基于多核SVM的AdaBoost心力衰竭死亡率评估模型

A Mortality Predicting Model for Heart Failure Patients Based on AdaBoost with Multi-kernel SVM

【目的】心力衰竭简称心衰,是一种复杂的临床综合征,具有高发病率、高死亡率和预后效果不佳等显著特点,是各类心脏疾病发展的终末期,严重危害人类健康。因此,对心衰患者进行早期的预后评估研究至关重要,可以最大程度地帮助患者生存。【方法】提出一种基于多核支持向量机(multi kernel support vector machine,MK-SVM)和自适应提升算法(adaptive boosting,AdaBoost)的心力衰竭死亡率评估模型(MK-SVM-AdaBoost).该算法利用MK-SVM将特征映射到高维空间,并依据AdaBoost算法将基本分类器进行集成,实现死亡率的精确预测。同时,将合成少数过采样技术(synthetic minority oversampling technique,SMOTE)和Tomek links欠采样技术相结合的混合抽样方法引入到预测模型中,减轻不平衡数据集对模型性能的影响。【结果】在收集于白求恩医院的小型心衰数据集上进行心衰患者30 d内死亡率预测实验。实验结果表明,MK-SVM-AdaBoost模型的准确率和召回率分别达到了85.63%和86.33%,优于现有方法,ROC曲线下与坐标轴围成的面积(area under curve,AUC)和其微观平均值(micro-mean AUC,MiA-AUC)分别达到了91.00%和92.00%,表明提出的模型具有良好的稳定性。【结论】提出的模型具有较高的准确率和稳定性,可以为医生的临床决策提供一定的参考。今后课题将继续对数据集进行扩充,并对分级预警进行研究,以便对患者进行更有效的评估。

【Purposes】Heart failure is a complex clinical syndrome with significant features such as high morbidity,high mortality,and poor prognosis.It is the terminal stage in the development of all types of heart disease and seriously threatens human health.Therefore,early prognostic assessment studies of heart failure patients are crucial to help the survival of patients.【Methods】A heart failure mortality assessment model(MK-SVM-AdaBoost)based on Multi Kernel Support Vector Machine(MK-SVM)and Adaptive Boosting(AdaBoost)algorithm is pro-posed.The algorithm utilizes MK-SVM to map features into a high-dimensional space and in-tegrates basic classifiers on the basis of the AdaBoost algorithm to achieve accurate mortality pre-diction.Meanwhile,a hybrid sampling method combining Synthetic Minority Oversampling Technique(SMOTE)and Tomek links under-sampling technique is introduced into the prediction model to alleviate the impact of unbalanced datasets on model performance.【Findings】Experi-ments were performed on a small heart failure dataset collected from Bethune Hospital for mor-tality prediction in heart failure patients within 30 days.The experimental results show that the accuracy and recall of the MK-SVM-AdaBoost model reach 85.63%and 86.33%,respectively,which are better than thase of the existing methods.The Area Under Curve(AUC)under the ROC curve enclosed with the axes and its micro-mean(MiA-AUC)reach 91.00%and 92.00%,respectively,which indicates that the proposed model has good stability.【Conclusions】The pro-posed model has high accuracy and stability,and can provide some reference for the clinical deci-sion-making of doctors.In the future,the dataset will be expaned and the graded warnings will be studied for more effective assessment of patients.