Investigating cognitive neuroscience theories of human intelligence: A connectome-based predictive modeling approach

Abstract

Central to modern neuroscientific theories of human intelligence is the notion that general intelligence depends on a primary brain region or network, engaging spatially localized (rather than global) neural representations. Recent findings in network neuroscience, however, challenge this assumption, providing evidence that general intelligence may depend on system-wide network mechanisms, suggesting that local representations are necessary but not sufficient to account for the neural architecture of human intelligence. Despite the importance of this key theoretical distinction, prior research has not systematically investigated the role of local versus global neural representations in predicting general intelligence. We conducted a large-scale connectome-based predictive modeling study (N = 297), administering resting-state fMRI and a comprehensive cognitive battery to evaluate the efficacy of modern neuroscientific theories of human intelligence, including spatially localized theories (Lateral Prefrontal Cortex Theory, Parieto-Frontal Integration Theory, and Multiple Demand Theory) and recent global accounts (Process Overlap Theory and Network Neuroscience Theory). The results of our study demonstrate that general intelligence can be predicted by local functional connectivity profiles but is most robustly explained by global profiles of whole-brain connectivity. Our findings further suggest that the improved efficacy of global theories is not reducible to a greater strength or number of connections, but instead results from considering both strong and weak connections that provide the basis for intelligence (as predicted by the Network Neuroscience Theory). Our results highlight the importance of considering local neural representations in the context of a global information-processing architecture, suggesting future directions for theory-driven research on system-wide network mechanisms underlying general intelligence.

Keywords: cognitive neuroscience; connectome; fMRI; individual differences; intelligence; network neuroscience theory.

FIGURE 1

Functional regions included in the Lateral PFC connectivity map, colored by intrinsic connectivity network. Purple, ventral attention; orange, frontoparietal; red, default mode

FIGURE 2

Functional regions included in the Parieto‐Frontal Integration Theory (P‐FIT) connectivity map, colored by intrinsic connectivity network. Green, dorsal attention; purple, ventral attention; orange, frontoparietal; red, default mode

FIGURE 3

Functional regions included in the Multiple Demand (MD) Theory connectivity map, colored by intrinsic connectivity network. Green, dorsal attention; purple, ventral attention; orange, frontoparietal; red, default mode

FIGURE 4

N‐fold cross‐validated performance for connectome‐based predictive modeling (CPM) predictions of g using functional edges from Lateral Prefrontal Cortex (PFC) Theory. (a) CPM results predict g from PFC edges at r = .25 (400 ROI, p < .05). (b) Permutation test significance (p ₁₀₀₀ = .008) suggests a significant association between edges used for prediction and psychometric g at 400 ROI and p < .05.

FIGURE 5

N‐fold cross‐validated performance for connectome‐based predictive modeling (CPM) predictions of g using functional edges from frontoparietal cortex. (a) CPM results predict g from Parieto‐Frontal Integration Theory edges at r = .25 (200 ROI, p < .01). (b) Permutation test significance (p ₁₀₀₀ = .02) suggests a significant association between edges used for prediction and psychometric g at 200 ROI and p < .01.

FIGURE 6

N‐fold cross‐validated performance for connectome‐based predictive modeling (CPM) predictions of g using functional edges from the cortical extended Multiple Demand (MD) network. (a) Connectome‐based predictive modeling results predict g from extended MD network at r = .18 (200 ROI, p < .01). (b) Permutation test significance (p ₁000 = .008) suggests a reliable association between edges used for prediction and psychometric g at 200 ROI and p < .01.

FIGURE 7

N‐fold cross‐validated performance for connectome‐based predictive modeling (CPM) predictions of g using functional edges from the Process Overlap Theory (POT). (a) CPM results predict g from right‐tailed process overlap edges at r = .11 and p ₁₀₀₀ = .01 (200 ROI, p < .01). (b) The comparable connectome‐based predictive modeling analysis predicts intelligence from left‐tailed (i.e., weak) edges at r = .15 and p ₁₀₀₀ = .01 (200 ROI, p < .01), providing evidence that is inconsistent with the predictions of POT.

FIGURE 8

N‐fold cross‐validated performance for connectome‐based predictive modeling (CPM) predictions of g based on network neuroscience theory (NNT). (a) CPM results predict g based on NNT at r = .25. (b) Permutation test significance (p ₁₀₀₀ = .005) suggests a significant association between edges used for prediction and psychometric g at 200 ROI and p < .05.

FIGURE 9

Connectome‐based predictive modeling (CPM) validation tests demonstrate that larger feature spaces do not produce a more reliable prediction of g. (a) Larger input features spaces in N‐fold CPM do not result in more accurate predictions of g. (b) Relaxing the p‐value threshold to include additional (non‐random) significant edges reduces accuracy when predicting g. Resulting CPM predictions of g become nonsignificant at p = .034.