关键词: 缺血性卒中, 溶栓药物, 人工智能算法, Wide&Deep模型, 精准治疗

Abstract: Background The thrombolytic effect for ischemic stroke (IS) is affected by complex factors, such as acute onset of stroke, short therapeutic time window, various individual patient factors, treatment model, types and doses of medicines as well as mode of administration. To identify the influencing factors of thrombolytic effect, most existing studies adopt statistical methods, while rare studies use artificial intelligence (AI) -based algorithms.Objective To establish models using AI-based algorithms for IS patients based on the real-world data including general patient characteristics, medication model and recovery effects, to achieve precise individualized thrombolytic treatment and provide data support for clinical prescription decisions.Methods A retrospective design was used. The clinical information of IS patients (n=55 621) was extracted from the Yidu Cloud scientific research big data server system of the Second Affiliated Hospital of Dalian Medical University from January 1, 2001 to December 31, 2021, among whom 1 855 with complete information were enrolled according to the inclusion criteria. Thrombolysis effect was evaluated by comparing the National Institutes of Health Stroke Scale (NIHSS) score measured at admission and discharge, and those with an improvement in the NIHSS score by ≥4 points and <4 points were assigned to neurological improvement group (n=1 236) , and control group (n=619) , respectively. Factors possibly associated with post-IS thrombolytic effect (including general patient characteristics, medication indicators, examination indicators, test indicators, and treatment methods) were obtained by summarizing the factors suggested separately by three neurology experts with a senior title, and reviewing relevant guidelines and literature, then were screened using univariate analysis, and the identified ones were treated by dimensionality reduction using principal component analysis (PCA) . Models of Logistic, support vector machine (SVM) , C5.0 decision tree arithmetic, classification and regression tree (CART) , deep neural network (DNN) , and Wide&Deep, were built and compared to find the one with the best performance in predicting thrombolytic effect, then to determine its parameters. Then by use of two randomly generated two numbers, 7 and 11, the 1 855 patients were randomly assigned to three datasets, training (n=1 113, for building and practicing models to discover rules) , validation (n=371, for adjusting model parameters) , and test (n=371, for evaluating the generalization ability of the final model) . Feature engineering was used to construct a simplified model and evaluate its accuracy. The clinical information of IS patients (n=3 925) was extracted from the Yidu Cloud scientific research big data server system of Dalian Central Hospital for external verification of the model.