An explainable artificial intelligence framework for weaning outcomes prediction using features from electrical...

Area of Science:

• Critical Care Medicine
• Medical Informatics
• Biomedical Engineering

Background:

• Prolonged mechanical ventilation (PMV) poses risks including ventilator-associated pneumonia and diaphragmatic injury, potentially complicating weaning.
• Current weaning outcome prediction methods may be limited by reliance on ventilator data.

Purpose of the Study:

• To develop and validate a machine learning (ML) framework for predicting weaning outcomes in PMV patients.
• To utilize electrical impedance tomography (EIT) data, independent of ventilator parameters, for enhanced prediction accuracy.
• To improve the interpretability of ML models in clinical settings.

Main Methods:

• Analysis of EIT data from 58 PMV patients.
• Feature extraction, standardization (min-max), and selection (Boruta) for ML model development.
• Data balancing (SMOTE) and comparison of ten ML algorithms, with hyperparameter tuning via Leave-One-Out cross-validation.
• Model interpretability using SHAP and LIME methods.

Main Results:

• ML models incorporating SMOTE balancing showed significant improvements in AUC, specificity, and precision (p < 0.05) compared to unbalanced data.
• The optimal model, XGBoost, achieved high performance metrics: AUC=0.862, sensitivity=0.923, specificity=0.800, accuracy=0.889, precision=0.923, f-score=0.923.
• Decision Curve Analysis and calibration curves confirmed the model's clinical generality and reliability.
• SHAP and LIME provided global and individual sample-level model interpretations.

Conclusions:

• An EIT-based weaning outcome prediction model offers a ventilator-independent solution, broadening applicability across diverse clinical scenarios.
• The proposed comprehensive ML framework, enhanced with SHAP and LIME, significantly improves the interpretability and clinical utility of weaning outcome predictions.