Addressing gaps in pediatric AI training data
Researchers have developed a method for adapting artificial intelligence (AI) models trained on adult radiology images for pediatric use. The approach aims to address the limited availability of children's medical imaging data for AI development.
Only about 17% of U.S. Food and Drug Administration-authorized AI medical devices are labeled for use with children. Most models are trained on adult data, which may not translate well to infants and children.
Strict privacy regulations governing pediatric medical records present another obstacle, as data tend to remain at the institution where treatment occurred. Children account for less than 1% of public medical imaging datasets.
Vishwanatha Rao, an MD-PhD student at the University of Pennsylvania and research fellow at Harvard, highlighted the consequences of this data gap. "Because it’s so much harder to get data for children, this transformative technology is disproportionately being developed to help adults. Children are being left behind."
Training across hospitals without pooling data
Rao and colleagues adapted an AI model trained on adult X-rays for pediatric use through an approach that allowed multiple hospitals to contribute data without transferring sensitive information between institutions.
The study included nearly 3,000 chest radiographs from 30 neonatal intensive care units (NICUs) across 20 countries and territories. To simulate a setting in which patient data never leave the originating hospital, the researchers trained the model sequentially using data from each institution.
The training process involved:
- 30 participating NICUs across 20 countries and territories.
- 50 chest radiographs contributed by each hospital for model training.
- Sequential training without centralized pooling of patient data.
The resulting model was designed to assist with assessing breathing tube placement by identifying the location of the endotracheal tube tip.
Model performance across participating sites
The adapted model outperformed adult models trained on pediatric data from a single hospital at 29 of the 30 sites. This result held despite differences in image quality and characteristics. The adapted model also performed better when tested on images from a hospital whose data had been used to train the single-institution models, although the adapted model had never seen those images.
The fully trained AI model identified the endotracheal tube tip with an average error of 0.67. For comparison, existing models adapted for children using data from a single institution had an average error of 1.29 cm, while adult models had an average error of 4.3 cm.
Potential applications beyond pediatrics
The researchers suggest that the framework could be used to adapt other adult AI models for pediatric applications. It may also be applicable to other patient populations that are underrepresented in AI training datasets.
Rao described the approach as a way to make greater use of existing adult data while maintaining privacy protections. "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," he said. "It could usher in a whole new generation of AI models that work for kids, adapted through this broad methodology."
Source: Harvard Medical School
