Interpretable machine learning model based on routine metabolic laboratory indices to identify advanced chronic kidney disease.

INTRODUCTION

Early identification of advanced chronic kidney disease (CKD), a condition accompanied by profound metabolic and endocrine disturbances, is essential for timely nephrology referral and intervention. However, widely used risk equations often require albuminuria or repeated measurements that are not consistently available in routine clinical practice.

METHODS

We retrospectively analyzed adult patients from three different departments affiliated to one university, including two independent hospitals and a clinic department. Routinely collected demographic, clinical, and metabolic laboratory variables were used to develop machine learning models for distinguishing preserved kidney function (CKD G1-2) from advanced stages (G3a-5). Five algorithms were trained and internally validated in a development cohort, followed by external validation in an independent cohort. Model performance was assessed by discrimination, calibration, and interpretability using feature importance and SHAP (Shapley Additive Explanations).

RESULTS

Among 308 patients in the development cohort and 52 in the external cohort, the Gradient Boosting classifier achieved the best discrimination (AUC = 0.972 internally; 0.965 externally) with good calibration. Urea, kidney disease type, phosphorus, albumin, and lipid-related parameters-reflecting systemic metabolic dysregulation-emerged as key contributors to model predictions.

DISCUSSION: An interpretable Gradient Boosting model leveraging routinely measured metabolic laboratory data accurately identifies advanced CKD and captures clinically meaningful metabolic patterns associated with disease severity, supporting its potential integration into electronic health records for risk stratification and identification of advanced CKD among patients with established CKD in specialist care.