Brain structural signatures in schizophrenia patients and their siblings: evidence for an intermediate phenotype from an interpretable machine-learning multicohort analysis

Introduction: Schizophrenia (SZ) is one of the most disabling diseases worldwide. Yet, its pathophysiology remains poorly understood despite significant brain structural changes. Furthermore, SZ is highly heritable; thus, first-degree siblings (SIB) may present intermediate characteristics between healthy controls (HC) and SZ. Whether this hypothesis is reflected in brain structure remains an open question. Interpretable machine-learning methods may help disentangle this complexity.

Aim: To use interpretable machine-learning methods to identify key features of the structural brain signature of subjects with schizophrenia, their siblings, and healthy controls, and to assess whether signatures differ.

Methods: We used brain sMRI data from 793 HC, 720 patients with SZ, and 244 SIB from 16 sites across the United States, Japan, and France [1–5] . Data were preprocessed using FreeSurfer and region-wise gray matter volumes (GMV) were extracted using the Desikan-Killiany atlas for cortical areas and FreeSurfer’s standard subcortical segmentations. We trained four machine-learning models (Support Vector Machine, XGBoost, Multi-layer Perceptron, Random Forest) on the GMV of HC and SZ using Python’s SciKit-Learn and a nested cross-validation approach. Site effects were corrected using ComBat, and age, sex, and intracranial volume were corrected using residualization. Models were evaluated using balanced accuracy (BAC). The best-performing model was applied to GMV from siblings, and decision scores were extracted from the test sets of HC, SZ, and SIB. Decision score distributions were compared using Mann-Whitney test with Benjamini-Hochberg correction, and effect sizes were estimated via rank-biserial correlation. Finally, the 20 most important features from SIB and SZ samples were extracted using SHAP.

Results: XGBoost was the best-performing model (72% BAC). SHAP showed that the most relevant areas for well-classified SZ were the bilateral lentiform nuclei, hippocampi, ventricular areas, temporal, and frontal areas, which aligns with previous research. Average decision scores were 0.61 for SZ, 0.41 for SIB, and 0.37 for HC. All pairwise comparisons (HC vs SZ, HC vs SIB, SIB vs SZ) between decision score distributions showed significant differences (p<0.05). Rank-biserial correlation values showed a small effect between HC vs SIB (r = 0.02) and larger effects between HC vs SZ (r = 0.57) and SIB vs SZ (r = 0.48). SHAP showed that areas relevant in siblings classified as SZ were similar to those in well-classified SZ subjects, though with smaller model contributions. Interestingly, the right insula was relevant for SIB but not for SZ, whereas the left choroid plexus was relevant for SZ but not for SIB.

Conclusions: We built a machine-learning model able to distinguish between HC and SZ using GMV data from a large multicohort sample. Applied to siblings, decision scores fell between HC and SZ, albeit closer to HC. Interpretable machine-learning methods showed similar relevant areas between well-classified SZ subjects and SIB classified as SZ. Further research should focus on whether SIB classified as SZ differ from SIB classified as HC in non-structural features such as neuropsychological, social, or environmental characteristics. Together, these findings highlight the potential of interpretable machine-learning approaches to characterize individuals along a continuum of familial risk and identify structural brain signatures in at-risk populations

References

1. Fiorito AM, Kheireddine A, Han H, Morel‐Prieur C, Nicolas Oriol, Schneider FC, et al. Emotion Processing in Schizophrenia: Insights From a Brain Imaging Study Comparing Patients, Siblings, and Healthy Controls. Hum Brain Mapp. 2026 Feb 15;47(3):e70437. doi:10.1002/hbm.70437
2. Repovš G, Barch DM. Working Memory Related Brain Network Connectivity in Individuals with Schizophrenia and Their Siblings. Front Hum Neurosci. 2012;6. doi:10.3389/fnhum.2012.00137
3. Wang L, Alpert KI, Calhoun VD, Cobia DJ, Keator DB, King MD, et al. SchizConnect: Mediating neuroimaging databases on schizophrenia and related disorders for large-scale integration. NeuroImage. 2016 Jan;124:1155–67. doi:10.1016/j.neuroimage.2015.06.065
4. Tamminga CA, Ivleva EI, Keshavan MS, Pearlson GD, Clementz BA, Witte B, et al. Clinical Phenotypes of Psychosis in the Bipolar-Schizophrenia Network on Intermediate Phenotypes (B-SNIP). Am J Psychiatry. 2013 Nov;170(11):1263–74. doi:10.1176/appi.ajp.2013.12101339
5. Tanaka SC, Yamashita A, Yahata N, Itahashi T, Lisi G, Yamada T, et al. A multi-site, multi-disorder resting-state magnetic resonance image database. Sci Data. 2021 Aug 30;8(1):227. doi:10.1038/s41597-021-01004-8