Predictive Model Study On Prognosis And Serious Complications Of Acute Ischemic Stroke Based On Interpretable Machine Learning

Stroke is one of the major diseases that cause death and disability worldwide;nearly 5.5 million people died from this disease in 20216.The mortality rate of stroke patients in China is higher than in Western countries such as Britain and America,and it is the first cause of death and disability in domestic adults.Acute ischemic stroke(AIS)refers to the sudden decrease or stop of the blood flow of the local vascular supply area of the cerebral tissue,resulting in ischemia and hypoxia of the cerebral tissue in the responsible vascular supply area,leading to the necrosis and softening of the cerebral tissue,and accompanied by the clinical symptoms and signs of the corresponding parts,such as hemiplegia,aphasia,and other neurological deficits.Timely and accurate prognosis prediction is of great significance for guiding clinical treatment decisions and the functional recovery of patients after AIS.Therefore,it is of great clinical significance to provide an interpretable and robust prediction model for the prognosis of AIS patients,evaluate the long-term prognosis and the probability of serious complications in the short term of patients in the early stage,guide and assist clinicians in making individualized treatment decisions for AIS patients,and timely conduct an early active intervention to improve the prognosis of them.In this study,AIS patients,who underwent MT and successful recanalization was achieved,was taken as the research object and clinical information and image quantitative features(manually quantitative features and radiomic features)were included.The prediction targets were futile recanalization(FR)and serious complications,including malignant cerebral edema(MCE)and cerebral hernia(CH).To construct a prediction model of prognosis and serious complications of AIS based on machine learning,and focus on the intserpretability of the output results of the models aiming to provide clinicians with an accurate and reliable auxiliary prediction tool to improve the prognosis of patients with AIS.Materials and Methods1.Research ObjectThe study recruited 110 patients with confirmed AIS and large vessel anterior circulation occlusion who underwent MT and successful recanalization was achieved modified Thrombolysis in Cerebral Infarction(mTICI)score 2b-3,in the Department of Neurology at Nanfang Hospital between June 2016 and November 2019.The clinical data and the first non-contrast CT image within 24 hours after the Mechanical Thrombectomy(MT)of patients were collected.2.Clinical Data(1)General information;(2)90-day mRS score,MCE,and CH;(3)NIHSS score and GCS score at admission;(4)Infarction site and TOAST classification;(5)Multiple points of time and treatment modes from the onset of stroke to the completion of MT;(6)Blood biochemical examination results before and after MT.3.Imaging Data(1)The hyperdensity volume and its proportion in the responsible blood supply area;(2)The location,volume,maximum cross-sectional area and proportion in the responsible blood supply area of the hyperdensity;(3)ASPECTS;(4)Radiomic features of infarction.4.Statistics MethodsThe statistical analysis of data in this study was completed by SPSS 25.0 and R Studio 4.0.3.Normal distribution data were expressed by mean ± standard deviation,and inter-group comparison was performed by students t-test;Non-normal distribution data were expressed by the median(interquartile interval),and inter-group comparison was conducted by Mann-Whitney U test The grade data was expressed by counting(percentage),and the inter-group comparison was performed by Chi-square test or Fisher’s exact probability method.The performance of each classification model was evaluated by the area under the curve of receiver operating characteristic(AUROC),decision curve analysis(DCA),sensitivity,specificity and accuracy,and the AUC of the ensemble machine learning model and each base model were compared by Delong test.Then compare the ensemble machine learning model based on clinicalimaging fusion features with the ensemble machine learning model based on clinical features constructed in the first part All the tests were two-sided,and statistical significance was set at p<0.05.5.Models ConstructionThe dataset were randomly divided into training sets and test sets according to the ratio of 7:3.After the completion of data preprocessing,based on clinical features and clinical-image fusion features respectively,combined with multiple machine learning algorithms to construct base models,and then fused them into an optimal ensemble machine learning model through the ensemble algorithm,to construct an interpretable risk prediction model of FR,MCE,and CH.The algorithms used in this study include support vector machine(SVM),random forest(RF),XGBoost(Extreme Gradient Boosting),K-Nearest Neighbor(KNN)classification algorithm,and Gradient Boosting Machine(GBM).Finally,the prediction model with excellent performance was fused using the Stacking method The feature contribution degree in the fusion model was calculated using the Shapley Additive Expansion(SHAP)algorithm.ResultsPart Ⅰ:Construction of prediction model for prognosis and serious complications of acute ischemic stroke based on clinical baseline characteristics.1.For AIS patients who have undergone mechanical embolectomy(MT)and achieved successful recanalization,the NIHSS score and GCS score at admission and the postoperative D-dimer level are related to the occurrence of FR.Age,the NIHSS score and GCS score at admission,TOAST,postoperative D-dimer,white blood cell and neutrophil count were related to the occurrence of MCE;The time from the onset to femoral artery puncture,and the postoperative D-dimer,white blood cell and neutrophil counts were related to the occurrence of CH.2.In the three prediction tasks of FR,MCE,and CH,the AUC values of the ensemble machine learning model based on clinical characteristics,LR-Stacking model,were 0.802,0.765,and 0.797,respectively.Part Ⅱ:Construction and evaluation of prediction model for prognosis and serious complications of acute ischemic stroke based on clinical-image fusion features1.For AIS patients who have undergone MT and achieved successful recanalization,the volume and their proportion in the responsible blood supply area of hypodensity and hyperdensity,the maximum area of hyperdensity,hyperdensity in the cerebral parenchyma or subarachnoid space,ASPECTS score and Radiomic score(Rad-score)are related to the occurrence of FR in the quantitative image features extracted from the first CT image within 24 hours after MT.The volume and their proportion in the responsible blood supply area of hypodensity and hyperdensity,the maximum area of hyperdensity,the appearance of hyperdensity in the subarachnoid space,the ASPECTS score at admission and after MT,and Rad-score is related to the occurrence of MCE.The volume and proportion in the responsible blood supply area of hypodensity and hyperdensity,the maximum area of hyperdensity,the ASPECTS score at admission and after MT,and the Rad-score are related to the occurrence of CH.2.In the three prediction tasks of FR,MCE,and CH,the AUC of the LR-Stacking model based on clinical and image fusion features were 0.938,0.889,and 0.986,respectively.Considering the results of the Delong test and DCA analysis,the performance and overall benefit of the LR-Stacking model based on clinical and image fusion features are higher than that of the LR-Stacking model based on clinical features.3.The results of SHAP analysis showed that the hypodensity volume and its proportion in the responsible blood supply area were significant risk factors affecting the poor prognosis of patients and had a strong predictive effect on FR,MCE,and CH.The text shows the importance and influence of other clinical features and quantitative imaging features on the prediction of poor prognosis and serious complications of the model and suggests the individualized high-risk factors of patients through case analysis.ConclusionThis study is aimed at AIS patients who have undergone MT and achieved successful recanalization.Through univariate analysis,we found the high-risk risk factors of poor prognosis and serious complications of the patients.Combined with clinical characteristics and quantitative characteristics of CT images,multiple interpretable ensemble machine learning prediction models were constructed,which can well complete the prediction tasks of FR,MCE,and CH.It can help clinical early identify high-risk populations with poor prognosis and the individualized risk favtors of each patient are given to benefit the corresponding AIS patients,which has good clinical promotion value.