XGboost model in predicting readmission of patients with ischemic stroke recurrence within 90 d

Abstract

Objective To explore the predictive efficacy of XGboost model in predicting risk of relapse and re-admission within 90 d in patients with ischemic stroke, and provide basis for early screening and prevention of high-risk population with ischemic stroke. Methods The clinical data of 6070 primary ischemic stroke patients admitted to our hospital from January 2007 to July 2017 were retrospectively collected. XGboost model and multivariate Logistic regression model were utilized to screen out the influencing factors of relapse and re-admission within 90 d in patients with ischemic stroke. A predictive model was set up. Receiver operating characteristic (ROC) curve was drawn and compared. Sensitivity, specificity and Youden index were calculated and compared to evaluate the prediction performance of XGboost model. Results During the observation period, a total of 520 patients with relapsed ischemic stroke were observed within a period of 90 d, and the incidence density was 8.57%. Multivariate Logistic regression analysis showed that length of first hospital stay, hypertension, pulmonary infection, neutrophil percentage, red blood cell distribution width (variable coefficient), and alkaline phosphatase level were independent influencing factors for re-hospitalization within 90 d of ischemic stroke, (OR=1.016, P=0.000, 95%CI: 1.008-1.025; OR=4.598, P=0.000, 95%CI: 3.717-5.687; OR=1.452, P=0.025, 95%CI: 1.048-2.012; OR=1.013, P=0.006, 95%CI: 1.004-1.022; OR=1.161, P=0.000, 95%CI: 1.090-1.237; OR=1.003, P=0.023, 95%CI: 1.000-1.005). Analysis of importance of risk factors for re-admission of ischemic stroke using XGboost model showed that the top 6 factors were hypertension, red blood cell distribution width, direct bilirubin, length of hospital stay, pulmonary infection, and alkaline phosphatase, and the corresponding importance scores were 32, 20, 19, 18, 15 and 14, respectively. ROC curve analysis results indicated that the area under the ROC for re-admission for XGboost model was 0.792 (95%CI: 0.717-0.762), which was improved by 5% as compared with that for multivariate Logistic regression model (0.739 [95%CI: 0.764-0.818]). The sensitivity was 89.30% and the Youden index was 0.444 for XGboost model, which were significantly higher than those for multivariate Logistic regression model (77.3%, 0.405). Conclusions XGboost model is superior to multivariate Logistic regression model in predicting recurrence and re-admission of first ischemic stroke patients within 90 d. This model is suitable for prediction and early diagnosis of re-admission of ischemic stroke, which is of great clinical value. Key words: XGboost model; Multivariate Logistic regression model; Predicting model; Ischemic stroke; Readmission; Risk factor