关键词: 脑卒中, 吞咽功能障碍, 卒中相关性肺炎, 决策树法, 预测模型

Abstract:

Objective To identify risk factors for stroke-associated pneumonia (SAP) among stroke (cerebrovascular accident, CVA) patients with swallowing dysfunction using early quantitative swallowing assessment indices, and to build a decision-tree model for predicting SAP. Methods From January 2022 to December 2023, 80 CVA patients with swallowing dysfunction treated at our hospital were enrolled and divided into an SAP group (n=29) and a non-SAP group (n=51). Clinical and laboratory data were collected. Univariate analysis and multivariable logistic regression were used to identify independent predictors of SAP. A decision-tree model was constructed using the CHAID (chi-square automatic interaction detection) algorithm. Model performance was evaluated by receiver operating characteristic (ROC) curves (AUC), sensitivity, and specificity; the DeLong test compared AUCs between models. Results Significant between-group differences were observed in Standardized Swallowing Assessment (SSA) scores, presence of an abnormal swallowing-evoked cough reflex, WST (Watanabe Water Swallow Test) grading, presence of an indwelling gastric tube, and neutrophil-to-lymphocyte ratio (NLR) (all P<0.05). Multivariable logistic regression identified the following independent risk factors for SAP: higher SSA score (OR=1.555; 95%CI:1.190-2.033; P=0.001), abnormal swallowing-evoked cough reflex (OR=10.036; 95%CI: 1.889-53.306; P=0.007), WST grade V (OR=7.499; 95%CI: 1.124-50.030; P=0.037), presence of an indwelling gastric tube (OR=5.814; 95%CI: 1.117-30.263; P=0.036), and elevated NLR (OR=3.031; 95%CI: 1.378-6.669; P=0.006). Using the CHAID algorithm, the decision-tree models elected five explanatory variables: SSA score, NLR, presence of an indwelling gastric tube, abnormal swallowing-evoked cough reflex, and WST grade. The tree comprised four levels with six terminal nodes, and SSA score emerged as the strongest predictor. The decision-tree model achieved an AUC of 0.925(95%CI: 0.844-0.972), with sensitivity 79.30% and specificity 92.20% for predicting SAP in CVA patients with dysphagia. By comparison, the multivariable logistic regression modelyielded an AUC of 0.847(95%CI:0.749-0.918), sensitivity 69.00%, and specificity 86.30%. The difference in AUCs between the two models was statistically significant (DeLong test: Z=2.022, P=0.043). Conclusion SSA score, NLR, presence of an indwelling gastric tube, abnormal swallowing-evoked cough reflex, and WST grade are independent predictors of SAP in CVA patients with swallowing dysfunction. The CHAID decision-tree model based on early quantitative swallowing assessment demonstrates strong predictive performance and may be useful for early identification and prevention of SAP.

Key words: stroke, swallowing dysfunction, stroke-associated pneumonia, decision tree method, predictive modeling