Development and internal validation of machine learning in predicting prognosis of acute kidney injury patients in...

Machine learning models for predicting acute kidney injury (AKI) prognosis have primarily been developed in resource-rich settings, with limited validation in resource-limited environments. This study applied machine learning techniques to predict in-hospital mortality and major adverse kidney events within 28 days (MAKE-28) among critically ill patients with AKI in Southeast Asia.

Data were derived from the Southeast Asia AKI cohort, a prospective multicenter study of critically ill patients. Demographic, clinical, and laboratory variables collected at ICU admission were analyzed. Logistic regression, random forest, and extreme gradient boosting (XGBoost) were used to develop prediction models, with recursive feature elimination applied for feature selection.

Of 6993 ICU patients, 1650 individuals with AKI were included for analysis. Of these, 778 (47.1%) died during hospitalization and 1204 (73.9%) experienced MAKE-28. The three models demonstrated comparable performance in predicting MAKE-28 and hospital mortality (AUC 0.73-0.76 for MAKE-28 outcome and AUC 0.71-0.75 for hospital mortality). Discrimination ability was moderate, and all machine learning approaches outperformed conventional clinical scores. No difference in performance was observed between logistic regression and more complex machine learning models.

Machine learning models using routinely available clinical variables may offer useful prognostic information for AKI outcomes in resource-limited settings and outperform traditional scoring systems. External validation is required to confirm generalizability and support clinical implementation.