A recent study reveals that artificial intelligence models in radiology are predominantly trained on adult data, leaving pediatric patients underserved. Researchers propose a method to adapt these models for children while maintaining data privacy.
In a groundbreaking study published on October 13, 2023, in Nature Biomedical Engineering, researchers from Harvard Medical School have highlighted a significant disparity in the application of artificial intelligence (AI) in medical imaging, particularly in pediatric care. The study, led by Vishwanatha Rao, a member of Pranav Rajpurkar’s biomedical informatics lab, emphasizes that current AI models are primarily developed using adult data, which consequently limits their effectiveness for infants and children.
Vishwanatha Rao remarked, “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.” This statement underscores a critical issue within the medical community, as children represent less than 1 percent of public medical imaging datasets, highlighting the challenges faced in integrating AI technology into pediatric healthcare.
The Privacy Challenge
A significant barrier to creating robust AI models for pediatric patients is the stringent privacy regulations surrounding medical records of newborns and children. Such records are often restricted to the healthcare institutions where the patients receive treatment, limiting access to data that could be utilized to enhance AI training. The Health Insurance Portability and Accountability Act (HIPAA) and other privacy regulations complicate the sharing of sensitive medical information, which is crucial for the development of effective AI models tailored for younger patients.
Rao and his colleagues developed a novel approach that enables AI models trained on adult data to be adapted for pediatric use without the sensitive data ever leaving its originating institution. This method ensures that while the AI model can be refined using various datasets, the actual data remains secure within the hospital’s systems. The researchers believe this approach could potentially serve as a blueprint for future developments in medical AI, particularly in fields where patient data privacy is a significant concern.
Research Methodology
The research team began their work with an AI model previously trained to assess the proper placement of endotracheal tubes—a critical procedure in managing respiratory issues, particularly in newborns. The accuracy of this placement is essential; misplacement can lead to serious complications. While several AI tools have received FDA clearance for use in adults, none currently exist for infants.
To illustrate the effectiveness of their method, the researchers collected a dataset comprising 2,965 chest X-rays from 30 neonatal intensive care units across 20 countries and territories. They trained their AI model sequentially on each institution’s data, allowing the model to learn from a diverse array of cases without compromising patient privacy. This innovative method facilitated the use of local data while ensuring compliance with privacy regulations, a critical step in the advancement of AI in healthcare.
The results were promising. The newly trained model demonstrated a significantly higher accuracy compared to existing models that were adapted for pediatric use from a single institution. In 29 out of 30 sites, the new model outperformed those based solely on pediatric data from individual hospitals. The model was able to pinpoint the tip of the endotracheal tube with an average error of 0.67 cm, compared to 1.29 cm for pediatric models and 4.30 cm for adult models. These findings not only validate the researchers’ approach but also highlight the potential for improved patient outcomes through the use of tailored AI applications.
Implications for Pediatric Healthcare
Only 17 percent of AI medical devices authorized by the FDA are labeled for use in children, a statistic that underscores the urgent need for advancements in this area. The underrepresentation of children in medical AI training datasets poses a significant risk, as treatments developed without considering pediatric physiology may not be effective or safe. While the study specifically focused on endotracheal tube placement, the researchers believe their methodology could be extended to adapt other adult-trained AI models for pediatric applications.
Rao stated, “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. It could usher in a whole new generation of AI models that work for kids, adapted through this broad methodology.” This statement reflects the potential for the research’s implications to extend beyond the immediate focus on respiratory care, potentially impacting various areas of pediatric medicine.
Furthermore, the framework developed by the research team is not limited to pediatric applications alone; it holds potential for addressing underrepresentation in AI training data across various patient populations. By adapting methodologies for different demographics, the healthcare field can work towards more equitable AI applications across all age groups and patient categories.
Research Team and Support
Emma Chen and Agustina Saenz served as co-first authors of the study, with contributions from a diverse group of researchers, including Oishi Banerjee, Henrik Marklund, and others. The research was supported by the Biswas Family Foundation’s Transformative Computational Biology Grant in collaboration with the Milken Institute, with additional support for team members from Junior Orsi and Barco. This diverse collaboration highlights the importance of interdisciplinary efforts in advancing medical technology.
As the medical field continues to explore the integration of AI technology, this study serves as a pivotal step toward ensuring that advancements in AI are accessible to all patient populations, particularly vulnerable groups like children. The findings point to a future where AI can be safely and effectively utilized across a broader spectrum of healthcare scenarios, ultimately improving outcomes for the youngest and most vulnerable patients.
