This retrospective cross-sectional study was conducted on a dataset of 341 hospitalized patients with confirmed COVID-19 infection. The data were extracted from chest CT scans and clinical records. The primary objective was to evaluate the predictive value of gynecomastia, hepatic steatosis (fatty liver), and epicardial fat thickness for clinical outcomes, specifically ICU admission as the primary endpoint.

All variables were derived from routine CT imaging performed during hospitalization, consistent with previous studies demonstrating the utility of opportunistic CT biomarkers in COVID-19 patients [3], [4].

The dataset included the following variables:

I. Age (years) II. Gynecomastia (right, left, bilateral) III. Hepatic steatosis (fatty liver; binary) IV. Epicardial fat thickness (continuous, mm) V. Retro-mammary fat thickness (continuous, mm) VI. ICU admission (binary outcome) The primary outcome variable was ICU admission, representing severe disease requiring intensive care support.

A secondary variable (mortality/expire) was available but was not used in the final modeling phase in this analysis.

Gynecomastia, hepatic steatosis, and epicardial fat were assessed using non-contrast chest CT scans, which have been validated in previous literature for opportunistic body composition analysis.

I. Hepatic steatosis was evaluated using liver attenuation values on CT, which correlate strongly with histopathological fat deposition [2].

II. Epicardial fat thickness was measured as a surrogate marker of visceral adiposity using axial CT images [6].

III. Gynecomastia was defined based on retroareolar glandular tissue thickness on CT imaging.

Data preprocessing included the following steps:

I. Missing value handling: Median imputation was applied using SimpleImputer.

II. Feature scaling: Standardization was performed using StandardScaler to normalize feature distributions. III. Train-test split: Data were split into:

 80% training set  20% testing set: Stratified sampling was applied to preserve class distribution.

(These preprocessing steps are essential for ensuring model stability and reproducibility in clinical ML pipelines.)

The dataset exhibited significant class imbalance in ICU admission outcome. To address this, Synthetic Minority Oversampling Technique (SMOTE) was applied only on the training set to avoid data leakage. SMOTE has been widely validated for improving classifier performance in imbalanced medical datasets. The effect of SMOTE was evaluated visually using:

 Bar plots of class distribution  Pie charts before and after resampling

Multiple supervised ML algorithms were implemented and compared:

 Random Forest  Gradient Boosting Machine (GBM)  Extra Trees Classifier  Logistic Regression  Support Vector Machine (SVM)  Decision Tree  XGBoost Classifier

Tree-based ensemble methods were included due to their strong performance in structured clinical datasets and ability to capture nonlinear relationships [7].

A unified pipeline was constructed using imblearn.Pipeline to ensure reproducibility:

Pipeline steps:

I. Median imputation II. Standard scaling III. SMOTE oversampling IV. Classification model This design ensures that preprocessing is strictly applied within cross-validation boundaries, preventing information leakage.

Model performance was assessed using multiple metrics:

 Accuracy  Precision  Recall (Sensitivity)  Specificity  F1-score  ROC-AUC  Precision-Recall AUC  Matthews Correlation Coefficient (MCC)  Negative Predictive Value (NPV)

These metrics were selected to provide a robust evaluation in the presence of class imbalance, where accuracy alone is insufficient.

To interpret model predictions, feature importance and SHAP analysis were applied.

I. Feature importance was extracted from tree-based models.

II. SHAP values were computed using:

 TreeExplainer for tree-based models  LinearExplainer for logistic regression  KernelExplainer for non-linear models

SHAP allows quantification of feature contributions at both global and individual levels, improving interpretability in clinical ML [8].

In addition to classical ML models, a fully connected neural network was implemented:

Architecture:

 Input layer  Dense layer (64 neurons, ReLU)  Dropout (0.3)  Dense layer (32 neurons, ReLU)  Dropout (0.3)  Output layer (Sigmoid activation)

The model was trained using:

 Adam optimizer  Binary cross-entropy loss  Early stopping (patience = 10)  Validation split = 20%

Deep learning performance was compared with classical models using ROC-AUC and accuracy.

All analyses were performed in Python using Google Colab. The following libraries were used:

 Scikit-learn  XGBoost  Imbalanced-learn  SHAP  TensorFlow/Keras  Matplotlib and Seaborn

Statistical significance was not the primary endpoint; instead, predictive performance and model interpretability were emphasized.

A total of 341 hospitalized COVID-19 patients were included in the final analysis. All variables derived from CT imaging and clinical records were complete, with no missing data observed across features.

The primary outcome variable (ICU admission) showed a significant class imbalance, with 35 ICU cases (10.3%) and 306 non-ICU cases (89.7%).

As in Fig. 1, histograms demonstrated heterogeneous distributions across demographic and imaging-derived variables, particularly epicardial fat thickness and retro-mammary fat measurements, indicating substantial inter-individual variability.

The correlation structure between variables is illustrated in Fig. 2, where moderate associations were observed among adiposity-related features. In contrast, gynecomastia showed weak correlations with metabolic imaging biomarkers.

Boxplot analysis presented in Figure 3 revealed the presence of outliers in epicardial fat and retro-mammary fat thickness, reflecting variability in visceral adiposity distribution.

Given the significant imbalance in ICU outcomes, SMOTE was only applied to training dataset.

