Understanding the neurophysiology of individual differences in fluid intelligence (Gf) is a key goal of intelligence research. Many existing brain imaging studies have linked Gf to the structure and function of frontal and parietal brain regions, however more recent research points at the importance of dynamic interactions between brain regions for explaining Gf. Here, we investigated correlations between connectivity patterns in the MEG data and Gf. Moreover, in concord with models of female connectome to be more distributed and frontally-oriented than male connectome, we searched for differences between females and males in connectome patterns predicting Gf. We analyzed MEG data of 89 young healthy people recorded during three consecutive resting state sessions. The data was a part of the openly available dataset from the Human Connectome Project.
The connectivity patterns were assessed with the method of coherence, which measures the synchronization of two time series for a given frequency band. We used coherence for 16 frequency bands (ranging from 2 to 60 Hz) to generate connectome graphs, comprising the 152 MEG channels as the graph nodes, and the coherence values between the node pairs as the graph edges. In order to reduce the complexity of the resulting graphs, principal component regression (PCR) was used, which is based on principal component analysis (PCA). PCR is a standard linear regression, which regresses the dependent variable on a set of principal components of the independent variables. This approach helps to cope with the multicollinearity problem and reduces the dimensionality of the parameter space. The fit of the connectome-based PCR model was tested by comparing the predicted and actual Gf values of participants, the latter values assessed with four cognitive tests (including a simplified variant of the Raven as well as spatial and memory tasks).
We trained PCR on 100 principal components of our connectome. The resulting model achieved Spearman correlation of rho = .52 (p < .0001) between predicted and actual Gf. Then we analyzed the role of certain types of connections in predicting Gf, by forcing the model to use only such connections. The model restricted to fronto-frontal edges achieved significantly higher (p = .041) correlation for females (rho = .42) than for males (rho = .13). In contrast, its restriction to parieto-parietal connections gave significantly stronger (p = .020) prediction for males (rho = .40) than females (rho = -.01). Moreover, for males most predictive were the high gamma frequencies, while for females the alpha frequencies mattered (p = .0004). Finally, for males most predictive were the short to average length connections, while for females the average to long connections mattered (p = .045). Overall, these data suggest that the patterns of dynamic interactions between brain regions may substantially contribute to intelligence. They implicate that connectome modeling may help in understanding Gf. The results also support theoretical models assuming that male and female brains are differently organized functionally (yet scoring comparably on the Gf tests), with males relying on relatively local, gamma-based processing in the parietal cortex, while females using more distributed, alpha-dependent networks encompassing also the frontal cortex.
