Abstract:

Objective To establish a 1-year functional outcome prediction model for new-onset acute ischemic stroke (AIS) patients based on machine learning algorithms, to provide reference for related research and clinical work. 

Methods This study was based on the data of new-onset AIS patients from China national stroke registry (CNSR) database. Based on machine learning [CatBoost model, XGBoost model, GBDT model, randomized forest model] and traditional logistic regression model, the 1-year poor prognosis (mRS≥3) prediction models for new-onset AIS patients were constructed. According to the ratio of 7:3, the patients were randomly divided into training set and test set. The training set was used for model training and parameter optimization, and the test set was used to evaluate the prediction value of the models. The evaluation indicators were mainly the AUC in the discrimination index and the Brier score in the calibration index. 

Results A total of 8230 eligible patients were included, with a mean age of 64.4±12.8 years old and 3113 females (38.7%), and 2360 patients with 1-year poor prognosis. Multivariate analysis showed that aging, female, mRS≥3 before stroke onset, NIHSS score at admission and discharge, limb dysfunction, history of peripheral vascular disease, blood glucose at admission, blood lipidregulating drugs (with medications at discharge) , antiplatelet drugs (1-year medication compliance) were predictors for 1-year poor prognosis. The AUC of Catboost, XGBoost, GBDT, random forest and logistic regression models for predicting 1-year functional prognosis of new-onset AIS patients were 0.857 (0.850-0.864), 0.856 (0.850-0.863), 0.856 (0.848-0.864), 0.853 (0.846-0.859) and 0.846 (0.837-0.855), respectively. The prediction performance of machine learning-based prediction models were all superior than that of logistic regression model (Catboost vs . logistic, P =0.0130, XGBoost vs . logistic, P =0.0133, GBDT vs . logistic, P =0.0229, random forest vs . logistic, P =0.0429), and the calibration of each model was good. 

Conclusions The 1-year functional prognosis prediction models of new-onset AIS patients based on machine learning algorithm had high predictive value, and the Catboost model has the best prediction effect.

Key words: Ischemic stroke; 1-year functional prognosis; Prediction model; Machine learning

摘要：

目的 建立基于机器学习算法的新发急性缺血性卒中（acute ischemic stroke，AIS）患者1年预后的预测模型，为相关研究和临床工作提供借鉴。 

方法 以中国国家卒中登记（China national stoke registry，CNSR）数据库中的新发AIS患者为研究对象，通过logistic回归模型确定进入模型的预测因子，分别基于机器学习[CatBoost模型、XGBoost模型、梯度提升决策树（gradient boosted decision trees，GBDT）模型、随机森林模型]和传统logistic回归模型构建新发AIS患者1年预后不良（mRS≥3分）的预测模型。将入组患者按照7∶3的比例随机分为训练集和测试集。训练集用于模型训练和参数优化，测试集用于评价模型效果。评价各模型预测预后不良效能 的区分度指标主要是AUC，校准度指标为Brier分数。 

结果 共纳入8230例新发AIS患者，平均64.4±12.8岁，其中女性3113例，1年预后不良患者2360 例。针对训练集的logistic回归结果显示，年龄、女性、卒中前mRS≥3分、入院和出院时NIHSS、肢体运 动功能障碍、周围血管病史、入院血糖、调节血脂药物（出院带药）、抗血小板药（1年服药依从）可 作为1年预后不良的预测因子。Catboost、XGBoost、GBDT、随机森林及logistic回归模型预测新发AIS患者1年功能预后的AUC分别为0.857（0.850～0.864）、0.856（0.850～0.863）、0.856（0.848～0.864）、 0.853（0.846～0.859）和0.846（0.837～0.855）。Catboost（P =0.0130）、XGBoost（P =0.0133）、GBDT （P =0.0229）和随机森林（P =0.0429）等机器学习模型的预测效能均优于logistic回归模型；所有预测模型的校准度均良好。 

结论 基于机器学习算法建立的新发AIS患者1年功能预后预测模型具有较高的预测价值,其中 Catboost模型的预测效果最好。

关键词: 缺血性卒中; 1年功能预后; 预测模型; 机器学习