for improved patient screening and management.