Prediction of Clinical Outcome of Acute Ischemic Stroke Patients with Hyperglycemia Based on Machine Learning Model

doi:10.3969/j.issn.1673-5765.2022.07.009

Abstract

Abstract:

Objective To establish the prediction models of prognosis of acute ischemic stroke patients with hyperglycemia based on machine learning, and to compare the prediction performance of traditional logistic model and machine learning model.

Methods This study included the patients from the China National Stroke Registration Ⅲ. The baseline information including patients' demographic characteristics, medical history, laboratory tests, head magnetic resonance imaging results and stroke etiology classification were collected case report forms. The cases were divided into the training set (3325 patients) and test set (369 patients) using stratified 10-fold cross-validation. Poor clinical outcome was defined as a modified Rankin score of 3-6 at 3-month follow-up. Machine learning methods such as random forest model, GBDT (Gradient Boosted Decision Trees) model, XGBoost (eXtreme Gradient Boosting) model, and traditional logistic model were used to construct the 3-month poor prognosis prediction models. The area under the receiver operating characteristic curve (AUC) was used to evaluate the degree of discrimination, and the Brier score was used to evaluate the degree of calibration.

Results A total of 3694 acute ischemic stroke patients with hyperglycemia were included, with an average age of 62.4±10.4 years and 2408 males (65.2%). There were 585 patients (15.8%) with poor prognosis at 3 months. The AUCs of logistic model, random forest model, GBDT model and XGBoost model for predicting 3-month poor prognosis were 0.843 (0.814-0.872), 0.847 (0.823-0.871), 0.845 (0.819-0.871), 0.848 (0.820-0.876), respectively. The sensitivity of logistic model, random forest model, GBDT model and XGBoost model were 0.373 (0.340-0.405), 0.679 (0.629-0.728), 0.426 (0.383-0.468), 0.634 (0.583-0.686), respectively. Although the AUC of the machine learning model was higher than that of the logistic model, the difference was not statistically significant (P>0.05). The sensitivity of the machine learning model was better than that of the logistic model (all P<0.05), and the calibration of all models were good (0.094-0.138).

Conclusions The traditional logistic model and machine learning model have high predictive value in predicting 3-month poor prognosis of acute ischemic stroke patients with hyperglycemia, and there is no significant difference in discrimination. The results of this study need to be validated in a larger sample size cohort.

Key words: Ischemic stroke; Functional prognosis; Machine learning; Prediction model

摘要：

目的建立基于机器学习的血糖异常急性缺血性卒中患者的预后预测模型，比较传统logistic模型与机器学习模型的预测效能。

方法以中国国家卒中登记研究Ⅲ（China national stroke registration study III，CNSR-Ⅲ）血糖异常急性缺血性卒中患者为研究对象，采用病例报告表收集患者的人口学信息、既往病史、实验室检查、头颅影像学检查、卒中病因分型等临床资料。采用分层10折交叉验证划分训练集（3325例）和测试集（369例），基于随机森林、梯度提升决策树（gradient boosted decision trees，GBDT）、极致梯度提升（eXtreme Gradient Boosting，XGBoost）等机器学习方法和传统logistic回归方法构建3个月功能预后不良（mRS≥3分）的预测模型。采用ROC的AUC评价效度，Brier分数评价校准度，同时综合F1分数、准确率、灵敏度及特异度等指标评价不同模型的预测效果。

结果共纳入3694例血糖异常的急性缺血性卒中患者，平均年龄62.4±10.4岁，男性2408例（65.2%），3个月预后不良585例（15.8%）。logistic回归、随机森林、GBDT和XGBoost模型预测患者3个月预后不良的AUC分别为0.843（0.814～0.872）、0.847（0.823～0.871）、0.845（0.819～0.871）、0.848（0.820～0.876），灵敏度分别为0.373（0.340～0.405）、0.679（0.629～0.728）、0.426（0.383～0.468）、0.634（0.583～0.686）。机器学习模型的AUC有高于logistic回归模型的趋势，但差异没有统计学意义，机器学习模型的灵敏度较传统logistic回归模型好（均P＜0.05），所有模型的Brier分数提示校准度均良好（0.094～0.138）。

结论传统logistic回归模型与机器学习模型对血糖异常的急性缺血性卒中患者3个月预后不良均有较高的预测价值，且区分度没有显著差异。本研究结果有待应用于更大样本量的队列进行验证。

关键词: 缺血性卒中; 功能预后; 机器学习; 预测模型

YANG Jialei, CHEN Siding, MENG Xia, JIANG Yong, WANG Yongjun. Prediction of Clinical Outcome of Acute Ischemic Stroke Patients with Hyperglycemia Based on Machine Learning Model[J]. Chinese Journal of Stroke, 2022, 17(07): 730-736.

