Scientists at the March of Dimes Prematurity Research Center (PRC) at Stanford have created an Artificial Intelligence/Machine Learning (AI/ML) algorithm to identify early preterm babies at highest risk of some of the most serious complications of prematurity — from the same heel prick routinely given to babies after birth to rule our rare disorders.
The findings, published recently in the journal Science Translational Medicine, are striking: the ML algorithm can analyze data from a routine blood spot taken from a preterm baby and estimate that baby’s risk of developing four serious health problems. One is a chronic lung disease called bronchopulmonary dysplasia, another is a type of brain bleeding called intraventricular hemorrhage, the third is a dangerous intestinal condition called necrotizing enterocolitis, and the last is an eye disease called retinopathy of prematurity that can threaten vision. Together, they are among the most common and serious complications faced by extremely preterm infants.
Focusing on blood metabolites, or signals related to metabolism — one of the most revealing biological processes that gives insight into health and disease — the algorithm looks for patterns associated with the four complications and scores babies born between 22 and 29 weeks based on their metabolic risk. Babies with the strongest signals toward each disease pattern are flagged as high risk, and, if the model was available clinically in the future, could be immediately referred for closer monitoring, said Stanford PRC investigator and senior study author Dr. Nima Aghaeepour.
“This model is a powerful early warning sign that has the potential to get at-risk newborns attention and personalized care much faster than existing tools,” said Dr. Aghaeepour, “which could save their lives.”
Dr. Aghaeepour cautioned that the model, which requires further validation and testing, is not diagnostic, and is intended to be used a risk stratification tool, prompting vulnerable babies to be assessed further and potentially diagnosed and treated.
“This isn’t a matter of relying on the algorithm,” he said. “It’s about using it to get perspective earlier and then to confirm a diagnosis.”
Dr. Aghaeepour said the predictions from the model were consistently stronger than those from models based only on standard clinical variables such as gestational age, birthweight, sex, Apgar scores, and clinician judgment, which represent the traditional clinical barometers for assessing risk for prematurity-related complications.
Those indicators, Dr. Aghaeepour and other PRC investigators say, are useful, but miss the variabilities and differences in every baby that inform a distinct risk profile and health trajectory. For example, he said, the study showed that while some babies born early developed one or more of the four complications, others born at the same time did not, reinforcing previous findings from the PRC that challenge the traditional definition of prematurity and push the field into an era of precision medicine, or tailored treatment to each baby based on that baby’s biology.
“I’ve said this before, but prematurity is not one size fits all,” he said, “and this research is more evidence of it. It brings us closer to precision medicine for every mother and every newborn, not only so they can get faster and more appropriate care, but so we can make research breakthroughs from the nuances of their biological profiles.”
Dr. Aghaeepour’s team created the “metabolic health index” algorithm by linking newborn screening blood spot tests from 13,536 preterm babies born between 22 and 29 weeks in California from 2005 to 2010 with the future clinical outcomes of those same babies, getting insight into the metabolic patterns that most often result in BPD, IVH, NEC, and ROP. Once trained, the algorithm was validated on another cohort of 3,299 preterm babies from Canada and performed well in predicting babies that went on to develop the complications, Dr. Aghaeepour said.
“Essentially, it compressed dozens of metabolite measurements into a single score for each newborn, and newborns with the highest scores ended up having the worst outcomes.”
Blood metabolites that acted as the strongest red flags for the algorithm, resulting in a high index score, included long-chain and short-chain acylcarnitines and several amino acid ratios, such as ornithine/citrulline and phenylalanine/tyrosine — all signs of potential metabolic dysfunction.
Because the blood collected for newborn screening is meant to detect specific diseases associated with errors in metabolism, the existing blood analysis is enough for the algorithm to use for its own metabolic inquiry, requiring no additional blood or analysis to be taken from preterm babies at birth.
“This data is already collected from almost every newborn in the U.S.,” Dr. Aghaeepour said. “It’s all work we’re doing anyway, it just the AI model is giving us more value out of what we already have.”
When the PRC team added clinical variables on top of metabolites, the algorithm’s prediction accuracy improved to more than 85%, suggesting little overlap between the two, and strengthening the argument for using the algorithm as an addition to existing risk tools.
Dr. Aghaeepour said the future looks bright for the algorithm, with more validation planned in more recent birth cohorts, integration into clinical workflows, and studies to see whether use of the model can change clinical management or outcomes.
“A best-case scenario in five years is that this becomes a risk layer added to newborn screening to guide monitoring and early interventions,” he said, “but that will of course require further trials and regulatory review.”