As demonstrated in Fig. 4, SMOTE successfully balanced the dataset, increasing the minority ICU class to match the majority class distribution. The proportional changes before and after resampling are further These transformations ensured improved model learning stability without introducing data leakage.

Seven supervised ML models were trained and evaluated for ICU prediction.

The overall quantitative performance of all models is summarized in Table 1.

As shown in Table 1, the XGBoost classifier achieved the best overall performance in terms of accuracy (0.884), F1-score (0.333), and MCC (0.276), demonstrating superior predictive capability compared to other models.

ExtraTrees also showed strong performance, particularly in ROC-AUC (0.729), indicating robust discriminative ability.

To improve interpretability, feature importance analysis and SHAP-based explanations were applied.

As shown in Fig. 6, ExtraTrees feature importance analysis identified epicardial fat thickness, age, and retro-mammary fat as the most influential predictors.

SHAP summary analysis presented in Fig. 7 confirmed that epicardial fat and hepatic steatosis had the strongest global impact on ICU prediction.

Global feature contribution based on SHAP values is further illustrated in Fig. 7, reinforcing the dominant role of visceral adiposity markers in disease severity prediction.

A feedforward neural network was trained to evaluate nonlinear deep learning performance.

As shown in Fig. 8, the deep learning model achieved moderate discriminative performance, with lower ROC-AUC compared to tree-based ensemble models. Despite acceptable training accuracy, generalization performance remained limited, likely due to small dataset size and class imbalance.

The most important finding of this study is the consistent predictive contribution of epicardial fat thickness across all top-performing models, especially XGBoost and ExtraTrees. This aligns with previous evidence suggesting that epicardial fat acts as an active metabolic and inflammatory tissue, closely associated with cardiometabolic risk and systemic inflammation [5], [6].

Similarly, hepatic steatosis showed moderate predictive relevance. Prior studies have demonstrated that Non-Alcoholic Fatty Liver Disease (NAFLD) is associated with systemic metabolic dysregulation and increased inflammatory burden, both of which may worsen COVID-19 outcomes [2], [9].

Although gynecomastia was not among the top-ranked predictors in feature importance plots, SHAP analysis indicated a non-linear and context-dependent contribution. This supports previous findings suggesting that gynecomastia may reflect underlying endocrine and metabolic imbalance rather than acting as an isolated pathological entity [4], [10].

Among all evaluated models, XGBoost achieved the best overall performance, particularly in accuracy (0.884) and F1-score (0.333), followed by ExtraTrees with the highest ROC-AUC (0.729). This superiority of gradient-boosting and ensemble tree-based models is consistent with prior studies in medical tabular datasets, where nonlinear interactions and feature heterogeneity dominate model structure [7].

The better performance of tree-based methods compared to Logistic Regression and SVM suggests that the relationships between CT-derived biomarkers and ICU admission are highly nonlinear and interaction dependent.

The relatively poor performance of Logistic Regression (ROC-AUC ≈ 0.49) confirms that linear assumptions are insufficient for modeling such complex physiological interactions.

The deep learning model demonstrated lower generalization performance (ROC-AUC ≈ 0.41–0.53 across runs) despite moderate training accuracy. This is likely due to:

I. Limited dataset size (n = 341), which is insufficient for stable neural network training II. Class imbalance, even after SMOTE III. Tabular data structure, which is typically less suitable for deep learning compared to tree-based models These findings are consistent with previous literature showing that deep learning does not necessarily outperform gradient boosting methods in structured clinical datasets [11].

The ICU outcome was highly imbalanced (approximately 10% positive cases). Application of SMOTE improved minority class representation during training. However, as shown in confusion matrices (Fig. 8), sensitivity for ICU prediction remained limited across most models.

This suggests that while SMOTE improves learning balance, it does not fully resolve the inherent difficulty of predicting rare clinical outcomes, particularly in small datasets.

SHAP-based interpretation provided important insights into model decision-making. Epicardial fat thickness, age, and hepatic steatosis consistently ranked among the most influential predictors.

The SHAP summary plot demonstrated that higher epicardial fat values were associated with increased ICU risk, supporting the hypothesis that visceral adiposity contributes to COVID-19 severity through inflammatory and cardiometabolic pathways.

Feature importance analysis from tree-based models confirmed simillar ranking patterns, reinforcing model interpretability consistency.

The observed associations are consistent with previous CT-based COVID-19 severity studies. For example, Aksoy et al. [10] reported correlations between hepatic density, gynecomastia, and adiposity-related CT biomarkers in COVID-19 patients, suggesting shared metabolic pathways.

Similarly, Aslan et al. [3] and Kazemi et al. [12] demonstrated that incidental gynecomastia is common in COVID-19 CT populations, although its prognostic value remains unclear.

Our findings extend these results by integrating multiple CT-derived biomarkers into a unified ML framework for ICU prediction.

This study highlights the potential of incidental CT findings as predictive biomarkers in COVID-19 triage. Since CT scans are already widely performed in hospitalized patients, extracting additional quantitative biomarkers such as epicardial fat and liver attenuation can provide:

 Early risk stratification  Improved ICU resource allocation  Objective support for clinical decision-making The integration of XAI methods (SHAP) further enhances clinical interpretability, which is essential for adoption in medical settings.

I. Multi-center external validation II. Integration of clinical laboratory biomarkers with imaging data III. Use of larger datasets for deep learning optimization IV. Prospective validation in real-world ICU triage settings