AI Travels Between Hospitals to Bring Adult-Trained Medical Models to Children

Artificial intelligence has transformed many corners of adult medicine, particularly in radiology, where algorithms now help radiologists detect disease, prioritize urgent cases, and catch errors that human eyes might miss. Yet one of the most vulnerable groups of patients — newborns and children — has remained largely outside this technological revolution. The reason is deceptively simple: the data needed to train medical AI models is far harder to obtain for children than for adults, and strict privacy rules keep the limited pediatric records that do exist locked inside individual hospitals. A new study published on October 13 in Nature Biomedical Engineering by researchers at Harvard Medical School and collaborators around the world now offers a way around this barrier, showing that adult-trained AI models can be adapted to work in infants using data from dozens of hospitals without any sensitive patient information ever leaving the institution that collected it.

The scale of the disparity is striking. According to Vishwanatha Rao, a member of Pranav Rajpurkar’s biomedical informatics lab at Harvard Medical School and co-first author of the study, children account for less than 1 percent of public medical imaging datasets. Medical records on newborns and children are subject to stringent privacy protections and tend to remain within the institution where the child is treated, which makes it nearly impossible to assemble the large, diverse datasets needed to train robust AI systems. “Because it’s so much harder to get data for children, this transformative technology is disproportionately being developed to help adults,” Rao said. “Children are being left behind.” The consequence is that advances that could speed diagnoses and reduce errors in neonatal and pediatric care remain out of reach for the very patients who arguably need them most.

The Harvard-led team chose a clinically critical test case to demonstrate their approach: the placement of endotracheal tubes, the breathing tubes inserted into the airway to treat respiratory problems. Correct positioning of these tubes is crucial, and there is little margin for error, especially in newborns whose airways are tiny and whose bodies tolerate mistakes poorly. A tube placed too shallow may fail to ventilate the lungs effectively, while one inserted too deep can damage tissue or ventilate only one lung. Prior studies have shown that AI models can quickly flag potential problems and shorten the time it takes radiologists to review and correct tube placement in adults, and multiple AI tools for this task have already received clearance from the U.S. Food and Drug Administration for adult use. Not one of them, however, is cleared for infants.

To build their pediatric model, the researchers started with an AI system that had been trained on adult chest radiographs to assess whether breathing tubes were positioned correctly. They then collected 2,965 chest x-rays from 30 neonatal intensive care units spanning 20 countries and territories. The key methodological innovation lay in how this distributed data was used. Rather than pooling the images into a central repository — which would violate the privacy constraints that make pediatric data so scarce in the first place — the researchers trained the model sequentially on each institution’s data, one hospital at a time. In this scheme, the model itself travels from site to site, learning from each hospital’s images locally, while the data stays in-house. This proof-of-principle design simulates a real-world scenario in which no sensitive information ever crosses institutional boundaries.

The results were remarkable, even to the researchers themselves. Despite wide differences in image quality and imaging characteristics across the participating neonatal intensive care units, the newly adapted model outperformed adult models trained on pediatric data from a single hospital at 29 of the 30 sites. Crucially, this advantage held even when the model was tested on images from a hospital it had never encountered during training. “We tested the model on a hospital it had never seen before, and it did better than a model trained on scans from that specific hospital — that’s a pretty powerful result,” said Rao, who is an MD-PhD student at the University of Pennsylvania. “And compared to the original adult model, it’s leaps and bounds better.”

The quantitative gains were substantial. The fully trained model was able to identify the location of the tip of the endotracheal tube — the measurement used to gauge whether the tube is too shallow or too deep — with an average error of 0.67 centimeters. By comparison, existing models adapted to children using data from a single institution averaged errors of 1.29 centimeters, and the original adult models were off by an average of 4.30 centimeters. In a clinical setting where millimeters matter, the difference between an error of more than four centimeters and one of less than seven-tenths of a centimeter is the difference between a tool that is unusable in newborns and one that could meaningfully support bedside decision-making in the neonatal intensive care unit.

The broader context underscores why this work matters. Only 17 percent of FDA-authorized AI medical devices are labeled for use in children, a statistic that reflects the same data scarcity that motivated the study. The sequential, privacy-preserving training framework demonstrated here is not tied to the specific task of tube placement. The researchers believe it could be generalized to adapt other adult-trained AI models for pediatric use, addressing what they describe as critical gaps in the deployment of pediatric AI. “This is a general framework that shows a cheap, privacy-preserving approach to take advantage of existing adult information and get it to work on children,” Rao said. “It could usher in a whole new generation of AI models that work for kids, adapted through this broad methodology.”

Senior author Pranav Rajpurkar, associate professor of biomedical informatics in the Blavatnik Institute at Harvard Medical School, framed the achievement in terms of equitable access to emerging technology. “Our goal is to figure out how to make medical AI models work for everyone, including vulnerable populations where data can’t easily be shared,” he said. “Now we’ve come up with a method which would allow for every institution to keep its data within its walls.” The approach also extends beyond pediatrics. Any patient population that is underrepresented in AI training data — whether because of rarity, privacy restrictions, or institutional silos — could in principle benefit from the same strategy of sequentially adapting models across many data-holding sites without centralizing sensitive records.

The study, whose co-first authors include Emma Chen and Agustina Saenz alongside Rao, drew on the collaborative effort of dozens of clinicians and researchers across the 20 countries and territories whose neonatal intensive care units contributed images. The work was supported by the Biswas Family Foundation’s Transformative Computational Biology Grant in Collaboration with the Milken Institute, with additional support for co-author Lucas Bijnens from Junior Orsi and Barco. The international breadth of the collaboration is itself part of the method’s promise: because the model learns sequentially from many institutions rather than one, it absorbs the natural variation in equipment, technique, and patient characteristics that would otherwise cause an AI system to fail when deployed outside its original training environment.

For the field of medical AI, the study offers both a technical demonstration and a conceptual shift. Instead of treating the scarcity of pediatric data as an insurmountable obstacle, the researchers showed that the knowledge embedded in adult-trained models can serve as a foundation, refined iteratively by small, locally held pediatric datasets at each participating hospital. “A lot of what research is meant to accomplish is showing that something is possible,” Rajpurkar said. “We’ve shown there’s a system by which you can tune these models without sensitive data leaving a hospital.” If the framework generalizes as the team hopes, the children who have been left behind by the AI revolution in medicine may finally begin to catch up — not by loosening privacy protections, but by designing algorithms that respect them from the start.

Subject of Research: A privacy-preserving sequential training method for adapting adult-trained medical AI models to pediatric imaging data

Article Title: Children are left out of medical AI – a new method could help close the gap

Article References: Children are left out of medical AI – a new method could help close the gap. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: medical AI, pediatrics, radiology, neonatal intensive care, endotracheal tube placement, privacy-preserving machine learning, chest x-rays, Harvard Medical School, Nature Biomedical Engineering, FDA-cleared AI devices, health equity, biomedical informatics