016 - Automated Learning and Early Recognition Technology (ALERT): A Novel Neonatal AKI Risk Stratification Tool
Saturday, April 26, 2025
2:30pm – 4:45pm HST
Publication Number: 16.3660
Steven Dick, Nationwide Children's Hospital, Columbus, OH, United States; Sven Bambach, Nationwide Children's Hospital, Columbus, OH, United States; Francis Wilson, Yale School of Medicine, New Haven, CT, United States; Laura Rust, Nationwide Children's Hospital, Columbus, OH, United States; Shama Patel, Nationwide Children's Hospital, Columbus, OH, United States; Jacqueline Magers, Nationwide Children's Hospital, Columbus, OH, United States; Steven W. Rust, Nationwide Children's Hospital, Columbus, OH, United States; John Spencer, Nationwide Children's Hospital, Columbus, OH, United States; Jonathan L. Slaughter, Nationwide Children's Hospital and The Ohio State University, Columbus, OH, United States; Tahagod Mohamed, Nationwide Children's Hospital, Columbus, OH, United States
Resident Physician Nationwide Children's Hospital Columbus, Ohio, United States
Background: Neonatal acute kidney injury (AKI) is underrecognized due to a lack of evidence-based tools that aid clinicians in risk identification. Neonatal AKI carries significant morbidity and mortality and therefore reliable AKI risk prediction and identification can assist providers in intervening before kidney injury occurs to improve patient outcomes. Objective: To identify evidence-based and reliable neonatal AKI predictors and develop a machine learning-based neonatal AKI prediction model that can identify risk before kidney injury occurs. Design/Methods: Data from the electronic health records of 5400 neonates admitted to Nationwide Children's Hospital's level IV neonatal intensive care unit from 2017-2021 were used. Neonates who required dialysis were excluded. AKI was defined via the neonatal modification of KDIGO criteria (based on serum creatinine or urine output). Established neonatal AKI risk factors from the literature were entered in various machine learning algorithms to model the relationship between the predictors and AKI development in the next 48 hours. The least absolute shrinkage and selection operator (LASSO) method and Extreme Gradient Boosting (XGBoost) were employed to develop a predictive model. The area under the curve (AUC) was used to evaluate performance. Results: Based on a univariate analysis of risk factors at the patient-level, the most predictive were invasive respiratory support (odds 4.82), hypotension requiring vasopressor in the preceding day (odds 3.95), clinical concern for sepsis (odds 3.13) and treatment with nephrotoxic medications (odds 2.2), Table 1. Overall, the models predict neonatal AKI development up to 48 hours before changes in urine output and serum creatinine occur. XGBoost algorithm outperformed LASSO with AUC of 0.899 (0.893, 0.901) vs 0.779 (0.755, 0.801) respectively.
Conclusion(s): Predictive modeling of neonatal AKI is achievable by using combination of clinical data and machine learning algorithms to allow prompt AKI risk recognition. Reliable and early recognition of neonatal AKI risk is the first step in effective management of AKI which will allow rapid intervention on modifiable factors before kidney injury occurs. This would not only bolster neonatal AKI recognition, but also aid in proactively addressing it before it manifests.
