Research on functional prognosis prediction model of non-cardiac ischemic stroke based on machine learn-ing,thromboelastography and white matter lesions
10.3969/j.issn.1002-0152.2024.12.004
- VernacularTitle:基于机器学习联合血栓弹力图和脑白质病变的非心源性缺血性卒中功能预后预测模型研究
- Author:
Min XIA
1
;
Guoxiang HUANG
;
Jianli WANG
;
Nengwei YU
;
Daizong WU
Author Information
1. 成都中医药大学附属第三医院(西区)脑病科(成都 611730)
- Publication Type:Journal Article
- Keywords:
Non-cardiac ischemic stroke;
Thromboelastography;
White matter hyperintensities;
Prediction model;
Machine learning;
Neural network
- From:
Chinese Journal of Nervous and Mental Diseases
2024;50(12):726-734
- CountryChina
- Language:Chinese
-
Abstract:
Objective To explore the role and value of thromboelastography(TEG)combined with white matter hyperintensity(WMH)in predicting the functional prognosis of patients with non-cardiogenic acute ischemic stroke(AIS)through machine learning.Methods This study included 130 patients with non-cardiogenic AIS from August 2022 to February 2024.General clinical data,TEG and WMH information of all patients were collected.Three months later,functional outcomes were followed up using the modified Rankin scale(mRS),with an mRS score of≥2 indicating a poor prognosis.The prediction models were divided into four feature sets according to different ranges of predictors:set A(general clinical data+TEG indicators+WMH score),set B(general clinical data+TEG indicators),set C(general clinical data+WMH score),and set D(general clinical data).For each feature set,three machine learning algorithms,traditional logistic regression(LR)model,random forests(RF),neural network(NNET),and K-nearest neighbors(KNN),were used to construct models for predicting the 3-month neurological function outcome of patients with non-cardiogenic AIS.Bootstrap resampling internal validation was used to compare the performance of prediction models.Results The training and testing of the model were performed on 130 patient samples,and the AUC value and its confidence interval of the model were corrected by the 0.632+method(optimism correction).For the LR,NNET,and KNN models,the corrected AUC values of feature set A were significantly better than those of feature set D(DeLong test,P<0.05).For all models,the corrected AUC value of feature set A was higher than that of other feature sets.For feature set A,the corrected AUC value(0.830)of the NNET model was higher than that of other models.Among the 19 features of feature set A,six features with important associations with functional prognosis were selected including National Institute of Health stroke scale(NIHSS)score,stroke history,small artery occlusion subtype,periventricular white matter hyperintensities(PWMH)score,and TEG indicators maximum amplitude(MA)and LY30.Conclusion Combining TEG indicators and WMH information on the basis of general clinical data can significantly improve the accuracy of predicting poor functional prognosis in patients with non-cardiogenic AIS.The prediction models established by machine learning-based NNET and KNN algorithms have high predictive value.
