A Dynamic Risk Trajectory Theory-Guided Prediction Model for Post-Extubation Dysphagia in Mechanically Ventilated ICU Patients: A Clinical Risk Stratification Tool

<title>Abstract</title> <p>Objective Post-extubation dysphagia (PED) is a prevalent and high-risk complication among mechanically ventilated ICU patients, increasing risks of aspiration, aspiration pneumonia, and prolonged hospital stay. Conventional one-off static swallowing assessments fail to reflect cumulative physiological risks during ICU admission, a gap addressed by dynamic risk trajectory theory. This study screened routine clinical monitoring indicators reflecting progressive disease risk and constructed and validated two predictive models (logistic regression and XGBoost machine learning) to support dynamic, individualised bedside risk stratification. Methods A retrospective cohort study was conducted with 271 adult intubated ICU patients who achieved successful extubation from June 2023 to May 2025 at a tertiary hospital. Based on the 10-events-per-predictor variable rule, the sample size satisfied the requirement of at least 50 PED positive events for five candidate predictors. Participants were randomly split into a modelling group (n = 190, 70%) and an independent external validation group (n = 81, 30%). Longitudinal proxy indicators reflecting illness progression, ventilation duration, airway intervention, and gastrointestinal function were collected. Univariate analysis and forward stepwise binary logistic regression were applied to identify independent predictors. Model performance was comprehensively evaluated via ROC curves, 1000-bootstrap internal validation, calibration curves, Hosmer–Lemeshow test, and decision curve analysis (DCA). An XGBoost machine learning model was developed for performance comparison. Model interpretability was assessed using SHAP (SHapley Additive exPlanations) values. All model development and reporting strictly followed the TRIPOD statement. PED was diagnosed via the Gugging Swallowing Screen (GUSS), with a total score &lt; 10 defined as dysphagia per previously validated ICU cutoff criteria. Results Five independent risk factors for PED were identified: APACHE II score ≥ 15, indwelling nasogastric tube, mechanical ventilation duration ≥ 72 h, gastric residual volume retention, and tracheotomy (all P &lt; 0.05). The logistic regression model yielded an AUC of 0.718 (95% CI 0.629–0.807, P &lt; 0.001); bootstrap-corrected AUC was 0.712 without obvious overfitting. External validation AUC reached 0.756, with sensitivity 89.5%, specificity 85.0%, and accuracy 81.4%. Calibration tests and DCA demonstrated favourable calibration and substantial clinical net benefit for both models. SHAP analysis revealed that higher APACHE II scores, prolonged ventilation, and presence of indwelling gastric tube contributed most substantially to elevated PED risk. Notably, tracheotomy presented a negative regression coefficient (β=−0.487) but an adjusted odds ratio &gt; 1 (OR = 1.401) after adjusting for confounding clinical variables, confirming its independent predictive effect on PED. Conclusion The prediction model incorporating five routine bedside clinical indicators achieves stable discrimination and calibration for PED risk stratification under the guidance of dynamic risk trajectory theory. The concise logistic regression model is suitable for routine bedside screening, while the XGBoost model with SHAP-based interpretability supports precise high-risk patient sorting for hospital intelligent management systems. This clinical tool enables continuous dynamic risk monitoring and early targeted swallowing rehabilitation, reducing PED-related adverse clinical events. The single-centre design and use of cumulative proxy variables rather than serial multi-timepoint longitudinal measurements are the primary limitations; multi-centre prospective studies with repeated daily indicator collection are warranted to further validate and optimise the model.</p>