Machine Learning-Based Multimodal Framework for Early Risk Stratification of Metabolic Dysfunction-Associated Steatotic Liver Disease Using Salivary and Voice Biomarkers

This preprint presents a machine learning-based multimodal framework for early risk stratification of Metabolic Dysfunction-Associated Steatotic Liver Disease (MASLD) using salivary and voice biomarkers. The study integrates biochemical salivary biomarkers with acoustic voice features and applies machine learning approaches, including XGBoost, Random Forest, LightGBM, and TabNet, for multimodal risk prediction. Explainable artificial intelligence techniques, including SHAP analysis, are used to investigate the contribution of individual biomarkers and features. The framework is intended to explore a non-invasive, accessible approach for early MASLD risk assessment and provides a computational basis for future validation using prospectively collected paired clinical data.