分类树模型在缺血性脑卒中危险因素筛查中的应用研究
本研究采用整群抽样纳入2017年1月至12月桂林医学院附属医院858例临床资料完整的缺血性脑卒中患者(IS组),并以同期性别、年龄匹配的844例健康体检者作为对照(健康对照组)。比较两组代谢特征后,用分类树模型构建缺血性脑卒中风险预测模型,并通过增益图、指数图、误分类概率风险值及受试者工作特征曲线(ROC)评价模型应用价值。结果显示,IS组BMI、空腹血糖、甘油三酯、总胆固醇、LDL-C均显著高于健康对照组,HDL-C显著降低,高血压、吸烟、饮酒比例显著升高。分类树模型共4层17个节点,筛出高血压、空腹血糖、HDL-C、LDL-C和吸烟5个解释变量,模型误分类概率风险值为0.291,AUC为78.0%(95%CI 75.9%-79.9%,P<0.001),敏感度62.5%、特异度79.4%。作者认为分类树模型可较好地预测缺血性脑卒中危险因素,最重要的危险因素为高血压、高血糖、高LDL-C和吸烟。
要点
- 分类树模型最重要的缺血性脑卒中危险因素为高血压、高血糖、高LDL-C和吸烟,模型AUC为78.0%(95%CI 75.9%-79.9%,P<0.001)。
- 高血压人群缺血性脑卒中发生率为62.6%,显著高于非高血压人群的35.2%。
- 非高血压人群中,空腹血糖≥7.0 mmol/L者缺血性脑卒中概率为71.1%,显著高于空腹血糖<7.0 mmol/L者的28.3%。
- 高血压且HDL-C≥1.04 mmol/L人群中,空腹血糖≥7.0 mmol/L者缺血性脑卒中概率为78.6%,显著高于空腹血糖<7.0 mmol/L者的46.7%。
- 模型敏感度为62.5%(95%CI 59.1%-65.7%),特异度为79.4%(95%CI 76.5%-82.1%),误分类概率风险值为0.291。
适用人群与场景
- 纳入2017年1月至12月桂林医学院附属医院临床资料完整的缺血性脑卒中患者858例(IS组),以及同期性别、年龄匹配的健康体检者844例(健康对照组)。
诊断标准与关键阈值
- 分类树模型变量赋值:缺血性脑卒中无=0、有=1;BMI<24.0 kg/m²=0、≥24.0 kg/m²=1;空腹血糖<7.0 mmol/L=0、≥7.0 mmol/L=1;甘油三酯<2.26 mmol/L=0、≥2.26 mmol/L=1;总胆固醇<6.22 mmol/L=0、≥6.22 mmol/L=1;LDL-C<4.14 mmol/L=0、≥4.14 mmol/L=1;HDL-C<1.04 mmol/L=0、≥1.04 mmol/L=1;高血压无=0、有=1;吸烟无=0、有=1;饮酒无=0、有=1。
研究设计与关键结果
- 本研究为整群抽样病例对照研究,纳入缺血性脑卒中患者858例和健康对照844例,比较两组代谢特征并构建分类树预测模型。
- 与健康对照组比较,IS组BMI(25.34±3.70 vs. 24.24±3.10 kg/m²)、空腹血糖(6.79±2.89 vs. 5.73±1.17 mmol/L)、甘油三酯(1.62±1.06 vs. 1.44±1.06 mmol/L)、总胆固醇(4.70±2.73 vs. 4.35±0.79 mmol/L)、LDL-C(3.18±0.94 vs. 2.73±0.73 mmol/L)均显著升高(均P<0.01),HDL-C显著降低(1.12±0.33 vs. 1.35±0.36 mmol/L,P<0.01),高血压、吸烟、饮酒比例显著升高(69.0% vs. 41.9%、23.1% vs. 16.8%、19.2% vs. 13.4%,均P<0.01)。
- 分类树模型共4层17个节点:第一层为高血压,第二层为空腹血糖和HDL-C,第三层为HDL-C和空腹血糖,第四层为LDL-C和吸烟;筛出高血压、空腹血糖、HDL-C、LDL-C和吸烟5个解释变量。
- 第一层:高血压人群缺血性脑卒中发生率为62.6%,显著高于非高血压人群的35.2%。
- 第二层:高血压且HDL-C≥1.04 mmol/L人群缺血性脑卒中发生率为53.6%,低于HDL-C<1.04 mmol/L人群的78.5%;非高血压人群中,空腹血糖≥7.0 mmol/L者缺血性脑卒中概率为71.1%,显著高于空腹血糖<7.0 mmol/L者的28.3%。
