Prediction of stroke-associated pneumonia risk in stroke patients based on interpretable machine learning
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Stroke-associated pneumonia (SAP) is a frequent complication of stroke, characterized by its high incidence rate, and it can have a severe impact on the prognosis of patients. The limitations of current clinical treatment measures underscore the critical need to identify high-risk factors promptly to decrease the incidence of SAP.
Objective
To analyze the risk factors of SAP in stroke patients, construct a predictive model of SAP based on the SHAP interpretable machine learning method, and explain the important variables.
Methods
A total of 763 stroke patients admitted to the Second Affiliated Hospital of Anhui University of Traditional Chinese Medicine from July 1, 2023, to May 31, 2024, were selected and randomly divided into the model training set (n=457) and model validation set (n=306) according to the ratio of 6:4. Firstly, the included data were sorted out, and then Lasso regression was used to screen the included characteristic variables. Based on the tidymodels framework, Using decision tree (DT), logistic regression, extreme gradient boosting (XGBoost), support vector machine (SVM), The classification model was constructed by five machine learning methods, including SVM and LightGBM. The grid search and 5-fold cross validation were used to optimize the hyperparameter optimization strategy and the performance index of the model. The predictive performance of the model was evaluated by the area under the receiver operating curve (AUC), calibration curve, and decision curve analysis (DCA), and we used Shapley additive explanation (SHAP) to account for the model predictions and provide interpretable insights.
Results
The incidence of SAP in this study was 31.72% (242/763). Six variables were selected by Lasso regression, including nasogastric tube use, age, ADL score, Alb, Hs-CRP, and Hb. The model with the best performance in the validation set was the XGBoost model, with an AUC of 0.926, an accuracy of 0.914, and an F1 score of 0.889. Its calibration curve and DCA showed good performance. SHAP algorithm showed that ADL score ranked first in importance.
Conclusion
The model constructed using XGBoost has good prediction performance and clinical applicability, which is expected to support clinical decision-making and improve the prognosis of patients.