Children Are Still Missing From the Imaging AI Data That Will Shape Their Care
A new Nature analysis warns that children remain underrepresented in public medical imaging datasets, raising concerns about whether AI tools trained on those data will perform safely in pediatric care. The finding underscores a recurring problem in health AI: the populations most in need are often the least represented in the training data.
Public imaging datasets have become the backbone of many radiology AI systems, but they often reflect the easiest data to collect rather than the patients who most need them. If children are systematically underrepresented, then performance claims made on adult-heavy datasets can give a false sense of generalizability.
That matters because pediatric imaging is not just a smaller version of adult care. Anatomy, disease prevalence, imaging protocols, and acceptable error thresholds all differ, which means models that appear robust in adults may be unreliable in children even when they look clinically polished.
The broader issue is governance, not just data volume. Hospitals, researchers, and vendors need to treat dataset composition as a safety metric, with explicit reporting on age coverage, scanner diversity, and downstream validation in pediatric cohorts.
This is also a warning for policymakers and procurement teams. As imaging AI moves deeper into routine workflows, buyers should ask not only whether a model is accurate, but whether it was built for the patients they actually see.