Development and interpretation of a machine learning risk prediction model for post-stroke depression in a Chinese population

Abstract

Current evidence for predictive models of post-stroke depression (PSD) risk based on machine learning (ML) remains limited. The aim of this study is to develop a superior predictive model based on ML algorithms for PSD in the Chinese population. We retrospectively gathered data from 507 patients who were hospitalized for ischemic stroke between June 2022 and August 2023 at a medical facility in China. The patients were separated into a training group (70%) and a validation group (30%). After selecting the core predictors of PSD using LASSO regression dimension reduction, six machine learning (ML) algorithms were used to statistically model the risk prediction of PSD. Based on five-fold cross-validation, we evaluated the predictive performance of the developed model. Shape Additive Explanation (SHAP) analysis was then used to interpret the best-performing model. PSD developed in 158 (31.16%) of the 507 eligible patients. Frontal lobe lesion, number of lesions, ALB, NIHSS, PSQI, and MMSE were considered significant predictors of PSD. The AUC(0.941), accuracy (0.876), sensitivity (0.822), specificity (0.899), F1 score (0.802), and average precision (AP) value (0.858) of XGBoost model were superior to other ML models. The XGBoost model offers an interpretable PSD prediction tool using key clinical indicators, though external validation is needed to confirm generalizability.