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

Ensuring the long-term patency of arteriovenous fistulas (AVFs) is essential for patients undergoing hemodialysis, yet existing monitoring methods often lack the accessibility and interpretability required for routine point-of-care use. In this study, we performed a comparative evaluation of classical machine learning (ML) models versus a state-of-the-art Vision Transformer (ViT) deep learning (DL) architecture for automated detection of AVF stenosis using non-invasive acoustic recordings. Our approach employed a comprehensive pipeline combining expert-designed acoustic features-such as Mel-frequency cepstral coefficients (MFCCs) and peak amplitude-with anatomical metadata to train a range of classical classifiers. These models were compared against a ViT trained on Mel-spectrogram representations of the same recordings. The results show that classical ML models, whether applied to individual anatomical sites or using universal models informed by anatomical context, consistently outperformed the ViT deep model. At the patient level, both approaches achieved comparable performance, with an F1 score of 0.91. Importantly, integrating Explainable AI (XAI) through SHAP analysis demonstrated that classical models base their predictions on physiologically meaningful features-such as elevated signal energy and spectral shifts-that reflect the hemodynamic turbulence associated with stenosis. By combining high precision with interpretability, classical ML offers a clinically reliable framework for early-stage AVF monitoring, with the potential to enhance long-term vascular access outcomes.