Brain Network Fingerprints Under High Working Memory Load Distinguish Multiple Sclerosis Patients

Multiple sclerosis is best known for its effects on movement and sensation, but for more than half of those living with the disease, cognition takes a quiet hit as well. Working memory—the brain’s scratchpad for holding and manipulating information in the moment—is among the most commonly affected functions, and it can erode early in the disease course.

Researchers at Iran University of Science and Technology recruited twenty-two right-handed participants: eight patients with relapsing–remitting multiple sclerosis, diagnosed by a neurologist using the McDonald criteria, and fourteen healthy controls. The patient group, seven women and one man, had a mean age of about 35 years and had been diagnosed for an average of roughly 32 months. While lying in a 3-Tesla MRI scanner, each participant performed a visual Persian version of the n-back task, a classic working memory challenge in which letters flash on a screen and the participant must indicate when the current letter matches the one seen one, two, or three steps earlier.

The task had four levels of difficulty, though the simplest, a 0-back control condition in which participants simply responded whenever a designated target letter appeared, was set aside and not used in the analysis. The remaining three conditions formed the core of the experiment: in the 1-back task, participants pressed a key when a letter repeated after a single intervening letter; in the 2-back and 3-back versions, the match had to be recognized after two or three different letters had intervened. Each testing session contained four blocks of sixty seconds, with each difficulty level appearing once in random order across three sessions, giving the researchers ninety brain volumes for each task condition per participant.

From these scans, the team constructed a functional connectivity network for every individual under every task condition. The brain was parceled into 116 regions using the automated anatomical labeling atlas, and the average activity within each region was extracted and standardized. Pearson correlation coefficients between all pairs of regions served as the connections, or edges, of the network, yielding a 116-by-116 matrix for each person and each cognitive load. Because such matrices are fully connected and dense with redundant information, the researchers applied a proportional thresholding procedure, preserving only the strongest connections across a wide range from 5 to 80 percent in steps of 5 percent, and converting the surviving links into a binary, unweighted, undirected graph.

On these networks the team computed six local graph-theoretic measures, each producing a value for every one of the 116 brain regions: degree, which counts how many connections a node has; clustering coefficient, which captures how densely a node’s neighbors are interconnected; betweenness centrality, which reflects how often a node lies on the shortest paths through the network; eigenvector centrality, which rewards connections to other well-connected nodes; page-rank centrality, a related measure derived from a random-walk process; and k-coreness centrality, which identifies nodes belonging to densely interconnected cores. All measures were calculated with the Brain Connectivity Toolbox.

Classification then proceeded with careful attention to the small sample. Three algorithms were compared: support vector machines with linear, polynomial, and radial basis function kernels, k-nearest neighbors with k set to five, and decision trees. Feature selection relied on the non-parametric Wilcoxon rank-sum test, which the authors report outperformed mutual information and Fisher score approaches in their data. Crucially, the entire pipeline was embedded within a leave-one-subject-out cross-validation scheme: one participant was held out, features were ranked and selected using only the remaining subjects, the classifier was trained on the reduced feature set, and the model was then evaluated on the untouched individual. This procedure was repeated until every participant had served once as the test case, ensuring that no information from the test subject leaked into feature selection or model training.

The headline result came from the highest cognitive load. Using degree, k-coreness centrality, and eigenvector centrality extracted from the 3-back networks, a support vector machine classified patients and controls with 95.5 percent accuracy and an F1-score above 0.9 within the studied cohort. A permutation test with one thousand label shuffles confirmed that this accuracy significantly exceeded chance at p below 0.001. The authors note that the standard deviations reported alongside their accuracy figures reflect variability in subject-level prediction outcomes across the leave-one-subject-out evaluation, not variability across repeated runs, an important distinction for interpreting the precision of the estimate.

The pattern across difficulty levels was not strictly monotonic. The 3-back network proved the most informative overall, achieving accuracy above 80 percent and F1-scores above 0.7 across the local features, and it reached sensitivity and specificity above 80 percent for nearly all of them. Yet the 1-back condition showed comparable discriminative power for certain features, most notably betweenness centrality, which achieved an optimal F1-score of 0.94 in the 1-back network when half of the strongest connections were retained. The 2-back condition, by contrast, performed poorly and yielded few discriminative features. The authors suggest this non-monotonic pattern may reflect differences in experimental paradigms or inter-individual variability in how cognitive resources are recruited, and they point to prior work in healthy individuals showing that 2-back and 3-back activation maps need not differ even when their functional connectivity patterns diverge.

