S2428 Uncovering Hidden Phenotypes and Referral Bias in MASLD Screening: A Combined Latent Class and Explainable Machine Learning Approach

Introduction: MASLD risk stratification in patients with Type 2 Diabetes Mellitus (T2DM) is often guided by fibrosis indices such as FIB-4. However, substantial heterogeneity in clinical profiles and referral patterns may obscure high-risk patients. We hypothesized that latent phenotypes exist within the T2DM population and that nonclinical factors influence specialist referral decisions. Methods: We retrospectively analyzed 1,149 adult patients with T2DM from a multi-provider outpatient network in 2021. In Phase 1, we used Latent Class Analysis (LCA) on clinical and demographic data (e.g., age, sex, BMI, A1c, FIB-4, aspartate aminotransferase, alanine aminotransferase, platelets, provider type, race) to identify hidden patient phenotypes. In Phase 2, we trained an XGBoost classifier to predict referral to GI/Hepatology, and used SHapley Additive exPlanations (SHAP) to interpret feature contributions across LCA-defined classes. Results: LCA identified 4 distinct patient phenotypes with varying MASLD risk and demographic profiles. One class, characterized by older men with elevated FIB-4 and A1c, showed high referral rates. In contrast, a phenotype dominated by younger women with moderate FIB-4 was under-referred despite risk factors. The XGBoost model achieved strong performance (AUC = 0.81), and SHAP analysis revealed that provider type and race were among the top 5 predictors of referral, alongside FIB-4, age, and A1c. When stratified by LCA class, disparities emerged: race and provider type had outsized influence in lower-referral phenotypes, suggesting systemic bias. Conclusion: Combining LCA and explainable ML highlights important gaps in MASLD referral practices. Clinical risk alone does not determine referral, and subgroup analyses reveal inequities by race and provider type. These findings underscore the need for targeted interventions that align referral decisions with evidence-based risk rather than provider or demographic heuristics.