S2743 Enhancing Fibrosis Risk Stratification in T2DM Patients Using Machine Learning: A Real-World Outpatient Study

Introduction: Noninvasive fibrosis indices such as the FIB-4 score are valuable tools for identifying patients with metabolic dysfunction-associated steatotic liver disease (MASLD) at risk for advanced fibrosis. However, screening uptake in outpatient populations with Type 2 diabetes mellitus (T2DM) remains inconsistent. We aimed to develop and compare machine learning (ML) models to predict elevated fibrosis risk and determine the most important clinical predictors in a real-world setting. Methods: We retrospectively analyzed data from 1,149 adult patients with T2DM seen in outpatient clinics across a multi-provider network in 2021. The primary outcome was elevated fibrosis risk, defined as a FIB-4 score >1.3. Predictor variables included demographics (age, sex, race), BMI, hemoglobin A1c, and provider type. We evaluated the performance of 3 predictive models—logistic regression, random forest, and XGBoost—using 5-fold cross-validation. Feature importance was derived from the final XGBoost model. Results: Of the 1,149 patients, 14.1% had an FIB-4 score >1.3. Model comparison showed that XGBoost achieved the highest discriminatory performance (AUC = 0.812), outperforming random forest (AUC = 0.798) and logistic regression (AUC = 0.774). The top predictors of elevated FIB-4 included older age, higher BMI, elevated A1c, and male sex. Provider type and race also contributed substantially to model performance. Conclusion: Machine learning models, particularly XGBoost, demonstrate strong potential for improving fibrosis risk stratification in high-risk T2DM patients. These models may complement FIB-4 scoring by leveraging routinely available clinical data to prioritize patients for MASLD screening and referral. Prospective validation is warranted to guide implementation.