At a glance
Researchers adapted an adult X-ray AI model for newborns with data from 30 hospitals, using a method designed to keep each hospital’s data in place.
The new model outperformed adult models trained on pediatric data from a single hospital at 29 of 30 sites, even when tested on that hospital’s own images.
This approach could be generalized to adapt other adult-trained AI models for use in children without needing large amounts of pediatric data to be added to public imaging datasets.
In the field of radiology, AI models have the potential to improve imaging analysis, speed up diagnoses, and catch errors that might otherwise be missed. But the vast majority of these models are trained on adult data and don’t translate well to infants and children, keeping these advances out of reach of neonatal and pediatric health.
“Because it’s so much harder to get data for children, this transformative technology is disproportionately being developed to help adults,” said Vishwanatha Rao, a member of Pranav Rajpurkar’s biomedical informatics lab at Harvard Medical School. “Children are being left behind.”
A large part of the reason for the gap is that medical records on newborns and children are subject to strict privacy regulations and tend to remain within the institution where the child is a patient, leaving data less accessible to train robust AI models. According to Rao and colleagues, children represent less than 1 percent of public medical imaging datasets.
Rao led and is co-first author on a new study, published Oct. 13 in Nature Biomedical Engineering, showing that models trained on adult radiographs can be adapted for infants using data from multiple hospitals without sensitive information ever needing to leave each institution. The model travels, but the data stays in-house.
The team believes its proof-of-principle approach could be extended to adapt other adult-trained AI models for pediatric use, addressing critical gaps in the deployment of pediatric medical AI.
“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,” said senior author Rajpurkar, associate professor of biomedical informatics in the Blavatnik Institute at HMS. “Now we’ve come up with a method which would allow for every institution to keep its data within its walls.”
Training sequentially
The researchers started with a model that had been trained on adult data to assess whether breathing tubes have been placed correctly. Breathing tubes, or endotracheal tubes, are crucial for treating respiratory issues, and there is little margin for error, especially in newborns.
Studies have shown that AI models can quickly flag potential issues and shorten the time it takes for radiologists to review and correct tube placement. Multiple AI tools have already received FDA clearance for this task in adults, but none of them are cleared for infants.
Rajpurkar and colleagues collected 2,965 chest X-rays from 30 neonatal intensive care units across 20 countries and territories. To simulate a scenario in which the data never leaves the hospital, the researchers trained the model sequentially on each institution’s data.
Despite differences in image quality and characteristics, the new model outperformed adult models trained on pediatric data from a single hospital at 29 of 30 sites, even when tested on that hospital’s own images, which the new model had never seen.
The fully trained model was able to identify the location of the tip of the endotracheal tube — used to gauge whether the tube is too shallow or too deep — with an average error of 0.67 cm. Existing models adapted to children with data from a single institution averaged errors of 1.29 cm, and adult models were off by an average of 4.30 cm.
“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.”
Medical AI for all
Only 17 percent of FDA-authorized AI medical devices are labeled for use in children. While this particular model was trained on endotracheal tube placement, the approach could be generalized to adapt other adult models for pediatric use, the researchers said.
“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.”
The method isn’t limited to pediatrics, either. The same framework could potentially be applied to any patient population that is underrepresented in AI training data.
“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.”
Authorship, funding, disclosures
Emma Chen and Agustina Saenz are co-first authors of the study. Additional authors include Oishi Banerjee, Henrik Marklund, Lucas Bijnens, Sung Eun Kim, Evan Zucker, Susan Shelmerdine, Jeremy Palacio, Kent Ryan Kleinschmidt, Sathvik Suryadevara, Sri Sai Dinesh Jaliparthi, Jiaxi Lu, Rithvik Akula, Mario Scala, Julian F. Niedermair, Nikolaus Stranger, Sebastian Tschauner, Georg M. Schmölzer, Maria Jose Veloza Vega, Juan Sebastián Osorio Valencia, Oscar Jaramillo Robledo, Oscar Julian López Uribe, Saif Afat, Thomas Küstner, Soroosh Tayebi Arasteh, Daniel Truhn, Kelsey L. Pomykala, Jens Kleesiek, Naela Himayati Afifah, Rozan M. Irfan, Reza Shahriarirad, Amirhossein Erfani, Israel Amirav, Moria Be’er, Yasuhito Mitsuyama, Daiju Ueda, Alain S. Abi-Ghanem, Safaa Al Zakleet, Mossad Abdelhak Shaban Mohamed, Wan Irfan Bin W. Mustapha, Naila Nizami, Rafeah Khan, Khalid Al-Kharazi, Zafar Iqbal, Abdullah Al Shahrani, Amna Kashgari, Tijana Prodanovic, Nikola Prodanovic, Lim Shi Wei Desmond, James Thomas Patrick Decourcy Hallinan, Phumza Nongena, Melissa Louise Cooke, Montserrat Bret-Zurita, Almudena Alonso-Ojembarrena, Alexandra Nemes, Yung-Chieh Lin, Wei-Ting Lin, Wen-Hui Fang, Chin Lin, Atchara Mahayosnond, Sira Sriswasdi, Piyapong Khumrin, Suphakris Tanmuk, Didem Arman, Serdar Cömert, Serdar Alan, Umit Ayse Tandircioglu, David Andrew Mong, Esther Chung, Michael Sadler, Morgan P. McBee, and Richard H. Jones.
This work was supported by the Biswas Family Foundation Transformative Computational Biology Grant in collaboration with the Milken Institute. Bijnens was supported by Junior Orsi and Barco.
