Machine learning-assisted prediction of 5-year mortality in chronic kidney disease: the KoreaN cohort study for Outcome in patients With Chronic Kidney Disease (KNOW-CKD).

BACKGROUND

Mortality prediction models for patients with non-dialysis chronic kidney disease (CKD) remain limited despite their clinical importance. While machine learning (ML) offers the potential to improve prediction accuracy, its "black-box" nature has hindered clinical adoption. This study aimed to develop and validate an interpretable ML model for predicting 5-year all-cause mortality in patients with non-dialysis CKD and to deploy it as a user-friendly web-based risk stratification tool.

METHODS

We analyzed 1,858 patients (94 deaths) from a prospective cohort of non-dialysis CKD patients. Several ML algorithms including CatBoost were trained and compared with conventional logistic regression. SHapley Additive exPlanations (SHAP) analysis was employed to identify key prognostic features and ensure model interpretability. The final simplified model was externally validated in an independent cohort of 348 CKD patients.

RESULTS

The CatBoost model demonstrated superior performance (area under the curve [AUC], 0.813; 95% confidence interval [CI], 0.737-0.888; p = 0.04) compared to logistic regression (AUC, 0.747; 95% CI, 0.655-0.840). A simplified model using only the top-5 SHAP-ranked features-age, estimated glomerular filtration rate, serum albumin, spot urine protein-to-creatinine ratio, and serum total calcium-maintained robust predictive accuracy in the external validation cohort (AUC, 0.795). Notably, SHAP visualization revealed a U-shaped relationship for serum calcium, identifying hypocalcemia as a significant and under-recognized mortality risk factor.

CONCLUSION

The CatBoost-based ML model accurately predicts 5-year mortality in non-dialysis CKD patients using five readily available clinical parameters. We expect that the clinical implementation of this model may offer a practical method for early risk stratification and personalized management.