Development and external validation of an interpretable machine learning model for predicting prolonged postoperative ICU length of stay in coronary artery bypass grafting patients using MIMIC-IV 3.1 and eICU-CRD 2.0.

BACKGROUND

Prolonged postoperative intensive care unit (ICU) length of stay (LOS) after coronary artery bypass grafting (CABG) drives resource use yet remains difficult to anticipate at the 24-hour ICU evaluation moment when extended monitoring beyond 72 h is first considered. We developed and externally validated an interpretable machine-learning decision-support calculator for this deployment moment.

METHODS

Adult CABG patients were identified from MIMIC-IV 3.1 (n = 6,919; 7:3 stratified split for development) and eICU-CRD 2.0 (n = 5,972; external validation). The outcome was prolonged ICU LOS (> 3 days). Elastic Net plus Boruta selected eight bedside features collected within the first 24 h of ICU admission: 24-hour fluid intake, Charlson Comorbidity Index (CCI), Sequential Organ Failure Assessment (SOFA) score, Simplified Acute Physiology Score II (SAPS-II), Glasgow Coma Scale (GCS), vasopressor use, congestive heart failure, and atrial fibrillation. Nine machine-learning algorithms were compared by 10-fold cross-validation paired t-tests with Bonferroni and Benjamini-Hochberg correction. Calibration metrics included Hosmer-Lemeshow test, Integrated Calibration Index (ICI), expected-to-observed ratio, intercept, and slope (full Methods). SHapley Additive exPlanations (SHAP) provided per-patient feature attribution.

RESULTS

CatBoost was selected as the deployed model. On the MIMIC-IV internal test set (n = 2,076), the area under the receiver-operating-characteristic curve (AUC) was 0.7739 (95% confidence interval 0.7379-0.8099), Hosmer-Lemeshow p = 0.224, calibration slope 0.973. On the eICU-CRD external cohort, AUC was 0.6452 (95% CI 0.6311-0.6602), calibration slope 0.998, ICI 0.023. At the prevalence-anchored threshold of 0.30, sensitivity was 0.55, specificity 0.65, positive predictive value 0.40, and negative predictive value 0.77. Decision Curve Analysis showed positive net benefit over treat-all and treat-none across t = 0.20-0.40. SHAP top-3 features were 24-hour fluid intake, CCI, and atrial fibrillation.

CONCLUSIONS

The model provides decision support at the post-CABG 24-hour ICU evaluation moment with modest discrimination and well-calibrated probabilities at deployment. The deployed online calculator renders per-patient feature contributions transparent at the bedside via SHAP; prospective validation is required before clinical deployment.