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

High-dimensional medical data hinder predictive modeling because of noise and the curse of dimensionality, making robust feature selection (FS) essential. We propose a model-agnostic feature selection method that aggregates three explainers-SHAP, LIME, and layer-wise relevance propagation (LRP)-into a consensus feature ranking and then reorders it using true-positive (TP)-specific attributions. Using the MIMIC-IV database, we analyzed 19,567 mechanically ventilated patients and trained a multilayer perceptron to predict extubation success. On the test set, the proposed XAI FS achieved an AUROC of 0.8527 [Formula: see text] and a sensitivity of [Formula: see text], performing comparably to conventional statistical, wrapper-based, and single-explainer FS methods while prioritizing a compact, highly interpretable, and clinically plausible feature subset. The key predictors closely matched established extubation readiness criteria, supporting their practical relevance. In additional experiments on seven other medical prediction datasets, the framework improved performance on most classification tasks, suggesting potential generalizability beyond extubation. These findings indicate that TP-focused, multi-explainer FS can provide robust, interpretable models and establish reliable consensus feature rankings in high-risk ICU decision support.