Individual Differences in Brain Activity Related to Human Intelligence

Post by Elisa Guma

What's the science?

One of the characteristics of humans is the ability to perform cognitively challenging tasks in varied situations. This adaptability is typically used as a measure of general intelligence. However, there is still debate over whether general intelligence should be studied as one general ability, or as a mixture of many distinct psychological processes. For more than three decades, neuroscientists have tried to tie this question to the brain's functional architecture. It has been proposed that to meet task demands, specific brain regions become transiently active and that these brain regions are part of heavily overlapping dynamic functional networks. This week in Nature Communications, Soreq and colleagues sought to investigate how different cognitive tasks and performance on those cognitive tasks relate to dynamic brain activity.

How did they do it?

The authors relied on an established intelligence test which included twelve cognitive tasks. The data they used had online test responses from over 60,000 participants. They also obtained functional magnetic resonance imaging (fMRI) data from healthy young adults as they performed the exact same test. These data enabled them to identify the brain regions found to be consistently active for all 12 tasks ('domain-general regions’). To identify these brain regions, the authors relied on a classic yet often underused conjunction analysis method. These regions are considered task agnostic since they are active for all tasks and contain very little or no task-specific information. Areas outside the 'domain general' mask include the regions in which task-specific information is stored. The authors then compared how similar tasks were using either a cortical mask (a selection of particular brain regions) containing general and specific regions or the domain-general mask. By applying the same approach to the online behavioural data, they could compare brain and psychometric (the behavioural test data) task similarities.

Next, they wanted to see how distinct tasks were. In other words, they asked: Can we tell that a person is doing a particular task only from their neural information? The authors examined the relationships between neuronal information based on brain activity (the relative local metabolic requirements of each brain region) and dynamic functional connectivity (how different brain regions interact in specific time windows). To quantify brain activity, the authors assessed all voxels contained within the ‘domain general’ mask, and to quantify dynamic functional connectivity the authors parcelled the 'domain general' mask into distinct regions For the cortical mask, the authors used an unbiased parcellation set based on resting-state fMRI data. Using these mined features (i.e. brain activity and connectivity), the authors performed a machine-learning-based classification analysis with the 12 cognitive tasks. 

In response to the reviewers' comments, the authors were asked to confirm that their ability to correctly classify between the different tasks was not influenced by the motor or visual differences between the tasks but rather was a function of the different cognitive abilities required of each task. The authors accomplished this by aggregating the various tasks into three meta-labels (i.e. cognitive, visual, and motor) and conducting another classification analysis, this time comparing the actual task meta-label assignment to permuted ones. Furthermore, following the reviewers' criticism, the authors examined whether neuronal information can predict how well an individual will perform on the intelligence test. Specifically, they calculated each individual's task performance principal component analysis (PCA) and overall classification accuracy from the brain (i.e., how well we can correctly predict from your brain what tasks you are performing).

What did they find?

The authors first replicated the existence of a task-active (i.e., domain-general) network that showed activity across all 12 tasks. This network consisted of areas including the parietal, visual, and motor areas. The authors then demonstrated that cortical and domain-general similarities are highly correlated with psychometric similarity. This suggests that they all rely on the same mixtures of the underlying physiological and neurobiological systems that actually perform the tasks. In addition, the slight difference between the domain-general mask and the cortical mask suggests that the domain-general mask is not entirely homogeneous.

Classification of tasks based on either brain activity or connectivity from domain-general regions yielded impressive model performance with 37% and 43.1% accuracies which are 5 times better than chance (8%) for a balanced 12-tasks classification problem. The same learning algorithm, however, using cortical information, produced dramatic improvement, achieving 49% and 69% accuracy. Connectivity and activity complemented each other - the model trained on both measurements outperformed independent models (with 74%). 

Then the authors showed that the actual assignment of tasks into different cognitive and motor meta-labels significantly outperformed permuted assignments that maintained the same imbalance. It is possible that we decode these tasks by a mixture of networks that carry out both cognitive and motor functions. Additionally, they demonstrated a link between the accuracy of the models at the individual level and the performance of the tasks being classified. Tasks from higher-performing individuals were classified more accurately. Finally, they demonstrated that they could accurately predict individual performance and that this ability was mainly due to connectivity between different prominent resting state networks, such as default mode and dorsal attention.

What's the impact?

This study used a combination of machine learning techniques applied to fMRI and psychometric data to show that performance on cognitive tasks could accurately predict dynamic brain connectivity networks and that more similar cognitive tasks activated more overlapping networks. Further, better performance on cognitive tasks was related to the ability to activate more specific dynamic networks and to flexibly switch between them. The data presented here provide interesting information about how human intelligence and brain activity may be related. Finally, this framework may be applied to clinical data in hopes of identifying markers for quantifying disease pathologies.

Soreq E al. Neuroimaging evidence for a network sampling theory of individual differences in human intelligence test performance. Nature Communications (2021). The original scientific publication here.