An intriguing methodological detail concerns thresholding. Except for k-coreness centrality, the best 3-back results emerged at low density thresholds, hinting that under high cognitive load even a small fraction of the strongest connections carries the discriminative information separating the groups. Visual inspection of the group-averaged connectivity matrices reinforced this picture: as load increased from 1-back to 3-back, healthy controls displayed progressively more homogeneous and structured connectivity, while patients showed more heterogeneous patterns with pronounced variability and localized increases in connectivity strength in specific regions under higher load.

The regions driving classification were also of interest. Among those distinguishing the groups were the left amygdala, left postcentral gyrus, left superior temporal region, parts of the cerebellum, the left middle orbital frontal region, right Heschl’s gyrus, and the right cuneus. Regional analyses flagged cerebellar areas, corresponding to regions 91 through 108 in the atlas, as showing comparatively lower uncorrected p-values, with the left cerebellum appearing across both the 1-back and 2-back networks for all graph measures, and the left middle orbital frontal region standing out in the 3-back network. The authors are careful to frame these regional findings as relative importance within a classification framework rather than definitive region-wise statistical confirmation, and because no correction for multiple comparisons was applied, they describe them as exploratory.

This caution is consistent with the study’s broader limitations, which the authors discuss candidly. The sample was small and imbalanced, with eight patients and fourteen controls, which limits stability and generalizability and heightens susceptibility to overfitting in machine-learning analyses. The high dimensionality of the feature space relative to the number of subjects means the feature selection process may be unstable, and the specific features identified could vary in other samples. Cross-validation and F1-score reporting partially mitigate the imbalance, and embedding feature selection within each cross-validation fold reduces, though does not eliminate, the risk of false positives. The authors also note that binarizing the networks discards information about connection weights, potentially limiting sensitivity to subtle connectivity differences, and that anatomical atlases like the one used may not align perfectly with functional boundaries.

Within these constraints, the findings sit sensibly in a wider literature. More than half of MS patients experience cognitive impairment over the course of the disease, with deficits in attention, processing speed, episodic memory, and executive functions that can appear early and progress over time. Working memory dysfunction is particularly prominent, and pathological work points to white matter damage as a key contributor. Previous neuroimaging studies in MS have reported altered functional connectivity in both supratentorial regions and the cerebellum with associations to cognitive performance, and early-stage connectivity changes have been interpreted as compensatory mechanisms that may eventually give way as pathology advances. Earlier machine-learning efforts have also distinguished patients from controls, including a study using graph features from a different working memory task that achieved 81.25 percent sensitivity, and resting-state work reaching 85.7 percent accuracy. The present results, with their emphasis on task-dependent network reconfiguration, extend this line by showing that the conditions under which networks are measured matter: difficulty modulates how separable the groups are.

The authors position the framework as a technical blueprint for task-based network analysis and, cautiously, as a potential auxiliary tool rather than a diagnostic in its own right. They emphasize that the reported performance reflects discriminative capability within this dataset rather than definitive clinical applicability, and that validation in larger, balanced, independent cohorts with stronger statistical control will be essential. They also suggest future extensions, including weighted or directed graphs, effective connectivity approaches, and functional atlases, to capture richer network dynamics. For now, the study offers a measured demonstration that the topology of the working brain, examined through the mathematics of graphs, carries a readable signature of the changes that multiple sclerosis inscribes on cognition.

Subject of Research: Graph-theoretic analysis of task-based fMRI functional brain networks to classify working memory impairment in multiple sclerosis patients.

Article Title: Investigating functional brain networks in multiple sclerosis patients: a graph theory approach for evaluating working memory impairment

Article References: Yousefi, S., & Daliri, M. R. (2026). Investigating functional brain networks in multiple sclerosis patients: a graph theory approach for evaluating working memory impairment. Brain Informatics, 13(1), Article 37. https://doi.org/10.1186/s40708-026-00324-y

Image Credits: AI Generated

DOI: 10.1186/s40708-026-00324-y

Keywords: multiple sclerosis, working memory, functional connectivity, graph theory, fMRI, machine learning, n-back task, support vector machine, cerebellum, orbitofrontal cortex, brain networks, cognitive impairment