Feature selection methodology based on explainable AI: application in predicting extubation success in the intensive...

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Machine Learning

Background:

• High-dimensional medical data present challenges for predictive modeling due to noise and the curse of dimensionality.
• Robust feature selection (FS) is crucial for developing reliable predictive models in healthcare.

Purpose of the Study:

• To develop and evaluate a novel, model-agnostic feature selection (FS) method using explainable AI (XAI) techniques.
• To improve the interpretability and robustness of predictive models for high-dimensional medical data.

Main Methods:

• Proposed a consensus feature ranking by aggregating SHAP, LIME, and LRP explainers.
• Reordered features using true-positive (TP)-specific attributions.
• Applied the method to predict extubation success in 19,567 mechanically ventilated patients using the MIMIC-IV database and a multilayer perceptron.

Main Results:

• The XAI FS method achieved an AUROC of 0.8527 and comparable sensitivity to conventional methods.
• The selected feature subset was compact, interpretable, and clinically plausible, aligning with established extubation readiness criteria.
• The framework demonstrated potential generalizability, improving performance on most classification tasks across seven additional medical datasets.

Conclusions:

• True-positive focused, multi-explainer FS provides robust and interpretable models for high-risk ICU decision support.
• This approach establishes reliable consensus feature rankings, enhancing clinical utility.
• The method shows promise for broader application in medical predictive modeling.