Explainable machine learning for postoperative respiratory failure prediction in open-heart surgery patients - a
Riliang Ma1, Hong Wang2, Chengmei Lv1
1Department of Anesthesiology, The People's Hospital of Guangxi Zhuang Autonomous Region, Nanning, Guangxi, China.
BMC Medical Informatics and Decision Making
|March 14, 2026
Summary
An interpretable machine learning model accurately predicts postoperative respiratory failure (PRF) in open-heart surgery patients within 24 hours of ICU admission. Key predictors include ionized calcium, vasopressor score, and ScvO₂, aiding early risk stratification.
Area of Science:
- Medical Informatics
- Critical Care Medicine
- Machine Learning
Background:
- Postoperative respiratory failure (PRF) is a critical complication following open-heart surgery, leading to increased mortality and extended ICU stays.
- Existing machine learning (ML) models for PRF prediction often use incomplete data and lack transparency.
- There is a need for interpretable ML models for early PRF prediction using initial ICU data.
Purpose of the Study:
- To develop an interpretable machine learning model for the early prediction of postoperative respiratory failure (PRF).
- To utilize data from the first 24 hours of ICU admission for PRF risk stratification.
- To enhance clinical decision-making through an understandable and accurate predictive tool.
Main Methods:
- Analysis of the MIMIC-IV database for patients undergoing open-heart surgery with cardiopulmonary bypass (CPB).
- Exclusion of patients with preoperative respiratory failure or substantial missing data; imputation of remaining missing values.
- Feature selection using LASSO regression, comparison of eight ML models, and interpretation of the optimal model with Shapley Additive exPlanations (SHAP).
Main Results:
- The Gradient Boosting Machine (GBM) model achieved the highest performance (AUROC 0.808, AUPRC 0.369).
- Key predictors identified by SHAP analysis included minimum ionized calcium, vasopressor score, and central venous oxygen saturation (ScvO₂).
- The model demonstrated balanced sensitivity (0.703) and specificity (0.776) with a Youden's index of 0.479.
Conclusions:
- The interpretable GBM model offers a promising approach for early PRF risk stratification after open-heart surgery.
- SHAP analysis provides insights into the influence of hemodynamic and metabolic factors on PRF prediction.
- The model improves clinical understanding and supports informed decision-making for managing patients at risk of PRF.
Related Concept Videos
01:22Cardiopulmonary Resuscitation II: ACLS Airway Management
898
Airway management is a key skill in emergency and critical care settings, as maintaining a clear airway is essential for adequate oxygenation and ventilation.Head Tilt-Chin Lift TechniqueThe head tilt-chin lift maneuver is an essential technique primarily used in patients without suspected cervical spine injuries. To perform this maneuver, one hand is placed on the patient’s forehead, and gentle pressure is applied backward to tilt the head. The fingertips of the other hand are positioned...
898
01:28Cardiomyopathy VII: Pre and Post Operative Nursing Management
413
Patients with hypertrophic cardiomyopathy (HCM) and left ventricular outflow tract (LVOT) obstruction who remain symptomatic despite optimal medical therapy may undergo a septal myectomy (Morrow procedure). This procedure involves excising a portion of the hypertrophied septum below the aortic valve using a heart-lung machine to improve blood flow through the LVOT. Effective preoperative and postoperative nursing management ensures successful patient outcomes, minimizes complications, and...
413