An interpretable machine learning framework for retrospective attribution of hospitalization costs in patients with chronic kidney disease

Hospitalization costs vary substantially among patients with chronic kidney disease (CKD), yet the factors underlying this heterogeneity are not fully understood. We developed an interpretable machine learning framework to retrospectively characterize differences in hospitalization costs among patients with CKD and externally evaluated its performance. The study included 3,017 hospitalizations from one tertiary hospital and 2,708 hospitalizations from another tertiary hospital for external validation. Patients were analyzed separately according to kidney transplant status, including non-transplant ( n = 2,435) and transplant ( n = 582) groups. Multiple machine learning and conventional statistical models were compared using temporal holdout testing, external validation, and sensitivity analyses. LightGBM performed best in the non-transplant group (R² = 0.7060), whereas XGBoost performed best in the transplant group (R² = 0.6235). External validation yielded R² values of 0.5921 and 0.5513, respectively. Model interpretation identified length of stay, treatment-related features, age, and selected laboratory measures as prominent contributors to predicted costs, with nonlinear patterns and substantial patient-level heterogeneity. This framework provides a transparent approach for retrospective hospitalization cost attribution and case-mix analysis in CKD.