Reduced Feature Sets Match Full Models for ICU Mortality Prediction with Chest X-Ray

A small set of four clinical and radiographic features predicted ICU mortality with an AUC of 0.96 (accuracy 0.82), matching a 74-feature radiomic model. Cobb angle, bilateral infiltrates, and pleural effusions were identified as consistent, robust predictors.

• In Experiment Set-1, reducing 12 clinical and radiographic features down to 4 still yielded an AUC of 0.96 for mortality prediction (Subspace KNN classifier).
• Across multiple feature selection methods, Cobb angle, bilateral infiltrates, bilateral pleural effusion, and unilateral pleural effusion were consistently identified as stable mortality predictors.
• The study was limited to a single-center cohort of 500 patients with an internal test set; external multicenter validation was not performed.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.