Publications

Feeding critically ill patients through an IV (parenteral nutrition) is one of the most error-prone processes in intensive care. This review surveys how AI is beginning to change that by predicting complications like cholestasis and feeding intolerance, generating real-time nutrient recommendations, and reducing the variability that comes from different clinicians making different calls for similar patients. The findings point toward a future where AI-guided nutritional therapy can make ICU feeding safer and more personalized, particularly for neonates whose nutritional needs change rapidly day to day.

Doctors write thousands of clinical notes about newborns in the NICU, but most AI tools can only use structured data like lab values. This study built NeonatalBERT, an AI language model trained to read clinical notes and estimate a baby's risk for serious complications. Validated across two major hospitals, it outperformed traditional risk scores. By extracting richer clinical context from unstructured text, models like NeonatalBERT could improve AI-driven TPN recommendations by incorporating information that structured data alone would miss.

Premature babies often depend on IV nutrition (TPN) to survive, but prescribing it is slow, subjective, and the single largest source of medication error in NICUs. This study trained an AI model on a decade of TPN prescriptions from over 9,000 patients across two hospitals. The result, TPN2.0, identified 15 standardized formulas that matched expert-level quality, and in a blinded test, physicians actually rated its recommendations higher than current best practice. When past prescriptions deviated from what TPN2.0 would have recommended, babies had significantly higher rates of serious complications like necrotizing enterocolitis and sepsis.

Prematurity is the leading cause of death in children under five, yet the tools used to assess risk in newborns are limited. This study built a deep learning model that mines both maternal and infant health records to create a data-driven morbidity index, predicting which babies are most likely to develop serious NICU complications starting as early as pregnancy. It outperformed existing clinical risk scores across two hospital systems. This morbidity index can be incorporated into future AI models to not only predict complications but also optimize interventions like TPN to improve outcomes.

Many babies who end up in the NICU were born prematurely, and identifying at-risk pregnancies earlier could improve outcomes. This study used wearable devices to track physical activity and sleep in pregnant women, then applied AI to find patterns linked to preterm delivery. The results revealed previously unknown associations between maternal activity and prematurity risk. Earlier identification of at-risk pregnancies could help clinicians prepare NICU resources and nutritional plans, including TPN readiness, before the baby arrives.