Prediction of prolonged intensive care unit stay after coronary artery bypass grafting: development and internal validation of preoperative and 24-hour models in a Colombian cohort

Abstract Background Prolonged intensive care unit (ICU) stay after coronary artery bypass grafting (CABG) drives morbidity and resource use. Existing prediction models were derived mainly in elective, high-income-country cohorts, and it is unclear whether EuroSCORE II identifies these patients. We developed and internally validated two models to predict ICU stay ≥ 7 days, one for use before surgery and one 24 hours after ICU admission and compared them with EuroSCORE II. Methods Retrospective cohort of adults undergoing isolated or combined CABG at a university hospital in Bogotá, Colombia (2,600 m above sea level), between 2019 and 2026. Predictors were pre-specified. Model 1 included ten preoperative predictors; Model 2 added cardiopulmonary bypass duration, intraoperative transfusion, cardiogenic shock, PaO 2 /FiO 2 on ICU admission and first-day chest-tube drainage to the Model 1 linear predictor. Missing data were multiply imputed. Internal validation used bootstrapping, followed by uniform shrinkage. We assessed discrimination, calibration, decision curves and temporal transportability, and derived a point score. Results Of 915 patients, 131 (14.3%) stayed in the ICU for ≥ 7 days; 88.7% of operations were urgent in-hospital procedures. The area under the receiver operating characteristic curve (AUC) was 0.73 (95% CI 0.68–0.78) for Model 1 and 0.76 (95% CI 0.70–0.80) for Model 2 (optimism-corrected 0.70 and 0.73), compared with 0.68 (95% CI 0.63–0.73) for EuroSCORE II. Concomitant cardiac procedures (odds ratio 3.52), older age, reduced left ventricular ejection fraction, renal dysfunction and extracardiac vascular disease were the main preoperative predictors; cardiogenic shock and chest-tube drainage added information at 24 hours. Both models had greater net benefit than recalibrated EuroSCORE II at thresholds of 10–30%. The 24-hour points score separated low- (7.1%), intermediate- (18.5%) and high-risk (47.3%) groups. Baseline risk varied across calendar periods (observed-to-expected ratio for Model 2, 0.58–1.60). Conclusions Preoperative and early postoperative data predicted prolonged ICU stay after CABG with moderate discrimination. The 24-hour model discriminated better than EuroSCORE II, whereas the advantage of the preoperative model was modest. The models may support ICU capacity planning and the early targeting of preventive care, but they require external validation and local recalibration before clinical use.