基于机器学习的原发性高血压心血管风险预后模型

Prognostic model of cardiovascular risk in essential hypertension based on machine learning

目的筛选原发性高血压心血管风险的中西医预后危险因素,构建基于血管功能、中医证候的原发性高血压预后的最优预测模型。方法以前期建立的高血压队列人群中原发性高血压患者为研究对象,采集人口社会学资料、病情病史特征、实验室指标、血管功能及心功能检查指标、中医证候等相关基线指标,随访心血管风险的发生情况。使用比例风险模型单因素、多因素分析及共线性诊断初步确定高血压心血管风险的中西医预后模型的纳入变量。筛选的病例按7∶3的比例随机划分为训练集和测试集,基于训练集使用随机森林算法建立原发性高血压的预后预测模型,利用测试集评价预后模型的预测效能,分别使用决策树、随机森林、支持向量机和人工神经网络算法建立原发性高血压的预后预测模型,以测试集评估并对比4种预后预测模型的预测效能,评估并建立具有较好预测能力的模型。结果纳入985例病例中有284例出现心血管风险。COX回归单因素、多因素分析及共线性诊断确定18个变量纳入预后模型变量。变量包括:一般资料〔病程、性别、早发心血管病家族史、体重指数(BMI)、饮食习惯〕、实验室指标〔糖代谢异常、脂代谢异常、血同型半胱氨酸(Hcy)〕、血管功能指标〔平均踝臂压指数(ABI)、平均动脉压、颈股脉搏波传导速度(cfPWV)、血流介导的血管舒张功能(FMD)〕、中医证候(头晕、头痛、气虚血瘀证、阴虚阳亢证、肝肾阴虚证、阴阳两虚证)。基于这18个建模变量,分别通过决策树、随机森林、支持向量机和人工神经网络算法构建4个原发性高血压的预后预测模型。通过混淆矩阵评估4种算法对训练集的数据解析能力,发现基于相同变量的情况下,人工BP神经网络的错误率最低(19.1%),其次为支持向量机(24.2%),决策树(28.7%)和随机森林(28.7%)并列最差。测试集带入模型,支持向量机错误率最低(26.5%),其次为随机森林(28.2%)和决策树(28.9%),错误率最高的是人工BP神经网络(30.9%)。因而,基于相同变量情况下,4种模型中支持向量机的预测效能最好,其次为随机森林和决策树,预测效能最差的是人工BP神经网络。结论基于血管功能、中医证候构建原发性高血压模型心血管风险预后模型有较好的应用,使用机器学习可以对高血压心血管风险进行初步判定。该模型构建中支持向量机的预测效能较好。

Objective To screen the prognostic risk factors of traditional Chinese and Western medicine for the cardiovascular risk of essential hypertension(EH),evaluate and establish a prognosis prediction model for EH,with a view to construct the optimal prognosis of EH based on machine learning of vascular function and traditional Chinese medicine(TCM)syndromes.Methods The previously established hypertension cohort with EH patients were selected,and the demographic and sociological data,medical history and characteristics,laboratory indicators,vascular function and cardiac function examination indicators were obtained.TCM syndromes and other relevant baselines were collected.The follow-up was conducted to observe the cardiovascular risk.The proportional hazard model single factor analysis,multivariate analysis and collinearity diagnosis were used to initially determine the included variables of the prognosis model of Chinese and Western medicine for the cardiovascular risk of EH.The selected cases were randomly divided into training set and test set according to the ratio of 7∶3.Based on the training set,the random forest algorithm was used to establish the prognosis prediction model of EH,and the test set was used to evaluate the prediction performance of the prognosis model.Tree,random forest,support vector machine and artificial neural network algorithm were adopted to establish the prognosis prediction model of EH,evaluate and compare the prediction performance of the four prognosis prediction models with the test set,evaluate and establish a model with better prediction ability.Results 985 cases were completed and 284 patients had cardiovascular risk.The single factor,multifactor analysis and colinear diagnosis of COX regression determined 18 variables incorporated into the prognostic model variables.Variables include:general data(course,sex,family history of early cardiovascular disease,BMI),laboratory indicators(abnormal glucose metabolism,abnormal lipid metabolism,Hcy),vascular function indicators(average ABI,average arterial pressure,cfPWV,FMD),TCM syndrome(dizziness,headache,Qi deficiency blood stasis syndrome,Yin deficiency and Yang hyperactivity syndrome,liver and kidney Yin deficiency syndrome,Yin and Yang deficiency syndrome).Based on these 18 modeling variables,four prognostic models of EH were constructed by decision tree,random forest,support vector machine and artificial neural network algorithm.The confusing the matrix was used to evaluate the data resolution of the training set of four algorithms,the error rate of artificial BP neural networks was the lowest(19.1%)based on the same variable,followed by that of support vector machines(24.2%);that of decision tree and random forests were the worst(both 28.7%).The test set was introduced into the model,with the lowest error rate of the support vector machine(26.5%),followed by random forests(28.2%)and decision trees(28.9%),and the highest error rate was the artificial BP neural network(30.9%).Therefore,based on the same variables,the prediction performance of support vector machines was best in the four models,followed by random forests and decision trees,and the worst prediction performance was artificial BP neural networks.Conclusions The cardiovascular risk prognostic model based on vascular function and TCM syndrome has a good application.Machine learning can be used to determine the cardiovascular risk of EH.The prediction performance of support vector machine is better in this model.