No Normal Brain: How Demographic Exclusion Undermines Neuroimaging AI Validity

Comprehensive Summary

Neuroimaging AI tools are increasingly used to make clinical decisions, yet the brain scan datasets used to train these tools are overwhelmingly drawn from a narrow slice of humanity, raising serious concerns about whether these systems work reliably for most of the world's population. The author analyzed demographic data from over 55,000 participants across five major neuroimaging datasets, comparing their composition against global population statistics and reviewing published studies that reported AI performance differences across demographic groups. The findings reveal striking imbalances: White participants make up 52 to 94% of these datasets despite representing only about 16% of the global population, and 85 to 100% of participants come from North America or Europe, regions that together represent only 15% of people worldwide. These imbalances translate into real performance gaps, with seizure detection AI showing 27% lower sensitivity in adolescent females and dementia screening tools producing twice as many false positives in Black participants compared to White participants. The author argues that this is not merely a fairness issue but a fundamental scientific validity problem, because AI systems trained on such skewed data learn a distorted definition of what a normal brain looks like and may misclassify normal variation in underrepresented groups as disease.

Outcomes and Implications

This work is directly relevant to any clinician who relies on or is likely to rely on AI-assisted neuroimaging interpretation, as the documented performance gaps mean that current tools may systematically fail the very patients who face the highest disease burden. The practical implication is that AI tools should not be deployed in diverse clinical populations without first being validated in those populations, and that the currently available tools showing the largest performance disparities should be used with caution outside the demographic groups they were trained on. The author does not provide a specific timeline for correction but calls on funding agencies to strengthen requirements for demographic representation in AI training datasets as a near-term institutional priority.