Back to 2026 Abstracts
Multi-centeric Predictors of In-Hospital Mortality after CABG Using A ML-Driven Phenotyping Approach
Ali Homaei1, Vivian Wang1, Justin S. Heidel1, Shahab Masoumi1, Asishana A. Osho2, Farhad R. Nezami1, Sari Aranki1, Borami Shin1 1Cardiac Surgery Revision, Brigham and Women's Hospital, Boston, MA 2Cardiothoracic Surgery, Mass General Hospital, Boston, MA
Background: Although in-hospital mortality after isolated CABG is generally low, patient-level risk varies substantially. We hypothesized that machine learning combining mortality prediction with patient phenotyping could improve risk stratification and identify actionable subgroups for preoperative optimization and operative planning. Methods: We analyzed 9,265 consecutive isolated CABG procedures performed from 2013–2025 at two institutions using a harmonized STS database. In-hospital mortality occurred in 123 patients (1.33%). An XGBoost model using 61 preoperative and perioperative variables was compared with L2-penalized logistic regression. Latent profile analysis identified patient phenotypes, which were added to the gradient-boosted model. Performance was evaluated using 5-fold cross-validation, and feature importance was assessed with SHAP and multivariable-adjusted odds ratios. Results: In-hospital mortality was 1.33% (123/9,265). XGBoost outperformed logistic regression (AUROC, 0.852 vs 0.837) with excellent calibration (Brier score, 0.015). Intraoperative transfusion was the strongest predictor of mortality (adjusted OR, 5.08; 95% CI, 3.43–7.53), followed by recent myocardial infarction, anemia, thrombocytopenia, and leukocytosis. Three preoperative phenotypes showed an approximately 10-fold mortality gradient: low risk (55%; 0.55%), intermediate risk (40%; 1.85%), and high risk (5%; 5.41%). The high-risk group accounted for 22% of all deaths. Conclusion: Machine learning enabled accurate mortality prediction and identified a small high-risk CABG phenotype accounting for disproportionate mortality. Intraoperative transfusion was the strongest modifiable predictor of death, supporting preoperative anemia optimization and transfusion-sparing strategies.
Back to 2026 Abstracts
