A deep neural network using audio files for detection of aortic stenosis

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Aortic stenosis (AS) is a prevalent valvular heart disease, yet many cases go undiagnosed.
• Cardiac auscultation is a key screening tool for AS, but requires significant expertise.
• Undiagnosed AS poses a significant public health challenge.

Purpose of the Study:

• To evaluate the accuracy of a deep neural network (DNN) in detecting AS from audio recordings.
• To compare the AI algorithm's diagnostic performance against cardiologists, residents, and medical students.
• To explore the potential of AI-assisted auscultation in primary care settings.

Main Methods:

• A deep neural network (DNN) was trained using preprocessed audio files from 100 AS patients and 100 controls.
• The DNN's diagnostic accuracy was assessed using a test dataset of 40 patients.
• Performance metrics included sensitivity, specificity, and F1-score, with comparisons to human expert evaluations.

Main Results:

• The DNN achieved a sensitivity of 0.90, specificity of 1, and an F1-score of 0.95 for AS detection.
• The AI's F1-score (0.95) was comparable to that of experienced cardiologists (0.94).
• Residents and medical students showed lower F1-scores (0.88) compared to the DNN and cardiologists.

Conclusions:

• Deep learning-guided auscultation demonstrates accuracy comparable to cardiologists in predicting significant AS.
• AI-assisted auscultation shows promise as a tool to support general practitioners in diagnosing AS.
• This technology could improve early detection rates of aortic stenosis in non-specialist settings.