Interpretable Machine Learning for Chronic Kidney Disease Diagnosis: A Gaussian Processes Approach

Abstract

Chronic Kidney Disease (CKD) is a global health issue impacting over 800 million people, characterized by a gradual loss of kidney function leading to severe complications. Traditional diagnostic methods, relying on laboratory tests and clinical assessments, have limitations in sensitivity and are prone to human error, particularly in the early stages of CKD. Recent advances in machine learning (ML) offer promising tools for disease diagnosis, but a lack of interpretability often hinders their adoption in clinical practice. Gaussian Processes (GP) provide a flexible ML model capable of delivering predictions and uncertainty estimates, essential for high-stakes medical applications. However, the integration of GP with interpretable methods remains underexplored. We developed an interpretable CKD classification model to address this knowledge gap by combining GP with Shapley Additive Explanations (SHAP). We assessed the model's performance using three GP kernels (Radial Basis Function, Matern, and Rational Quadratic). The results show that the Rational Quadratic kernel outperforms the other kernels, achieving an accuracy of 98.75%, precision of 100%, sensitivity of 97.87%, specificity of 100%, and an F1-score of 98.51%. SHAP values indicate that haemoglobin and specific gravity are the most influential features. The results demonstrate that the Rational Quadratic kernel enhances predictive accuracy and provides robust uncertainty estimates and interpretable explanations. This combination of accuracy and interpretability supports clinicians in making informed decisions and improving patient management and outcomes in CKD. Our study connects advanced ML techniques with practical medical applications, leading to more effective and reliable ML-driven healthcare solutions.