【摘要】 目的 为预测未接受再通治疗的急性缺血性卒中（NR-AIS）患者早期神经功能恶化（END）的发生，通过分析NR-AIS患者基线资料及WORSEN评分构建与验证机器学习预测模型。方法 回顾性收集2023年1月10日—2025年12月1日期间符合标准的天津市环湖医院神经内科收治的NR-AIS患者的临床资料与实验室指标，并引入WORSEN评分。训练集与测试集通过7∶3进行划分。≤20%的缺失数据采用K近邻（KNN）进行插补；训练阶段引入合成少数类过采样技术（SMOTE）过采样以缓解类别偏态。数据预处理后，特征筛选通过最小绝对收缩和选择算子（LASSO）回归完成。在此基础上构建支持向量机（SVM）、随机森林（RF）等多种预测模型。受试者工作特征曲线下面积（AUC）与Brier评分用于评估模型的区分能力与校准精度，并采用沙普利加和解释（SHAP）框架解释各变量的贡献方向及其对预测输出的影响。结果 共纳入825例患者，其中209例发生END。LASSO回归筛选后纳入5种关键变量：美国国立卫生研究院卒中量表（NIHSS）评分、WORSEN评分、低密度脂蛋白（LDL）、总胆固醇及年龄。测试集SVM模型的AUC达到0.928(95%CI:0.893～0.958),Brier评分为0.108，整体表现优于其他预测模型。SHAP结果显示，WORSEN评分和基线NIHSS评分是预测END的最主要驱动因素。结论 基于入院基线临床资料与实验室指标构建的机器学习模型对预测NR-AIS患者是否发生END具有一定效果。SVM模型在研究中展现出较为理想的判别与校准能力。更多还原

【Abstract】 Objective To predict the occurrence of early neurological deterioration(END) in patients with non-reperfusion acute ischemic stroke(NR-AIS), a machine learning model was developed and validated by analyzing baseline data and WORSEN scores from NR-AIS patients. Methods We retrospectively retrieved clinical data and laboratory results for patients with NR-AIS admitted to the Department of Neurology at Tianjin Huanhu Hospital between January 10, 2023, and December 1, 2025, who met the inclusion criteria, and incorporated the WORSEN score. The training set and test set were split in a 7:3 ratio. Missing data(≤20%) were imputed using the K-nearest neighbors(KNN) method; SMOTE oversampling was applied during the training phase to mitigate class skewness. After data preprocessing, feature selection was performed using LASSO regression. Based on this, multiple predictive models, including Support Vector Machines(SVM) and Random Forests(RF), were constructed. The Area Under the Curve(AUC) and Brier score were used to evaluate the models’ discriminatory power and calibration accuracy, and the SHAP framework was employed to interpret the contribution directions of each variable and their impact on the prediction output. Results Among 825 patients, 209 experienced an END. After screening using LASSO regression, five key variables were identified:NIHSS score, WORSEN score, low-density lipoprotein,(LDL), total cholesterol, and age.The AUC of the SVM model on the test set reached 0.928(95% CI:0.893-0.958), with a Brier score of 0.108, demonstrating overall superior performance compared to other predictive models. SHAP results indicated that the WORSEN score and baseline NIHSS score were the primary drivers for predicting END. Conclusion Machine learning models constructed using baseline clinical and laboratory data at admission demonstrate some effectiveness in predicting whether patients with NR-AIS will experience END. The SVM model exhibited relatively ideal discriminatory and calibration capabilities in this study.更多还原