Integrating multiple MRI modalities improves cognitive ability prediction

Predicting cognitive abilities from brain imaging has long been a central goal in cognitive neuroscience. While machine learning has modestly improved predictions using brain MRI data, most studies rely on a single MRI modality. Narun Pat and colleagues integrated multiple MRI modalities through a technique called stacking.

The method combines structural MRI (e.g., cortical thickness), resting-state and task-based functional connectivity, and task-evoked blood-oxygen-level-dependent (BOLD) contrasts to build a more robust neural marker of cognitive function. The authors analyzed data from 2,131 participants aged 22 to 100 from three large-scale MRI datasets in the US and New Zealand. Across the three datasets, stacking consistently and significantly improved predictions of cognitive test scores collected outside the scanner. To assess whether stacking could capture stable cognitive traits, the authors applied the method to the Dunedin Multidisciplinary Health and Development Study. Using brain imaging at age 45, the model predicted childhood cognitive scores (ages 7, 9, and 11) with a .52 Pearson's correlation-indicating a substantial degree of predictive accuracy. Stacking also tackled a major challenge in MRI-based models: test-retest reliability-the stability of individual rankings over time. The improved consistency suggests that stacking enables MRI data to more reliably capture enduring individual differences in cognitive ability than models using a single MRI modality.

Finally, the researchers assessed the generalizability of stacking by training on one dataset and testing on a separate, independent dataset. Due to differences in task protocols, the authors were unable to include several key MRI modalities-most notably, task-evoked BOLD contrasts. Still, the model achieved above-chance predictive performance, with a .25 Pearson's correlation. Although this was lower than the within-dataset performance, the correlation nonetheless demonstrated a meaningful degree of cross-sample applicability. According to the authors, the study sets a valuable benchmark for how stacking can strengthen the use of brain MRI as a reliable and robust neural marker of cognitive function.