Abstract WP038: Development Of Diagnostic Models For Anterior Circulation Large Vessel Occlusion Stroke Using Machine Learning Methods

BACKGROUND: The early diagnosis and identification of anterior circulation large vessel occlusion (LVO) stroke are important. The aim of this study is to develop diagnostic prediction models for anterior circulation LVO stroke using traditional methods and machine learning methods based on the Chinese population. METHODS: Acute ischemic stroke patients who presented to Beijing Tsinghua Changgung Hospital (China) within 24 hours of symptom onset and underwent cranial vascular imaging were enrolled between March 2021 and January 2024. Data on variables were collected including baseline clinical manifestation, demographic data, vital parameters, medical history, and laboratory values. The univariate/multivariate analysis and the LASSO method were used for feature selection to establish different variable combinations. Diagnostic models were constructed using Logistic regression (LR) and 4 machine learning (ML) methods (SVM, XGBoost, AdaBoost,LightGBM). The logistic model was evaluated by the area under curve (AUC) of receiver operating characteristic (ROC) , calibration curve, and decision curve analysis (DCA).The machine learning models were evaluated by ROC-AUC, precision, recall, and F1-score. The SHapley Additive exPlanations (SHAP) approach was introduced to analyze the relative importance of features. RESULTS: A total of 735 acute ischemic stroke patients were included. The LR model demonstrated strong discrimination in both training and test sets (AUC: 0.8783 and 0.8655, respectively), along with good calibration and clinical utility. The derived scoring scale also showed superior performance (AUC: 0.869) compared to existing international scales (FAST -ED, RACE, CPSSS, EMSA). XGBoost model achieved the highest discrimination among ML models. The optimized XGBoost model yielded an AUC value of 0.8956 (training set) and 0.8878 (test set), alongside good precision value(0.8597), recall value(0.8416), and F1-score (0.8513). The SHAP approach shows that upper limb weakness, disturbance of consciousness, and oculomotor disorder were of high importance in the optimal model. CONCLUSIONS: In this study, we developed new diagnostic models for anterior circulation large vessel occlusion stroke in Chinese population. The XGBoost model demonstrates the best performance, suggesting strong potential for future clinical application.