Objective To predict the in-hospital mortality risk among patients with ischemic stroke combined by dysphagia using interpretable machine learning methods, so as to provide more evidence-based support for the prognosis prediction of patients with ischemic stroke combined by dysphagia.

Methods Medical record of 308 patients diagnosed with ischemic stroke combined by dysphagia in the Medical Information Mart for Intensive Care Ⅳ (MIMIC-Ⅳ) (2.0) in the United States were retrospectively analyzed. Features of the research data were screened based on the least absolute shrinkage and selection operator, and which were randomly divided into a training set and a test set at a ratio of 7∶3. Then ten models, including logistic regression, random forest, K-nearest neighbor, linear discriminant analysis, naive bayes (NB), neural network, quadratic discriminant analysis, recursive partitioning tree, extreme gradient boosting tree, and support vector machine, etc. were constructed. The predictive effect was measured by calculating the area under the curve (AUC) of receiver operating characteristics. In addition, the calibration curve and Brier score were used to evaluate the calibration degree of the model, and the decision curve was drawn to reflect the clinical net benefit. The Shapley additive explanation method was used to analyze the interpretability of the black box model and explore the important decision-making factors.

Results The NB model in the test set showed better predictive ability compared with other models (AUC=0.85, 95%CI: 0.83‒0.88). After interpretability analysis, it was found that blood urea nitrogen (BUN), age, sequential organ failure assessment, bicarbonate, chloride, and hypertension were important risk factors for in-hospital mortality in patients with ischemic stroke combined by dysphagia.

Conclusion The comprehensive performance of the NB model is better than that of the other nine models in predicting the risk of in-hospital mortality in patients with ischemic stroke combined by dysphagia. The interpretability of the model can help clinicians better understand the reasons behind the results and take further reasonable intervention measures for risk factors to improve the survival probability of patients.