关键词: 卒中后认知障碍, 机器学习, 极端梯度提升, 预测模型, 可解释性

Abstract: Objective To develop and validate a machine learning prediction model for post-stroke cognitive impairment （PSCI） and to identify key predictive factors. Methods We retrospectively enrolled 1070 patients with acute ischemic stroke admitted to the Department of Neurology at the Affiliated Lianyungang Hospital of Xuzhou Medical University from January 2020 to August 2023, and randomly divided them into a training set （n=749） and a validation set （n=321） in a 7∶3 ratio. LASSO regression was used for feature selection. Based on selected variables, seven machine learning models were built： Logistic Regression （LR）, Decision Tree （DT）, Random Forest （RF）, Support Vector Machine （SVM）, eXtreme Gradient Boosting （XGBoost）, Light Gradient Boosting Machine （LightGBM）, and K-Nearest Neighbors （KNN）. Model performance was evaluated using the area under the receiver operating characteristic curve （AUC）, calibration curves, and decision curve analysis （DCA）. The SHapley Additive exPlanations （SHAP） method was applied for interpretability of the optimal model. Results LASSO regression identified 11 core predictive variables： atrial fibrillation, speech clarity impairment, depression, TOAST classification, medial temporal lobe atrophy, white matter hyperintensities, multiple lesions, years of education, age, the National Institutes of Health Stroke Scale （NIHSS） score, and infarct volume. The XGBoost model demonstrated optimal and stable predictive performance, with an AUC of 0.863 in the training set and 0.854 in the validation set, differing by only 0.009. Calibration curves yielded Brier scores of 0.141 and 0.155 for the training and validation sets, respectively. DCA showed that the XGBoost model provided a higher clinical net benefit within the threshold probability range of 20%-70%. SHAP analysis revealed that infarct volume, depression, multiple lesions, age, and NIHSS score were the top predictors of PSCI, while years of education had a protective effect. Conclusion The XGBoost-based PSCI prediction model demonstrates good discrimination, calibration, and clinical utility, providing a data-driven decision-support tool for early risk stratification and clinical management of PSCI.

Key words: post-stroke cognitive impairment, machine learning, eXtreme Gradient Boosting, prediction model, interpretability