Technical Summary: Machine Learning-Based Prediction of IABP Need in HFmrEF Patients Post-CABG

1. Problem Statement

Heart failure with mildly reduced ejection fraction (HFmrEF), defined by a left ventricular ejection fraction (LVEF) of 40%–49%, represents a distinct and clinically heterogeneous phenotype. While patients in this "gray zone" share characteristics with both heart failure with reduced ejection fraction (HFrEF) and preserved ejection fraction (HFpEF), their specific perioperative risks following coronary artery bypass grafting (CABG) are not fully understood.

Current clinical practice often lacks a comprehensive risk stratification tool specifically for HFmrEF patients. Consequently, clinicians may underestimate the risk of postoperative low cardiac output syndrome (LCOS) and the subsequent need for intra-aortic balloon pump (IABP) support in this population. The study addresses the need to identify preoperative and intraoperative predictors to enable early risk stratification, allowing for optimized myocardial protection and timely preparation for mechanical circulatory support.

2. Methodology

Data Source and Cohort

• Study Design: Retrospective single-center study conducted at Beijing Anzhen Hospital.
• Population: 712 adult patients with HFmrEF (LVEF 40%–49%) who underwent isolated CABG between 2019 and 2022.
• Exclusion Criteria: Preoperative IABP use, concomitant non-coronary cardiac surgeries (e.g., valve replacement), and missing data exceeding 30%.
• Primary Endpoint: The requirement for postoperative IABP support, defined by cardiac index < 2.0 L/(min·m²), mean arterial pressure < 60–70 mmHg, or manifestations of LCOS/cardiac failure.
• Data Split: The dataset was randomly divided into a training cohort (70%) and a validation cohort (30%). A total of 214 patients (30.00%) were selected as the testing set.

Feature Engineering and Selection

• Variables: 30+ features including demographics, comorbidities, echocardiographic parameters (LVEF, LV dimensions), laboratory values (lactate, creatinine, troponin, etc.), and surgical data (operative time, blood loss, graft types).
• Handling Missing Data: Variables with >30% missingness were excluded. For variables with <30% missingness, multiple imputation using the random forest algorithm was applied.
• Class Imbalance: The training set exhibited a class imbalance (10.8% IABP requirement). The Synthetic Minority Oversampling Technique (SMOTE) was applied to the training set to balance the classes.
• Feature Selection Strategy: A parallel analytical approach was employed to ensure robustness:
  1. LASSO Regression: To handle multicollinearity and shrink coefficients.
  2. Boruta Algorithm: A wrapper-based method using random forests to identify all relevant features by comparing them against shadow features.
  3. Intersection: Only variables selected by both methods were retained for the final model.

Model Development

• Algorithms: Eight machine learning algorithms were trained and compared: Logistic Regression (LR), K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Random Forest (RF), XGBoost, Gradient Boosting Decision Tree (GBDT), AdaBoost, and Decision Tree.
• Optimization: Bayesian optimization was used for hyperparameter tuning across all models.
• Validation:
  • Internal: 10-fold cross-validation within the training cohort.
  • External: Evaluation on the independent validation cohort (30% split).
• Performance Metrics: Area Under the Receiver Operating Characteristic Curve (AUC), sensitivity, specificity, accuracy, F1-score, Kappa statistic, Decision Curve Analysis (DCA), and Precision-Recall (PR) curves.
• Interpretability: The best-performing model was interpreted using SHapley Additive exPlanations (SHAP) to quantify feature contributions.

3. Key Results

Baseline Characteristics and Predictors

• Incidence: 10.8% (77/712) of patients required postoperative IABP.
• Significant Predictors: The intersection of LASSO and Boruta selection identified seven key predictors:
  1. Preoperative lactate level
  2. Intraoperative use of the Left Internal Mammary Artery (LIMA)
  3. Intraoperative blood loss
  4. Preoperative Systolic Blood Pressure (SBP)
  5. Preoperative Diastolic Blood Pressure (DBP)
  6. Left Ventricular Ejection Fraction (LVEF)
  7. Operative time
• Statistical Findings: Patients requiring IABP had significantly higher preoperative lactate (P < 0.001), lower SBP and DBP (P < 0.05), longer operative times (P = 0.002), greater blood loss (P < 0.001), and lower rates of LIMA use (P = 0.017).

Model Performance

• Best Model: XGBoost demonstrated the superior predictive performance.
  • Internal Validation (10-fold CV): Mean AUC of 0.821 (SD: 0.072).
  • Internal Validation (Training Set): AUC of 0.836 (SD: 0.068).
  • External Validation (Test Set): AUC of 0.733.
• Comparison: XGBoost outperformed other models (Random Forest AUC 0.830, AdaBoost 0.795, Logistic Regression 0.751) in terms of AUC, F1-score (0.578), and Kappa statistic (0.480) in the training set.
• Clinical Utility: Decision Curve Analysis (DCA) indicated that the XGBoost model provided the highest net clinical benefit across most threshold probabilities compared to "treat all" or "treat none" strategies.
• Calibration: The model showed good calibration with a mean calibration error of 0.116.

Interpretability (SHAP)

• SHAP analysis confirmed that preoperative lactate, SBP, DBP, LVEF, operative time, blood loss, and LIMA use were the most influential features.
• Force plots illustrated how specific feature values (e.g., high lactate, low SBP) drove individual predictions toward the positive class (IABP requirement).

4. Key Contributions and Claims

• Novelty: This study addresses a gap in the literature where corresponding prediction models for HFmrEF patients remain scarce and evidence regarding IABP requirement in this specific population is limited.
• Methodological Rigor: The study employed a robust feature selection strategy by intersecting results from LASSO regression and the Boruta algorithm, balancing statistical significance with clinical plausibility.
• Clinical Utility: The authors claim the model offers a practical tool for early risk stratification. By identifying high-risk patients preoperatively, clinicians can:
  • Enhance intraoperative myocardial protection strategies.
  • Proactively prepare for mechanical circulatory support (IABP) and enable timely preparation rather than reacting to postoperative deterioration.
  • Shift management from "passive postoperative rescue" to "proactive preoperative prevention" of adverse hemodynamic events.
• Accessibility: All predictors used in the model are derived from routine clinical examinations and standard perioperative assessments, ensuring the model is accessible for real-world application without requiring complex or specialized testing.
• Economic Implication: The authors suggest the model's risk assessment capability could aid in health economics and insurance planning by facilitating individualized risk coverage schemes.

5. Limitations and Future Directions

The authors acknowledge the following limitations:

• Single-Center Bias: Data was derived from a single tertiary center, which may overrepresent complex cases due to referral patterns, potentially limiting generalizability.
• Retrospective Nature: The study is retrospective, and while multiple imputation was used for missing data, this relies on statistical assumptions.
• Need for Validation: The authors emphasize that external validation using multicenter prospective cohorts is necessary to confirm the model's robustness.
• Future Work: The paper suggests integrating advanced myocardial viability assessments (e.g., PET, cardiac MRI) to better understand the impact of hibernating myocardium in this population and to further refine individualized therapeutic decision-making.