A Comparative Study on Prediction of Postoperative Acute Heart Failure in Patients with Gynecologic Malignancies under Small-Sample Constraints

Technical Summary: A Comparative Study on Prediction of Postoperative Acute Heart Failure in Patients with Gynecologic Malignancies under Small-Sample Constraints

Problem Statement
Postoperative acute heart failure (AHF) is a severe complication in patients undergoing surgery for gynecologic malignancies, with an incidence ranging from 3.2% to 8.7%. Early risk prediction is critical for improving patient outcomes but remains challenging due to the "small-sample, class-imbalanced" nature of clinical datasets. Traditional risk assessment tools (e.g., CHADS2, CHA2DS2-VASc) were developed for established cardiovascular populations and fail to account for surgery-specific factors like operative duration and intraoperative blood loss. Furthermore, conventional statistical models often struggle with the low event rates and high dimensionality of such datasets, leading to issues with stability and overfitting.

Methodology
The study utilized a retrospective cohort of 1,049 patients from The Second Hospital of Hebei Medical University (January 2021–December 2023), comprising 100 AHF cases (9.53%) and 949 non-AHF cases. The methodology followed a three-stage workflow:

1. Data Preprocessing and Feature Engineering:

   • Variables: 19 clinical variables were analyzed, including baseline demographics, comorbidities, preoperative laboratory indices (e.g., LVEF, D-dimer), and perioperative parameters (e.g., operation duration, fluid infusion).
   • Handling Imbalance and Skewness: The dataset was split 7:3 into training and independent test sets. Skewed variables (e.g., D-dimer, intraoperative blood loss) underwent logarithmic transformation, while continuous variables were standardized via Z-score normalization.
   • Correlation Analysis: Spearman's rank correlation revealed generally low multicollinearity, with the exception of a strong correlation between weight and BMI.

2. Multi-Source Feature Selection Strategy:
   To address the limitations of single-algorithm selection in small-sample settings, the authors proposed an Integrated Feature Selection (IFS) framework. Five individual methods were employed to generate candidate subsets:

   • Filter Methods: Mutual Information (MI).
   • Wrapper Methods: Recursive Feature Elimination (RFE) using SVM.
   • Embedded Methods: Random Forest (RF), XGBoost, and Lasso Regression.
   • Integration: A weighted rank aggregation strategy combined the importance scores from all five methods to construct a final Integrated Feature Subset (FS6), selecting the top 15 features.

3. Model Development and Comparison:
   Seven machine learning algorithms were trained and compared using the generated feature subsets:

   • Linear Models: Logistic Regression (LR), Elastic Net (EN).
   • Tree-Based Models: Decision Tree (DT), Gradient Boosting Decision Tree (GBDT), XGBoost, CatBoost.
   • Ensemble Learning: A Stacking model (Base learners: RF, XGBoost, CatBoost; Meta-learner: LR).
   • Evaluation: Models were evaluated using 5-fold stratified cross-validation and an independent test set. Metrics included Accuracy, Precision, Recall, and F1-score to address class imbalance.

Key Results

• Feature Selection Performance: The integrated feature subset (FS6) demonstrated superior robustness compared to subsets generated by single algorithms (FS1–FS5). Under the GBDT model, FS6 achieved an Accuracy of 1.00, Precision of 1.00, Recall of 0.96, and an F1-score of 0.98. It significantly reduced false negatives compared to other subsets.
• Model Performance: Among the seven algorithms, GBDT achieved the best overall performance using FS6, balancing high sensitivity (Recall 0.96) with perfect precision (1.00).
  • Ensemble Superiority: Nonlinear ensemble models (GBDT, XGBoost, Stacking) significantly outperformed linear models (LR, EN). Linear models exhibited high precision but critically low recall (0.08–0.13), indicating a failure to capture complex nonlinear relationships and a high rate of missed diagnoses.
  • Key Predictors: The analysis identified Left Ventricular Ejection Fraction (LVEF), operative duration, and postoperative fluid volume as the most critical predictors, alongside platelet count (PLT) and hemoglobin (HGB).
• Clinical Framework: The study established a risk stratification framework based on the GBDT model, categorizing patients into low, intermediate, and high-risk groups to guide tiered clinical management (routine monitoring vs. proactive intervention).

Significance and Claims
The paper claims to provide a feasible methodological framework for early risk identification of postoperative AHF under small-sample and imbalanced data constraints. Its primary contributions are:

1. Methodological Innovation: It demonstrates that feature-level ensemble learning (integrating multiple selection algorithms) combined with nonlinear models (specifically GBDT) significantly improves prediction stability and performance in small-sample medical datasets compared to single-method approaches.
2. Clinical Relevance: The model identifies clinically interpretable key variables (cardiac function, surgical trauma, volume load) that align with established pathophysiological mechanisms, supporting the biological plausibility of the predictions.
3. Translational Potential: The proposed framework offers a structured approach for converting machine learning outputs into actionable clinical decision support, enabling individualized risk stratification.

Limitations
The authors modestly acknowledge that the study is a single-center retrospective analysis with a limited sample size, which may constrain external generalizability. Additionally, the proposed risk stratification framework has not yet undergone prospective clinical validation, and future work is required to confirm its feasibility in real-world settings.