The algorithm uses information in preemies’ electronic medical records to predict which nutrients they need and in what quantities. The AI tool was trained on data from almost 80,000 past prescriptions for intravenous nutrition, which was linked to information about how the tiny patients fared.
Using AI to help prescribe IV nutrition could reduce medical errors, save time and money, and make it easier to care for preemies in low-resource settings, the researchers said. IV nutrition, also known as total parenteral nutrition, is the only way to feed preemies who are born before their digestive systems are mature enough to absorb nutrients.
“Right now, we come up with a TPN prescription for each baby, individually, every day. We make it from scratch and provide it to them,” said senior study author Nima Aghaeepour, PhD, associate professor of anesthesiology, perioperative and pain medicine and of pediatrics. “Total parenteral nutrition is the single largest source of medical error in neonatal intensive care units, both in the United States and globally.”
Not only is the process error-prone but it also makes it difficult for doctors to know if they’ve gotten the formula right. There’s no blood test to measure whether a preemie received the right number of calories each day, for example, and unlike full-term babies, preemies don’t necessarily cry when they are hungry and show contentment when they are full.
“Nutrition is one of the areas of neonatal intensive care where we are weakest,” said study coauthor David Stevenson, MD, a neonatologist and the Harold K. Faber Professor in Pediatrics.
“We can’t approximate what the placenta is doing,” he said.
A slow process
About 10% of babies are born prematurely, meaning at least three weeks before their due dates. Babies born more than about eight weeks early are not ready to absorb nutrients through their intestines and require IV feeding. In addition, some preemies experience gastrointestinal complications of early birth and need IV nutrition while the gut heals.
At present, IV nutrition is prescribed daily for these patients on an individual basis. Patients need macronutrients, the molecular building blocks of protein, fat and carbohydrates; micronutrients such as vitamins, minerals and electrolytes; and medications such as heparin, which is added to the IV preparation to reduce risk of blood clots. The current prescriptions are based on factors such as the baby’s weight, stage of development and the results of their lab work.
Providing these prescriptions requires input from six experts working together over a multihour process: A neonatologist or pharmacist writes each prescription, which is checked by a dietitian for nutrient composition and by a second pharmacist for safety. The prescription goes to a compounding pharmacy, where it is prepared, then to the neonatal intensive care unit, where one nurse gives the IV and a second nurse double-checks that each patient receives the correct preparation.
“It’s a high-risk drug because it is a mixture of many different things,” said study co-author Shabnam Gaskari, PharmD, executive director and chief pharmacy officer at Stanford Medicine Children’s Health. “If we had manufactured, ready-to-use TPNs, that would be very beneficial. I think it would be safer for patients.”
Toward standard formulas
The researchers wondered if they could use AI to help provide hospitals with manufactured, ready-to-use nutrient formulas.
Their AI algorithm was trained on 10 years of electronic medical record data from the neonatal intensive care unit at Lucile Packard Children’s Hospital Stanford, including 79,790 prescriptions for IV nutrition from 5,913 premature patients. The algorithm also had access to information about patients’ medical outcomes, enabling it to find subtle patterns that connected nutrient levels to babies’ health. Although the doctors had not always gotten each prior prescription exactly right, the volume of data helped overcome that problem, enabling the algorithm to learn in a general way about what works for babies in different medical situations.
“This is a strength of AI: Sometimes imperfect data is good enough as long as you have a lot of it,” Aghaeepour said.
After training on the decade of patient data, the algorithm grouped similar nutrient prescriptions to determine how many standard formulas would meet all patients’ nutrition needs, and what would go into each.
“We wondered: What if we make three standard formulas, or 10, or 100?” Aghaeepour said. “It turns out that with 15 distinct formulas for IV nutrition, what you are recommending is pretty similar to what the physicians, pharmacists and dietitians would have done anyway. But then these 15 AI-based formulas can be used to significantly improve speed and safety.”
Further, the researchers showed that the AI algorithm could use data from patients’ electronic medical records to predict which of the 15 formulas each baby might need, and it could adjust the recommendations each day, as patients grew and their medical condition changed. So, the algorithm might recommend that a specific baby needed formula No. 8 for five days, then formula No. 3 for a week, then formula No. 14 for a few days, and so on.
To test how this approach would stack up against real prescriptions, the research team created a test for 10 neonatologists: The doctors were shown clinical information for past patients, alongside the IV nutrition prescriptions they had actually received and the prescriptions the algorithm would recommend. Doctors were not told which prescription was which; they were asked which they thought was better. Doctors consistently preferred the AI-generated prescriptions to the real prescriptions.
The researchers also used AI to scan the electronic medical records from past patients, looking for instances where the patient’s actual nutrition prescription was quite different from what the AI would have recommended. For those patients, risk of mortality, sepsis and bowel disease were significantly higher than for patients whose prescriptions matched what the AI would have recommended, they found.
The team also validated the AI model using real data from the University of California, San Francisco (including 63,273 nutrition prescriptions from 3,417 patients) and found that the model did a good job of predicting nutrient needs for this population, too.
Steps to implementation
The next step will be to run a randomized clinical trial in which some patients receive nutrient prescriptions using the manual method, others receive AI-recommended prescriptions and the researchers see how each group fares.
Assuming the system is implemented, the team plans to have doctors and pharmacists continue to check the AI recommendations and adjust the prescriptions if necessary.
“The AI recommendation is based on whatever information has been added to a patient’s electronic medical record, so if something is missing from the record, the recommendation won’t be accurate,” Gaskari said. “We need a clinician to look at it and review.”
But once the prescription has medical approval, one of the 15 standard nutrient formulas, kept on a hospital shelf, could be given to the patient immediately, she added.
Using standard formulas would also make IV nutrition more accessible and less expensive, as it would no longer require the large expert team now involved, nor access to a compounding pharmacy. This could have benefits for hospitals in lower-income countries or other low-resource settings.
“This reflects our hope for how AI will enhance medicine: What it’s going to do is make doctors better and make top-notch care more accessible,” Stevenson said. “Hopefully, it will also give our physicians more time to do the things computers can’t do, such as spending time with babies and their families, listening to them, and providing comfort and reassurance.”
Scientists from the University of Southern California Keck School of Medicine and Children’s Hospital of Los Angeles contributed to the research.
This work was supported by the National Institutes of Health (grant R35GM138353), the National Center for Advancing Translational Sciences (grant UL1TR001872), the Eunice Kennedy Shriver National Institute of Child Health and Human Development (grant R42HD115517), the Burroughs Wellcome Fund, the March of Dimes, the Alfred E. Mann Foundation, the Stanford Maternal and Child Health Research Institute through Stanford’s SPARK Translational Research Program, Stanford High Impact Technology Fund, and Stanford Biodesign. This project was also supported by the National Center for Advancing Translational Sciences, National Institutes of Health, through UCSF Clinical & Translational Science Institute.
# # #
About Stanford Medicine
Stanford Medicine is an integrated academic health system comprising the Stanford School of Medicine and adult and pediatric health care delivery systems. Together, they harness the full potential of biomedicine through collaborative research, education and clinical care for patients. For more information, please visit med.stanford.edu.
END
