Machine Learning-Based Prediction of ICU Admission in COVID-19 Patients Using CT-Derived Biomarkers

Abstract

Coronavirus disease 2019 (COVID-19) exhibits highly variable clinical outcomes ranging from mild symptoms to critical illness requiring Intensive Care Unit (ICU) admission. Early identification of high-risk patients remains essential for optimizing resource allocation. Recently, Computed Tomography (CT)-derived biomarkers such as gynecomastia, hepatic steatosis, and epicardial fat thickness have been proposed as potential indicators of systemic metabolic dysfunction associated with disease severity.  This study aimed to develop and evaluate Machine Learning (ML) and deep learning models for predicting ICU admission in hospitalized COVID-19 patients using CT-derived imaging biomarkers, including gynecomastia, fatty liver, and epicardial fat thickness. A total of 341 hospitalized COVID-19 patients were retrospectively analyzed. Clinical and CT-derived variables were extracted, including age, gynecomastia subtypes, hepatic steatosis, epicardial fat thickness, and retro-mammary fat measurements. Multiple ML models (Random Forest, Gradient Boosting, ExtraTrees, Logistic Regression, Support Vector Machine (SVM), Decision Tree, and XGBoost) were trained using a pipeline incorporating imputation, scaling, and SMOTE-based oversampling. Model performance was evaluated using accuracy, sensitivity, specificity, F1-score, ROC-AUC, PR-AUC, and Matthews Correlation Coefficient (MCC). Shapley Additive Explanations (SHAP) analysis was applied for model interpretability. Among all models, XGBoost achieved the best overall performance (accuracy = 0.884, F1-score = 0.333), followed by ExtraTrees with the highest ROC-AUC (0.729). Epicardial fat thickness and hepatic steatosis were consistently identified as the most important predictors across feature importance and SHAP analyses. Deep learning models demonstrated inferior generalization performance (ROC-AUC ≈ 0.41–0.53), likely due to limited dataset size and class imbalance. CT-derived metabolic imaging biomarkers, particularly epicardial fat thickness and hepatic steatosis, provide meaningful predictive value for ICU admission in COVID-19 patients. ML models, especially ensemble methods, outperform deep learning approaches in structured clinical datasets. Explainable Artificial Intelligence (XAI) enhances clinical interpretability and supports the integration of imaging biomarkers into predictive decision-support systems.