Review: Brain models explain 10–40% of cognitive variation

A review of brain-based predictive models finds that imaging data can explain roughly 10% to 40% of the variation in cognitive ability across youth, adulthood and older age. The analysis, published in Communications Psychology, concludes that predictive performance is broadly consistent across the lifespan, but its magnitude depends strongly on the mental ability being targeted, the brain features measured, the way data are combined and the algorithm selected. The review focuses on large, predominantly public datasets to reduce the inflated estimates often associated with small-sample research, although it cautions that apparent accuracy can fall substantially when models are tested in new cohorts.

The central finding is that brain-based models perform best when they predict broad, composite measures of general cognition rather than narrow abilities such as working memory, processing speed or cognitive flexibility. Among young people aged 8 to 21, models explained 11% to 39% of variance in general cognitive ability. In adults aged 22 to 37, estimates ranged from less than 1% to 36%, while older cohorts aged 55 to 89 yielded cross-sectional values of approximately 5% to 40%. The strongest within-age results clustered around 36% to 40%, suggesting that high predictive accuracy is possible at different life stages when models are trained and tested within the same developmental period. This is a review of published prediction studies, not a new experiment or a single population-wide test of one model.

Longitudinal results show that measured brain characteristics can carry information about later cognition. In youth, structural and functional models explained up to 17.5% of variance in general cognitive performance two years later. Studies of older adults found models accounting for about 11% to 36% of variance at one year, 19% to 28% at two years and 20% to 25% at three years. These intervals do not establish a consistent decline in explained variance as follow-up lengthened. The results instead describe associations between measured brain features and subsequent cognitive scores; they do not demonstrate that the brain features cause later differences or that one scan can reliably forecast an individual’s future cognitive decline.

The cognitive target is a major source of variation. General intelligence was more predictable than specific functions in both young and adult samples. In youth, one set of models explained 39% of variance in general cognition, compared with 38% for crystallized ability and 26% for fluid intelligence. In adults, crystallized abilities were generally predicted more accurately than fluid abilities, with reading-related outcomes reaching 16% to 22% explained variance. Working memory and processing speed were often minimally predictable, at around 2% to 4%. An exception appeared in older adults, where processing speed reached 17% to 19% explained variance, compared with 7% to 8% for semantic memory or executive function. The review therefore identifies predictability as dependent on both the cognitive domain and the population in which it is measured.

Imaging modality produced another clear pattern. Functional connectivity generally outperformed diffusion measures, while anatomical features offered moderate predictive value. Resting-state functional connectivity alone typically explained 4% to 10% of variance in general cognition, although longer scans raised the estimate to about 20%. Task-based connectivity performed better: working-memory connectivity predicted roughly 28% of general cognitive variance, compared with 8% for resting-state connectivity, and one study obtained 36% from a combination of working-memory, language and relational-processing tasks. In youth, task-based connectivity also outperformed resting-state approaches, but the reverse pattern emerged in older adults. There, structural features generally predicted cognition better than resting-state connectivity, and models based on volumes of the hippocampus, entorhinal cortex and lateral ventricles explained 20% to 35% of variance at baseline and up to three years ahead.

Combining data did not always guarantee an improvement. Some multimodal models performed as well as their strongest functional component alone, and adding grey-matter volume, white-matter integrity and resting-state fluctuation to a structural model did not beat the structural model by itself. Yet one model combining task activation, resting-state connectivity and several structural MRI measures explained 32% of variance in general cognition, compared with 7% for a single modality. The review favors hierarchical stacking, in which separate models are built for each modality and their outputs are combined in a final model, because this can improve performance, reduce dimensionality and clarify each modality’s contribution. Task selection also matters: language-derived connectivity was the strongest predictor of oral reading, whereas working-memory connectivity was better for picture vocabulary. Simply including more tasks did not guarantee higher accuracy, making strategic selection important.

Algorithm choice affected performance, but the most complex method was not consistently necessary. Deep-learning systems sometimes outperformed traditional approaches when applied to structural connectome data, yet elastic net and ridge regression often achieved comparable accuracy with lower computational costs and greater interpretability. Elastic net was particularly consistent across populations and outcomes. In older cohorts, random forests performed well for processing speed and some longitudinal predictions, while kernel ridge regression was superior for quantitative reasoning in one comparison. These differences favor matching the algorithm to the target, population and study design rather than assuming that a more elaborate artificial-intelligence system is inherently superior. Cross-validation remains standard for estimating accuracy, but external testing on an independent dataset is essential for demonstrating that performance transfers beyond the sample in which a model was developed.

Confounding variables pose a major challenge to interpretation. Sex, age, education and socioeconomic circumstances are associated with both brain characteristics and cognitive performance, so a model may exploit information that is not specific to brain-cognition relationships. The review describes large differences in the optimal task used to predict fluid intelligence for males and females, underscoring the need to test performance within demographic groups. More fundamentally, adding baseline cognitive scores and grey-matter volume can raise explained variance to 70% to 81%, while demographic information alone explains 31% to 38% of cognitive variation in some older-adult datasets and may add little beyond what simpler clinical assessments provide. Brain models can remain valuable for identifying neural markers, but their added value must be demonstrated beyond demographic and cognitive information alone.

Several changes could make future models more useful. Repeated cognitive testing across days or weeks could capture the natural fluctuation of executive functions, which vary with fatigue, stress and environmental demands and are often represented in datasets by a single assessment. Standardized, domain-based assessments such as the Research Domain Criteria framework could provide measures that are consistent across diagnostic and developmental boundaries. Models should also use age-appropriate brain parcellations, because adult-derived divisions of the cortex may fail in developing brains. Clinical deployment remains premature: although imaging models could eventually support estimates of post-stroke cognitive recovery or trajectories of decline, the field must first show that they add meaningful and reproducible information beyond less expensive measures in independent populations.

The review itself did not analyze new data, and its comparisons are constrained by differences in datasets, cognitive tasks, inclusion criteria, quality control and analysis pipelines. Many studies used overlapping cohorts, including the Human Connectome Project, the Adolescent Brain Cognitive Development Study and UK Biobank, and most relied on internal cross-validation rather than external validation. The few cross-dataset studies produced lower effect sizes, meaning the headline range may represent upper-bound estimates. Large resources also disproportionately represent Western, educated, industrialized, rich and democratic populations, narrowing the conclusions that can be drawn about global applicability. Finally, successful prediction is not the same as explanation: a model can forecast a cognitive score without revealing which neural processes cause it. The review therefore positions brain-based prediction as a promising measurement tool whose scientific and clinical value will depend on rigorous testing, transparent confounding control, harmonized measures and longitudinal evidence.

Subject of Research: Predictive accuracy of brain-based models for cognition across youth, adulthood and older age

Article Title: Effect sizes for brain-based predictive models of cognition across the lifespan

Article References: Christensen, E., Wiersch, L., Wang, M., McPherson, B., Shearer, H., Noble, S., Lam, K., Poline, J. B., & Dhamala, E. (2026). Effect sizes for brain-based predictive models of cognition across the lifespan. Communications Psychology, 4(1), Article 143. https://doi.org/10.1038/s44271-026-00528-2

Image Credits: AI Generated

DOI: 10.1038/s44271-026-00528-2

Keywords: brain imaging, cognitive prediction, neuroimaging, machine learning, functional connectivity, crystallized ability, fluid intelligence, aging, predictive modeling, brain–cognition relationship, Effect, sizes