Toward a “treadmill test” for cognition: Improved prediction of general cognitive ability from the task activated brain

Abstract

General cognitive ability (GCA) refers to a trait‐like ability that contributes to performance across diverse cognitive tasks. Identifying brain‐based markers of GCA has been a longstanding goal of cognitive and clinical neuroscience. Recently, predictive modeling methods have emerged that build whole‐brain, distributed neural signatures for phenotypes of interest. In this study, we employ a predictive modeling approach to predict GCA based on fMRI task activation patterns during the N ‐back working memory task as well as six other tasks in the Human Connectome Project dataset ( n = 967), encompassing 15 task contrasts in total. We found tasks are a highly effective basis for prediction of GCA: The 2 ‐ back versus 0 ‐ back contrast achieved a 0.50 correlation with GCA scores in 10‐fold cross‐validation, and 13 out of 15 task contrasts afforded statistically significant prediction of GCA. Additionally, we found that task contrasts that produce greater frontoparietal activation and default mode network deactivation—a brain activation pattern associated with executive processing and higher cognitive demand—are more effective in the prediction of GCA. These results suggest a picture analogous to treadmill testing for cardiac function: Placing the brain in a more cognitively demanding task state significantly improves brain‐based prediction of GCA.

Brain Connectivity Patterns in Children Linked to Neurocognitive Abilities

“…The utility of imaging-based markers for psychological traits and abilities depends heavily on their applicability to new datasets collected at heterogenous sites with different subject characteristics and scanners, and this study confirms that strong generalizability is possible. Third, most previous neuroimaging studies exclusively examined a single aspect of cognition, such as a single neurocognitive ability 1,4,40 , or a general factor of cognitive ability 3,36,41 . This study is among the first to shed light on how brain connectivity contributes to the general factor of neurocognition, several specific factors, as well as a number of individual neurocognitive abilities.…”

“…We also produced resting state connectomes for 5,937 youth who met stringent neuroimaging quality control standards; these connectomes capture tens of thousands of functional connections between hundreds of brain regions. We next applied a multivariate approach predictive modeling approach, brain basis set (BBS) 35,5,36 , to build neuromarkers for neurocognitive abilities from whole-brain functional connectivity patterns. To guard against identifying spurious relationships and to provide evidence of generalizability, we coupled BBS with leave-one-site-out cross-validation, in which we construct neuromarkers in all sites except one, test the marker at the held-out site, and repeat until each site is held out.…”

“…Constructing a Brain Basis Set (BBS) BBS is a validated multivariate method that uses principal components dimensionality reduction to produce a basis set of components that are then associated with phenotypes 4,35 . We select the top 250 components for our basis set based on previous work showing that 50-100 components per 1000 subjects captures most meaningful variance without overfitting 35,36 .…”

“…To maintain strict separation between training and test datasets, regression coefficients for the covariates learned from the training sample are applied to the test sample, and the variance they explain is subtracted away. This procedure, described in detail in our previous publication 36 , yields an estimate of the contribution of brain components alone in predicting test subject P factor scores, excluding the contribution of the nuisance covariates. Note that by employing leave-one-site-out, members of twinships and sibships are never present in both training and test samples.…”

Validating Dynamicity in Resting State fMRI with Activation-Informed Temporal Segmentation

“…We performed correlation analysis between subject-specific dFC feature vectors, averaged across the four resting state sessions, and several neuro-relevant phenotypes. Specifically, we consider ten cognitive metrics: a general factor of intelligence (G; generated from a bifactor model as described in Sripada et al, 2020 ), processing speed (generated from factor modeling of three NIH Toolbox tasks as described in (Sripada et al, 2019) ), the five facets of personality given by the Revised NEO Personality Inventory (openness to experience, conscientiousness, extraversion, agreeableness, and neuroticism), and the three dimensions of psychopathology given by the Adult Self Report Scale (Internalizing, Attention Problems, Externalizing). We report the ten strongest phenotype correlations in Table 3, along with corresponding uncorrected and false discovery rate (FDR) corrected p-values (Benjamini & Hochberg, 1995) .…”