Adding Explainability to Machine Learning Models to Detect Chronic Kidney Disease

Abstract

Chronic Kidney Disease is a common term for multiple heterogeneous diseases in the kidneys. It is also known as Chronic Renal Disease. Chronic kidney disease (CKD) has a gradual loss of glomerular filtration rate (GFR) over three months. The patient does not observe any significant symptoms in the earlier stage of CKD, and it is not identifiable without clinical tests like urine and blood tests. Patients with CKD would have a higher chance of developing heart disease. CKD is a progressive and often irreversible process of renal function decline, which may reach an endpoint of end-stage renal failure, requiring renal replacement therapy. It is critical to diagnose progressive CKD at an early stage and predict patients prone to developing the disease further for timely therapeutic interventions. As such, researchers have expended enormous efforts in the development of novel biomarkers that may identify subjects with early CKD at risk of progression. In this study, we have developed an explainable machine learning model to predict chronic kidney disease by implementing an automated data pipeline using the Random Forest ensemble learning trees model and feature selection algorithm. The explainability of the proposed model has been assessed in terms of feature importance and explainability metrics. Three explainability methods; LIME, SHAP, and SKATER have been applied to interpret the developed model and to compare the explainability results using Interpretability, Fidelity, and Fidelity-to-Interpretability ratio as the explainability metrics.