New Review Paper Examines Role of Artificial Intelligence Models in NICU Mortality Prediction
May 7, 2026
Chelsea Bitler, DO, MS, a dual fellow in neonatal-perinatal medicine and clinical informatics, UPMC Newborn Medicine Program and Division of Clinical Informatics, was the first author of a recent review article titled, “Neonatal Artificial Intelligence and Machine Learning Mortality Prediction Modeling: A Systematic Review for Risk Adjustment in the Neonatal Intensive Care Unit.” The review was published in Seminars in Fetal and Neonatal Medicine.
Christopher M. Horvat, MD, MHA, associate professor, critical care medicine and pediatrics, and director, clinical informatics, Department of Critical Care Medicine, was the paper’s senior author.
About the Review
The review from Dr. Bitler and colleagues examines whether current artificial intelligence and machine learning models can be used to support mortality risk adjustment when comparing outcomes across neonatal intensive care units.
Mortality is commonly used to compare NICU outcomes, but the comparisons depend on accounting for differences in gestational age, birth weight, and illness severity, among other variables. Risk adjustment models are intended to address that problem.
Dr. Bitler and colleagues reviewed 37 studies describing 242 artificial intelligence and machine learning–based mortality prediction models. The models vary widely in how they are constructed and evaluated.
Most are developed using data from a single center and are not tested on independent patient populations. Study populations differ in gestational age and clinical characteristics, and models use different input variables and time points for prediction. Evaluation methods are also inconsistent, with limited reporting of measures needed to assess how well models perform across different settings.
These differences prevent direct comparison between models and limit their use beyond the datasets in which they were developed. Performance reported in one study does not establish how a model will function in another NICU.
“We were interested in whether these models could be used for risk adjustment across NICUs, because that’s where they would have practical value, and what we found is that the issue isn’t a lack of models or even performance within individual datasets. It’s that they’re being developed in different populations, using different inputs, and evaluated in different ways, with very limited validation outside the original setting. When you put that together, you don’t have something that can be compared or applied across institutions, and that’s what would be required for benchmarking,” Dr. Bitler says.
Reference
Bitler CK, Bertoni CB, King BC, Hooven TA, Horvat CM. Neonatal Artificial Intelligence and Machine Learning Mortality Prediction Modeling: A Systematic Review for Risk Adjustment in the Neonatal Intensive Care Unit. Semin Fetal Neonatal Med. 2025 Feb; 31(1): 101688.
Learn More
