Comparative Evaluation of Feature Selection Strategies for ICU Mortality Prediction Using Chest Radiographic, Radiomic, and Demographic Features.

Background/Objectives: This study investigated the impact of feature type, feature dimensionality, and feature selection strategies on intensive care unit (ICU) mortality prediction using first-day portable frontal chest radiographs and demographic data. In addition, the study evaluated whether comparable predictive performance could be achieved using reduced, clinically interpretable feature sets. Methods: A total of 500 patients were included, comprising 400 cases for model development and 100 independent cases for testing. Two complementary experimental frameworks were evaluated. In Experiment Set-1, a clinically interpretable feature set consisting of 12 radiographic and demographic variables was analyzed. In Experiment Set-2, a high-dimensional feature space consisting of 74 radiographic, radiological, radiomic, and image-derived features was investigated using six feature selection methods: ANOVA, Chi-Square, Kruskal-Wallis, MRMR, ReliefF, and Shapley-based importance analysis. Machine learning models were evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, accuracy, and F1-score. Results: In Experiment Set-1, feature selection reduced the number of predictors from 12 to 4. Using the complete 12-feature set, the Bagged Trees classifier achieved the highest performance (AUC = 0.95). Following feature reduction, the Subspace KNN classifier achieved an AUC of 0.96 with an accuracy of 0.82, indicating that satisfactory predictive performance could be achieved using a substantially smaller feature set. In Experiment Set-2, comparison of six feature selection strategies demonstrated that MRMR and Kruskal-Wallis produced the highest-performing feature subsets. Cobb angle, bilateral infiltrates, bilateral pleural effusion, and unilateral pleural effusion were consistently identified across all feature selection methods as stable predictors associated with ICU mortality. These predictors were considered stable because they were consistently identified across multiple independent feature selection methods despite differences in the underlying selection algorithms. Conclusions: The findings suggest that appropriate feature selection strategies may reduce model complexity while retaining clinically relevant predictive information. The consistent identification of a small group of radiographic features across multiple feature selection methods in the 74-feature experimental framework suggests that these variables represent robust and clinically meaningful predictors of ICU mortality. Further external validation using larger multicenter datasets is required before clinical implementation.