A comparison of clinical, lesion-based and connectome-based models of post-stroke depression: a prospective...

Post-stroke depression (PSD) is a frequent and disabling complication of ischemic stroke. Although clinical risk factors of PSD are well-established, the added predictive value of lesion and connectome-based brain-symptom mapping methods remains unclear. This study aimed to compare the predictive performance of clinical, lesion based, and network-based models of 3-months PSD in patients with minor ischemic stroke. We conducted a prospective longitudinal study including 263 patients with recent ischemic stroke. Clinical, radiological, and psychometric data were collected at baseline and at 3-month follow-up. Structural lesions were segmented on acute MRI and used to generate disconnectome maps and compute graph-theory metrics. Depressive symptoms severity at 3 months were measured using the CES-D scale. We developed six predictive models (clinical, radiological, gray-matter lesion, white matter disconnection, functional network disconnection, and topological graph features) using LASSO regression. Models performances were compared using the coefficient of determination (R²) and their combinations further examined through hierarchical regression. The clinical model demonstrated the best individual predictive performance (R² = 23 %), identifying female sex, low cognitive status, poor functional outcome, and socioeconomic deprivation as significant predictors. Imaging-based models alone showed limited predictive power (R² < 6 %), but identified significant associations for frontal and cerebellar lesions, somatomotor network disconnections, and altered network topology. A hierarchical model combining clinical, anatomical, and topological features significantly improved prediction (R² = 31.6 %, p < 0.001), outperforming all individual models. In patients with minor ischemic stroke, PSD is most accurately predicted by clinical factors. However, combining lesion and network level neuroimaging features with clinical variables significantly enhances predictive accuracy. This multimodal approach supports the development of integrative, personalized models for PSD risk stratification and early intervention.