Explainable machine learning models for identifying significant fibrosis in adults with metabolic dysfunction-associated steatotic liver disease

Significant fibrosis is a key determinant of long-term outcomes in metabolic dysfunction-associated steatotic liver disease (MASLD). We aimed to develop explainable machine-learning models using routine clinical and laboratory variables and assess their performance in temporal external validation. This cross-sectional diagnostic prediction study used National Health and Nutrition Examination Survey (NHANES) 2017–March 2020 data for model development and internal validation, and NHANES 2021–2023 data for temporal external validation. MASLD was defined as controlled attenuation parameter ≥ 274 dB/m plus at least one cardiometabolic risk factor. Significant fibrosis was defined as liver stiffness measurement ≥ 8.0 kPa. Six machine-learning algorithms were evaluated. The fitted pipelines were applied unchanged to the external cohort. The development cohort included 3,202 adults with MASLD, including 555 (17.3%) with significant fibrosis. In the held-out test set, the random forest model achieved an AUC of 0.795 (95% CI, 0.757–0.833), outperforming the best conventional score (NAFLD fibrosis score, AUC 0.709). The external cohort included 2,107 adults, including 459 (21.8%) with significant fibrosis. External AUCs ranged from 0.709 to 0.753; the logistic regression classifier had the highest AUC (0.753; 95% CI, 0.728–0.778), while random forest achieved 0.741 (95% CI, 0.714–0.766). Random forest showed higher discrimination than all four conventional fibrosis scores in paired external comparisons (all P < 0.001). SHAP analysis highlighted adiposity, glycemic status, liver enzymes, and platelet count as leading contributors. The random forest model showed higher discrimination than conventional fibrosis scores for identifying prevalent significant fibrosis in both internal and temporal external validation. It may help prioritize further fibrosis assessment.