New AI Method OSCAR Reveals Hidden Brain Network Reorganization in Tasks and Psychosis

Neuroscientists have introduced a new computational tool that can spot brain regions whose wiring patterns quietly shift when people perform a task or when illness takes hold, even when those shifts are invisible to standard analytical approaches. The method, called OSCAR — short for one-class SVM-based connectome anomaly recognition — treats the question of how brain organization changes as an anomaly detection problem. It first learns what a region’s connectivity to the rest of the brain typically looks like in a reference state, such as quiet rest in healthy people, and then flags any region whose multivariate connectivity profile deviates significantly in a target condition, whether that condition is a demanding cognitive task or a clinical disorder.

The study, which validated OSCAR across three distinct cognitive tasks and a patient cohort with early psychosis, addresses a long-standing gap in neuroimaging analysis. Activation-based analyses, the traditional workhorse of task fMRI, identify where local blood-oxygen signals rise or fall during a task, but they say nothing about how those local changes alter a region’s embedding within large-scale networks. Conventional univariate functional connectivity analyses, in turn, capture changes in individual connections one at a time, missing multivariate differences that only become visible when a region’s entire connectivity pattern is considered together. Existing multivariate methods either compare conditions directly rather than against a learned reference distribution, or operate at a voxel level, limiting their generalizability. OSCAR fills this middle ground by using a sophisticated multivariate density estimation technique — a one-class support vector machine — applied to region-to-whole-brain connectivity patterns, and by delivering a p value for each brain region.

To test the approach, the researchers applied OSCAR to data from the Amsterdam Open MRI Collection, which includes resting-state scans and a Gender-Stroop task, and to two task datasets from the MeMoSLAP consortium: an object-location memory task and a novel-word-learning task. They also analyzed resting-state data from patients with early psychosis drawn from the Human Connectome Project for Early Psychosis. Throughout, OSCAR was benchmarked against a standard multivariate analysis of variance based on permutation testing, known as perMANOVA, which tests whether the mean connectivity pattern — the centroid — differs between two groups. The conceptual distinction matters: perMANOVA is sensitive to the dispersion of data points around group means, but it cannot detect deviations in the overall distribution of the data, whereas OSCAR is built precisely to catch such deviations from a learned reference.

The Gender-Stroop results were striking. Out of 232 brain regions, OSCAR identified 110 regions whose whole-brain connectivity differed significantly between rest and the task, while perMANOVA found 56, with 32 regions shared between the two methods. The three most significantly different regions according to OSCAR were the right-hemisphere precuneus, the right anterior globus pallidus, and part of the left-hemisphere somatomotor A network corresponding to the postcentral gyrus. Crucially, only OSCAR detected connectivity differences in the insula and the anterior cingulate cortex, both well established as involved in conflict monitoring during the Stroop task, as well as in the caudate, parts of the thalamus, and the inferior frontal gyrus. Several other regions known from the Stroop literature, such as the precuneus, posterior cingulate gyrus, precentral gyrus, and lateral prefrontal cortex, were identified by both methods.

Spatial correspondence between the two methods was on average low, with nearest-neighbor distances ranging from 0 to 50 millimeters and a mean of 13.5 millimeters. To contextualize the findings, the team calculated the Euclidean distance between each identified region’s centroid and the nearest activation peak from the corresponding task-activation contrasts. For the negative Gender-Stroop contrast, correct responses relative to incorrect ones, the two methods performed similarly, with five of 56 regions each located more than 14 millimeters from the nearest activation peak. But for the positive contrast, incorrect relative to correct, OSCAR-identified regions were significantly closer to the activation peaks than perMANOVA regions, a difference confirmed by a Mann-Whitney U test with a p value of 0.00349. Only 2 of 56 OSCAR regions, or 3.6 percent, sat more than 14 millimeters from the nearest peak, compared with 11 of 56 perMANOVA regions, or 19.6 percent.

The object-location memory task told a similar story. OSCAR identified 48 regions with deviating connectivity compared with rest, perMANOVA found 37, and the methods shared 18. The most pronounced state-dependent changes according to OSCAR appeared in the right-hemisphere visual peripheral network: the extrastriate inferior region corresponding to the lingual gyrus, the striate calcarine region corresponding to the intracalcarine cortex, and the occipital pole. Both methods captured parts of the ventral visual stream necessary for object identification, as well as the precuneus, retrosplenial cortex, middle frontal gyrus, insula, precentral gyrus, and lateral occipital complex, all tied to visual spatial learning and object recognition. Only OSCAR, however, detected connectivity changes in the bilateral pallidum and left thalamus, regions previously implicated in object-location memory, along with the frontal orbital cortex, which may reflect feedback-based learning and decision-making, and parts of the left inferior and middle temporo-occipital cortices involved in object processing. Notably, the OSCAR-identified regions largely overlapped with a general linear model analysis from a recent publication on the same data.

