Abstract
OBJECTIVES: Acute kidney injury (AKI) is a common and serious condition associated with prolonged hospitalization, chronic kidney disease, and increased mortality. Early prediction of AKI offers an opportunity to mitigate these adverse outcomes, yet existing models often fail to generalize to the emergency department (ED) setting, particularly for patients discharged directly to the community. We sought to develop and validate machine learning models that predict new or progressive AKI within 72 hours of ED departure, addressing challenges related to missing outcome data for discharged patients. METHODS: This retrospective, multicenter study of adult patients from 5 EDs within a large health-care system included adult patients with at least 1 serum creatinine measurement during their ED visit. AKI was defined using Kidney Disease Improving Global Outcomes serum creatinine-based criteria, and prediction models relied on demographic, clinical, and laboratory data routinely collected during ED care. Extreme gradient boosting algorithms were trained using 4 approaches to handle missing outcome data: incomplete case exclusion, negative outcome assumption, multiple imputation, and inverse probability weighting. Model performance was evaluated via 10-fold cross-validation and external temporal validation using area under the receiver operating characteristic curve, precision, recall, calibration curve analyses, and measurement of diagnostic performance across a range of risk thresholds. RESULTS: A total of 1,124,017 ED visits between 2017 and 2024 were included in the study; 5.7% (22,093) met AKI progression outcome criteria. The models demonstrated robust predictive performance for any new or progressive AKI (area under the receiver operating characteristics curve, 0.81-0.82) and severe AKI (area under the receiver operating characteristics curve, 0.87-0.88) across validation cohorts. Inverse probability weighting provided a reliable and consistent method for handling missing outcome data, ensuring accurate risk estimates for both hospitalized and discharged patients. Models performed similarly across diverse subgroups and ED sites. CONCLUSION: Machine learning models trained on routinely collected ED data can provide reliable early predictions of AKI progression, supporting actionable clinical decision making for a broad spectrum of patients. This study advances the real-world usability of such models by expanding their applicability to discharged patients and by enabling estimation of ongoing kidney risk, irrespective of AKI status on arrival.