Prediction of Post-Stroke Depression Using Inflammatory Markers and Functional Status: A Machine Learning Approach

Introduction Post-stroke depression (PSD) is a common neuropsychiatric complication that adversely affects recovery and quality of life in stroke survivors. Inflammation has been increasingly recognized as a potential contributor to PSD. However, effective tools for early prediction of PSD remain limited. This study aimed to identify key inflammatory and functional indicators associated with PSD and to construct a clinically applicable predictive model. Methods We conducted a prospective cohort study including 228 patients with acute ischemic stroke. Five variables—C-reactive protein (CRP), serum amyloid A (SAA), interleukin-6 (IL-6), homocysteine (HCY), and Activities of Daily Living (ADL) score—were selected based on univariate analysis and clinical relevance. Three machine learning models (logistic regression, random forest, and XGBoost) were constructed to predict PSD. Model performance was evaluated using area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, and F1-score. The best-performing model was further assessed by decision curve analysis (DCA). Results The logistic regression model achieved the best overall performance, with an AUC of 0.757, accuracy of 80.9%, and specificity of 97.8%. Random forest yielded the highest sensitivity (52.2%) but comparable AUC (0.761). XGBoost demonstrated lower performance after tuning (AUC: 0.717). DCA showed that the logistic regression model provided a clear net clinical benefit across a wide range of threshold probabilities (0.1–0.6), supporting its use in clinical decision-making. Conclusion A logistic regression model based on CRP, SAA, IL-6, HCY, and ADL accurately predicted post-stroke depression and showed good clinical utility. It may aid in early risk identification and targeted intervention in stroke survivors.