Can Deep Learning Models Acquire Specialist Knowledge? Identifying Carotid Bruits in Type 2 Diabetes Using an Electronic Stethoscope

The present study explores the potential of artificial intelligence as a substitute for trained clinician identification of a carotid bruit from multiple auditory recordings from an electronic stethoscope in 98 people with type 2 diabetes, 48 of whom had a bruit. We employed various deep networks, including ResNet and VGGish, alongside tailor-made deep learning models for this purpose. Our findings indicated that in scenarios such as this, where the data show high variability and yet are limited, the highest model accuracy achieved was 67.8%. This suggests the inadequacy of AI-driven systems in accurately detecting carotid bruits in the present participant sample. Detailed results for each deep learning and machine learning approach are presented separately, and the dataset has been established as a benchmark for future studies. Given the complexity observed with the current data volume, additional studies with larger sample sizes are warranted.