S2893 Comparing Machine Learning to Conventional FIB-4 Thresholding for Fibrosis Risk Stratification: A Decision Curve Analysis Approach

Introduction: Current fibrosis screening strategies in Type 2 Diabetes Mellitus (T2DM) patients rely on FIB-4 thresholding, yet may not optimize clinical decision-making. Decision Curve Analysis (DCA) evaluates the net benefit of predictive models across a range of referral thresholds. We compared the clinical utility of machine learning (ML) vs traditional models in identifying patients at risk of advanced fibrosis. Methods: We included 1,149 adult patients with T2DM from a multi-provider outpatient network. Elevated fibrosis risk was defined as FIB-4 >1.3. Predictors included demographics, BMI, A1c, and provider type. We trained logistic regression and XGBoost models using 5-fold cross-validation and calculated net benefit at thresholds ranging from 0.01–0.50. DCA was conducted to compare each model to “refer-all” and “refer-none” strategies. Results: XGBoost outperformed logistic regression in net benefit across nearly all thresholds. For example, at a threshold probability of 20%, net benefit was 0.094 (XGBoost) vs 0.068 (logistic). Both models outperformed “refer-all” and “refer-none” strategies across realistic threshold ranges (10–30%). The highest gain in clinical utility was observed between thresholds of 15–25%, aligning with typical screening practices. Conclusion: Machine learning models such as XGBoost offer greater net clinical benefit compared to standard FIB-4 cutoff-based approaches. DCA provides a robust framework for optimizing referral thresholds and supports ML implementation in outpatient MASLD screening strategies.