When a healthy human brain transitions from quiet rest to a demanding mental task, its large-scale communication networks flexibly reorganize to meet the challenge. A new neuroimaging study published in Nature Communications indicates that this adaptive shifting is dampened, or “flattened,” in people with psychiatric diagnoses. The findings also hint that these network shifts may line up more closely with a person’s particular mix of symptoms than with a formal diagnosis.
To study how the brain coordinates information, neuroscientists examine functional brain networks, which are groups of distinct regions across the outer cerebral cortex and deeper structures that work together. Using functional magnetic resonance imaging (fMRI), a scan that tracks blood flow changes tied to neural activity, scientists estimate functional connectivity by measuring how closely activity in different brain regions rises and falls in sync. Rather than operating as a fixed circuit board, the brain acts as a dynamic system that continuously adjusts these connections to respond to changing demands.
Many brain imaging studies rely on static resting-state scans that measure average connectivity while a person lies quietly in the scanner. However, a 2016 neuroimaging study of healthy young adults demonstrated that brain networks actively reorganize from this baseline during tasks, shifting differently for simple motor actions than for complex working memory challenges. Similarly, a 2021 study in the Journal of Neuroscience co-authored by the new study’s lead author, Carrisa Cocuzza, indicated that although task-related connectivity shifts are subtle, these individual-specific adjustments improve predictions of brain activity across two dozen cognitive conditions.
Connecting these network shifts to mental illness is complicated because traditional psychiatric diagnoses often group together patients with very different symptom combinations, while people with distinct diagnoses like depression and schizophrenia can share problems such as low motivation, fatigue, or difficulty planning. Even individuals without a clinical diagnosis can experience mild or temporary versions of these difficulties, prompting researchers to view mental health through broad dimensions of functioning rather than rigid categories.
Supporting this view, an earlier neuroimaging study of children, adolescents, and adults found that brain connectivity patterns were more consistent within broad symptom categories, such as inward-focused internalizing problems versus outward-focused externalizing behaviors, than between them across development.
Earlier studies have also shown that time-varying connectivity during rest separates patients with schizophrenia, bipolar disorder, or ADHD from healthy peers better than static connectivity measures. Building on this work, Carrisa V. Cocuzza of Rutgers University and Yale University, senior author Avram J. Holmes of Rutgers, and their colleagues investigated whether whole-brain network reconfigurations across resting and task scans could map individualized symptom profiles across health and psychiatric illness.
The researchers analyzed data from 219 adults aged 18 to 70 (average age 36.2 years; 123 women, 93 men, and 3 nonbinary individuals) from the Transdiagnostic Connectome Project, a public dataset previously collected at Yale University and McLean Hospital by several of the same authors. The sample included 134 participants across 11 primary diagnostic groups, including depression, anxiety, bipolar disorder, PTSD, schizophrenia or schizoaffective disorder, substance use disorder, ADHD, OCD, panic disorder, and eating disorders, alongside 85 participants with no psychiatric history. To prevent data leakage, where information from test participants slips into the training process, the team split the sample into a discovery set of 176 participants and a held-out validation set of 43 participants reserved strictly for final testing. Both the data and analysis code were posted publicly, and the authors framed their scan-comparison models as an exploratory proof-of-concept.
To profile participants’ functioning, the team analyzed 110 clinical, behavioral, and cognitive measures from interviews, self-report scales, and cognitive tests. Using a clustering algorithm that groups measures based on how similarly scores vary across people, the researchers organized the assessments into four dimensions: internalizing (such as depression, anxiety, and neuroticism), externalizing (such as irritability and extraversion), cognition (such as intelligence and cognitive test performance), and social or reward functioning (such as relationship experiences and novelty seeking).
Each label reflected only the single most common theme in its cluster, matching between 20% of the measures in the social/reward cluster and 40% in the internalizing cluster, and the authors caution against reading too much into the names. By testing whether each participant’s score on each dimension stood out relative to the rest of the cohort, the researchers assigned a binary pattern across the four domains, yielding 16 possible combinations termed “symptom fingerprints.”
Participants also completed six fMRI scans across three mental states: three resting-state runs, two runs of a Stroop cognitive control task requiring them to suppress an automatic reading response to name a word’s ink color, and one run of an Emotional Faces task involving fearful expressions. For each scan, the researchers measured activity correlations across 434 cortical, subcortical, and cerebellar regions organized into 21 functional networks. They then used a machine learning decomposition technique to extract underlying network building blocks and track how strongly the most efficient, well-segregated building block was expressed across the six scans.
Comparing network patterns across the six scans in the 176-person discovery set, the researchers found that the 108 participants with psychiatric diagnoses showed less variation from scan to scan than the 68 non-diagnosed participants. Using a similarity score ranging from -1 (completely opposite) to 1 (identical) across all 15 scan pairs, participants without a diagnosis averaged 0.91, whereas those with a diagnosis averaged 0.95, indicating statistically significant greater flattening when tested against 10,000 shuffled null models. Descriptive comparisons across diagnostic groups indicated that flattened dynamics occurred across conditions, appearing most pronounced in panic disorder, eating disorders, OCD, and ADHD, though each of these groups included only two to four people.
