When Nima Aghaeepour and team created an AI model that forecasts a baby's NICU complications before birth, he thought clinicians at Lucile Packard Children's Hospital would be thrilled.
Instead, he got a practical question.
“If I tell somebody that this patient, who isn’t even born yet, is going to be at risk of necrotizing enterocolitis three months after birth, I figured everybody would love me,” said Aghaeepour, PhD, professor of pediatrics and biomedical data science at Stanford Medicine. “Turns out that in 99% of cases, they say, ‘Well, what do you want me to do about that?’”
That lesson completely reshaped how his clinical research team at Stanford University approaches AI for the hospital’s highest acuity patients.
“Predicting the future is not enough,” he said. “Even if it shows you something no human could ever see, it doesn’t make these models useful. They have to be actionable.”
From prediction to prescription
After that response, Aghaeepour’s team stopped leading with premature risk taxonomies and started with a smaller question: What can a neonatologist actually modify in the NICU? The list is narrow: Ventilator settings, nutrition, and a handful of medications and procedures. So the team worked backward from those actions instead of forward from the prediction.
They started with nutrition. For babies in the NICU, getting the right mix of total parenteral nutrition (TPN) is slow and subjective. A dietitian, a pharmacist, and a neonatologist negotiate each bag, and the pharmacy compounds them one at a time.
Aghaeepour’s team trained a model on a decade of the hospital’s TPN orders — nearly 80,000 of them — to find out how much of that variation was necessary. The results surprised even the team.
“If you ask the experts, they’ll tell you every patient is unique, that every composition has to be made individually,” Aghaeepour said. “But our AI was telling us that 15 is enough.”
Fifteen standardized formulas covered nearly every patient. They could be mass-produced centrally instead of compounded one at a time, which cut costs and reduced the pharmacy burden.
And it led to a simple, direct action for clinicians.
“Forget about dashboards, forget about predicted risk,” Aghaeepour said. “All I’m going to tell you is to use bag number three. And bag number three is already sitting there. You just take it.”
Getting buy-in from clinicians
The idea of relying on a computer to choose their patients’ nutrition was not readily accepted by the clinicians.
Aghaeepour let the evidence do the arguing.
The team ran a blinded study that set an individualized, human-designed TPN next to one of the AI’s 15 formulas and asked neonatologists to review both without knowing which was which.
“If they didn’t know which one came from the AI, they liked the AI solution better in the vast majority of cases,” Aghaeepour said.
In a formal evaluation, clinicians rated the AI recommendations higher than current best practice. That result, not a sales pitch, is what moves intensive-care physicians, Aghaeepour said.
“These are quantitative, risk-averse physicians who want the best for their patients. It’s not an emotional argument for them,” he said. “If you put stringent evidence in front of them and show them what the science says, they’re happy to use it.”
What’s next
The team is now applying the same actionable framing to ventilator management.
“We have models that can predict the next five minutes of vital signs with good accuracy and can notify staff through a smart watch that something is about to happen in that room,” Aghaeepour said.
Instead of a parent running down a nurse when the ventilator beeps, the nurse will already be in the room ready to make the proper adjustments before it ever sounds.
Ultimately, Aghaeepour believes AI will handle all the number crunching in the NICU and free clinicians to do the truly human work. While AI can out-perform any human on data, it can’t intubate a baby or sit with frightened parents.
“Anything that is computational and quantitative, AI is going to win every time,” he said, “but anything with a physical or emotional component stays with the humans.”
The future of effective AI
Aghaeepour’s lab collaborates with hospitals that want to test his team’s models on their own patients. The standing invitation, he said, is real: His email is on the internet, and he is always looking for clinical partners.
His advice for other children’s hospitals is to invest in data.
The questions that matter most in neonatology take years or decades to answer, and no model can find what a hospital never records. He points to newborn metabolic screens, genetic testing, and high-fidelity vital-sign data as the raw material worth investing in now.
“That infrastructure will help with not just today’s problems, but generations to come,” he said.
The challenges of building AI for high-acuity settings are vast, but so are the rewards.
“It’s one thing to have AI help optimize billing codes or take notes in family practice,” he said. “Using it to make real-time decisions in intensive care or the operating room is a completely different game. Every second matters. But that’s also where the impact is, because you’re changing a patient’s health trajectory.”
