AI-assisted detection of high-pitched bruits in arteriovenous fistulas using a digital stethoscope

Abstract

Background:

The arteriovenous fistula (AVF) is the preferred vascular access for patients undergoing hemodialysis, and early identification of complications such as stenosis or dysfunction is essential to preserve access patency and reduce morbidity.

Method:

AVF bruit recordings were collected from 65 patients across 12 dialysis centers in Europe and Asia using a digital stethoscope connected to the medical record of the patients. A deep learning model was developed to detect high-pitched bruits—an acoustic marker commonly associated with AVF stenosis. Expert-annotated recordings served as the reference standard for supervised training and evaluation.

Results:

Mean age of patients was 68, and the average blood flow during the dialysis session was 352 ml/min. The model demonstrated excellent performance on independent testing datasets, achieving a sensitivity of 97.1%, specificity of 73.8%, and an overall accuracy of 82%. The area under the receiver operating characteristic curve (ROC-AUC) was 94%, reflecting strong discriminative ability. The model showed excellent calibration. Model performance across different experimental retraining folds indicates a stable and reliable training process.

Conclusion:

The integration of this deep learning tool into clinical workflows could provide clinicians with a sensitive, objective, and time-efficient method for detecting high-pitched bruits which may be used in combination with other clinical signs for the detection of AVF complications such as stenosis. Implemented through a low-cost phono angiography protocol requiring minimal training, this approach has the potential to support earlier interventions and improve outcomes in the hemodialysis population.