Abstract:

Objective To construct a prediction model for aspiration risk in stroke patients，identify the optimal model，and provide a predictive tool for patients at risk of aspiration. Methods A systematic review of literature on aspiration in stroke patients was conducted to identify risk factors，which were then incorporated into a risk factor questionnaire developed with expert consultation. Data from 515 stroke patients at a tertiary hospital in Zhejiang Province were collected. Core factors were determined through univariable and multivariable analyses. Models were constructed using logistic regression，extreme gradient boosting，random forest，support vector machine，and artificial neural network. Model performance was evaluated using area under the curve，Brier score，accuracy，and 10-fold cross-validation. Results Fourteen independent aspiration risk factors were identified，including age，Glasgow Coma Scale score，and feeding position. The extreme gradient boosting model demonstrated optimal overall performance，AUC=0.8071，Brier score=0.123，F1=0.5994 in test set，with good stability after 10-fold cross-validation（AUC=0.792）. Core predictors included feeding position（semi-recumbent），age，and NIHSS score. Conclusion Extreme gradient boosting accurately identifies stroke patients at high risk for aspiration. Core factors provide clear targets for clinical intervention. Future multicenter external validation is needed to enhance generalization capability.

Key words: Stroke, Aspiration, Risk Prediction Model, Machine Learning, Logistic Regression, Extreme Gradient Boosting Algorithm, Nursing Care