Acoustic-based Stenosis Detection for Dialysis Patients using Explainable Machine Learning.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiovascular Monitoring

Background:

• Long-term patency of arteriovenous fistulas (AVFs) is critical for hemodialysis patients.
• Current AVF monitoring methods often lack point-of-care accessibility and interpretability.
• Automated stenosis detection can improve vascular access outcomes.

Purpose of the Study:

• To compare classical machine learning (ML) models against a Vision Transformer (ViT) deep learning (DL) architecture for automated AVF stenosis detection.
• To evaluate the performance and interpretability of different AI models using non-invasive acoustic recordings.
• To determine the clinical reliability of AI-driven AVF monitoring.

Main Methods:

• Acoustic recordings from AVFs were analyzed using expert-designed features (e.g., MFCCs, peak amplitude) and anatomical metadata.
• Classical ML classifiers were trained and compared with a ViT model trained on Mel-spectrograms.
• Explainable AI (XAI) using SHAP analysis was employed to interpret model predictions.

Main Results:

• Classical ML models consistently outperformed the ViT deep learning model in detecting AVF stenosis.
• Both classical ML and ViT achieved comparable patient-level performance with an F1 score of 0.91.
• XAI revealed that classical models utilized physiologically relevant acoustic features indicative of stenosis.

Conclusions:

• Classical ML models provide a precise and interpretable framework for early-stage AVF stenosis detection.
• The findings suggest classical ML is a clinically reliable tool for enhancing long-term vascular access monitoring.
• This approach has the potential to significantly improve outcomes for hemodialysis patients.