A Trustworthy Model with Uncertainty Management for Predicting Vascular Access Dysfunction in Hemodialysis Patients

Vascular access dysfunction is a prevalent and critical complication among hemodialysis patients, particularly in Taiwan, which has the highest proportion of dialysis patients globally. Early diagnosis and effective management are essential for improving patient outcomes. However, traditional diagnostic approaches, such as routine surveillance and fixed blood flow thresholds, often fail to reliably identify patients requiring timely surgical intervention. To address these challenges, this study proposes a trustworthy AI system utilizing an uncertainty-aware, tree-based machine learning framework to improve the assessment of vascular access dysfunction. The proposed framework incorporates advanced uncertainty management techniques to address both aleatoric uncertainty (data uncertainty) originating from inherent data variability, and epistemic uncertainty (model uncertainty) arising from model limitations. Aleatoric uncertainty is systematically quantified using a multipass perturbation strategy that simulates sample variability to capture the distribution of potential outcomes, while epistemic uncertainty is mitigated using ensemble methods. By leveraging the multipass perturbation strategy, the framework generates calibrated uncertainty estimations. To increase the reliability and trustworthiness of clinical decision-making, the framework assigns low-confidence (high-uncertainty) predictions to an ‘Uncertain’ category, thereby achieving a near-zero leakage rate, which is crucial in high-stakes clinical scenarios. Additionally, an extended confusion matrix and novel uncertainty metrics are introduced to comprehensively evaluate model performance. The system has been validated using real-world hospital datasets, demonstrating its deployable AI potential with superior predictive accuracy, sensitivity, and robustness compared to traditional methods such as the Kidney Disease Outcomes Quality Initiative (KDOQI) guidelines.