Lack of children in public medical imaging data points to growing age bias in biomedical AI

Stanley Bryan Z. Hua, Nicholas Heller, Ping He, Alexander J. Towbin, Irene Y. Chen, Alex X. Lu, Lauren ErdmanmedRxiv 2025 · DOI

Abstract

Artificial intelligence (AI) is rapidly transforming healthcare, but its benefits are not reaching all patients equally. Children remain overlooked with only 17% of FDA-approved medical AI devices labeled for pediatric use. In this work, we demonstrate that this exclusion may stem from a fundamental data gap. Our systematic review of 181 public medical imaging datasets reveals that children represent just under 1% of available data, while the majority of machine learning imaging conference papers we surveyed utilized publicly available data for methods development. Much like systematic biases of other kinds in model development, past studies have demonstrated the manner in which pediatric representation in data used for models intended for the pediatric population is essential for model performance in that population. We add to these findings, showing that adult-trained chest radiograph models exhibit significant age bias when applied to pediatric populations, with higher false positive rates in younger children. This work underscores the urgent need for increased pediatric representation in publicly accessible medical datasets. We provide actionable recommendations for researchers, policymakers, and data curators to address this age equity gap and ensure AI benefits patients of all ages.

Introduction

While AI is revolutionizing medical imaging, its benefits are not distributed equally. Currently, only 17% of FDA-approved medical AI devices are labeled for pediatric use. This disparity is not merely a regulatory hurdle but a fundamental data problem. In medical AI research, public data is the engine of innovation; however, if that engine is fueled almost exclusively by adult data, the resulting models will inevitably fail when applied to children.

Children are not merely "small adults"; they possess distinct anatomy and physiology, such as different heart-to-chest ratios and bone growth patterns. Without representative data, AI models risk producing "algorithmic bias," leading to higher false-positive rates and potential over-treatment in younger populations.

Methodology

To quantify and demonstrate this bias, the researchers employed a three-pronged approach:

  1. Literature Survey: An analysis of papers from the Medical Imaging with Deep Learning (MIDL) 2023 and 2024 conferences to determine how often researchers use public datasets and whether they report patient ages.
  2. Systematic Review: An audit of 181 publicly available medical imaging datasets from major repositories (e.g., TCIA, Stanford AIMI, UK BioBank) and challenges to calculate the actual proportion of pediatric data.
  3. Bias Case Study: Training a state-of-the-art neural network (ConvNeXt-B) to detect cardiomegaly (enlarged heart) using four large adult chest X-ray datasets, then evaluating these models on a pediatric-specific dataset (VinDr-PCXR) to measure age-related error rates.
Overview of the study's three-pronged approach: scoping data use, reviewing public datasets, and evaluating age bias through a cardiomegaly case study.

The Pediatric Data Gap

The study revealed a staggering "pediatric data gap." Despite children making up a significant portion of the population and roughly 4% of all imaging exams in the US, they represent less than 1% of the patients in public medical imaging datasets. Out of 489,743 patients across 117 datasets with age information, only 4,558 were children.

Furthermore, a "metadata blackout" persists in the research community. 85% of the conference studies surveyed failed to report patient ages at all. Even when data is repackaged into new datasets or foundation models, age information is often discarded, making it impossible to audit for age bias. For many imaging modalities, a single dataset accounts for the vast majority of all available pediatric data (e.g., 70% of pediatric X-ray data comes from one bone age challenge), creating a fragile research ecosystem.

Algorithmic Bias: A Case Study in Cardiomegaly

The case study on cardiomegaly provides a clear example of how data exclusion leads to clinical harm. Children naturally have larger hearts relative to their chest size compared to adults. Because the AI models were trained only on adults, they interpreted the normal anatomy of young children as a pathological condition.

Quantitatively, there was a strong negative correlation (Pearson r = -0.928) between age and false-positive rates. The younger the child, the more likely the adult-trained model was to incorrectly diagnose them with an enlarged heart. Standard techniques like histogram matching failed to fix these high false-positive rates, proving the bias is rooted in anatomical differences rather than image quality.

Conclusion and Recommendations

The scarcity of pediatric data translates directly into algorithmic harm. To address this age equity gap, the authors propose several actionable steps:

  • Standardized Reporting: Researchers should be required to report age distributions in their datasets.
  • Incentivizing Pediatric Data: Policymakers and funding bodies should prioritize the curation and release of AI-ready pediatric datasets.
  • Targeted Evaluation: AI developers should perform stratified evaluation across age groups to identify potential failures before models reach clinical settings.

By addressing these data gaps, the medical AI community can ensure that the benefits of innovation reach patients of all ages.

Summary updated

Abstract

Artificial intelligence (AI) is rapidly transforming healthcare, but its benefits are not reaching all patients equally. Children remain overlooked with only 17% of FDA-approved medical AI devices labeled for pediatric use. In this work, we demonstrate that this exclusion may stem from a fundamental data gap. Our systematic review of 181 public medical imaging datasets reveals that children represent just under 1% of available data, while the majority of machine learning imaging conference papers we surveyed utilized publicly available data for methods development. Much like systematic biases of other kinds in model development, past studies have demonstrated the manner in which pediatric representation in data used for models intended for the pediatric population is essential for model performance in that population. We add to these findings, showing that adult-trained chest radiograph models exhibit significant age bias when applied to pediatric populations, with higher false positive rates in younger children. This work underscores the urgent need for increased pediatric representation in publicly accessible medical datasets. We provide actionable recommendations for researchers, policymakers, and data curators to address this age equity gap and ensure AI benefits patients of all ages.
BibTeX
@misc{hua2025lack,
  author = {Stanley Bryan Z. Hua and Nicholas Heller and Ping He and Alexander J. Towbin and Irene Y. Chen and Alex X. Lu and Lauren Erdman},
  title = {{Lack of children in public medical imaging data points to growing age bias in biomedical AI}},
  year = {2025},
  doi = {10.1101/2025.06.06.25328913},
  howpublished = {medRxiv},
}
Figure