Automated Model Generation for Arteriovenous Shunt Stenosis Detection Using Machine Learning Algorithms

This study develops automated machine learning algorithms to assess hemodialysis access stenosis using auscultation signals, providing objective second opinions for early detection. Auscultation signals were recorded from three sites for arteriovenous fistulas (AVFs) and four for arteriovenous grafts (AVGs), with blood pressure and Doppler ultrasound data as inputs. The models were developed using automated machine learning techniques, which identify the best combination of feature selectors and classifiers from a predefined search space. In the three-class classification of forearm arteriovenous fistula and forearm arteriovenous graft, the models, relying primarily on time-domain and frequency-domain features, with minimal contribution from heart rate and blood pressure, achieved F-scores of 0.851 and 0.825, respectively.