The novel-word-learning task revealed an instructive discrepancy. OSCAR identified 18 regions with deviating connectivity, perMANOVA found 17, and only three regions were shared. Both methods detected the left superior temporal gyrus, a core region for phonological processing, speech perception, and lexical access, and both picked up changes in extra-striate superior regions that may reflect the visual component of the stimulus. OSCAR additionally identified the medial prefrontal cortex, which supports integrating new information into existing schemas as required when linking pseudowords to familiar objects, and the bilateral fusiform gyrus, demonstrated to be involved in lexical learning, along with the globus pallidus and orbitofrontal cortex. Yet OSCAR did not detect the inferior frontal gyrus, left supplementary motor area, or insula, all of which reliably show strong activation effects in previous studies. An additional parcel-wise GLM analysis on the same dataset confirmed significant task-evoked activity in 20 regions, including those three. The authors suggest that task-evoked activity in these regions may not translate into widespread connectivity reconfiguration; changes may be confined to a few specific connections, remaining undetected when focusing on region-to-whole-brain connectivity profiles, or may be too transient to produce sustained changes when connectivity is aggregated across time.

Applied to the clinical cohort, OSCAR identified 10 regions with deviating resting-state connectivity in patients with early psychosis compared with healthy controls, while perMANOVA found only three, all of which OSCAR also identified. The most significant regions for OSCAR were the right ventral anterior thalamus, the left dorsal anterior thalamus, and the right posterior globus pallidus. Beyond the thalamic regions shared with perMANOVA, OSCAR detected differences in additional parts of the thalamus, the nucleus accumbens, the globus pallidus, and the ventromedial prefrontal cortex, all known to be affected in psychosis. The thalamus has long been associated with the condition and is believed to play a key role in its pathophysiology, affecting sensory, cognitive, and sleep processes. The authors note that across all three tasks, particularly the novel-word-learning task, several identified regions from both methods were situated substantially farther than 10 millimeters from activation peaks, supporting the premise that connectivity reconfiguration and task-evoked activity are only partially overlapping phenomena.

The researchers are careful about the method’s boundaries. Regions with strong task activation but little large-scale connectivity reorganization, such as the inferior frontal gyrus in the word-learning task, can escape OSCAR’s detection, and the authors emphasize that OSCAR is not an alternative to traditional activation analyses but a complementary lens. Other limitations include multi-site data acquisition, which introduces potential scanner-related variability that standardized preprocessing with fMRIPrep reduces but cannot eliminate; the choice of the Schaefer 200-region parcellation, which can influence which effects are detectable; and the fact that OSCAR identifies which regions deviate but not with which specific networks those alterations occur. The resting-state connectomes used to train the model were derived from scans of 6 to 8 minutes, a common length, though previous research indicates reliability improves substantially with 9 to 13 minutes or longer. The tasks analyzed were also designed to maximize group-level effects rather than individual differences. Even so, the authors anticipate that OSCAR will serve researchers aiming to map condition-specific connectivity changes, generate hypotheses about compensatory or maladaptive network reconfiguration, and provide region-level readouts comparable across tasks, datasets, and cohorts — including studies of causal perturbation approaches such as transcranial magnetic stimulation, deep brain stimulation, and transcranial direct current stimulation, where localized interventions shift the network embedding of both targeted and downstream regions.

Subject of Research: OSCAR method for detecting multivariate functional connectivity deviations in task fMRI and early psychosis

Article Title: Multivariate connectivity deviations reveal task-evoked and clinical network reorganization

Article References: Kröll, J.-P., Abdelmotaleb, M., Kocataş, H., Müller, V., Paas, L., Meinzer, M., Flöel, A., Eickhoff, S., & Patil, K. R. (2026). Multivariate connectivity deviations reveal task-evoked and clinical network reorganization. iScience, 29(11), Article 117959. https://doi.org/10.1016/j.isci.2026.117959

Image Credits: AI Generated

DOI: 10.1016/j.isci.2026.117959

Keywords: neuroscience, functional connectivity, OSCAR, one-class SVM, fMRI, brain networks, early psychosis, Stroop task, connectome, perMANOVA, anomaly detection, thalamus