Privacy-Preserving Learning Model Adapts Adult Radiology AI for Neonatal Care

Privacy restrictions limit access to children’s medical records, and pediatric images account for less than 1% of public medical imaging datasets. This scarcity encourages reliance on models trained primarily on adult data, which may not perform well in infants and children. A new study explores a way to bridge this divide by adapting adult-trained radiograph models for neonatal use while keeping hospital data local.

Harvard Medical School researchers developed a privacy-preserving continual learning approach to adapt artificial intelligence radiograph models trained on adult data for neonatal care. The model was originally designed to assess whether endotracheal tubes had been placed correctly in adults. These tubes help support breathing in patients with respiratory problems, and precise placement is particularly important in newborns, making this task a focus for adapting the model to neonatal imaging.

To enable this adaptation without sharing sensitive patient records, the approach allows the model to move between institutions while patient data remain within each hospital. Researchers simulated this process by training the model sequentially on data from individual institutions, allowing it to learn from multiple sites. This strategy was designed to overcome privacy barriers that often restrict access to pediatric medical records for artificial intelligence development.

The approach was evaluated in an international multicenter study involving 2,965 chest X-rays from 30 neonatal intensive care units across 20 countries and territories. After sequential training, the fully trained model identified the endotracheal tube tip location with an average error of 0.67 cm. This compared with an average error of 1.29 cm for existing models adapted to children using data from a single institution and 4.30 cm for adult models, indicating improved localization through the multicenter learning approach.

Despite differences in image quality and characteristics, the continually trained model outperformed adult models trained on pediatric data from a single hospital at 29 of 30 sites. This was observed even when the comparison model was tested on images from the same hospital and the continually trained model had not seen those images. The study, titled “Repurposing adult AI radiograph models for neonatal care through continual learning: an international multicentre study,” was published in Nature Biomedical Engineering.

Only 17% of FDA-authorized artificial intelligence medical devices are labeled for pediatric use, according to the article. Although this work focused on endotracheal tube placement, the researchers stated that the framework could potentially be applied to other adult models and to other underrepresented patient populations. They described the study as a proof-of-principle approach for adapting medical artificial intelligence without sensitive data leaving a hospital.

“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“. And compared to the original adult model, it's leaps and bounds better.,” said Vishwanatha Rao, a member of Pranav Rajpurkar’s biomedical informatics lab at Harvard Medical School and an M.D.-Ph.D. student at the University of Pennsylvania.

“A lot of what research is meant to accomplish is showing that something is possible. We've shown there's a system by which you can tune these models without sensitive data leaving a hospital,” said Pranav Rajpurkar, associate professor of biomedical informatics in the Blavatnik Institute at Harvard Medical School.

Related Links
Harvard Medical School