Machine learning-based predictive model of postoperative acute heart failure among elderly hip fracture patients

Technical Summary: Machine Learning-Based Predictive Model of Postoperative Acute Heart Failure Among Elderly Hip Fracture Patients

Problem Statement
Hip fractures in the elderly, particularly those aged 80 and above, are associated with high mortality, disability, and significant healthcare costs. A critical complication in this population is postoperative acute heart failure (AHF), which occurs in 6% to 20% of cases generally, but this study highlights a significantly higher incidence (42.6%) in the ultra-elderly subgroup (≥80 years). Current clinical risk assessment relies heavily on physician experience and routine laboratory indicators, often failing to capture the complex, nonlinear pathophysiological interactions inherent in this heterogeneous population. Conventional binary logistic regression models struggle to adequately model these nonlinear relationships, leading to limited predictive performance and generalizability. There is a lack of standardized, individualized quantitative prediction systems for precise perioperative prevention and management in this specific demographic.

Methodology
This retrospective study enrolled 661 patients aged ≥80 years with hip fractures who underwent surgical treatment at the Third Hospital of Hebei Medical University between January 2020 and December 2022. Patients with pre-existing chronic heart failure or acute heart failure at admission were excluded.

• Data Collection: Clinical data included demographics, imaging (LVEDD, LVEF), comorbidities (hypertension, diabetes, coronary heart disease, ACCI score), surgical parameters (time to surgery, anesthesia type, blood loss), and laboratory indicators (hemoglobin, albumin, creatinine, CRP, BNP, etc.).
• Variable Selection: A dual-feature selection strategy was employed using LASSO regression (with 10-fold cross-validation) and the Boruta algorithm. This process screened 36 candidate variables down to nine core predictors: admission BNP, preoperative BNP, diabetes mellitus, coronary heart disease, admission hemoglobin, admission albumin, serum creatinine, admission CRP, and hematocrit.
• Model Development: Six predictive models were constructed: one conventional binary logistic regression (LR) and five machine learning (ML) algorithms—Random Forest (RF), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), Naive Bayes (NB), and Gradient Boosting Machine (GBM).
• Optimization and Validation: Hyperparameters for all ML models were tuned using Bayesian optimization. The dataset was split into training (70%) and validation (30%) sets. Model performance was evaluated using discrimination (AUC, DeLong test), calibration (Brier score, calibration curves), and clinical utility (Decision Curve Analysis, Clinical Impact Curves).
• Interpretability: To address the "black-box" nature of ML, SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) algorithms were applied to analyze global feature importance and local individual sample attribution.

Key Results

• Incidence: Among the 661 enrolled patients, 282 (42.6%) developed postoperative acute heart failure.
• Feature Importance: The nine selected variables were identified as independent risk factors. SHAP analysis revealed that preoperative BNP was the most significant predictor (mean absolute SHAP value: 0.735), followed by admission albumin (0.515) and admission hemoglobin (0.479). Interestingly, admission BNP showed a lower predictive contribution (0.063) compared to preoperative BNP, suggesting that preoperative interventions may have mitigated the initial risk reflected by admission levels.
• Model Performance: The XGBoost model demonstrated superior overall performance compared to the other models and the conventional logistic regression.
  • Validation Set AUC: XGBoost achieved an AUC of 0.859 (95% CI: 0.805–0.912), significantly outperforming the logistic regression model (AUC 0.813) and other ML models (RF: 0.832, LightGBM: 0.825).
  • Calibration and Utility: XGBoost exhibited the lowest Brier score, indicating optimal calibration, and demonstrated a higher clinical net benefit across a wider threshold range in Decision Curve Analysis.
• Interpretability: SHAP dependence plots confirmed nonlinear relationships, such as the positive correlation between increasing preoperative BNP and heart failure risk, and the negative correlation between albumin/hemoglobin and risk. LIME analysis successfully provided personalized risk attributions for individual cases, validating the model's decision-making logic at the micro-level.

Significance and Claims
The authors claim that this study fills a gap in existing literature by focusing specifically on the ultra-elderly (≥80 years) subgroup, a population with distinct physiological deterioration and high comorbidity burdens. The primary contribution is the development and validation of an XGBoost-based predictive model that outperforms traditional logistic regression in identifying high-risk patients for postoperative AHF.

The study emphasizes the value of combining rigorous feature selection (LASSO + Boruta) with advanced machine learning and explainable AI (SHAP + LIME). This approach not only improves prediction accuracy but also deciphers the "black-box" mechanism, offering clinicians a transparent tool for risk stratification. The authors assert that the identified predictors (particularly preoperative BNP, albumin, and hemoglobin) provide actionable insights for early identification and the formulation of individualized perioperative management strategies, potentially improving medical safety and prognosis for this vulnerable population.

The paper acknowledges limitations, including its single-center retrospective design and the need for multi-center prospective validation to confirm generalizability and practical clinical effectiveness. However, it posits that the established workflow of "accurate prediction – mechanism interpretation – individualized clinical intervention" offers a robust framework for future research and clinical application.