- 第三层:非高血压且空腹血糖<7.0 mmol/L人群中,HDL-C≥1.04 mmol/L者缺血性脑卒中发生率为21.8%,低于HDL-C<1.04 mmol/L者的48.7%;高血压且HDL-C≥1.04 mmol/L人群中,空腹血糖≥7.0 mmol/L者缺血性脑卒中概率为78.6%,显著高于空腹血糖<7.0 mmol/L者的46.7%。
- 第四层:非高血压、空腹血糖<7.0 mmol/L且HDL-C≥1.04 mmol/L人群中,LDL-C≥4.14 mmol/L者缺血性脑卒中发生率为53.8%,高于LDL-C<4.14 mmol/L者的19.0%;非高血压、空腹血糖<7.0 mmol/L且HDL-C<1.04 mmol/L人群中,吸烟者缺血性脑卒中发生率为76.9%,高于不吸烟者的39.1%;高血压、HDL-C≥1.04 mmol/L且空腹血糖<7.0 mmol/L人群中,LDL-C≥4.14 mmol/L者缺血性脑卒中概率为72.5%,高于LDL-C<4.14 mmol/L者的44.4%。
- 分类树模型增益图显示增益值从0%迅速升至100%后趋于稳定;指数图显示指数值在100%以上移动方向保持稳定后迅速降至100%;模型误分类概率风险值为0.291,缺血性脑卒中患者危险因素正确率为70.90%;AUC为78.0%(95%CI 75.9%-79.9%,P<0.001),敏感度为62.5%(95%CI 59.1%-65.7%),特异度为79.4%(95%CI 76.5%-82.1%)。
证据类型、等级与局限
- 证据类型:整群抽样病例对照研究,构建并评价分类树预测模型。
- 局限:原文未报告局限性。
对床旁实践的改变
- 作者认为分类树模型可较好地预测缺血性脑卒中危险因素,最重要的危险因素为高血压、高血糖、高LDL-C和吸烟。
提炼说明
- 本条依据论文摘要整理;摘要未涉及「处置流程与药物剂量」「红旗征与不能漏的情况」「指南或共识出处」,故未列出这些栏目;全文请点「阅读原文」。
英文原文(摘要)· 点击展开
OBJECTIVE: To construct a prediction model for the risk of ischemic stroke (IS) by classification tree model, and evaluate its application value.
METHODS: By cluster sampling, 858 IS patients with perfect clinical data from January to December 2017 in the Affiliated Hospital of Guilin Medical College (IS group) were enrolled, and 844 health checkups matched with the gender and age of IS patients in the same period were enrolled as controls (healthy control group). The metabolic characteristics of the two groups were compared and analyzed. The classification tree model was used to construct the prediction model of the risk of IS, and the gain diagram, index chart, risk value of misclassification probability and receiver operating characteristic curve (ROC) were used to evaluate the application value of the model.
