In a prospective cohort of 573 cardiac surgery patients at Guangdong Provincial People's Hospital (Sept 2024–July 2025), 23.03% developed postoperative dysphagia. A gradient boosting machine (GBM) model best predicted risk (test AUC 0.851; 95% CI 0.781–0.908), with top drivers including age, Diabetes mellitus (DM), chronic lung disease, CKD, AF, and airway/tube and sedation durations.

Why It Matters To Your Practice

• Dysphagia after cardiac surgery is common and clinically consequential (aspiration pneumonia, malnutrition, longer LOS), but risk is often recognized late.

• Machine learning-based stratification could help target early swallow screening and prevention resources to the patients most likely to benefit.

Clinical Implications

• Consider pre-op and early post-op workflows that flag higher-risk patients (older age, DM, chronic lung disease, CKD, AF) for earlier bedside swallow evaluation and diet precautions.

• Modifiable perioperative signals in the model (tracheal intubation time, gastric intubation time, sedative drug use duration) point to potential levers for prevention and quality improvement.

• For implementation, pair model output with clear action pathways (e.g., automatic speech-language pathology consult thresholds; aspiration precautions; nutrition escalation).

Insights

• Among eight approaches (LR, AdaBoost, CatBoost, XGBoost, SVM, NB, KNN, GBM), GBM performed best on discrimination and was supported by calibration and decision-curve analysis.

• Use of SHAP improves interpretability by showing how individual features contribute to a given patient’s predicted risk—key for clinician trust and auditability.

• Risk appears to be a mix of baseline vulnerability (age/comorbidity) and care-course exposures (airway/tube time and sedation duration), reinforcing that prevention may be partly operational.

The Bottom Line

• About 1 in 4 cardiac surgery patients developed dysphagia in this prospective cohort.

• A validated GBM model (AUC 0.851) could support earlier, individualized screening and prevention—especially for patients with DM and other high-risk features.