Prediction of Arteriovenous Access Dysfunction by Mel Spectrogram-based Deep Learning Model

Area of Science:

• Nephrology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Early detection of arteriovenous (AV) access dysfunction is critical for maintaining vascular access patency in hemodialysis patients.
• Predicting AV access malfunction necessitates timely vascular management to prevent complications.

Purpose of the Study:

• To employ deep learning models for predicting AV access malfunction requiring further vascular management.
• To evaluate the efficacy of different deep learning architectures in detecting AV access dysfunction.

Main Methods:

• A prospective cohort study involving hemodialysis patients with AV fistulas or grafts.
• Weekly recording of AV access bruit sounds from three distinct sites using an electronic stethoscope.
• Conversion of audio signals to Mel spectrograms for training deep learning models, including CNN, CRNN, and ViT-GRU.

Main Results:

• The Convolutional Neural Network (CNN) model demonstrated superior performance in the test set with an F1 score of 0.7037 and AUROC of 0.7112.
• The Vision Transformers-Gate Recurrent Unit (ViT-GRU) model showed high out-of-fold prediction accuracy but limited generalization in the test set.
• The CNN model successfully predicted malfunctioning AV access requiring intervention within 10 days.

Conclusions:

• A CNN model utilizing Mel spectrograms of AV access bruit sounds can effectively predict malfunction.
• This AI-driven approach can serve as a valuable screening tool for identifying high-risk AV access requiring intervention.