机器学习预测出血性脑卒中功能预后  

Machine Learning Predicts Prognosis of Intracerebral Hemorrhage

作　　者：陈琳 CHEN Lin(School of Mathematics and Statistics,Chongqing Technology and Business University,Chongqing 400067,China)

机构地区：[1]重庆工商大学数学与统计学院,重庆400067

出　　处：《数学建模及其应用》2024年第4期53-63,共11页Mathematical Modeling and Its Applications

摘　　要：出血性脑卒中是中国居民死亡和残疾的主要原因之一,预测功能预后并识别其影响因素,对临床治疗有指导意义.本文利用出血性脑卒中高维小样本数据,建立机器学习模型预测患者的预后三级分类:在22项临床指标上使用3种特征选择方法(US、 RFE-RF、 RFE-NB),在84项影像指标上使用3种降维方法(PCA、 MDS、 UMAP),比较不同特征选择和降维组合下7种机器学习分类器(LASSO/Ridge/ENet惩罚三项逻辑回归、 SVM、 RF、 XGBoost、 ANN)的预测表现(Accuracy、 F1、 Kappa、 HUM);然后使用Wilcoxon符号秩检验比较分类器的性能差异.结果发现:不考虑特征选择和降维时(基准),RF准确率最高;降维效果MDS最优,然后PCA优于UMAP;特征选择效果RFE-RF优于US和RFE-NB;LASSO和XGBoost经过特征选择或降维后预测准确率多数表现为上升;对优于基准RF准确率的组合进行Wilcoxon符号秩检验,UMAP+US+LASSO的Macro-F1和Weighted-F1优于基准RF,MDS23+SVM、 US+SVM、 RFE-RF+SVM和RFE-NB+SVM的HUM优于基准RF;临床数据3种特征选择方法共同选择的变量有低压、高低压比和糖尿病史.Intracerebral Hemorrhage(ICH) is one of the main causes of death and disability in Chinese residents.To predict the functional prognosis and identify its influencing factors is of guiding significance for clinical treatment.In this paper,seven machine learning models were established to predict the three-level prognosis of patient by using high-dimensional and small-sample ICH data.Three feature selection methods(US,RFE-RF,RFE-NB)were used on 22 clinical indicators,and three dimensionality reduction methods(PCA,MDS,UMAP)were used on 84 imaging indicators.Compare the prediction performance(Accuracy,F1,Kappa,HUM)of different combinations of feature selection and dimensionality reduction for seven machine learning classifiers(LASSO/Ridge/ENet penalized trinomial logit models,SVM,RF,XGBoost,ANN).Then,the Wilcoxon signed-rank test is used to compare classifier performance differences of the classifiers.Experiment results show that the RF is the highest accuracy without considering feature selection and dimensionality reduction(benchmark).In terms of the number of better than the benchmark RF accuracy,the effectiveness of dimensionality reduction was ranked as MDS,PCA and UMAP,and RFE-RF outperformed US and RFE-NB.The prediction accuracy of LASSO and XGBoost is mainly increased after feature selection or dimensionality reduction.The Wilcoxon signed rank test was performed on the combinations that outperformed the benchmark RF prediction performance.The Macro-F1 and Weighted-F1 of UMAP+US+LASSO were superior to the benchmark RF.The HUM of MDS23+SVM,US+SVM,RFE-RF+SVM and RFE-NB+SVM is better than that of the benchmark RF.The common variables selected by the three feature selection methods of clinical data are low pressure,high-low pressure ratio and diabetes history.

关 键 词：出血性脑卒中 预后 机器学习 惩罚三项逻辑回归 特征选择 降维 

分 类 号：O29[理学—应用数学]