Clinical neuroscience needs an algorithmic fairness perspective

Large-scale, data-driven models are transforming clinical neuroscience, with normative modeling now enabling lifespan characterization of brain structure and function from tens of thousands of individuals across dozens of sites. These models are increasingly applied to psychiatric and neurological conditions, including psychosis, autism, mood disorders, Alzheimer’s disease, and Parkinson’s disease. Yet as big data and machine learning become the norm, the field must remain self-critical about how these models are built and interpreted.

This talk addresses the reference class problem: the challenge that a patient’s estimated risk or “normative” deviation depends on which comparison group they are assigned to, and that the dimensions used to define these groups (age, sex, race, socioeconomic status, and countless others) are neither neutral nor obvious choices. Drawing on an audit of widely used brain-structure normative models, the talk demonstrates that racial disparities emerge in individual deviation scores, even when overall model performance metrics show no group differences, and even when race is excluded from training. Self-reported race can still be predicted from deviation patterns alone, revealing bias encoded in how models represent “normal.”

The talk argues that big data does not equal diverse, representative data, and that ignoring the reference class problem does not make it disappear. It closes with a call for greater transparency, more diverse data curation, and conceptual clarity about research questions and target populations before deploying machine learning in clinical settings — positioning algorithmic fairness not as a side concern, but as central to responsible clinical translation of neuroimaging research.