CKD-M2 study: 2-year mortality prediction tool for advanced kidney disease.

Area of Science:

• Nephrology
• Biostatistics
• Medical Informatics

Background:

• Existing models for predicting all-cause mortality in chronic kidney disease (CKD) often lack robust methodology and external validation.
• This study addresses the need for reliable prediction tools in advanced CKD stages (4-5).

Purpose of the Study:

• To enhance a previously validated Bayesian network-based tool for predicting 2-year all-cause mortality in patients with stage 4-5 CKD.
• To improve the tool's robustness and generalizability through an enlarged training dataset and a second external validation.

Main Methods:

• A Bayesian network model was trained on an expanded national dataset, incorporating data from prior external validation.
• Internal validation used 10-fold cross-validation; external validation was performed on the CERRENE cohort.
• Performance metrics included accuracy, AUC-ROC, sensitivity, and specificity; calibration was assessed using a calibration curve and Brier score.

Main Results:

• The prediction tool was developed using 1,061 patients and validated on 409 CKD stage 4-5 patients.
• Satisfactory internal and external validation performance was observed: accuracy (77.2% vs. 77.8%), AUC-ROC (0.76 vs. 0.74).
• External validation showed sensitivity of 54.2% and specificity of 82.3%, with acceptable calibration (Brier score = 0.132).

Conclusions:

• The updated Bayesian network prediction tool demonstrates satisfactory performance in both internal and external validation.
• Further national and international external validations are required before widespread clinical implementation.
• The tool shows promise for improving risk stratification and clinical decision-making in advanced CKD.