Predicting Post-Stroke Depression Risk in Elderly Patients Based on Machine Learning: A Retrospective Cohort Study Integrating Neuroimaging and Psychosocial Variables.

Objective: To construct a machine learning-based risk prediction model for post-stroke depression (PSD) in elderly stroke patients by integrating neuroimaging and psychosocial variables, and to improve prediction accuracy for individualized prevention. Methods: A retrospective cohort study included 691 elderly (≥60 years) stroke patients from Xiangyang Central Hospital (2016-2023). Baseline clinical data, neuroimaging features (eg, lesion volume, CMBs), and psychosocial variables (eg, SSRS, HAMA-14, HAMD-17) were collected. Seven machine learning models were built; performance was evaluated via AUC, accuracy, and DCA. The optimal model was used to develop a nomogram. Results: The logistic regression (LR) model outperformed others, with AUC=0.88 in the test set. Seven independent predictors were identified: NIHSS score, lesion volume, CMB number, DWI range, Fazekas grade, SSRS score, and HAMA-14 score. The LR-derived nomogram showed good calibration and discriminated PSD from non-PSD effectively. Conclusion: The LR model with 7 predictors is accurate and stable for elderly PSD prediction. Its nomogram aids early high-risk identification, supporting personalized intervention, though single-center limitations require multi-center validation.