Assessing the generalizability of individual brain models

Cognitive neuroscience has traditionally focused on identifying the neural basis of psychological traits or state effects across large samples of participants. Recently, researchers have pushed towards providing more precise estimates of individual functional organization to better understand both psychological constructs as well as their supporting neural mechanisms. Despite growing interest, field-standard methods are poorly suited to these questions. In particular, they assume that aligning large-scale neuroanatomy enables directly mapping between individual’s neural activity patterns. However, variable structure-function correspondence across cortex—as well as degenerate neural population dynamics—suggests this assumption is misplaced. To address this gap, cognitive neuroscientists have turned to “deep phenotyping” studies, in which neural activity is repeatedly measured from a limited number of participants performing a large range of tasks. It remains unclear, however, how best to draw inferences from these datasets beyond the small number of originally-sampled participants. Such inferences will be essential to understand how fine-grained differences in individual function are associated with behavioral differences or disease states, particularly in populations where we cannot collect extensive measurements. In this project, we approach this problem through the application and extension of transfer learning methods. We first ask whether we can reconstruct some task activity patterns from other, related tasks (e.g., naturalistic movie-watching and classic, controlled psychological tasks examining constructs such as working-memory), which would help to expand the available information when only a limited number of measurements can be collected. We will then ask whether we can generalize learned relationships from one deep phenotyping cohort of participants to another similarly-sampled cohort. Together, these experiments will provide a first estimate of how comprehensive brain models may generalize across individuals. By examining successfully learned features, we will identify shared principles that organize neural activity and bridge between neuroanatomy and individual experience.