When the Doctor on Duty Changes What's in the IV Bag: How AI Is Helping Standardize Newborn Nutrition

What Two Stanford Studies Reveal About Variation in Newborn Feeding Formulas — and a New AI Tool Designed to Help

Premature and critically ill newborns often cannot eat by mouth or stomach tube right away, so hospitals feed them through a vein instead, using a specially mixed liquid nutrition called parenteral nutrition, or PN (sometimes called TPN, for "total parenteral nutrition"). A Stanford research team studied more than a decade of these prescriptions and found that, for certain nutrients, which doctor happened to be writing the order mattered more than the baby's own medical needs [1]. A related, earlier study by an overlapping Stanford and UCSF (University of California, San Francisco) team built and tested an artificial intelligence (AI) tool called TPN2.0, designed to make these prescriptions more consistent while still fitting each baby, and found that babies whose actual nutrition closely matched what the AI recommended tended to have fewer complications than babies whose nutrition diverged widely from it [2].

A Brief History: Why Feeding Through a Vein Is Both Essential and Tricky

PN has been used in newborn intensive care units (NICUs) for decades because many premature babies, especially those born many weeks early, have digestive systems too immature to handle milk feeds right away, yet still need calories, protein, sugar, fat, and minerals to grow and heal. Because getting the balance of PN wrong carries real risks, doctors have debated for years whether it is safer to give every baby a small number of pre-mixed, standardized nutrition formulas, or to custom-tailor every single prescription to that baby's own labs and growth — an official 2018 international pediatric nutrition guideline weighed both approaches and concluded that each has real trade-offs, with standardized formulas being safer and more consistent but custom prescriptions better able to track a rapidly changing baby [3]. Adding to the worry, a European study of newborn and pediatric intensive care units found that a meaningful share of these hand-written nutrition orders contained errors, some of them significant [4] — a reminder that even experienced clinicians can get a complex, individualized prescription wrong. The problem is that PN is not a simple, one-size-fits-all fluid: getting its balance wrong, or leaving a baby on it too long, has been linked in the medical literature to liver problems (a condition called cholestasis, essentially a backup of bile that can injure the liver) [5], to bloodstream infections from the IV line itself, and to a serious intestinal problem called necrotizing enterocolitis, or NEC. Even the specific type of fat (lipid) used in the IV nutrition mix matters: a study of extremely small, extremely premature babies found that the choice of fat emulsion was tied to how often cholestasis developed [6], which is part of why newer fat emulsions made from fish oil have become an active area of NICU research in their own right (https://inneosight.org/en-US/articles/smoflipid-neonatal2024). The two Stanford-led studies in this cluster sit directly inside that decades-old debate: the first study is essentially the "how big is this problem, really?" evidence, showing that a real hospital's decade of prescriptions had substantial doctor-to-doctor variation for several nutrients; the second study is the "here's one way to fix it" response, using AI trained on tens of thousands of real prescriptions to suggest a middle path between rigid standardization and unlimited customization.

What the Researchers Did and Found

In the first study, researchers looked at 54,464 individual nutrition prescriptions written for 5,052 babies by 652 different clinicians at Stanford's NICU between 2011 and 2021 ([1]). After accounting for each baby's age, weight, sex, and how many days old they were, the researchers found that changes in the three main "macronutrients" — protein, sugar, and fat — mostly tracked which babies were in the unit at a given time, not a real change in how doctors prescribed. But for salts and trace minerals (sodium, and several others), the picture flipped: which clinician wrote the order explained noticeably more of the difference between prescriptions than the year did — for sodium, roughly three times more. One mineral, selenium, was the exception that proves the rule: prescribing became far more consistent almost overnight in 2019, exactly when the hospital changed its official default dose — showing that a clear policy change, not just individual habit, is what actually makes practice consistent.

The second study, TPN2.0, is where AI enters the picture ([2]). It draws on a largely overlapping group of Stanford NICU babies as the first study, though the two research groups pulled slightly different, not identical, sets of records: this AI-development study used 79,790 prescriptions from 5,913 patients treated at Stanford between 2011 and 2022 to train its system, then tested that trained system on another 63,273 prescriptions from a completely different hospital, UCSF, to make sure it wasn't just memorizing one hospital's quirks. The system studies an enormous amount of information about each baby — lab results, growth measurements, fluid needs, and more — and boils it down into one of just 15 standardized nutrition "recipes" that a hospital pharmacy could pre-mix in batches ahead of time, saving time and reducing manual compounding errors, while still adjusting for the individual baby and allowing the doctor to override its suggestion. When researchers compared the AI's suggested recipes to what real expert doctors had actually prescribed, the two matched closely — a statistical correlation of 0.94 out of a possible 1.0 at Stanford and 0.91 at UCSF, both very strong agreement. In a blind test, ten experienced clinicians rated prescriptions without knowing which ones were real, which were AI-generated, and which were computer-generated nonsense as a comparison — and, on average, they rated the AI's suggestions (56 out of 100) even a bit higher than real current best-practice prescriptions (35 out of 100), and far higher than the nonsense prescriptions (20 out of 100). Just as importantly, the clinicians could not reliably tell the AI's suggestions apart from real orders, suggesting the tool wasn't producing anything a doctor would flag as unusual or unsafe.

