By Vijay Kumar MalesuReviewed by Benedette Cuffari, M.Sc.Dec 16 2024
A recent study explores the relationship between brain connectivity and intelligence, highlighting the value of interpretability in predictive modeling for deeper insights into human cognition.
Study: Choosing explanation over performance: Insights from machine. Image Credit: Christoph Burgstedt / Shutterstock.com
In a recent study published in PNAS Nexus, researchers provide new insights into human cognition by prioritizing interpretability in predictive modeling of intelligence from brain connectivity, rather than solely focusing on prediction performance.
Machine learning in neuroscience
Neuroscientific research on human cognition has evolved from focusing on single-variable explanatory studies to employing machine learning-based predictive modeling. This shift enables the analysis of relationships between behavior and multiple neurobiological variables to forecast behavior across diverse samples.
Intelligence, a key predictor of life outcomes such as health and academic achievement, has been extensively studied, with theories dividing it into fluid and crystallized components. Recent machine learning approaches have enhanced intelligence prediction using brain connectivity data. However, limited conceptual insights, reliance on specific intelligence measures, and methodological constraints highlight the need for further research to systematically explore predictive brain features.
About the study
The present study adhered to a rigorous methodology, with all analyses, sample sizes, and variables preregistered on the Open Science Framework. The primary analyses followed preregistered protocols, with additional post hoc analyses conducted to further explore brain connections most relevant for intelligence prediction.
Study participants were drawn from the Human Connectome Project (HCP) Young Adult Sample S1200, consisting of 1,200 individuals between 22 and 37 years of age. Informed consent was obtained in accordance with the Declaration of Helsinki and all procedures were approved by the Washington University Institutional Review Board.
After exclusions for missing data, cognitive impairment based on Mini-Mental State Examination (MMSE) scores of 26 and less, or excessive head motion, the final sample included 806 participants, 418 of whom were female and 733 right-handed. Measures of intelligence including general intelligence (gg), crystallized intelligence (gCgC), and fluid intelligence (gFgF) were estimated using bi-factor and exploratory factor analyses from cognitive test scores.
Functional magnetic resonance imaging (fMRI) data were collected during resting state and seven cognitive tasks to construct subject-specific functional connectivity (FC) matrices. Minimally pre-processed fMRI data underwent additional preprocessing steps, including nuisance regression, global signal correction, and removal of task-evoked activation, to improve connectivity estimates. Predictive modeling utilized feedforward neural networks, which incorporated five-fold cross-validation, hyperparameter optimization, and an out-of-sample deconfounding approach to control for covariates such as age, sex, and head motion.
Model interpretability was enhanced using layer-wise relevance propagation (LRP) to identify functional brain connections most critical for predictions. External replication was performed using two independent datasets from the Amsterdam Open MRI Collection (AOMIC), thereby ensuring clarity and generalizability of the results.
Statistical analyses included Pearson's correlation coefficients, error-based metrics, and nonparametric permutation tests to evaluate prediction performance and compare models across datasets.
Study results
FC and its role in predicting intelligence were investigated using data from the HCP. Herein, gg, gCgC, and gFgF components were estimated from 12 cognitive measures and showed significant intercorrelations.
Individual FCs were constructed from 100 cortical nodes during resting state and seven task states. Additionally, two latent FCs, one spanning resting and task states and another across task states only, were computed, which resulted in a total of ten cognitive states for analysis.
Prediction performance varied across intelligence components. To this end, gg predictions achieved the highest correlation between observed and predicted scores, followed by gCgC and gFgF. Task states also influenced prediction accuracy, with cognitively demanding tasks and latent FCs outperforming others.
Across brain networks, cognitive networks like the default mode, control, and attention networks provided the most predictive power, as they significantly outperformed networks like the somatomotor and limbic systems.
A systematic selection of functional brain connections demonstrated that intelligence prediction depends on the interaction between components, states, and networks. Models trained with all but one network showed minimal performance reductions, thus highlighting compensatory intelligence-relevant information distributed across brain networks.
The best predictions were obtained using 1,000 of the most relevant brain connections identified through stepwise LRP. These connections were widely distributed across cortical regions and varied between cognitive states.
Validation in the HCP lockbox sample confirmed the accuracy of the findings, with significant correlations in prediction performance for all intelligence components. External replication using data from the AOMIC demonstrated generalizability, though with lower effect sizes.
Key patterns of predictive performance across states and networks remained consistent. Furthermore, models based on the 1,000 most relevant connections significantly outperformed random selections.
Conclusions
Intelligence predictions were consistently better for gg and gCgC as compared to gFgF, thus highlighting systematic differences in their neural pathophysiology. These findings varied across cognitive states and networks, with cognitively demanding tasks and brain-wide connectivity outperforming resting-state measures.
Theory-driven models based on established intelligence frameworks provided meaningful insights but were outperformed by whole-brain models. Taken together, the current study identified about 1,000 highly predictive brain connections that form a distributed network spanning major functional systems.
Journal reference: Thiele, J. A., Faskowitz, J., Sporns, O., et al. (2024). Choosing explanation over performance: Insights from machine learning-based prediction of human intelligence from brain connectivity. PNAS Nexus. doi:10.1093/pnasnexus/pgae519