Machine-Learning Models Predict Biopsy-Defined MASH and Fibrosis

Key Takeaways

• Among patients screened for the MAESTRO resmetirom trials, Lean Model 1 identified biopsy-defined MASH with study-defined moderate-to-advanced noncirrhotic fibrosis with 74.5% sensitivity and 71.8% specificity.
• Full-model discrimination was strongest for cirrhosis, while Lean Model 1 discriminated all three biopsy-defined fibrosis categories.
• Stiffness measures and composite scores were influential predictors, and smaller models retaining stiffness showed similar discrimination on internal testing.
• Loomba and colleagues concluded that external and prospective validation remains necessary before clinical deployment.

In Loomba and colleagues’ machine-learning study of noninvasive tests and biopsy-defined MASH fibrosis, investigators retrospectively analyzed screening noninvasive measures and biopsy-derived fibrosis stages from 1,970 patients screened for the phase 3 MAESTRO-NASH and MAESTRO-NAFLD-1 resmetirom trials; for most patients in the F4 group, staging came from historic biopsies. 

Biopsy defined the F0/F1, F2/F3, and F4 fibrosis categories and a separate binary outcome: MASH with a nonalcoholic fatty liver disease (NAFLD) Activity Score (NAS) ≥4 and study-defined moderate-to-advanced noncirrhotic fibrosis. F1B was grouped with F2, so the F2/F3 category did not correspond to conventional histologic F2–F3 alone.

Full random-forest models used 37 clinical, laboratory, and imaging candidate predictors; Lean models used 21–25 predictors. Lean Models 1 and 2 included liver stiffness measurements, whereas Lean Model 3 did not. Internal four-fold cross-validation was repeated 100 times, yielding 400 test folds. 

For F0/F1, F2/F3, and F4, respectively, full-model mean one-versus-rest AUCs were 0.76, 0.75, and 0.94. Magnetic resonance elastography (MRE) and magnetic resonance imaging–proton density fat fraction (MRI-PDFF) led its fibrosis predictor ranking.

In the same stage order, Lean Model 1 had mean one-versus-rest AUCs of 0.83, 0.78, and 0.89. Vibration-controlled transient elastography (VCTE), FibroScan–aspartate aminotransferase score (FAST), Steatosis-Associated Fibrosis Estimator (SAFE), and Fibrosis-4 Index (FIB-4) were influential Lean fibrosis predictors. Differences between full and Lean results cannot be attributed to predictor selection alone.

For the distinct biopsy-defined MASH-with-F2/F3 outcome, AUCs were 0.75 for the full model and 0.79 for Lean Model 1. At the evaluated threshold, the full model had lower sensitivity and specificity than the Lean model. A FIB-4-plus-VCTE comparator discriminated this outcome less well; a descriptive reduced-marker Lean subset nearly matched the larger Lean set on internal testing. These comparisons do not establish clinical cutoffs.

More than 40% of magnetic resonance measures were missing and were not imputed. Full models therefore included fewer patients than Lean models, so differences may reflect both predictors and analyzed populations. Trial screening enriched the cohort for metabolic risks, and participants were predominantly White. Cirrhosis was relatively uncommon, with modest positive predictive values despite high negative predictive values. Internal cross-validation was not independent validation; discordant stiffness could affect an individual classification, and the models were not packaged for clinical deployment.

Loomba and colleagues concluded that noninvasive features, including stiffness, discriminated biopsy-anchored fibrosis categories and the separate MASH-plus-fibrosis outcome internally. External and prospective testing remains necessary to establish generalizability and clinical utility.

Clinician Questions

Why was F1B fibrosis included in the F2/F3 category for MASH prediction?

F1B and F2 were considered moderate fibrosis visible on routine hematoxylin-and-eosin staining. They differ in fibrosis location rather than amount, and biopsy sampling can make them difficult to distinguish.

Can an unusually high liver stiffness reading misclassify fibrosis stage in a MASH prediction model?

In an illustrative case, a patient with F0 fibrosis on biopsy and marked obesity had an isolated high VCTE stiffness measurement that shifted the prediction toward F4, although other measures did not support advanced fibrosis. The case does not establish a misclassification rate.

Recommended Reading

• For more on screening and staging MASLD: Managing MASLD in Patients with Obesity

This content is intended for healthcare professionals for educational purposes. The article and its image were developed with AI assistance and reviewed by members of the ReachMD team prior to publication. Please refer to the original source linked for the complete and most up-to-date information.