Fig. 7: Exploring feature redundancy using supervised distance.
From: Interpretable machine learning prediction of all-cause mortality
a The feature importance ranking of the BMI-related features in original models and reducing redundancy models (models using one weight-related feature and all non-weight-related features), and the AUC of the single-feature models controlling for age and gender. b The feature importance ranking of the selected laboratory features in original models and reducing redundancy models, and the AUROC of the single-feature models confounded by age and gender. c The AUROC of the models using the selected feature sets and minimum feature redundancy within the selected feature sets when running supervised distance-based feature selection. The purple dashed line shows the AUROC of the model trained on age and gender. The pink dashed line indicates the feature set we select for further analysis. d The SHAP summary plot for the gradient boosted trees trained on the selected 90 features for the 5-year mortality prediction.
