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

Abstract

Background

Prolonged mechanical ventilation (PMV) might cause ventilator-associated pneumonia and diaphragmatic injury, and may lead to worsening clinical weaning outcomes. The present study proposes a comprehensive machine learning (ML) framework for predicting the weaning outcomes of patients with PMV, without relying on ventilator data, by utilizing features from electrical impedance tomography (EIT).

Methods

EIT data from 58 patients with PMV were analyzed. Extracted EIT image features were standardized using the min-max method. The Boruta method was employed to select significant features for the ML model. To balance the data, the SMOTE method was utilized. Ten ML algorithms commonly used in clinical prediction were compared. The SHAP and LIME methods were used to explain the ML models. Feature selection, data balancing, and hyperparameter adjustment all adopt the Leave-One-Out cross-validation method to avoid overfitting.

Results

The area under the receiver operating characteristic (AUC), specificity, and precision of the ML model with SMOTE balance were significantly improved (p < 0.05) compared to unbalanced data. However, the sensitivity was reduced considerably (p = 0.02). The optimal ML model, extreme gradient boost (XGBoost), demonstrated excellent performance: AUC = 0.862, sensitivity = 0.923, specificity = 0.800, accuracy = 0.889, precision = 0.923, and f-score = 0.923. Decision Curve Analysis and calibration curve evaluation indicated that the model has high clinical generality and reliability. The SHAP and LIME methods enabled model interpretation at both the global and individual sample levels.

Conclusion

The weaning outcome prediction model based on EIT data does not rely on ventilator data, which is suitable for a broader range of weaning scenarios. We proposed a comprehensive ML framework for weaning outcome prediction and incorporated the SHAP and LIME methods, which significantly improved the interpretability of the model.