Endless data collection occurs in the neonatal intensive care unit (NICU) among physicians tasked with taking care of one of the most vulnerable populations: young babies. Researchers in the Aghaeepour Lab at Stanford Medicine and the Sirota Lab at the University of California, San Francisco (UCSF) aspire to use this data to develop artificial intelligence (AI) that guides interventions.
The scientists most recently developed TPN2.0, a new AI model that creates data-driven recipes of total parenteral nutrition (TPN), a life-saving treatment given to infants in the NICU. Current TPN recipe formulation requires multidisciplinary teams and is time-consuming and costly.
Led by anesthesiology professor Nima Aghaeepour, the Aghaeepour Lab uses machine learning for translational biology. The research study, which was published in Nature Medicine, was spearheaded by co-first authors Joe Phongpreecha, postdoctoral researcher Marc Ghanem, neonatologist Jon Reiss and UCSF associate professor of clinical informatics and digital transformation Tomiko Oskostsky.
Lab members initially considered spearheading a predictive AI model, but ultimately determined that a model which makes recommendations for clinical actions would be more beneficial, according to Aghaeepour.
“We quickly learned that if we just predict that this patient is going to have this or that morbidity, it’s not going to be particularly helpful to the care team,” Aghaeepour said. “There were colleagues at Stanford who initially reached out to me to talk about TPN and it’s been very helpful.”
Using data derived from the electronic health records of 5,913 patients, the researchers trained TPN2.0 to create 15 standardized formulas that would improve the safety and cost-effectiveness of the TPN prescription process. The interdisciplinary project required the researchers to consider perspectives from pharmacologists, physicians, nurses and dieticians alike.
“If you are a [computer science] person, and don’t have any collaboration with them, you just predict in whatever format the [received] data is in, which might not be relevant to people who are going to use it,” Phongpreecha said. “All of this is knowledge from the medical community that the [computer science] people, like me, have to adapt into the AI.”
Having worked mostly with machine learning in the context of chemical engineering during his doctoral studies, Phongpreecha had to adjust to more direct patient consequences when transitioning to AI healthcare applications.
Ghanem joined the project after completing medical training at Lebanese American University. Arriving at Stanford for postdoctoral studies in machine learning rather than residency, the physician-turned-researcher found a hub of endless opportunity that drove him deeper into AI modeling.
“Where I come from, there’s always a resource issue slowing you down,” Ghanem said. “And when you come to Stanford, they offer you everything.”
All clinical problems and lab values preceding TPN administration also influence model decision-making, making TPN2.0 a more “complicated” AI model than current models used in the NICU, according to Reiss.
Reiss and his colleagues rigorously trained TPN2.0 to produce clinically plausible prescription recommendations, then passed it onto UCSF colleagues for model validation, led by Oskotsky.
The team’s work has opened doors to new clinical AI interactions and “also methodologically opens up a way to implement different types of data and improve the overall health of the baby,” said Marina Sirota ’06 M.S. ’06 Ph.D. ’10, a pediatrics professor and leader of March of Dimes Prematurity Research Center at UCSF.
When considering the actual implementation of TPN2.0 into a clinical setting, parents are another key stakeholder. Working with an infant’s medical providers, parents may approach AI model implementation with both encouragement and skepticism.
“I could see some parents trusting the power of computers and data to make sure that what we’re giving their baby is the best,” said Lisa Bain ’01, associate chief of quality improvement of the NICU at Stanford Children’s Health, who also did not work on the study. “Parents who might not be as trusting of technology might want somebody with a lot of experience writing TPN recipes to look over their baby’s recipe and make sure that it’s accurate.”
Meera Sankar, neonatologist and medical director of the NICU at Lucile Packard Children’s Hospital, sees potential in TPN2.0 to transform medical provision at hospitals even outside the U.S., given appropriate model testing among patient populations like those in India, where she received medical training.
“There are several NICUs that are collaborating with some of the colleagues here in the U.S., so I think the way to approach it is to study in that particular population and validate the results,” Sankar said. “Just like how this study was done at Stanford and UCSF, I think it would involve doing a multi-center study to validate the information in the Indian population.”
Ultimately, the potential for implementing an AI model as a physician’s “assistant” excites Oskotsky, who formerly practiced as a clinician in emergency medicine.
“I remember being in the NICU [during clinical rotations] and thinking, ‘It would be amazing if we could have computers help us do this in a smarter fashion,’” Oskotsky said. “For me, to see this all evolve within a lifetime is really mind-blowing.”
This article has been updated to reflect that the study was published in Nature Medicine and to correct the spelling of Sirota Lab. The Daily regrets this error.