What This Means, and an Important Caution

Looking backward at outcomes, babies whose real prescriptions closely matched what TPN2.0 would have recommended had noticeably lower rates of death, liver injury (cholestasis), gut injury (NEC), and bloodstream infection than babies whose prescriptions diverged widely from the AI's suggestion — for instance, a survival analysis found roughly 80% survival among close-match babies versus roughly 50% among the most divergent group, in the context of cholestasis. These numbers are genuinely encouraging, but it is important to understand exactly what kind of evidence they are: this was a look back at existing records, not a planned experiment where babies were randomly assigned to follow the AI's advice or not. It is entirely possible that the sickest babies, who need the most unusual, individualized prescriptions, were also more likely to have worse outcomes for reasons that had nothing to do with the AI recommendation itself. The study's own authors are careful to call this finding "preliminary and associative," meaning it shows a pattern worth taking seriously, not proof of cause and effect — and they explicitly say a properly designed, prospective clinical trial is needed before anyone can say that following the AI's advice actually causes better outcomes.

What Researchers Are Working On Next

The natural next step, which the researchers themselves point to, is exactly that kind of prospective trial — following babies forward in time under carefully monitored conditions to see whether AI-guided prescribing genuinely improves outcomes, rather than just correlating with them. Researchers will also need to check how well a tool like TPN2.0 performs for babies from racial and ethnic groups that were underrepresented in the hospitals studied so far, since the tool is only as fair as the data it learned from. This kind of careful, staged validation — build a tool, test it against experts, then test it prospectively before trusting it with real decisions — is the same standard being applied across other AI tools being explored in newborn intensive care, an area of active research generally (https://inneosight.org/en-US/articles/agentic-ai-npm2026). Other related nutrition research covered elsewhere in this space includes efforts to use machine learning to manage blood phosphate levels in premature babies (https://inneosight.org/en-US/articles/phosphate-mlpreterm2026), studies of how quickly to start intravenous protein and amino acids in the first days of life (https://inneosight.org/en-US/articles/pro-vide-early-amino-acids2026), and research into safer fat emulsions for babies on long-term IV nutrition (https://inneosight.org/en-US/articles/smoflipid-neonatal2024) — all part of a broader push to make one of the NICU's most common, but most complex, treatments both safer and more consistent, without taking away a doctor's ability to make the right call for the baby in front of them [7]. For families whose babies are on IV nutrition right now, the key takeaway is reassuring: this is an area of active, careful research, existing safety checks (like the ones built directly into TPN2.0) are designed to catch problems before they reach a baby, and any AI tool being developed here is explicitly meant to support, not replace, the judgment of the doctors and pharmacists caring for their child [7]. That said, it is worth knowing that healthcare AI as a field has a mixed track record of turning promising research results into real, everyday clinical benefit once a tool leaves the study that built it [8], which is exactly why the careful, staged testing described above matters so much before any tool like TPN2.0 is used to guide real prescriptions. Just as importantly, a recent narrative review reminds readers that AI-based tools like TPN2.0 are best seen as one option alongside standardized and individualized approaches, not a finished replacement for either [9], and oversight bodies are actively working out how software like this should be regulated as it moves toward real clinical use [10].

References

  1. Phongpreecha T, Sin LWY, Gaskari S, Aghaeepour N. Analysis of factors driving variations in neonatal parenteral nutrition over a decade: A retrospective study. Nutrition in Clinical Practice. 2026 (published online 10 May 2026). doi:10.1002/ncp.70134 ↩
  2. Phongpreecha T, Ghanem M, Reiss JD, et al. AI-guided precision parenteral nutrition for neonatal intensive care units. Nature Medicine. 2025;31(6):1882–1894. doi:10.1038/s41591-025-03601-1 ↩
  3. Riskin A, Picaud JC, Shamir R, ESPGHAN/ESPEN/ESPR/CSPEN Working Group on Pediatric Parenteral Nutrition. ESPGHAN/ESPEN/ESPR/CSPEN guidelines on pediatric parenteral nutrition: Standard versus individualized parenteral nutrition. Clinical Nutrition. 2018;37(6 Pt B):2409–2417. doi:10.1016/j.clnu.2018.06.955 ↩
  4. Hermanspann T, Schoberer M, Robel-Tillig E, Härtel C, Goelz R, Orlikowsky T, Eisert A. Incidence and Severity of Prescribing Errors in Parenteral Nutrition for Pediatric Inpatients at a Neonatal and Pediatric Intensive Care Unit. Frontiers in Pediatrics. 2017;5:149. doi:10.3389/fped.2017.00149 ↩
  5. Żalikowska-Gardocka M, Przybyłkowski A. Review of parenteral nutrition-associated liver disease. Clinical and Experimental Hepatology. 2020;6(2):65–73. doi:10.5114/ceh.2019.95528 ↩
  6. Xie T, Mei H, Chen M, Xu Y, Ma X, Shi L, Chen Z. Association between mixed fatty acid emulsion and parenteral nutrition–associated cholestasis in extremely low-birth-weight infants: A retrospective cohort study. Journal of Parenteral and Enteral Nutrition. 2025;49(8):975–982. doi:10.1002/jpen.70012 ↩
  7. Beam K, Sharma P, Levy P, Beam AL. Artificial intelligence in the neonatal intensive care unit: the time is now. Journal of Perinatology. 2024;44(1):131–135. doi:10.1038/s41372-023-01719-z ↩
  8. Kelly CJ, Karthikesalingam A, Suleyman M, Corrado G, King D. Key challenges for delivering clinical impact with artificial intelligence. BMC Medicine. 2019;17(1):195. doi:10.1186/s12916-019-1426-2 ↩
  9. Majchrzyk M. Standardized, Individualized, or AI-Based Approach to Parenteral Nutrition in Neonatal Intensive Care Units: A Narrative Review. Cureus. 2026;18(3):e105328. doi:10.7759/cureus.105328 ↩
  10. Singh V, Cheng S, Kwan AC, Ebinger J. United States Food and Drug Administration Regulation of Clinical Software in the Era of Artificial Intelligence and Machine Learning. Mayo Clinic Proceedings: Digital Health. 2025;3(3):100231. doi:10.1016/j.mcpdig.2025.100231 ↩