July 21, 2025 — Imagine if when a soon-to-be mother arrived at a hospital in labor, her health care team could predetermine what challenges she might encounter with breastfeeding and develop a game plan to address any lactation complications, before she even gave birth.

Clinicians and engineers at the University of Florida College of Medicine have developed artificial intelligence models that they’re now applying to make this a reality for families delivering babies at UF Health Shands Hospital in Gainesville.

The Maximizing Initiatives for Lactation Knowledge, or MILK+, project combines the skills of physicians, nurses and AI engineers to identify mothers who are not producing enough breastmilk for their babies and increase lactation upon discharge for a particularly vulnerable population: premature babies in the neonatal intensive care unit. As the AI models continue to improve, the team hopes to make these insights common practice to supplement the care patients already receive.

“Breastfeeding has so many potential benefits for both mother and child, yet people have a misconception that it should come easily since it’s a natural process. In practice it can be extremely challenging, particularly for mothers of vulnerable patients such as those admitted to the NICU,” said project lead Helen Hu, M.D., a clinical assistant professor in the Department of Pediatrics.

Breastfeeding offers many health benefits to mother and child: Research has shown that for the mother, it can help with the postpartum recovery process and lower risk for conditions including high blood pressure, diabetes and breast and ovarian cancers.

While lactation initiation in Florida is above the national average at 85%, the duration of lactation falls significantly over time, according to data from the Florida Department of Health.

For the infant, maternal breastmilk promotes healthy weight gain and brain development, and for preterm infants in particular, it can decrease their chances of contracting illnesses. Though Hu said optimal benefits come from the infant’s mother’s milk, UF Health also has a supply of donor breastmilk available to preterm infants in the NICU if needed, which can also offer important nutritional benefits.

A mother’s ability to breastfeed and an infant’s access to maternal breastmilk is greatly impacted by a stay in the NICU.

“The first hour after birth is critical to starting milk production, and the first week or two afterward is important for establishing supply moving forward,” Hu said. “When a child is separated from the mother and brought to the NICU, we’re taking away that natural feedback loop.”

The NICU at UF Health Shands Children’s Hospital includes a lactation support committee composed of nurses, lactation consultants and counselors, occupational therapists and neonatologists who pursued special training to better support breastfeeding mothers.

The hospital has also been awarded the Baby-Friendly designation by Baby-Friendly USA, a global initiative of the World Health Organization and UNICEF that recognizes birthing facilities that implement specific breastfeeding procedures.

AI scientists and engineers in the UF College of Medicine’s Quality and Patient Safety initiative are working with Hu and her clinical colleagues to fine-tune two AI models they created in an earlier phase of the MILK+ project that helps health teams predict risk of breastfeeding challenges.

A prenatal model synthesizes patient information to determine and notify care teams of factors that may increase difficulties, such as preexisting conditions and demographic and socioeconomic information including their home ZIP code and financial barriers. A postnatal model uses data collected on breastmilk production while the baby is still at the hospital to predict whether the mother will be able to continue fully supplementing the infant’s diet through breastmilk upon discharge.

“Before the baby is born, we only have mom’s information, but it still tells us something,” said Tanja Magoc, Ph.D., an associate director of AI/QI team that developed the models. “Then once the baby is born, we have much more information. That guided our decision to have two different models, because we have two different intervention points.”

Data from more than 18,000 mothers who received care and delivered at the hospital during a nine-year period contributed to the development of the prenatal model, while data from more than 22,000 newborns captured over an eight-year period guided creation of the postnatal model.

“We looked at hundreds of different features from the mother and baby related to their demographics, medical history, medications they take and more, and the AI model helped us screen which pieces of information were important to the development of maternal breastmilk,” Magoc said. “It was a big collaboration with the clinical team and their expertise over the course of a year and a half to create models that came up with good predictors.”

“We looked at hundreds of different features from the mother and baby related to their demographics, medical history, medications they take and more, and the AI model helped us screen which pieces of information were important indicators of the development of maternal breastmilk,” Magoc said. “It was a big collaboration with the clinical team and their expertise over the course of more than a year to create models that came up with good predictors.”

Hu said both models offer very good predictions, with the postnatal model providing a 95% accuracy rate. The models also provide a list of the top 10 contributing factors to a mother’s ability to breastfeed, which she said will be refined as the models continue to synthesize data from larger patient sample sizes.

“We are still in early days in terms of interventions, but what the model has helped us do so far is work on a pathway for identifying patients who are most vulnerable for not being successful with lactation on admission and throughout their baby’s stay,” she said.

For those patients, she explained, the team is developing a standardized protocol to review a checklist that ensures support has been provided in a timely manner. If a patient is already flagged as being at high risk, the team is prompted to check in with them earlier, to review progress. These patients are also flagged to receive lactation consultant services.

More clinicians at UF Health Shands Children’s Hospital are continuing to gain access to the models in their practice, and the hope is to eventually expand them to other UF Health facilities around the state.

“It’s important to have good communication between the clinical side and the AI side when you’re developing a model like this, to be able to put things into context and really emphasize the things that are actually actionable,” Hu said.

She also said health care providers need to be careful not to let the AI models unfairly influence their medical judgment.

“We’re also very mindful that it’s possible for the model to have skewed judgments that we’re not fully aware of,” Hu said. “The model is only as good as the data it was trained on. We have certain patient populations that are frequently represented and other populations that are not, so keeping in mind that it’s still just a prediction is very important. A patient who comes in with a very high-risk score would receive the same support we would provide for anyone.”