S2840 Artificial Intelligence-Driven Multiclass Model for Early Detection and Prognostication of Hepatitis C Using EHR-Based Laboratory Data

Introduction: Accurate classification of liver disease stage remains essential for guiding treatment and monitoring progression. Traditional assessment relies on invasive biopsy or specialist interpretation of laboratory trends. We leveraged routinely collected demographic and biochemical data to develop and compare 5 machine-learning algorithms for automated assignment into 5 clinically relevant categories: donor, hepatitis, fibrosis, cirrhosis, and suspect. Methods: We assembled a structured dataset of demographic variables and 13 routinely collected laboratory biomarkers from electronic health records. We handled missing values with median imputation for continuous data and mode imputation for categorical data, then engineered alanine aminotransferase/aspartate aminotransferase (ALT/AST) and alkaline phosphatase (ALP)/ALT ratios before applying min–max normalization. To correct for imbalance across 5 clinically relevant categories (Donor; Hepatitis; Fibrosis; Cirrhosis; Suspect), we employed the Synthetic Minority Oversampling Technique. We split the data 80:20 into training and test sets, trained logistic regression, random forest, XGBoost, stacking ensemble (random forest + XGBoost), and feedforward neural network models using 5-fold cross-validation, and evaluated final performance on the held-out test set. We applied SHAP (SHapley Additive exPlanations) to interpret feature contributions. Results: The feedforward neural network achieved the highest accuracy at 95.1%, outperforming XGBoost (93.5%), random forest (92.7%), stacking ensemble (92.7%), and logistic regression (88.6%). Per-class F1 scores for the FNN reached 0.99 for Donor and 0.83 for Cirrhosis, with all classes maintaining precision and recall above 0.80. SHAP analysis identified AST, albumin, and bilirubin as the most influential predictors across classifications. Balanced performance on minority classes and clear differentiation among disease stages underscore the model’s promise for real-time detection and staging of hepatitis C–related liver disease. Conclusion: A feedforward neural network achieved the highest overall accuracy (95.1%) and maintained strong F1 scores across all classes, including minority categories. Interpretability via SHAP highlighted AST, albumin, and bilirubin as primary drivers. Embedding this model within electronic health record systems offers a scalable strategy for real-time screening and staging; prospective, multi-center validation will determine its clinical utility.