S2839 Multimodal Machine Learning Integration of Clinical Biomarkers and CT Radiomics for Precise Liver Lesion Classification

Introduction: Differentiating malignant from benign liver lesions remains challenging because imaging and laboratory tests often yield inconclusive results. Radiology may overlook subtle changes, and biochemical markers lack spatial context. Machine learning can integrate these data, but most existing models analyze imaging features or laboratory values separately. Methods: We combined clinical data from the Indian Liver Patient Dataset (ILPD) with radiomic features extracted from LiTS CT scans. After cleaning and standardizing variables—total bilirubin, albumin, alkaline phosphatase (ALP), age, and lesion characteristics—we applied Chi-Square testing and tree-based importance scoring to select the most informative predictors. Principal component analysis retained 95 percent of feature variance. Four classifiers (logistic regression, Random Forest, XGBoost, and a multilayer perceptron) underwent stratified 10-fold cross-validation. We evaluated accuracy, AUC, precision, recall, and F1 score and then formed an ensemble combining XGBoost and MLP outputs. Finally, SHAP values quantified each feature’s contribution to ensure interpretability. Results: The ensemble distinguished malignant from benign cases with 92.1% accuracy and an AUC of 0.947. Precision was 89.3%, recall 91.7%, and the F1 score 90.5%. In comparison, standalone XGBoost and MLP achieved AUCs of 0.931 and 0.925, respectively. Calibration analysis yielded a Brier score of 0.046, confirming reliable probability estimates. SHAP rankings highlighted lesion size, ALP, total bilirubin, and albumin as the top predictors: larger lesions and elevated ALP/bilirubin increased malignancy risk, while higher albumin reduced it. Feature contributions remained consistent across folds, demonstrating balanced integration of clinical and imaging data for accurate, interpretable classification. Conclusion: The ensemble classifier demonstrated robust discrimination—92.1% accuracy and 0.947 AUC—while maintaining tight calibration (Brier score 0.046) and transparent feature contributions. Radiomic indicators, led by lesion size, complemented biochemical predictors such as ALP, bilirubin, and albumin to generate a balanced risk estimate. These findings suggest that a multimodal model can enhance diagnostic confidence and potentially reduce unnecessary biopsies. Prospective validation in diverse patient populations will determine generalizability and guide integration into clinical workflows.