Back

Leveraging Machine Learning for Developing and Validating a Neonatal Acute Kidney Injury Prediction Model (NEPHRO): A Comprehensive Evidence-Based Neonatal AKI Risk Stratification Tool

Mohamed, T. H.; Bambach, S.; Spencer, J. D.; Rust, L.; Patel, S.; magers, j.; Neyra, J.; Wilson, F. P.; Ning, X.; Newland, J.; Rust, S.; Slaughter, J. L.

2025-06-13 nephrology
10.1101/2025.06.12.25329508 medRxiv
Show abstract

BackgroundAcute kidney injury (AKI) is a serious and common complication among critically ill neonates. Preventing or treating AKI early requires timely prediction, but current tools to forecast AKI in neonates are limited. We proposed that machine learning could help predict AKI by analyzing routinely collected clinical data. MethodsWe conducted a retrospective analysis of 8,059 critically ill neonates admitted to our level IV neonatal intensive care unit (NICU) from January 2017 to December 2021 for development and cross-validation (n=5,443, 68%), and from January 2022 to June 2024 (n=2,616, 32%) for temporally validating a neonatal AKI predictive model. Risk factors for model input were identified from the literature, and data were extracted from electronic health records. According to the neonatal modification of kidney disease: Improving Global Outcomes criteria, an increase in serum creatinine (SCr) and/or a decrease in urine output (UOP) defines neonatal AKI. We trained a machine learning model using a least absolute shrinkage and selection operator (LASSO) algorithm to develop and validate a predictive AKI model. The area under the receiver operating characteristic curve (AUROC) and F1-scores evaluated the models performance. FindingsAmong 206,220 NICU patient days, AKI occurred in 881 (11%) neonates. Using 27 potential AKI risk factors, LASSO identified key AKI predictors: fluid balance, hypotension requiring vasopressors, invasive ventilation, sepsis, surgical procedures, and congenital kidney and urinary tract anomalies. The model predicted the occurrence of a critical SCr increase or UOP decrease over the next 48 hours, with an AUROC of 0.814 (95% CI: 0.787-0.843) in development and 0.815 (0.795-0.834) in validation datasets. InterpretationWe developed a machine learning-based model that reliably predicted neonatal AKI before it became clinically apparent by conventional parameters. By identifying high-risk neonates earlier, timely interventions can be deployed to improve outcomes.

Matching journals

The top 5 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.