Using Clinical Data to Better Understand Risk in the NICU
August 20, 2026
Neonatal intensive care units generate continuous streams of physiologic and clinical data. Bedside monitors, ventilators, laboratory testing, and the electronic health record collectively capture far more information than clinicians can absorb or use at any single point in time.
Chelsea Bitler, DO, MS, a fourth-year dual neonatal-perinatal medicine and clinical informatics fellow at UPMC Children’s Hospital of Pittsburgh, is studying how those data can be used to improve risk prediction and risk adjustment in neonatal care. Her work focuses on developing computational approaches that can help clinicians and health systems interpret outcomes in the context of patient risk and, over time, identify patterns within clinical data that may be difficult to recognize at the bedside.
“In an ICU, we are collecting new data points every second, but we use only a fraction of them in our day-to-day care,” Dr. Bitler says. “As humans, we cannot fully absorb all of that continuous information. Machine learning and other analytic approaches give us ways to look for changes and patterns within those data that we may not otherwise recognize.”
A new award from the American Academy of Pediatrics (AAP) Section on Neonatal-Perinatal Medicine will support the next phase of that work. Dr. Bitler received the Marshall Klaus Neonatal-Perinatal Health Services Research Award for her project, “Exploring Federated Learning for the Development of a Contemporary Neonatal Mortality Risk Adjustment Model.” Christopher M. Horvat, MD, MHA, vice chair for Informatics and Digital Health and associate professor of critical care medicine, pediatrics, biomedical informatics, and clinical and translational science at the University of Pittsburgh, is her mentor on the project.
Bringing Neonatology and Clinical Informatics Together
Dr. Bitler entered medical school already intending to pursue neonatology. Her interest in clinical informatics developed later, during residency, through a quality improvement project focused on reducing inefficiencies in the transfer of infants with congenital heart disease from a delivery hospital to a children’s hospital. Working with a neonatal informatician on changes to the electronic health record introduced her to the specialty and initially drew her toward workflow improvement.
Her interests subsequently expanded into data analytics and prediction modeling. After beginning neonatal-perinatal medicine fellowship at UPMC Children’s, she began working with Dr. Horvat and added clinical informatics fellowship training. The combined four-year pathway allows her scholarly work in neonatal prediction modeling to span both fellowships.
Accounting for Differences in Patient Risk
Risk prediction can operate at different levels. An individual-patient model might estimate the likelihood that a particular infant will develop a complication or deteriorate. Dr. Bitler’s primary fellowship research has focused on the other end of that spectrum: population-level risk adjustment that can place outcomes across neonatal intensive care units, or across different periods within the same unit, into the context of how sick the patients were at baseline.
Raw outcome rates alone cannot provide that context. A unit caring for large numbers of extremely premature infants, surgical patients, or infants requiring extracorporeal membrane oxygenation has a fundamentally different patient population from a unit whose admissions primarily include infants with respiratory distress, suspected infection, or less complex prematurity. A simple comparison of mortality rates between those units does not account for those differences in case mix.
“If I just look at what percentage of babies died at one site compared with another, I may be comparing apples and oranges,” Dr. Bitler says. “Risk adjustment is a way to level that playing field. If we can account for how sick the babies were when they arrived, we can make a much more meaningful assessment of the outcomes we observed.”
The mortality model Dr. Bitler has been developing during fellowship uses information from the first 24 hours of admission, including maternal data, gestational age, birth weight, physiologic measures, and laboratory values, to estimate expected mortality based on an infant’s initial severity of illness. That expected rate can then be compared with observed mortality. The model has been evaluated within three hospitals in the UPMC system; work to assess the approach beyond a single health system is the next step.
The purpose of that comparison is quality improvement. Outcomes that differ from what would be expected based on patient risk can prompt closer examination of clinical practices and changes in care. Better-than-expected outcomes also can identify practices or approaches that may warrant further study and potentially inform care elsewhere.
