摘要：

目的: 基于真实世界临床数据，评估未加权机器学习模型对急性前循环大血管闭塞性缺血性卒中患者行机械取栓术后预后的预测效能，筛选最优模型，并评估类别加权策略与最优模型预测效能的差异。方法: 纳入2023年5月至2024年9月在天津市环湖医院行血管内机械取栓术的191例急性前循环大血管闭塞性缺血性卒中患者，收集其临床资料如入院时美国国立卫生研究院卒中量表（NIHSS）评分。同时回顾分析入院时头部平扫CT、多时相CT血管成像（mCTA）和CT灌注成像（CTP）资料，采用mCTA评估侧支循环状态；采用Alberta脑卒中计划早期CT评分（ASPECTS）基于平扫CT评估大脑中动脉供血区早期缺血性改变；采用CTP评估脑灌注状态，获得Mismatch体积、Tmax > 4 s体积、Tmax > 6 s体积、Tmax > 8 s体积、Tmax > 10 s体积。以术后90 d改良Rankin量表（mRS）评分作为预后评估指标（> 2分为神经功能预后不良）。采用最小绝对收缩和选择算子（LASSO）回归进行特征筛选，分别采用逻辑回归、随机森林、支持向量机、决策树、k近邻和极端梯度提升算法构建未加权预后预测模型。通过受试者工作特征曲线及曲线下面积（AUC）、校准曲线及Brier分数、决策曲线分析评估模型预测效能，筛选最优模型，采用Shapley加法解释对最优模型进行特征重要性分析；同时评估类别加权策略与该最优模型预测效能的差异。结果: 通过十折交叉验证最小偏差准则确定LASSO回归最优λ值为0.064，筛选出4个特征变量，即ASPECTS评分、Tmax > 10 s体积、入院时NIHSS评分及侧支循环不良。采用分层抽样将所有患者按7∶3比例随机分配至训练集（133例）和测试集（58例），基于上述6种机器学习算法及4个特征变量，建立未加权预测模型。在未加权模型中，除外过拟合的随机森林与极端梯度提升模型，Delong检验显示其余模型的AUC值两两比较差异无统计学意义（均P > 0.05）；但未加权支持向量机模型的Brier分数最低（0.16），提示其校准能力最强；在参考阈值概率15%~30% 范围内，未加权支持向量机模型的决策曲线最高，提示具有最佳临床适用性。类别加权与未加权支持向量机模型的AUC值、灵敏度、特异度、准确率、阳性预测值和阴性预测值比较差异无统计学意义（均P > 0.05）；但与未加权支持向量机模型相比，类别加权支持向量机模型的Brier分数较高（0.17对0.16），提示其校准能力减弱。结论: 在真实世界急性前循环大血管闭塞性缺血性卒中队列中，未加权支持向量机模型可以准确预测机械取栓术后神经功能不良结局，无需依赖类别加权，且该方法具有较高的临床转化潜力。

关键词: 缺血性卒中, 血栓切除术, 机器学习, 支持向量机, 预后

Abstract:

Objective: Based on real-world clinical data, the predictive efficacy of the unweighted machine learning (ML) models for the prognosis of acute ischemic stroke with large vessel occlusion in the anterior circulation (AIS-aLVO) patients after mechanical thrombectomy was evaluated. The optimal model was selected, and the impact of class-weighted strategies on the predictive efficacy of this model was assessed. Methods: A total of 191 patients with AIS-aLVO who underwent mechanical thrombectomy from May 2023 to September 2024 in Tianjin Huanhu Hospital were included. Collect their clinical data, such as pre-admission National Institutes of Health Stroke Scale (NIHSS) score, etc. Retrospectively analyze the brain non-contrast CT (NCCT), multi-phase CT angiography (mCTA) and CT perfusion (CTP) examinations of the patients upon admission. The mCTA was used to assess the collateral circulation status; the Alberta Stroke Program Early CT Score (ASPECTS) was used to evaluate the early ischemic changes in the middle cerebral artery (MCA) supply area based on the NCCT; the CTP was used to assess the cerebral perfusion status, and the Mismatch volume, Tmax > 4 s volume, Tmax > 6 s volume, Tmax > 8 s volume and Tmax > 10 s volume were obtained. The 90-day modified Rankin Scale (mRS) score after surgery was used as the prognostic evaluation index, and the score > 2 was determined as poor prognosis. The Least Absolute Shrinkage and Selection Operator (LASSO) regression was used for feature selection. Logistic regression (LR), random forest (RF), support vector machine (SVM), decision tree (DT), k-nearest neighbor (KNN), and eXtreme Gradient Boosting (XGBoost) algorithms were used to construct unweighted models. The predictive efficacy of the models was evaluated using the receiver operating characteristic (ROC) curve, area under the curve (AUC), calibration curve (Brier score), and decision curve analysis (DCA). The optimal model was selected, and the Shapley additive explanation (SHAP) method was used to analyze the feature importance of this model. At the same time, the impact of the class-weighted strategy on the predictive performance of the model was evaluated. Results: The optimal regularization parameter (λ = 0.064) of LASSO regression was determined by the ten-fold cross-validation minimum deviation criterion. Four feature variables were selected: ASPECTS score, Tmax > 10 s volume, pre-admission NIHSS score, and poor collateral circulation status. Stratified sampling was used to randomly allocate the subjects to the training set (n = 133) and the test set (n = 58), and unweighted models was established. In the unweighted model, except for the overfitting RF and XGBoost models, the Delong test showed that the pairwise comparison of the AUC values of the remaining models had no statistical significance (P > 0.05, for all); however, the unweighted SVM model had the lowest Brier score (0.16), and its calibration ability was the strongest. Within the 15%-30% threshold range, the DCA curve of the unweighted SVM model was the highest, suggesting the highest clinical applicability. There was no statistically significant difference in the AUC values, sensitivity, specificity, accuracy, positive predictive value and negative predictive value between the class-weighted and unweighted SVM models (P > 0.05, for all); however, compared with the unweighted SVM model, the Brier score of the class-weighted SVM model was higher (0.17 vs. 0.16), and its calibration ability was weakened. Conclusions: In a real-world cohort of AIS-aLVO cohort, the unweighted SVM model can accurately predict poor functional outcomes after mechanical thrombectomy without relying on class-weighted, and this method has high clinical translational potential.

Key words: Ischemic stroke, Thrombectomy, Machine learning, Support vector machine, Prognosis