RenalSense: Explainable Prediction of Chronic Kidney Disease.

Abstract

Chronic kidney disease (CKD) is a progressive disease, and often the symptoms are not detected until later stages, which makes it important to have clear and accurate prediction tools. This study aimed at creating a machine-learning model called RenalSense that explains the model and predicts incident CKD using demographic, behavioural, clinical and laboratory parameters. Multiple supervised algorithms, such as logistic regression, random forest, Extra Trees, and gradient boosting, were tested and evaluated, and a retrospective approach using predictive modelling was followed. The data processing steps included missing value imputation, categorical encoding, standardisation, removal of irrelevant features, and patient-level splitting to avoid information leakage. The accuracy, sensitivity, specificity, precision, F1 score, balanced accuracy, Matthews correlation coefficient, ROC-AUC, average precision and Brier score were used for assessing the model performance. All predictive performance was near chance, with logistic regression showing the highest ROC-AUC of 0.524. The other models made similar poor discrimination and calibration. Explanability analysis via permutation importance led to the identification of weight, nephrotoxic drugs, serum creatinine, diastolic blood pressure and water intake as the top five predictors, although these were very small in absolute importance value. The results show that the baseline variables available were not useful enough for predicting the future class of a patient with CKD. RenalSense thus illustrates the utility of the transparent evaluation of models, but is not intended to be used in clinical practice.