Conventional Preoperative Prediction of MASLD-Compatible Histologic Abnormality, Fibrosis, and Higher-Grade Steatohepatitis, and Exploratory Machine-Learning Classification of MASH Severity in Patients Undergoing Roux-en-Y Gastric Bypass

Abstract Introduction Metabolic dysfunction-associated steatotic liver disease (MASLD) is highly prevalent in patients with severe obesity undergoing metabolic/bariatric surgery. Noninvasive preoperative stratification of histologic liver abnormalities remains incomplete. We evaluated conventional regression models and a separate exploratory machine-learning analysis of MASH grade 1 versus grades 2/3. Methods This retrospective single-center cohort included patients undergoing Roux-en-Y gastric bypass between March 2000 and March 2025. Of 1,350 surgical patients, 745 underwent intraoperative wedge liver biopsy and 375 had complete data for conventional modeling. Conventional endpoints were MASLD-compatible histologic abnormality, any documented fibrosis, and higher-grade steatohepatitis, defined as biopsy-reported grades 2/3. The machine-learning analysis included 284 patients with steatohepatitis and used routine preoperative variables, ultrasound-reported steatosis, NFS, the AST/ALT ratio, and FIB-4. Model development used RandomizedSearchCV and repeated stratified cross-validation within the training set, followed by training-based threshold selection and held-out test evaluation. Results BMI and triglycerides were associated with MASLD-compatible histologic abnormality, BMI with any fibrosis, and age, BMI, triglycerides, and AST with higher-grade steatohepatitis. The latter conventional model had an apparent AUC of 0.737 and a mean cross-validated AUC of 0.692. L2-penalized logistic regression had the highest machine-learning test ROC-AUC point estimate (0.730; 95% CI, 0.585–0.864), with sensitivity 0.739 for MASH grades 2/3 at a threshold of 0.49. Test Brier score was 0.242; calibration was unsatisfactory (ECE 0.203; calibration slope 13.469), and decision-curve findings remained exploratory. Conclusion Routine preoperative information showed moderate discrimination for selected histologic outcomes. Both the conventional score and the machine-learning severity classifier require further validation before clinical implementation.