A recent study published in Nature Biomedical Engineering reveals a new method for adapting adult-trained AI models for pediatric use, addressing significant disparities in medical imaging for children.
On October 13, 2023, a team of researchers at Harvard Medical School announced a groundbreaking study that seeks to close the data gap in artificial intelligence (AI) applications for pediatric health. The study highlights that existing AI models in radiology predominantly focus on adult data, leaving children underserved in terms of advanced medical imaging analysis. Vishwanatha Rao, a member of the research team, emphasized that the technology’s current trajectory tends to favor adult patient populations, stating, “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.”
The challenge stems from strict privacy regulations surrounding medical records for infants and children. These records typically remain within the confines of the healthcare facility, which limits their availability for training robust AI models. According to Rao and his colleagues, children constitute less than 1 percent of public medical imaging datasets, underscoring the urgency for innovative solutions that respect patient privacy while enhancing healthcare delivery.
Innovative Approach to AI Training
In light of these challenges, the researchers developed a method to adapt existing adult-trained AI models for pediatric applications without compromising patient confidentiality. This approach, detailed in their study, enables the model to utilize data from multiple hospitals while ensuring that sensitive information remains securely within each institution. Rao explained that their goal is to create a framework that allows medical AI models to work for all populations, particularly those where data sharing is limited.
The researchers tested their method on a model trained to assess the correct placement of endotracheal tubes, which are critical for treating respiratory issues in newborns. The necessity for precise placement in this demographic makes the research particularly timely. While several AI tools have received FDA clearance for similar tasks in adults, none are yet authorized for use with infants.
To validate their approach, the team collected a dataset of 2,965 chest X-rays from 30 neonatal intensive care units (NICUs) across 20 countries and territories. They trained their AI model sequentially on data from each participating institution, demonstrating that the model could outperform existing adult models even when applied to data from hospitals it had never seen before. The new model achieved an average error of just 0.67 centimeters in identifying the position of the endotracheal tube tip, compared to 1.29 centimeters for models adapted from pediatric data at a single institution and 4.30 centimeters for adult models.
Implications for Pediatric Medicine
The significance of this research extends beyond the immediate findings. Only 17 percent of FDA-authorized AI medical devices currently include labeling for use in pediatric care. The study’s authors believe that their method could pave the way for a new generation of AI tools tailored for children, using existing adult data while protecting patient privacy. Rao articulated the broader applicability of their findings, stating, “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.”
Furthermore, the methodology developed in this study is not limited to pediatric applications alone. It has the potential to enhance AI training for any patient population that is typically underrepresented in existing medical datasets. Rajpurkar, the senior author of the study and an associate professor of biomedical informatics in the Blavatnik Institute at Harvard Medical School, remarked, “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.”
Collaborative Efforts and Future Directions
This research was collaboratively conducted by a diverse team of authors, including Emma Chen and Agustina Saenz as co-first authors. The study was supported by the Biswas Family Foundation’s Transformative Computational Biology Grant in Collaboration with the Milken Institute, with additional funding for some authors from various sources. The collaborative nature of this research underscores the importance of interdisciplinary efforts in addressing complex healthcare challenges.
As the field of medical AI continues to evolve, the findings from this study highlight the critical need for innovative approaches to ensure that advancements in technology benefit all patient populations, particularly the most vulnerable. By bridging the data gap through privacy-conscious methodologies, researchers hope to usher in a new era of AI applications that enhance diagnostic accuracy and improve health outcomes for children and beyond.
