MAgEC: Using Non-Homogeneous Ensemble Consensus for Predicting Drivers in Unexpected Mechanical Ventilation.

Area of Science:

• Medical informatics
• Machine learning in healthcare
• Clinical decision support

Background:

• Accurate prediction of patient healthcare risks is crucial for effective clinical decision-making.
• Existing feature importance algorithms often rely on homogeneous methods, potentially limiting the diversity of insights.
• Identifying key clinical drivers of critical events, such as mechanical ventilation, remains an important challenge.

Purpose of the Study:

• To introduce and explore the Model Agnostic Effect Coefficients (MAgEC) algorithm for feature importance extraction and risk prediction.
• To compare MAgEC's performance against established methods like SHAP (SHapley Additive exPlanations).
• To evaluate the stability and diversity of feature importances generated by MAgEC.

Main Methods:

• Development of a novel non-homogeneous consensus-based algorithm (MAgEC).
• Application of MAgEC to the MIMIC-III dataset for predicting unexpected mechanical ventilation.
• Validation using prediction accuracy and comparison of feature importances with SHAP.
• Analysis of feature importance stability under perturbations.

Main Results:

• MAgEC demonstrates accuracy in predicting mechanical ventilation.
• The algorithm provides feature importances comparable to SHAP.
• The non-homogeneous approach of MAgEC yields diverse feature importance insights.
• Feature importances show stability under different perturbations.

Conclusions:

• MAgEC offers a robust and novel approach for identifying critical clinical features in risk prediction.
• The algorithm's non-homogeneous nature enhances the diversity of extracted feature importances.
• MAgEC shows promise for improving clinical decision support systems and patient risk assessment.