Holistic compression model to predict post-stroke mood at 6 months

Introduction: Mood impairments are frequent after stroke yet remain difficult to predict using neuroimaging markers. While voxel-level analyses have struggled to identify reliable predictors, network-level approaches may offer improved sensitivity. Accordingly, we developed a pipeline to predict post-stroke mood status using whole-brain components derived from MRI sequences routinely acquired in clinical practice. Methods Three heterogeneous stroke cohorts (n _total = 672) were included, with mood evaluated six months post-stroke using the Hospital Anxiety and Depression (HAD) and Center for Epidemiologic Studies–Depression (CES-D) scales. For model training, diffusion-weighted signal from one cohort was compressed using principal component analysis (PCA), and components associated with mood scores were used in nonparametric regression models. Predictive performance was externally validated and compared with disconnectome-based models and models based on baseline clinical and demographic variables. Disconnectome-based models showed minimal predictive power across datasets (R² < 0.01). The imaging-based pipeline modestly outperformed clinical models for predicting CES-D scores (ΔAUC = 0.046, p < 0.001; ΔR² = 0.056), in the absence of baseline mood measures. For HAD outcomes, clinical models performed best for anxiety (AUC = 0.735) and depression (AUC = 0.795), with baseline mood status as the strongest predictor. The proposed pipeline achieved moderate predictive performance for HAD anxiety (AUC = 0.613) and depression (AUC = 0.708). Conclusion Overall, predictive performance remained limited across all approaches, highlighting the heterogeneity in neuropsychiatric predictions. This novel framework to capture global diffusion patterns without requiring advanced imaging sequences may provide complementary information for identifying neural components associated with neuropsychiatric dysfunction.