Key ML-based risk prediction studies in CABG.
| Study (year) | Population & setting | AI model(s) | Prediction target(s) | Validation | Performance metrics | Key predictors/features | Comparators | Key takeaways |
|---|---|---|---|---|---|---|---|---|
| Ma et al. [36] (2025)NOAF prediction | 2,994 CABG patients (2 centers, China) | Stacked ensemble (11 learners) | NOAF | Internal + external | AUC: 0.931F1: 0.797 | BNP, LVEDD, EF, BMI, NAR, LA diameter | CHA2DS2-VASc, HATCH, POAF | SHAP-informed bedside tool with superior accuracy |
| Akbulut et al. [37] (2025)POAF biomarkers (pilot) | 100 CABG patients (Turkey) | Not specified | POAF | Internal | Thresholds only | Mg 442 mmol/L, albumin < 29 g/L | None | First use of ML to identify lab cutoffs for POAF |
| Li et al. [38] (2025)AKI in elderly patients | 2,155 elderly CABG patients (China) | RF, XGBoost, LightGBM, etc. | AKI | Internal | RF AUC: 0.737 | Age, eGFR, UA, BNP, ALT, IABP use | None | Strong RF performance with explainability via SHAP |
| Jafarkhani et al. [39] (2025)ICU LOS (general) | 605 CABG patients (Iran) | RF + others | ICU LOS | Internal | R2: 0.28MSE: 1.64 | Intubation time, BMI, age, PRBCs, surgery time | None | Identifies modifiable contributors to ICU burden |
| Dong et al. [40] 2025GI bleeding (GIBCG) | 16,440 patients (4 centers + MIMIC-IV) (China) | Top model of 30 tested | GI bleeding | External | AUC: 0.848–0.851MIMIC: 0.781 | DAPT, PPI, anticoagulants, albumin | None | Web-based tool enables personalized prophylaxis |
| Yang et al. [41] 2025Prolonged ICU stay (IABP) | 236 IABP CABG patients (China) | 7 models (XGBoost best) | ICU stay > 14 days | Internal (train/val) | AUC: 0.92 (train); 0.73 (val) | Tracheotomy, albumin, Sv1, troponin T | None | Targeted model for IABP cohort; SHAP interpretation |
| Chen et al. [42] 2024Post-op stroke risk | 1,200 CABG patients (China) | RF (best of 6), LASSO | Stroke after CABG | Internal (70:30 split) | AUC: 0.901F1: 0.721 | Cr, IABP, ventilation, clamp time, COPD | None | Online tool; ranked features via SHAP |
| Xu et al. [43] 2024XCL ensemble mortality model | 4,764 patients (3 Chinese centers) | XGBoost + CatBoost + LightGBM | In-hospital mortality | Internal + external | AUC: 0.9145 | Composite model features | EuroSCORE II | Outperformed EuroSCORE II in all performance domains |
| Couto et al. [44] 2024Hospital LOS (Brazil) | 9,584 CABG patients (133 centers)Val: 2,627 | RF (best), XGBoost, Neural Net | Hospital stay duration | External | RMSLE: 0.412 (train); 0.454 (val) | Public hospital, emergency, HF, age | Poisson, NB, linear regression | National tool for capacity planning and benchmarking |
F1: F1 score (harmonic mean of precision and recall). CABG: coronary artery bypass grafting; AI: artificial intelligence; NOAF: new-onset atrial fibrillation; AUC: area under the receiver operating characteristic curve; BNP: B-type natriuretic peptide; LVEDD: left ventricular end-diastolic diameter; EF: ejection fraction; BMI: body mass index; NAR: neutrophil-to-albumin ratio; LA: left atrium; CHA2DS2-VASc: congestive heart failure, hypertension, age ≥ 75 (doubled), diabetes, stroke/transient ischemic attack (TIA)/thromboembolism (doubled), vascular disease, age 65–74, sex category; HATCH: hypertension, age ≥ 75, TIA or stroke, chronic obstructive pulmonary disease, heart failure (stroke risk score); POAF: postoperative atrial fibrillation; SHAP: SHapley Additive exPlanations; Mg: magnesium; ML: machine learning; AKI: acute kidney injury; RF: random forest; XGBoost: eXtreme Gradient Boosting; LightGBM: light gradient boosting machine; eGFR: estimated glomerular filtration rate; UA: uric acid; ALT: alanine aminotransferase; IABP: intra-aortic balloon pump; ICU: intensive care unit; LOS: length of stay; MSE: mean squared error; R2: coefficient of determination; PRBCs: packed red blood cells; GI: gastrointestinal; GIBCG: GI bleeding in CABG cohort; MIMIC-IV: Medical Information Mart for Intensive Care IV; DAPT: dual antiplatelet therapy; PPI: proton pump inhibitor; val: validation set; Sv1: S wave in lead V1; LASSO: least absolute shrinkage and selection operator; Cr: creatinine; COPD: chronic obstructive pulmonary disease; XCL: eXplainable Composite Learner; CatBoost: categorical boosting algorithm; EuroSCORE II: European System for Cardiac Operative Risk Evaluation II; Neural Net: neural network; RMSLE: root mean squared logarithmic error; HF: heart failure; NB: negative binomial.