Construction and comparison of five risk prediction models for delay in seeking medical care in patients with acute ischemic stroke based on machine learning

Yu Chuanshan; Zuo Xia; Ni Tingting; Liu Xingyu; Yu Hailong; Gu Zhie

doi:10.3969/j.issn.1009-6574.2026.03.003

Abstract：Objective To construct and compare five risk prediction models for delay in seeking medical care in patients with acute ischemic stroke(AIS)based machine learning,so as to provide support for early identification of high-risk patients in clinical practice.Methods Convenience sampling was used to select 380 AIS patients who visited the Emergency Department of Northern Jiangsu People's Hospital from November 2023 to March 2024 as study subjects.Using stratified sampling,patients were randomly divided into a training set(n=266)and a validation set(n=114)in a 7∶3 ratio.An additional 120 AIS patients who visited the Emergency Department of Northern Jiangsu People's Hospital from April to June 2024 were selected as the test set.Missing values were imputed using the mean method and random interpolation method.Single-factor analysis was employed to examine the factors influencing delay in seeking medical care.Feature selection was performed using the least absolute shrinkage and selection operator(LASSO)regression analysis.Data imbalance was addressed using a linear function normalization method,and five machine learning models were constructed,including Logistic regression,random forest,gradient boosting tree,bootstrap aggvegating(Bagging),and extreme gradient boosting(XGBoost).The correlation between variables was analyzed by drawing a heat map.Model performance was evaluated based on the area under the receiver operating characteristic curve(AUC),accuracy,sensitivity,specificity,F1 score,and decision curve.A variable feature importance map was generated for the optimal model.Results Single-factor and LASSO regression analysis of the training set collectively identified 16 risk factors.The heat map results indicated that correlations among variables were generally weak and did not exhibit significant multicollinearity.The AUC of the five machine learning models ranged from 0.825 to 0.945,with sensitivities between 0.810 and 0.967,specificities from 0.842 to 0.991,accuracies between 0.765 and 0.959,and F1 scores from 0.785 to 0.964.Among these,the XGBoost model emerged as the optimal model,achieving AUC of 0.946 for the validation set and 0.931 for the test set.The decision curve analysis further demonstrated its superior clinical net benefit across most threshold intervals,indicating robust stability.The top 10 risk factors ranked by importance were as follows:Stroke Pre-Hospital Delay Behavior Intention Scale scores,whether there were others around the patient at the time of onset,level of stroke cognition,smoothness of hospital green channels,mode of transport,time of onset,whether the patient knew to use the balance-eyes-face-arm-speech-time(BEFAST)method for stroke recognition,whether the patient experienced transfer or referral,monthly per capita household income and whether family members knew to use BEFAST for stroke recognition.Conclusions Among five machine learning algorithms,the XGBoost model demonstrates optimal predictive performance,aiding healthcare professionals in early identification of patients at high risk for delay in seeking medical care and providing a scientific basis for formulating precise intervention strategies.