MRI brain topology separates depression and predicts suicide risk in 154 patients

A cross-sectional MRI study of 154 adults with first-episode, drug-naïve major depressive disorder found that a machine-learning system combining white-matter connectivity with higher-order functional brain networks could distinguish patients from healthy controls with an area under the receiver operating characteristic curve of 0.882. The same multimodal features predicted measured suicide-risk scores in the depressed group with a correlation of 0.604, showing that information about brain organization was associated with both illness status and clinically assessed risk. The findings remain a research signal rather than a clinical test: the models were developed at one center, lacked an independent external cohort and cannot establish temporal relationships or causation.

Researchers from Yantai Yuhuangding Hospital and associated institutions recruited 154 people with major depressive disorder and 153 age-, sex- and education-matched healthy controls. Patients were diagnosed by qualified psychiatrists according to criteria for a major depressive episode and were experiencing their first episode without medication. MRI quality control and the availability of complete clinical assessments excluded 23 recruited patients and 14 controls.

The final groups had similar ages, sex distributions and educational attainment. The depressed group had substantially higher Hamilton Depression Rating Scale scores, lower Montreal Cognitive Assessment scores and elevated suicide-risk assessments. The researchers represented the brain using two complementary kinds of network.

Diffusion tensor imaging was used to estimate fractional anisotropy along reconstructed white-matter pathways, producing a 116-node structural connectome based on the Automated Anatomical Labeling atlas. Graph theory then quantified properties such as degree centrality, which measures a node’s relative importance within the network. Resting-state functional MRI supplied blood-oxygen-level-dependent time series from the same 116 regions.

Rather than relying only on conventional pairwise functional connectivity, which measures correlations between two regions, the study reconstructed interactions involving triplets of regions through persistent simplicial homology. This computational-topology method tracks how complex network structures emerge and disappear as connection strengths vary. For the higher-order analysis, the team calculated product time series for every pair and triplet of brain regions.

A coherence correction assigned different weights depending on whether the component signals had consistent signs, allowing fully coherent co-fluctuations to be distinguished from incoherent ones. At each time point, weighted triangles were organized into simplicial complexes subject to a closure rule: every face of an included triangle must also be present. Triangles violating that condition were recorded as “violating triangles,” and their proportion among all possible coherent triangles yielded a hyper-coherence index.

The analysis also produced a homological scaffold, a weighted network in which connection strength reflects how long one-dimensional loops persist during filtration. A hyper-complexity index summarized the overall distance between the persistent-homology diagram and an empty diagram. Compared with controls, participants with depression showed significant changes across several large-scale systems, including the default mode, visual, frontoparietal, salience/ventral-attention, sensorimotor, limbic, subcortical and cerebellar networks.

Structural degree centrality was reduced mainly in default-mode and visual regions and increased in frontoparietal and salience/ventral-attention regions. The functional analysis revealed more widespread violating triangles, particularly in medial prefrontal, cingulate, hippocampal and precuneus regions of the default mode network, as well as the insula, frontal cortex, basal ganglia, thalamus and cerebellum. Homological scaffold connectivity followed two patterns: stronger connections among frontal, limbic and default-mode areas and among cortical, subcortical and cerebellar areas, but weaker connectivity in parietal-cerebellar, sensorimotor-parietal, thalamic and frontostriatal pathways.

Four support-vector classifiers were evaluated under a repeated nested-validation design. This involved 50 repetitions of stratified 70/30 training-test splits, with 10-fold inner cross-validation used for feature screening, LASSO reduction and hyperparameter optimization. Data-dependent preprocessing was confined to each training set, reducing the risk that information from held-out participants influenced model development.

The structural-only model achieved mean outer-test accuracy of 66.9% and an AUC of 0.745, with sensitivity of 70.1% and specificity of 63.7%. The higher-order functional model performed better, reaching 75.9% accuracy and an AUC of 0.849, with 78.1% sensitivity and 73.9% specificity. A conventional pairwise-functional-connectivity model produced an AUC of 0.701, significantly below the higher-order model in a DeLong comparison after false-discovery-rate correction.

Fusing structural and higher-order functional features produced the strongest diagnostic results. The multimodal support-vector machine classified patients and controls with mean accuracy of 78.9%, sensitivity of 81.5%, specificity of 76.4% and an AUC of 0.882; the empirical 95% interval for the AUC was 0.80 to 0.95. Its AUC was significantly higher than that of the structural-only model, with a corrected p value of 0.028.

These are held-out results from the study’s repeated validation framework, not performance values from a separate external cohort. The comparison nevertheless indicates that the anatomical and higher-order functional descriptions carried complementary classification information, rather than one modality simply duplicating the other. The researchers also built a ridge-regression model to predict the Nurses Global Assessment of Suicide Risk score from fused imaging features.

The regularization parameter was selected within nested 10-fold cross-validation, and the model was evaluated by the relationship between predicted and observed scores. Predicted values correlated with measured suicide risk at r = 0.604, with a 95% interval of 0.499 to 0.694, while the coefficient of determination was 0.364, with an interval of 0.249 to 0.482. Root mean square error was 2.330 points and mean absolute error was 1.796 points.

Observed risk scores ranged from 1 to 16. The model therefore captured a substantial but incomplete portion of score variation and did not reproduce every patient’s clinical assessment exactly. Feature-weighting analysis indicated that the frontal, cerebellar and parietal lobes were major contributors to the risk model, a pattern broadly consistent with earlier imaging research implicating prefrontal and cerebellar abnormalities in suicidal ideation or self-injurious phenotypes in depression.

The findings should nevertheless be interpreted as associations within a cross-sectional dataset. The authors explicitly cautioned that the design cannot establish whether topological alterations precede depressive onset, contribute causally to suicide risk, or arise from illness or other factors, and that residual confounding cannot be excluded. The coarse anatomical atlas may also miss individualized network boundaries, while the single-center cohort raises questions about transportability across hospitals, scanners and populations.

Taken together, the study shows why brain activity may need to be considered as more than a collection of connections between pairs. Conventional functional connectivity performed less well than the higher-order topological model in this dataset, while adding structural white-matter information improved discrimination further. The combination also retained a moderate statistical association with clinician-rated suicide risk, highlighting the cerebellum and frontal systems as potential sources of predictive signal.

These results position multimodal brain topology as a candidate biomarker platform rather than a finished diagnostic test.

Subject of Research: Multimodal MRI brain-topology models for diagnosing major depressive disorder and predicting measured suicide risk

Article Title: Diagnostic and predictive value of structural and high-order functional brain topology in major depressive disorder

Article References: Wang, X., Wu, X., Shi, M., Che, K., Shen, X., Chu, T., Mao, N., & Wang, P. (2026). Diagnostic and predictive value of structural and high-order functional brain topology in major depressive disorder. Frontiers in Psychiatry, 17, Article 1964608. https://doi.org/10.3389/fpsyt.2026.1964608

Image Credits: AI Generated

DOI: 10.3389/fpsyt.2026.1964608

Keywords: major depressive disorder, brain topology, higher-order interactions, diffusion tensor imaging, resting-state fMRI, persistent homology, machine learning, suicide-risk prediction, multimodal neuroimaging, Diagnostic, predictive, value