Machine learning models of noninvasive tests to predict metabolic dysfunction–associated steatohepatitis and fibrosis stage
Rohit Loomba, Rebecca Taub, Dominic Labriola, Jan Priel, Krishna Padmanabhan, W. Ray KimBackground:
Fibrosis stage in metabolic dysfunction–associated steatohepatitis (MASH) is currently diagnosed by liver biopsy in clinical trials. Using noninvasive tests, no consensus method exists for staging fibrosis or diagnosing MASH. We aimed to develop machine learning (ML) methods to predict fibrosis stage and MASH with F2/F3 fibrosis using noninvasive biomarkers.
Methods:
Data from 1970 biopsy-confirmed patients screened for phase 3 resmetirom trials (MAESTRO-NASH, MAESTRO-NAFLD-1) were retrospectively analyzed. Outcomes were fibrosis stage and MASH with F2/F3 fibrosis [NAFLD Activity Score (NAS) ≥4 with moderate to advanced noncirrhotic fibrosis], modeled from 37 clinical, laboratory, and imaging predictors. Random forest models were developed, including a full model using all 37 predictors and “Lean” models using subsets of predictors routinely available in clinical practice. These models were evaluated by 4-fold cross-validation using AUC.
Results:
For three-class fibrosis staging (F0/F1 vs. F2/F3 vs. F4), one-versus-rest cross-validation mean AUCs were 0.76, 0.75, and 0.94 for the full model and 0.83, 0.78, and 0.89 for the best-performing Lean model. Liver stiffness measures (LSM) and MASH composite algorithms, such as the Fibrosis-4 Index (FIB-4) and FibroScan-AST, were among the most important predictors. For MASH with F2/F3 fibrosis, the full and best-performing Lean models achieved AUCs of 0.75 and 0.79, respectively. Parsimonious top-k models retaining LSM preserved much of this performance.
Conclusion:
ML models using routinely available noninvasive biomarkers discriminated fibrosis stage and MASH with F2/F3 fibrosis with strong performance, supporting patient selection for clinical trials or practice. Lean predictor models that incorporated LSM outperformed models limited to blood-based biomarkers or a combination of FIB-4 and LSM. These proposed ML models need additional external and prospective validation before deployment for clinical use.