Related Experiment Videos
Predicting extubation failure in patients with cardiorespiratory decompensation: development and validation of an
Ting Zhang1, Tianqi Lu1, Sha Yang1
1The First People's Hospital of Lin'an District, Hangzhou, Zhejiang, China.
Objective:
This study aimed to develop an interpretable machine learning model for predicting extubation failure in mechanically ventilated patients with combined respiratory failure and cardiac dysfunction.
Methods:
A retrospective cohort of 350 patients receiving invasive mechanical ventilation between January 2022 and January 2024 was analyzed. After rigorous data preprocessing ≤5% missing values were imputed using mean, median, or mode based on distribution type. The dataset was partitioned into training (n = 245) and testing (n = 105) sets at 7:3 ratio. Feature selection and hyperparameter tuning were synchronously optimized using an Improved Snow Geese Algorithm (ISGA)-driven Automated machine learning (AutoML) framework. Six supervised models [Logistic Regression (LR), Support vector machine (SVM), Adaptive Boosting (AdaBoost), eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), AutoML] were evaluated using six classification metrics [Accuracy, Sensitivity, Specificity, F1-score, Area Under the Receiver Operating Characteristic Curve (ROC-AUC), Area Under the Precision-Recall Curve (PR-AUC)] with five-fold cross-validation. Calibration curves and Brier scores assessed reliability, while SHAP analysis provided interpretability of predictors.
Results:
The ISGA-optimized AutoML framework (converging to LightGBM as the optimal algorithm) demonstrated superior performance on the independent test set, achieving a ROC-AUC of 0.9289 and a PR-AUC of 0.8892. Decision curve analysis confirmed clinical utility across threshold probabilities of 0.3-0.7, outperforming treat-all and treat-none strategies, and the calibration curve yielded a low Brier score of 0.114. Key predictors identified included brain Natriuretic Peptide (BNP), lactate, left ventricular ejection fraction (LVEF), troponin, and central venous pressure (CVP).
Conclusion:
The interpretable prediction tool accurately stratifies extubation failure risk using clinically available parameters, offering a prototype tool for personalized ventilation management.