RESULTS: Compared with the healthy control group, body mass index (BMI), fasting blood glucose (FPG), triglyceride (TG), total cholesterol (TC), low density lipoprotein cholesterol (LDL-C) in IS group were significantly increased [BMI (kg/m2): 25.34±3.70 vs. 24.24±3.10, FPG (mmol/L): 6.79±2.89 vs. 5.73±1.17, TG (mmol/L): 1.62±1.06 vs. 1.44±1.06, TC (mmol/L): 4.70±2.73 vs. 4.35±0.79, LDL-C (mmol/L): 3.18±0.94 vs. 2.73±0.73, all P < 0.01], high density lipoprotein cholesterol (HDL-C) was significantly decreased (mmol/L: 1.12±0.33 vs. 1.35±0.36, P < 0.01), and the proportion of hypertension, smoking and drinking were significantly increased (69.0% vs. 41.9%, 23.1% vs. 16.8%, 19.2% vs. 13.4%, all P < 0.01). By assigning values to each factor [IS: No = 0,Yes = 1; BMI: < 24.0 kg/m2 = 0, ≥ 24.0 kg/m2 = 1; FPG: < 7.0 mmol/L = 0, ≥ 7.0 mmol/L = 1; TG: < 2.26 mmol/L = 0, ≥ 2.26 mmol/L = 1; TC: < 6.22 mmol/L = 0, ≥ 6.22 mmol/L = 1; LDL-C: < 4.14 mmol/L = 0, ≥ 4.14 mmol/L = 1; HDL-C: < 1.04 mmol/L = 0, ≥ 1.04 mmol/L = 1; hypertension: No = 0,Yes = 1; smoking: No = 0,Yes = 1; drinking: No = 0,Yes = 1], a classification tree model was established to analyze the risk factors of IS. The classification tree model consisted of 4 layers and 17 nodes: the first layer was hypertension, the second layer was FPG and HDL-C, the third layer was HDL-C and FPG, and the fourth layer was LDL-C and smoking. There were five explanatory variables screened out in the model, including hypertension, FPG, HDL-C, LDL-C and smoking. The first layer of the tree showed that the incidence of IS in hypertensive population (62.6%) was significantly higher than that in non-hypertensive population (35.2%). The second layer of the tree showed that the incidence of IS in people with hypertension with HDL-C ≥ 1.04 mmol/L (53.6%) was lower than that in people with HDL-C < 1.04 mmol/L (78.5%). However, in the population without hypertension, the probability of IS occurrence in the population with FPG ≥ 7.0 mmol/L (71.1%) was significantly higher than that in the population with FPG < 7.0 mmol/L (28.3%). The third layer of the tree showed that the IS incidence of HDL-C ≥ 1.04 mmol/L (21.8%) was lower than that of HDL-C < 1.04 mmol/L (48.7%) in the population without hypertension and FPG < 7.0 mmol/L. However, in the population with hypertension and HDL-C ≥ 1.04 mmol/L, the probability of IS occurrence in the population with FPG ≥ 7.0 mmol/L (78.6%) was significantly higher than that in the population with FPG < 7.0 mmol/L (46.7%). The fourth layer of the tree showed that the IS incidence of people with LDL-C ≥ 4.14 mmol/L (53.8%) was higher than that of people with LDL-C < 4.14 mmol/L (19.0%) in the population without hypertension, FPG < 7.0 mmol/L and HDL-C ≥ 1.04 mmol/L. In the population without hypertension, the incidence of IS in smokers (76.9%) was higher than that in non-smokers (39.1%) of people with FPG < 7.0 mmol/L and HDL-C < 1.04 mmol/L. In the population with hypertension, the probability of IS occurrence in the population with LDL-C ≥ 4.14 mmol/L (72.5%) was higher than that in the population with LDL-C < 4.14 mmol/L (44.4 %) of people with HDL-C ≥ 1.04 mmol/L and FPG < 7.0 mmol/L. The gain diagram of IS classification tree model shown that the gain value increased rapidly from 0% to 100% and then tended to be stable. The index chart shown that the index value kept stable in the moving direction from above 100% and then dropped rapidly to 100%, indicating the model was very well. The risk value of misclassification probability of the classification tree model was 0.291, and the correct rate of risk factor for IS patients was 70.90%. The area under ROC curve (AUC) was 78.0% [95% confidence interval (95%CI) = 75.9%-79.9%, P < 0.001], the sensitivity was 62.5% (95%CI = 59.1%-65.7%) and the specificity was 79.4% (95%CI = 76.5%-82.1%).
CONCLUSIONS: Classification tree model can properly predict the risk factor of IS, and the most important risk factors are hypertension, hyperglycemia, high LDL-C and smoking.
参考来源
- Yao S, Li H, Liu K, et al. [Study on the application of classification tree model in screening the risk factors of ischemic stroke]. Zhonghua Wei Zhong Bing Ji Jiu Yi Xue. 2018. PMID: 30439320. doi:10.3760/cma.j.issn.2095-4352.2018.010.014
供执业医师抢救与诊疗参考,不替代临床判断;剂量请结合体重、肝肾功能和本院药品规格复核,以指南原文为准。