2026 ECNP Congress

PS02-2227

Neural correlates of comorbid depression in schizophrenia spectrum disorders: a multi-sample machine learning study

Sergio Mena Ortega — F. Coutts, G. Jacobs, F. Sequeira, C. Vetter, L. Hahn, L. Moles, N. Sirbu, L. Bryant, E. Meisenzahl, O. Howes, P. Psy-ShareD, P.A. Lalousis, N. Koutsouleris · IoPPN, Department of Psychosis, London, United Kingdom

2026 ECNP Congress

Background

Depression is a frequent and highly debilitating condition in schizophrenia spectrum disorders (SSD)​1–3​, associated with poor outcomes and elevated suicide risk. Despite its clinical relevance, the neurobiological substrates distinguishing SSD patients with comorbid depression remain poorly characterised​4​, and it is unclear to what extent they overlap with those observed in major depressive disorder (MDD). Identifying disorder-specific neural signatures of depression in SSD could improve clinical recognition and open new avenues for targeted intervention. We aimed to identify and cross-validate structural neural correlates of depression in SSD across large international databanks, and to assess the degree of shared pathophysiology with MDD.

Methods

We extracted baseline T1-weighted structural MRI scans from SSD patients across two independent international databanks: the Department of Psychosis Shared Database Initiative (DPSDI; N=586, 8 cohorts from North America, Europe and Asia; 38.56% with depression) and PsyShareD (N=338, 11 cohorts from Europe; 44.67% with depression). Depression status was harmonised across validated instruments (CDSS, HAM-D, MADRS, BDI-II, PANSS-D). For comparison, MDD patients and age-matched healthy controls were drawn from DPSDI (N=2,034, 6 cohorts from North America, Europe and Asia). Structural images were processed using voxel-based morphometry (CAT12 toolbox implemented in SPM12) to generate gray matter volume (GMV) and white matter volume (WMV) tissue maps, with ComBat correction applied for site and cohort effects. Support vector classifiers were trained using nested leave-cohort-out cross-validation (10 inner and outer folds and permutations). Between-disorder similarity of predictive patterns was assessed using distance correlation (dCor)–based permutation testing (N=10,000 permutations).

Results

CClassifiers trained in DPSDI achieved balanced accuracies of 60.3% (GMV; sensitivity=56.6%, specificity=63.9%), 62.0% (WMV; sensitivity=60.6%, specificity=63.3%), and 62.0% (stacked; sensitivity=59.7%, specificity=64.2%) for depression classification in SSD (all FDR-corrected p<0.01). Cross-databank validation in PsyShareD yielded balanced accuracies of 60.6% (GMV; sensitivity=57.6%, specificity=63.6%), 58.5% (WMV; sensitivity=60.9%, specificity=56.1%), and 60.5% (stacked; sensitivity=56.3%, specificity=64.7%; p<0.05–0.01). In parallel, MDD versus healthy control classifiers achieved balanced accuracies of 64.3% (GMV; sensitivity=69.7%, specificity=59.0%), 62.0% (WMV; sensitivity=51.9%, specificity=72.1%), and 66.0% (stacked; sensitivity=65.6%, specificity=66.3%; all p<0.01). Top predictors of depression in SSD included GMV reductions in cerebellar, limbic, and orbitofrontal regions. A weak but significant spatial similarity was found between depression-predictive patterns in DPSDI and PsyShareD (dCor=0.28, p=0.002, N=10,000 permutations), supporting cross-cohort reproducibility of the SSD-specific signature. No significant correspondence was found between SSD depression and MDD predictive patterns (dCor=0.16, p=0.14, N=10,000 permutations), suggesting that the structural underpinnings of depression differ meaningfully across these diagnostic categories.

Conclusion

This multi-databank machine learning study identifies reproducible structural brain alterations associated with comorbid depression in SSD, with moderate but geographically generalisable classification performance. The absence of significant pattern overlap with MDD suggests that the depression symptom dimension in SSD may involve partially distinct neurobiological substrates compared to primary affective disorders. Future research will explore the validation of models across disorders, and explore the prognostic validity of these neural correlates on the early prediction of future depressive episodes in SSD.

References

​​1.Mena, S. et al. AI-based prediction of depression symptomatology in first-episode psychosis patients: insights from the EUFEST and RAISE-ETP clinical trials. Psychol. Med. 55, e221 (2025). ​2.Mena, S. et al. Identifying Transdiagnostic Predictors of Depression Across Psychoses: Informing Stratified Antidepressant Treatments. Psychother. Psychosom. https://doi.org/10.1159/000551070 (2026) doi:10.1159/000551070. ​3.McGinty, J. & Upthegrove, R. Depressive symptoms during first episode psychosis and functional outcome: A systematic review and meta-analysis. Schizophr. Res. 218, 14–27 (2020). ​4.Alexandros Lalousis, P. et al. Transdiagnostic structural neuroimaging features in depression and psychosis: A systematic review. Neuroimage Clin. 38, 103388 (2023).