杨佳蕾, 陈思玎, 孟霞, 姜勇, 王拥军. 基于机器学习预测血糖异常急性缺血性卒中患者预后模型研究[J]. 中国卒中杂志, 2022, 17(07): 730-736.

References

[1] MA Q F，LI R，WANG L J，et al. Temporal trend and attributable risk factors of stroke burden in China，1990–2019：an analysis for the global burden of disease study 2019[J/OL]. Lancet Public Health，2021，6（12）：e897-e906[2022-01-18]. https://doi.org/10.1016/s2468-2667（21）00228-0.

[2] 《中国卒中报告》编写委员会. 中国卒中报告2020（中文版）（1）[J]. 中国卒中杂志，2022，17（5）：433-447.

[3] WILLIAMS L S，ROTICH J，QI R，et al. Effects of admission hyperglycemia on mortality and costs in acute ischemic stroke[J]. Neurology，2002，59（1）：67-71.

[4] MATSUMOTO K，NOHARA Y，SOEJIMA H，et al. Stroke prognostic scores and data-driven prediction of clinical outcomes after acute ischemic stroke[J]. Stroke，2020，51（5）：1477-1483.

[5] RAJKOMAR A，DEAN J，KOHANE I. Machine learning in medicine[J]. N Engl J Med，2019，380（14）：1347-1358.

[6] WANG Y J，JING J，MENG X，et al. The third China national stroke registry（CNSR-Ⅲ）for patients with acute ischaemic stroke or transient ischaemic attack：design，rationale and baseline patient characteristics[J]. Stroke Vasc Neurol，2019，4（3）：158-164.

[7] POLDRACK R A，HUCKINS G，VAROQUAUX G. Establishment of best practices for evidence for prediction：a review[J]. JAMA psychiatry，2020，77（5）：534-540.

[8] LIN W J，CHEN J J. Class-imbalanced classifiers for high-dimensional data[J]. Brief Bioinform，2013，14（1）：13-26.

[9] FRIEDMAN J H. Greedy function approximation：a gradient boosting machine[J]. Annals of statistics，2001，29（5）：1189-1232.

[10] ZHANG Z D，JUNG C. GBDT-MO：Gradient-boosted decision trees for multiple outputs[J]. IEEE T Neur Net Lear，2020，32（7）：3156-3167.

[11] CHEN T，GUESTRIN C. XGBoost：a scalable tree boosting system[C/OL]//KDD’16：Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining，2016，San Francisco，CA，USA. New York：ACM[2022-01-18]. https://dl.acm.org/doi/pdf/10.1145/2939672.2939785.

[12] LUNDBERG S M，LEE S I. A unified approach to interpreting model predictions[C/OL]//NIPS'17: Proceedings of the 31st International Conference on Neural Information Processing Systems，2017，Long Beach，CA，USA[2022-01-18]. https://dl.acm.org/doi/pdf/10.5555/3295222.3295230.

[13] HEO J N，YOON J G，PARK H，et al. Machine learning-based model for prediction of outcomes in acute stroke[J]. Stroke，2019，50（5）：1263-1265.

[14] ADAMS H P，DAVIS P H，LEIRA E C，et al. Baseline NIH stroke scale score strongly predicts outcome after stroke：a report of the trial of org 10 172 in acute stroke treatment（TOAST）[J]. Neurology，1999，53（1）：126-131.

[15] WEIMAR C，KONIG I R，KRAYWINKEL K，et al. Age and national institutes of health stroke scale score within 6 hours after onset are accurate predictors of outcome after cerebral ischemia：development and external validation of prognostic models[J]. Stroke，2004，35（1）：158-162.

[16] KOIKE T，KOIKE Y，YANG D，et al. Human apolipoprotein A-Ⅱ reduces atherosclerosis in knock-in rabbits[J/OL]. Atherosclerosis，2021，316：32-40[2022-01-18]. https://doi.org/10.1016/j.atherosclerosis.2020.11.028.