Development of a machine learning-based prediction model for postoperative delirium in frail elderly patients undergoing noncardiac surgery under general anesthesia.
Summary
In 2,089 frail older adults undergoing noncardiac surgery, an XGBoost model using 15 selected features predicted postoperative delirium with AUC 0.813, sensitivity 0.813, and specificity 0.793. SHAP interpretation highlighted MMSE, Charlson Comorbidity Index, and age as dominant predictors, supporting clinically actionable risk stratification.
Key Findings
- Among 2,089 frail adults, POD incidence was 16.52%.
- XGBoost outperformed seven ML algorithms with ROC-AUC 0.813; sensitivity 0.813 and specificity 0.793.
- SHAP identified MMSE, Charlson Comorbidity Index, and age as top predictors; decision-curve analysis suggested clinical utility and minimal overfitting.
Clinical Implications
Supports pre/intraoperative risk stratification for POD to target multicomponent prevention (e.g., delirium bundles, anesthetic choices), allocate monitoring, and inform consent and resource planning.
Why It Matters
Demonstrates a high-performing, interpretable ML model for a common, morbid perioperative complication in frail elders, aligning with precision perioperative care.
Limitations
- Single-center dataset; external validity may be limited despite reported external validation
- Potential residual confounding and need for prospective, real-time implementation studies
Future Directions
Prospective multicenter external validation, integration into EHR for real-time alerts, and testing model-guided prevention to reduce POD incidence.
Study Information
- Study Type
- Cohort
- Research Domain
- Prognosis
- Evidence Level
- III - Large single-center cohort with ML model development/validation and interpretability analyses
- Study Design
- OTHER