Abstract
BACKGROUND: Intraoperative acquired pressure injury (IAPI) represents a major risk in elderly cardiac surgery patients due to skin fragility and impaired perfusion. Reliable risk prediction is essential for prevention and personalized care. METHODS: Using data from 12,222 patients aged ≥60 years in the MIMIC-IV database, we developed machine learning models with 24 clinically relevant features. Seven classifiers were trained and integrated via a stacking ensemble. Internal evaluation employed stratified fivefold cross-validation with 10 repetitions and bootstrapped confidence intervals (2,000 resamples). Multiple imputations (m = 2000) quantified uncertainty from missing data. External validation was performed in 355 patients from Nanjing First Hospital, with resampling to address class imbalance. Calibration metrics and SHAP analysis with bootstrapping were used to assess reliability and interpretability. RESULTS: The ensemble model showed strong discrimination in both internal (AUC = 0.857) and external (AUC = 0.783) validation, with stable performance across imputations (mean AUC: 0.997 ± 0.004). Calibration analysis confirmed good agreement between predicted and observed risks. Key predictors as atrial fibrillation, lymphocyte count, and operative time were clinically plausible and robustly identified. CONCLUSIONS: This study presents an externally validated and interpretable ensemble learning framework for early IAPI risk prediction in elderly cardiac surgery patients. Incorporating multiple imputation, bootstrapped calibration, and SHAP-based robustness, this approach improves methodological transparency and offers a clinically actionable tool for perioperative risk stratification and prevention. CLINICAL TRIAL NUMBER: NCT05855954. Registration date: March 21, 2023. Registered on ClinicalTrials.gov. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1186/s12911-025-03325-9.