Abstract
Our study identified 13 important features and provided multiple ML models to enhance the risk stratification for ARDS after cardiac surgery. Using these predictors and ML models might provide a basis for early diagnostic and preventive strategies in the perioperative management of ARDS patients.
Introduction
Acute Respiratory Distress Syndrome (ARDS) remains a significant and life-threatening complication following cardiac surgery, characterized by acute lung inflammation and respiratory failure. Despite advances in perioperative care, ARDS mortality rates can reach up to 46% in severe cases. Early diagnosis is hindered by the limitations of the current gold standard, the Berlin definition, which often identifies the syndrome only after significant progression. Furthermore, traditional scoring systems like the Lung Injury Prediction Score (LIPS) or Sequential Organ Failure Assessment (SOFA) are frequently inadequate for the specific physiological challenges of cardiac surgery, such as the systemic inflammatory response induced by cardiopulmonary bypass (CPB). This study explores the application of tree-based ensemble machine learning (ML) models to provide more accurate, early-stage risk stratification using electronic health record (EHR) data.
Methods
This multicenter, retrospective cohort study utilized data from 1,996 patients across two major Chinese institutions: Nanjing First Hospital (Development Cohort, N=1,493) and Shanghai General Hospital (External Validation Cohort, N=503). To address the complexity of 126 perioperative variables, researchers employed a Sliding Windows Sequential Forward Feature Selection (SWSFS) method to identify the most predictive features while minimizing model error.
The study evaluated seven tree-based ensemble models, including Random Forest, XGBoost, LightGBM, AdaBoost, and Gradient Boosting Decision Tree (GBDT). Notably, a Deep Forest (DF) architecture was implemented. Unlike standard deep neural networks that require massive datasets, the DF model uses a multi-layered structure of random forests to achieve high representational learning on clinical-scale data. Model interpretability was ensured using SHAP (SHapley Additive exPlanation) values to visualize the impact of individual features on risk predictions.
Results
Model Performance
The Deep Forest model emerged as the superior predictive tool, achieving an Area Under the Curve (AUC) of 0.882 in the internal testing cohort and maintaining a robust AUC of 0.827 in external validation. In contrast, traditional clinical scores (LIPS and SOFA) performed significantly worse, with AUCs ranging between 0.613 and 0.708. The ML models also demonstrated excellent calibration, with Brier scores as low as 0.067, indicating a high degree of reliability in predicted probabilities.
Key Predictors
The feature selection process identified 13 critical perioperative variables. The top five predictors were Chronic Obstructive Pulmonary Disease (COPD), preoperative albumin, central venous pressure at T4 (CVP_T4), CPB time, and Left Ventricular Ejection Fraction (LVEF).
The high ranking of CPB and Aortic Cross-Clamping (ACC) times supports the hypothesis that pulmonary ischemia-reperfusion injury is a major trigger for post-surgical ARDS. Additionally, the importance of the Neutrophil-to-Lymphocyte Ratio (NLR) and preoperative albumin levels underscores the role of baseline inflammation and nutritional status in patient outcomes.
Conclusion
Tree-based machine learning models, particularly the Deep Forest architecture, offer a significant improvement over traditional scoring systems for predicting ARDS after cardiac surgery. By integrating complex perioperative data, these models provide a foundation for early diagnostic and preventive strategies. While limited by its retrospective nature, this study paves the way for prospective, real-time clinical decision support systems that could improve outcomes for high-risk surgical patients.
Summary updated