AI-based non-invasive prediction of liver fibrosis in a multicenter biopsy-proven MASLD cohort in Korea

초록

Background and aims: Fibrosis stage is the key determinant of longterm prognosis in metabolic dysfunction–associated steatotic liver disease (MASLD). Recent guidelines define “at-risk MASH” as biopsyproven disease with fibrosis stage ≥F2, underscoring the need for accurate non-invasive tests. This study aimed to develop an artificial intelligence (AI)–based model to predict liver fibrosis using routinely available clinical and laboratory variables from a Korean multicenter biopsy-proven MASLD cohort (STAR-K). Method: The STAR-K retrospective cohort comprises 1,000 patients with biopsy-proven MASLD enrolled from 10 tertiary hospitals in South Korea. The dataset was divided using 5-fold cross-validation for model training, validation, and testing. Predictive performance for significant (≥F2) and advanced fibrosis (≥F3) was evaluated and compared with conventional scores, including the fibrosis-4 (FIB-4) index and NAFLD fibrosis score (NFS). Machine learning (Random Forest) and deep learning (DL; FT-Transformer, TabM, ExcelFormer) models were developed using base variables (age, AST, ALT, platelet count, BMI, diabetes status, albumin, and AST/ALT ratio). Additional analyses incorporated extended variables (INR, GGT, uric acid, LDL, and HDL). In a subset of patients (n = 810), model performance was compared with liver stiffness measurement (LSM) by transient elastography (TE). Results: Under identical variable conditions, DL-based models showed superior performance compared with FIB-4 and NFS. The AUROC for predicting ≥F2 fibrosis was 0.7508 for FIB-4, 0.7587 for NFS, and 0.7947 for the DL model, while for ≥F3 fibrosis it was 0.7962, 0.8195, and 0.8446, respectively. Corresponding negative predictive values were also higher for the DL model (≥F2: 0.7901 vs. 0.7083 and 0.6937; ≥F3: 0.8791 vs. 0.8176 and 0.7827 for FIB-4 and NFS, respectively). Incorporation of extended variables resulted in similar AUROC and NPV compared with the base-variable DL model. However, ML and DL models did not demonstrate clear superiority over LSM-based TE for fibrosis prediction. Conclusion: In this multicenter biopsy-proven MASLD cohort, DLbased non-invasive models outperformed conventional clinical scores in predicting clinically meaningful fibrosis (F2–F3). These AIbased NITs may be useful for risk stratification and longitudinal monitoring in MASLD, although they currently serve best as complementary tools rather than replacements for TE.