Toward a statistical foundation for brain signatures of behavioral outcome

Abstract

AbstractBackgroundThe “brain signature of cognition” concept has garnered interest as a data‐driven, exploratory approach to better understand key brain regions involved in specific cognitive functions, with the potential to maximally characterize brain substrates of clinical outcomes. However, to be a robust brain phenotype, the signature approach requires a statistical foundation showing that model performance replicates across a variety of cohorts. Here, we outline a procedure that provides this foundation for a signature models of episodic memory and extend this to a measure of functional everyday memory.MethodUniversity of California, Davis (UCD) and ADNI provided imaging and memory data. For cognitive and ECog memory, we derived regional brain gray matter (GM) cortical thickness (CT) measures, testing for replicability (Fig. 1). Signature model generation used two steps in each cohort: 1) Randomly select 40 discovery subsets of size N = 400. In each subset, use voxelwise regression to compute the CT signature regions for outcome at three t‐values of association. 2) Compile frequency maps of region selection across all 40 discovery sets and designate regions with selection frequency >70% as “consensus” signature masks for each cohort. Model validation used regressions of outcomes on consensus masks: 1) Compare adjusted R2 fits of both consensus signature models in 40 discovery sets in each cohort (thus, symmetrically compare same and oppositely generated models). 2) In each full cohort, compare signature model fits with competing models.ResultSpatial replications produced strongly convergent consensus signature regions derived from UCD and ADNI (Fig. 2). Consensus model fits were highly correlated in 40 random subsets of each cohort (Fig. 3 top) indicating high replicability, with tight (∼80%) ratios of within/outside cohort model fits (Fig. 3 bottom), indicating low bias. In comparisons over each full cohort, signature models outperformed other models with one exception (Fig. 4).ConclusionMultiple random model generations, followed by consensus selection of regional brain substrates, produced signature models that both consistently replicated model fits to memory and outperformed other commonly used measures. Results support consensus signature models as robust biomarkers of cognition and everyday function across imaging cohorts with differing demographic composition (see Table).