Characterizing neuroimaging 'brain-behavior' model performance bias in rural populations

Modified Project Summary Section Nearly one-fifth of the Unites States population resides in a rural region, and approximately one-fifth of those residents suffers from a mental illness. While these rates of mental illness are similar to urban areas, individuals living in rural regions face a disproportionate burden of negative psychiatric outcomes. Modern advances in psychiatric research have focused on using machine learning and human neuroimaging to predict diagnoses and treatment outcomes. However, recent evidence suggests that machine learning models themselves may drive differences in health outcomes through performance differences. Specifically, clinical decision-making models created in predominantly one population group may demonstrate reduced generalizability in other population groups (e.g., poorer likelihood of choosing the correct treatment if patients are rural). Given that virtually all neuroimaging ‘brain-behavior’ predictive models in psychiatry research are generated from data collected in highly populated metropolitan areas, this study will evaluate ‘brain-behavior’ models for performance differences in rural populations. It will also investigate means of eliminating these performance differences that could create further health outcome gaps in rural populations. In Aim 1, I will use neuroimaging data from 9,811 individuals in the Adolescent Brain and Cognitive Development Study to create a ‘brain-behavior’ predictive model of cognition. In Aim 2, I will evaluate this model for urban-rural performance differences and pursue strategies to reduce these differences. This study will have important implications for understanding how algorithms in healthcare drive health outcome gaps and how we can reduce these gaps by designing models that perform the same across all populations.