Decision Tree Risk Prediction Model Of Hemorrhagic Transformation In Patients With Acute Ischemic Stroke

Objectives The decision tree algorithm is used to predict the risk of hemorrhagic transformation in patients with acute ischemic stroke,to find out the decision factors and to intervene the decision factors according to the decision path to reduce the occurrence of hemorrhagic transformation and improve the prognosis of the patients.Methods Patients with acute ischemic stroke admitted to the Affiliated Hospital of North China University of Science and Technology and Tangshan Municipal Workers' Hospital from January 2012 to June 2017 were enrolled.Hemorrhagic transformation group(HT group)and non-hemorrhagic transformation group(non-HT group)were divided according to whether hemorrhagic transformation occurred within 2 weeks after admission.Retrospectively collected patients' case information,including general conditions,past disease history,physical examination data,physio-biochemical examination data,imaging examination data,and treatment plan.Parameters with statistical significance in univariate analyses were selected as input variables combined with previous research results.Logistic regression model,RBF neural network model,and CART,QUEST,and C5.0 decision tree models were established with the ratio of training set and test set 7:3,and the prediction performance of the model was compared.Results 1 Univariate analysis showed that patients with hemorrhagic transformation had higher proportions of hypertension,diabetes,atrial fibrillation,history of cerebral infarction,history of antiplatelet drugs,large area cerebral infarction,leukoaraiosis,early low-density CT,thrombolytic therapy and anticoagulant therapy;in addition,their NIHSS score,leukocyte and PT-INR levels were significantly higher than those in the non-hemorrhagic transformation group,while albumin and triglyceride levels were significantly lower.2 Multivariate analysis showed that hypertension,diabetes,atrial fibrillation,history of cerebral infarction,NIHSS score,large area cerebral infarction,thrombolytic therapy,and leukocyte were risk factors for hemorrhagic transformation,while antiplatelet therapy and triglycerides were protective factors for hemorrhagic transformation.3 The five risk prediction models were used to predict 314 training set samples and 146 test set samples.Accuracy rates of Logistic regression model were 78.3% and 69.2%,sensitivity were 75.8% and 75.4%,specificity were 81.0% and 63.6%,Kappa index were 0.567 and 0.387,and AUC were 0.784 and 0.695,respectively.Accuracy rates of RBF neural network model were 72.6% and 74.7%,sensitivity were 87.6% and 88.4%,specificity were 56.9% and 62.3%,Kappa index were 0.448 and 0.500,and AUC were 0.719 and 0.754,respectively.Accuracy rates of CART decision tree model were 70.7% and 72.6%,sensitivity were 70.2% and 76.8%,specificity were 71.2% and 68.8%,Kappa index were 0.414 and 0.454,and AUC were 0.707 and 0.728,respectively.Accuracy rates of QUEST decision tree model were 76.8% and 76.0%,sensitivity were 85.1% and 84.1%,specificity were 68.0% and 68.8%,Kappa index were 0.533 and 0.524,and AUC were 0.765 and 0.764,respectively.Accuracy rates of C5.0 decision tree model were 96.5% and 80.1%,sensitivity were 98.1% and 82.6%,specificity were 94.8% and 77.9%,Kappa index were 0.930 and 0.603,and AUC were 0.965 and 0.803,respectively.4 In the training set,the AUC of C5.0 decision tree model was significantly higher than the other four models;in the test set,the AUC of C5.0 decision tree model was higher than QUEST decision tree model,CART decision tree model and Logistic regression model,but had no significant difference with RBF neural network models.Therefore,the prediction performance of C5.0 decision tree model was superior to Logistic regression model,neural network model,and CART and QUEST decision tree models,which was the optimal risk prediction model.5 C5.0 decision tree model generated 14 decision paths.The decision factors were NIHSS score,PT-INR,triglyceride,history of cerebral infarction,and antiplatelet therapy,etc,among which the NIHSS score was the most important decision factor.Conclusions 1 Hypertension,diabetes,atrial fibrillation,history of cerebral infarction,NIHSS score,large area cerebral infarction,thrombolytic therapy,and leukocyte were risk factors for hemorrhagic transformation,while antiplatelet therapy and triglycerides were protective factors for hemorrhagic transformation 2 The prediction performance of C5.0 decision tree model was superior to Logistic regression model,neural network model,and CART and QUEST decision tree models,which was the optimal risk prediction model.3 C5.0 decision tree model generated 14 decision paths.The decision factors were NIHSS score,PT-INR,triglyceride,history of cerebral infarction,and antiplatelet therapy,etc,among which the NIHSS score was the most important decision factor.Secondary prevention and treatment of hemorrhagic transformation in patients with acute ischemic stroke can be guided by decision factors and decision paths of C5.0 decision tree model.