Multimodal classification of mood disorders using structural MRI and polygenic risk scores: a machine learning ablation study in UK Biobank.

Objective: Mood disorder classification from neuroimaging alone remains limited by single-modality constraints and population-level heterogeneity. This study proposes a fusion-based machine learning framework combining structural brain MRI and polygenic risk scores (PRS) derived from genome-wide association studies, alongside demographic and imaging confounder data, to classify individuals into three categories: healthy controls (HC), mania/bipolar disorder (MBD), and major depressive disorder (MDD). Methods: Data were drawn from the UK Biobank under a retrospective, cross-sectional design. Three dataset configurations were structured as an ablation study: neuroimaging-only (D1; n=426), genetics-only (D2; n=264), and full multimodal fusion (D3; n=264). T1-weighted 3D MRI volumes were converted to grayscale, and the eight axial slices with the highest Shannon entropy per participant were selected, resized to 64×64 pixels, and reduced from 32,768 to 150 features via ANOVA F-test univariate selection. PRS for bipolar disorder and MDD were derived from Psychiatric Genetics Consortium GWAS summary statistics using LDpred, restricted to European ancestry participants. Six machine learning classifiers were trained under each configuration using an 80:20 stratified split with 10-fold cross-validation, with macro-averaged AUC as the primary metric. Results: XGBoost achieved the strongest discriminatory performance, attaining a macro-average AUC of 0.7647 and an MDD-specific AUC of 0.8175 on the neuroimaging-only configuration, representing peak performance across all conditions. The genetics-only configuration produced substantially lower macro-average AUC values across all classifiers (XGBoost: 0.6600), consistent with the modest individual-level effect sizes of common psychiatric genetic variants. The controlled same-sample comparison between D2 and D3 revealed that the addition of MRI features improved macro-average AUC by 0.052 (XGBoost: 0.6600 to 0.7121), with the most pronounced gain observed for MBD classification (AUC 0.7160 to 0.7840). Despite its smaller sample size, D3 achieved competitive performance, suggesting that PRS may provide complementary signal, although direct comparison with D1 is limited by differences in sample composition. Conclusion: These findings support the potential of interpretable, fusion-based machine learning for advancing diagnostic precision in mood disorder research, with multimodal integration of neuroimaging and genetic risk offering complementary signal beyond what either modality provides independently.