The 16 behavioral symptom fingerprints cut across traditional diagnostic labels. The profile featuring only internalizing symptoms consisted largely of people diagnosed with depression, whereas multi-dimensional profiles, such as those combining the internalizing and cognition dimensions or expressing all four dimensions at once, mixed participants across different diagnoses like PTSD as well as people without any psychiatric diagnosis.
Using a machine learning classifier tested on the 43 validation participants, brain network dynamics across the six scans predicted binary case-or-control status, primary diagnosis, and the 16 symptom fingerprints at rates above chance levels of roughly 50%, 8.3%, and 6.6%, respectively (the accuracy expected from guessing, estimated by shuffling the labels). Standard test statistics ranked binary case-or-control classification highest, but by the authors’ standardized measure of effect size, the values for classifying symptom fingerprints were more than five times larger than for case-or-control status and nearly twice as large as for primary diagnoses. However, the authors note that differing chance baselines and unquantified noise across the outcomes complicate direct comparisons.
To identify which brain systems drove fingerprint classification, the researchers digitally removed one functional network at a time before re-running the classifier on the validation group. Withholding systems involved in cognitive control, specifically part of the frontoparietal control network (regions in the frontal and parietal lobes that support goal-directed control) and the thalamus (a relay hub deep in the brain), as well as limbic, visual, and somatomotor networks, produced a statistically significant drop of up to 2% in classification accuracy. Removing a subsection of the default mode network, which typically supports inward focus and rest, slightly increased accuracy, though this change was not statistically significant; the authors suggest that default mode and frontoparietal systems may play complementary roles in shifting between easy and demanding mental states.
In an exploratory analysis comparing seven combinations of scans, including all six scans yielded the lowest fingerprint classification performance, suggesting that classification performance did not simply grow with more brain data, though the differences across the seven models were small and not formally tested against one another. Among these comparisons, the model combining resting-state and Stroop scans performed best, with regions in the frontoparietal network, the thalamus, and the striatum (a deep-brain structure involved in action and reward) acting as major communication hubs. The authors propose that how flexibly the brain transitions from rest to an effortful cognitive control state may be a key factor shaping transdiagnostic symptom profiles.
The finding that altered brain dynamics track symptom profiles aligns with an earlier neuroimaging study of healthy adults with psychotic-like experiences and an optical brain imaging study of young children with autism, both of which linked altered time spent in particular brain states to psychiatric symptoms: psychotic-like experiences in healthy adults, and greater symptom severity in children with autism.
Those earlier studies, however, examined subclinical psychosis during rest and moment-to-moment optical connectivity in children rather than fMRI shifts across multiple tasks. At the same time, the finding of flattened dynamics in psychiatric illness stands in tension with an earlier resting-state fMRI study of antidepressant users, which found that dynamic network reconfigurations varied with antidepressant use rather than with a depression or anxiety diagnosis alone. That difference may reflect how dynamics were measured, as the new study evaluated shifts between resting and task scans rather than fluctuations within a resting scan alone.
As with all research, there are some caveats to consider. Because brain scans and behavioral measures were collected during a single study period, the results cannot establish whether flattened network dynamics cause psychiatric symptoms or result from living with them. Many psychiatric conditions involve symptoms that cycle over time, and future longitudinal studies that also track rapid trial-by-trial shifts within tasks will be needed to test how network dynamics evolve alongside changing symptoms.
There are also sampling and statistical constraints to keep in mind. Although the full cohort included 219 people, the held-out validation set contained 43 participants, and several individual diagnostic groups included only a handful of people. Because the classification models operated against different chance baselines and potential noise levels, replicating these exploratory comparisons in larger, independent cohorts with additional cognitive tasks will be important to confirm how broadly the findings generalize.
The study did not report accounting for medication use, which an earlier study linked to differences in network dynamics, so some of the flattening seen in diagnosed participants could reflect treatment rather than illness. In addition, psychiatric symptoms and medications can be accompanied by differences in head movement, breathing, or alertness during scanning, and subtle scanner-site differences between Yale University and McLean Hospital could still influence model fit despite strict motion correction and standardized protocols. Even with these limitations, the study offers a promising framework for looking beyond traditional diagnostic boundaries to examine how flexible brain coordination relates to mental health.
The study, “Brain network dynamics reflect psychiatric illness status and transdiagnostic symptom profiles across health and disease,” was authored by Carrisa V. Cocuzza, Sidhant Chopra, Ashlea Segal, Loïc Labache, Rowena Chin, Kaley Joss, and Avram J. Holmes.
