AI might be coming to the NICU.
Using the medical records of thousands of preterm newborns, a team of scientists at the March of Dimes Prematurity Research Center (PRC) at Stanford has trained an Artificial Intelligence (AI) algorithm to predict which nutrients, and in what quantities, are best for individual preterm newborns as they grow by intravenous (IV) feeding.
The algorithm, detailed in a recently published paper in Nature Medicine, created 15 standard Total Parenteral Nutrition (TPN) recipes for preterm newborns in the neonatal intensive care unit (NICU) depending on different infants’ particular needs at different stages of development. TPN refers to the administering of IV fluid as the sole source of nutrition for a patient; many preterm newborns receive TPN.
By studying the medical records of preterm newborns, the contents of their TPN and their outcomes, the algorithm computed which TPN compositions would lead to the best outcomes for newborns, and importantly, when during a preterm newborn’s development they were needed. This led to a set of 15 distinct recipes, or TPN compositions. Those 15 recipes were subsequently validated on a second cohort.
“This precision medicine breakthrough may result in a significant improvement over the way TPN composition for preemies is decided on today,” said senior study author and Stanford PRC investigator Dr. Nima Aghaeepour.
“The algorithm’s main strength is that it reduces risk for babies given TPN that are not suitable for them,” he added.
Currently, the existing method to determine TPN formulation is time-consuming, costly and sometimes suboptimal. TPN formulations may be subject to errors during prescribing, mixing, labeling or administration processes. These could result in suboptimal nutrition for preterm newborns.
TPN is often given to newborns born prematurely, with or without breastmilk or formula, depending on the newborn and their health status. The TPN bags commonly contain nutrients such as fat, amino acids, copper, magnesium, multi-vitamins, sodium, zinc, glucose, potassium, and more—all ingredients whose correct dosages are critical to the health of preterm newborns.
In what became one of the paper’s key findings in support of the algorithm, the team asked the algorithm to match a group of past patients with one of the 15 TPN compositions it had previously decided on after scouring the medical records of 5,913 premature newborns born at California’s Lucile Packard Children’s Hospital Stanford from 2011 to 2022.
After the algorithm had paired the newborns with one of its 15 standard TPN recipes, the PRC team asked it to look at the actual TPN recipes those newborns received, as well as the outcomes of those newborns.
It found something striking: the patients whose actual prescriptions were significantly different from the AI-suggested prescriptions had a higher risk of sepsis, liver disease, Necrotizing Enterocolitis (NEC), and death.
“Collectively, these findings suggest that significant deviations from [the algorithm] are associated with increased odds of adverse outcomes, while close adherence [to the algorithm] could provide further benefits and reduce the likelihood of these complications,” the study’s authors wrote in the paper.
Plus, the authors reported that when 10 neonatologists, who were part of the study, were presented a choice between AI-suggested nutrition compositions for preterm newborns and the actual nutrient prescriptions those newborns received in the NICU, the neonatologists overwhelmingly preferred the AI-generated recipes—without knowing they were AI-generated.
Right now, the contents of TPN preparations for premature newborns are decided on by a multidisciplinary group of medical professionals in a time-consuming process that can differ between countries, states, and even hospitals in the same state. As a result of trying to provide the best outcomes, the process has a high degree of variability, and until now, lacked the bird’s eye view of the nearly 80,000 prescriptions from the 5,913 preterm newborns whose data it mined. By sifting through that data, the algorithm learned what mix of ingredients works best for individual newborns.
Dr. Aghaeepour said that TPN2.0, the technical name of the algorithm based on the current method, is just the first step in the right direction. He said additional study of larger and more granular preterm datasets promises to make the algorithm even more precise, further curating its equations to individual newborns.
“In the long run, we want to get to a truly personalized place, but right now, we need first to do things in a standard way, a safe way,” he said. “And we are definitely getting there.”
If further validated in additional studies, TPN2.0 could one day be on a path toward use in NICUs globally, he added, pairing preterm newborns with one of 15 custom TPN formulations based on their needs.