Prediction Of Hemorrhagic Transformation After Intravenous Thrombolysis In Acute Ischemic Stroke Based On Machine Learning

Background and objective:Intravenous recombinant tissue plasminogen activator improves outset in selected patients with acute ischemic stroke within 4.5 hours from onset.Symptomatic intracranial hemorrhage is one of the serious complications of intravenous rt-PA,and it is the main cause of neurological deterioration and even death.In this study,we explored the feasibility of XGBboost algorithm and residual neural network algorithm to predict the hemorrhagic transformation of patients after thrombolysis,and to identify the risk factors of hemorrhagic transformation,so as to provide evidence-based medicine for evaluating the risk of thrombolysis in patients.Methods:We retrospectively identified 603 patients who treated with intravenous thrombolysis therapy in the First Hospital of Jilin University from July 2015 to May 2018.There were 65(19.4%)patients with hemorrhagic transformation,486(80.6%)patients without hemorrhagic transformation.We select 20 patients respectively as test set.We upsampled patients with and without hemorrhagic transformation to a ratio of 2:1 to compensate for the imbalance of samples.We construct a three-layer convolutional neural network structure training model.XGBoost model defines a series of hyperparameters,including learning rate,leaf weight sum,maximum depth of tree,gamma value,L1 regularization parameters and L2 regularization parameters,and automatically optimizes them by Grid Search CV to find the optimal solution.The sensitivity,specificity,positive predictive value,negative predictive value,receiver operating characteristic(ROC)and area under the curves(AUC)were performed to evaluate the results.Results:1.The sensitivity,specificity,positive predictive value,nagative predictive value and accuracy of the residual neural network for predicting hemorrhagic transformation after intravenous thrombolysis were 85%?70%?73.9%?73.9%?82.3% and 77.5%,respectively.The AUC was 0.82.2.The sensitivity,specificity,positive predictive value,nagative predictive value and accuracy of the XGBoost for predicting hemorrhagic transformation after intravenous thrombolysis were 85%?90%?89.5%?85.7% and 87.5%,respectively.The AUC was 0.97.3.According to the importance of characteristic values to the classification results,Xgboost algorithm selected the three key characteristics,including baseline NIHSS,anti-platelet drugs and history of stroke.Conclusions:1.Compared with the residual network algorithm,the XGBoost algorithm can more accurately predict hemorrhagic transformation.The XGBoost machine learning algorithm is suitable for baseline screening of patients and has a certain guiding role in clinical thrombolytic treatment decision making.2.Xgboost algorithm found that baseline NIHSS,anti-platelet drugs and history of stroke were associated with hemorrhagic transformation.