Abstract: Objective To investigate the predictive value of quantitative parameters from high-resolution magnetic resonance vessel wall imaging (HR-VWI) for acute ischemic stroke (AIS). Methods The 120 patients with intracranial atherosclerosis who underwent both cranial CT and HR-VWI examinations at Second Affiliated Hospital of Nanjing Medical University from June 2021 to October 2024 and completed at least 6 months follow-up. These patients were divided into the training set (n = 84) and the testing set (n = 36) at a ratio of 7 ∶ 3. Feature selection was performed using the least absolute shrinkage and selection operator (LASSO), and the final predictive model was constructed with the XGBoost model. Model predictive performance was assessed using receiver operating characteristic (ROC) curves and area under the curve (AUC), and generalization ability was further evaluated through learning curves. Shapley additive explanation (SHAP) was used to calculate the global contribution of each feature variable to the model′s prediction. Results Among the 120 patients, 27 cases (22.50%) experienced AIS within the 6 -month follow-up period. Specifically, there were 19 cases (22.62%) in the training set and 8 cases (22.22%) in the testing set. Multivariate Logistic regression identified calcified plaque volume ratio (OR = 0.123, 95%CI: 0.039-0.393; P = 0.000), plaque enhancement ratio (OR = 1.130, 95%CI: 1.046-1.221; P = 0.002), plaque hemorrhage (OR = 9.519, 95%CI: 2.453-36.968; P = 0.008), and luminal stenosis ratio (OR = 1.106, 95%CI: 1.032-1.185; P = 0.004) as predictors of AIS. ROC curves showed the AUC of Logistic regression model was 0.950 and 0.812 in the training set and the testing set, with sensitivity of 0.947 and 0.750 and specificity of 0.800 and 0.857, respectively. Similarly, the XGBoost model achieved an AUC of 0.986 in the training set and 0.844 in the testing set, with a sensitivity of 0.947 and 0.875, and a specificity of 0.923 and 0.857, respectiveiy. Learning curves analysis further confirmed that both models exhibited good stability and generalization ability as the training sample size increased, with XGBoost showing slightly better overall predictive performance. SHAP analysis indicated that the calcified plaque volume ratio contributed the most to model prediction. Conclusions Machine learning models that integrate calcified plaque volume ratio, plaque enhancement ratio, plaque hemorrhage and lumen stenosis ratio can significantly improve the predictive performance for AIS, exhibit good generalization ability, and provide a potential reference tool for clinical risk assessment.

Key words: Ischemic stroke, Magnetic resonance imaging, Atherosclerosis, Machine learning, ROC curve

摘要： 目的 探讨高分辨率磁共振血管壁成像（HR-VWI）定量参数对急性缺血性卒中的预测价值。方法 纳入 2021 年 6 月至 2024 年 10 月在南京医科大学第二附属医院行头部 CT 平扫及 HR-VWI 检查，且完成至少 6 个月随访的 120 例颅内动脉粥样硬化患者，按照 7∶3 比例分为训练集（84 例）和测试集（36 例），采用最小绝对收缩与选择算子（LASSO）回归筛选特征变量，基于 XGBoost 算法构建预测模型，绘制受试者工作特征（ROC）曲线并计算曲线下面积评价模型的预测效能，并通过学习曲线评估模型的泛化能力，Shapley 加法解释（SHAP）计算各特征变量对模型预测的全局贡献值。结果 共 120 例患者中27 例（22.50%）发生急性缺血性卒中，其中训练集 19 例（22.62%），测试集 8 例（22.22%）。多因素 Logistic回归分析显示，斑块钙化体积比（OR = 0.123，95%CI：0.039 ~ 0.393；P = 0.000）、斑块强化率（OR = 1.130，95%CI：1.046 ~ 1.221；P = 0.002）、斑块出血（OR = 9.519，95%CI：2.453 ~ 36.968；P = 0.008）及管腔狭窄率（OR = 1.106，95%CI：1.032 ~ 1.185；P = 0.004）是急性缺血性卒中的预测因子。ROC 曲线显示，Logistic 回归模型在训练集和测试集的曲线下面积分别为 0.950 和 0.812，灵敏度为 0.947 和 0.750、特异度为 0.800和 0.857；XGBoost 模型在训练集和测试集的曲线下面积分别为 0.986 和 0.844，灵敏度为 0.947 和 0.875、特异度为 0.923 和 0.857。学习曲线显示，两种模型在增加训练样本量后均表现出良好的稳定性和泛化能力，其中 XGBoost 的整体预测性能略优。SHAP 值重要性分析，斑块钙化体积比是对模型预测贡献度最高的急性缺血性卒中预测因子。结论 整合斑块钙化体积比、斑块强化率、斑块出血及管腔狭窄率的机器学习模型可以显著提高急性缺血性卒中的预测效能，并具有较好的泛化能力，为临床风险评估提供潜在参考工具。

关键词: 缺血性卒中, 磁共振成像, 动脉粥样硬化, 机器学习, ROC曲线