The Architecture of Neurocognitive Abilities in the Developing Functional Connectome

• In Event: Using Connectomics and Machine Learning to Examine Brain Development

Abstract

Background: The architecture of cognitive abilities has been a hotly contested issue in psychology for over a century. Some theorists claim there is a single unified brain basis for performance across diverse cognitive tasks (Spearman, Jensen). Others argue the brain houses multiple distinct mechanisms that underpin domain-specific cognitive abilities, and psychometric G is nothing more than a statistical artifact (Thompson, Gould, Conway). This debate has particular relevance for developmental science. Cognitive abilities mature rapidly during childhood and adolescence, and a central aim for the field is to map behavioral and brain trajectories of these abilities—an aim that presupposes a basic understanding of their architecture. Recent years have seen the emergence of functional connectomics as a central method in neuroimaging, but the power of this approach has not yet been brought to bear to delineate the architecture of cognitive abilities. Here we introduce novel whole-brain multivariate connectomic methodology that sheds critical light on which cognitive abilities are “encoded” in the brain.

Methods: The sample consists of 5,937 9- and 10-year-olds in the Adolescent Brain Cognitive Development 21-site study who completed a comprehensive 11-task neurocognitive battery and had high-quality functional connectomes that capture brain-wide connectivity. In multivariate predictive modeling analysis, we build predictive models for G, as well as for 11 individual cognitive tasks controlling for G-related variance. In SEM analysis, we apply tests to each task-modulated connection of the connectome. These tests compare models in which the connection operates through a G-based “common pathway” that contributes to all 11 neurocognitive tasks or a “specific pathway” involving one or more individual tasks.

Results: In multivariate predictive modeling analysis with leave-one-site-out cross-validation, the correlation between predicted and actual G scores was 0.36 (14% R2cross-validated, pPERM<0.0001). Importantly, even after controlling for G, we could predict all 11 individual tasks (9 of 11 pPERM’s<0.0001), indicating substantial discriminative information in the connectome about individual cognitive tasks over and above shared G-related variance. In SEM analysis, we found G-related edges were the most prevalent single type of connection in the connectome (24.8%), followed by vocabulary (14.6%) and reading (12.3%), with the remaining tasks all under 10%. Given error profiles for SEM-based ascriptions (estimated from simulations), the observed distribution of connections ascribed to G and specific tasks statistically ruled out two hypotheses: 1) the connectome contains only G connections; and 2) the connectome contains only specific ability connections (p’s<0.0001).

Conclusion: Novel multivariate connectomic analyses jointly and convergently provide evidence against historically influential positions about the architecture of cognitive abilities. Our results instead best support a pluralist model in which “generalist” G-related connections co-exist with “specialist” task-specific connections in the 9- and 10-year old human functional connectome.

Author

• Chandra Sripada
  Presenting Author