Related Experiment Video
Updated: Aug 14, 2026
10:17
Guidelines and Experience Using Imaging Biomarker Explorer (IBEX) for Radiomics
Published on: January 8, 2018
Comparative Evaluation of Feature Selection Strategies for ICU Mortality Prediction Using Chest Radiographic,
Orhan Gok1, Türker Fedai Çavuş1, Omer Ozdemir2
1Department of Electrical and Electronics Engineering, Faculty of Engineering, Sakarya University, Sakarya 54050, Türkiye.
Diagnostics (Basel, Switzerland)
|August 13, 2026
Summary
This study shows that smart feature selection can simplify intensive care unit (ICU) mortality prediction models without losing accuracy. Key radiographic features like Cobb angle and pleural effusion consistently predict mortality, enabling more interpretable and efficient clinical tools.
Area of Science:
- Medical Informatics
- Radiology
- Machine Learning
Background:
- Intensive care unit (ICU) mortality prediction relies on complex data, often including radiographic and demographic information.
- Optimizing predictive models for clinical interpretability and efficiency is crucial.
Purpose of the Study:
- To investigate the impact of feature type, dimensionality, and selection strategies on ICU mortality prediction.
- To determine if reduced, clinically interpretable feature sets can achieve comparable predictive performance.
Main Methods:
- Two experimental frameworks were used: one with 12 interpretable features and another with 74 high-dimensional features.
- Six feature selection methods (ANOVA, Chi-Square, Kruskal-Wallis, MRMR, ReliefF, Shapley) were evaluated on the high-dimensional set.
- Machine learning models were assessed using AUC, sensitivity, specificity, accuracy, and F1-score.
Main Results:
- Feature selection reduced predictors from 12 to 4 in the interpretable set, with Subspace KNN achieving an AUC of 0.96.
- MRMR and Kruskal-Wallis were top-performing feature selection strategies in the high-dimensional set.
- Cobb angle, bilateral infiltrates, and bilateral/unilateral pleural effusion were identified as stable predictors of ICU mortality.
Conclusions:
- Appropriate feature selection can reduce model complexity while retaining predictive power for ICU mortality.
- A small set of radiographic features are robust and clinically meaningful predictors of ICU mortality.
- External validation on larger, multicenter datasets is needed for clinical implementation.