“The goal is to create a learning health system,” Dr. Bitler says. “You can learn from your own trends over time, but you can also begin comparing across units and asking where there may be something to learn from a high-performing center or where a change in practice needs a closer look.”
Developing a Multicenter Model for Research
Existing neonatal mortality risk-adjustment tools were developed more than two decades ago. As neonatal care practices and outcomes have changed, those models have become increasingly susceptible to calibration drift, in which predictions developed from an earlier patient population no longer correspond as closely to contemporary outcomes.
Dr. Bitler’s AAP Marshall Klaus Research Award project is designed to address both the need for a contemporary model and one of the practical obstacles to developing models across multiple health systems.
The project will use Epic Cosmos, a database containing information from more than 300 health systems, to develop a centralized machine-learning neonatal mortality risk-adjustment model and then evaluate whether a federated-learning approach can provide a viable alternative for multicenter model development.
Using Federated Learning to Break Down Research Barriers
Traditional multicenter model development generally requires participating institutions to pool patient-level data in a central location. Doing so introduces substantial privacy, data security, governance, institutional review, and data-use requirements. Federated learning changes where the model is trained. Patient-level data remain at each participating institution, while model parameters generated locally are shared and combined across sites. Updated parameters are returned to the participating institutions for additional training and evaluation without transferring the underlying patient records.
“The line-level patient data never leave the institution where they were collected,” Dr. Bitler says. “We can share the model parameters and performance measures without sharing the individual patient data. The question we are trying to answer is whether that approach can produce a model that performs comparably to the traditional centralized method.”
The project therefore has two potential outputs: a contemporary multicenter neonatal mortality risk-adjustment model and evidence about federated learning as a method for future collaborative modeling. Demonstrating that the distributed approach can perform acceptably compared with centralized model development could provide another option for multicenter studies in which sharing patient-level data presents substantial barriers.
Extending Risk Adjustment to Other NICU Outcomes
The same problem of separating changes in outcomes from changes in underlying patient risk occurs in other areas of neonatal care. Dr. Bitler has begun developing a risk-adjustment model for severe intraventricular hemorrhage (IVH) to support an ongoing quality improvement effort. Month-to-month IVH rates can be difficult to interpret when the gestational ages and baseline risks of the infants cared for during those periods differ substantially.
“You may have one month when many of the babies are 22 or 23 weeks and another when they are 29 or 30 weeks, and those groups carry very different baseline risks for IVH,” Dr. Bitler says. “If the rate changes, we need a way to understand whether we are seeing the effect of a change in care or the effect of a different patient population.”
Moving From Population Risk Assessment to the Individual Patient
Dr. Bitler’s longer-term research interests also extend from population-level risk adjustment to prediction for individual infants. The NICU presents a distinctive setting for that work because infants often remain in the unit through very different stages of illness and recovery. A critically ill infant receiving intensive support may be cared for only doors away from an infant who is medically stable and primarily working on feeding, growth, and maturation.
Subtle deterioration in those more stable infants may be difficult to recognize before a clinical complication becomes apparent. Continuous physiologic and other data could eventually provide additional signals that help identify those changes earlier.
“I am interested in continuing to work on risk stratification and risk adjustment, but I also want to explore individual-patient models that could help us detect deterioration earlier,” Dr. Bitler says. “There is so much information generated in the NICU. I want to understand how we can use more of it in ways that are clinically meaningful and that help us improve care for these babies.”
Learn More
References and Further Reading
- UPMC Physician Resources Article: New Review Paper Examines Role of Artificial Intelligence Models in NICU Mortality Prediction.
- 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. 2026 Feb; 31(1): 101688.
- Bitler CK, Rivera BK, Godavarthi S, Stehle CG, Smith CV, Halling C, Backes CH. Evaluating the evidence behind umbilical cord clamping practices in at-risk neonatal populations. Semin Perinatol. 2023 Jun;4 7(4): 151745.
