Machine Learning Prediction of Liver Fibrosis in Patients With Metabolic Dysfunction–Associated Steatotic Liver Disease

作者：Salvatore Petta, Grazia Pennisi, Ciro Celsa, Sofiane Messaoudi, Emmanuel Tsochatzis, Elisabetta Bugianesi, Masato Yoneda, Ming-Hua Zheng, Hannes Hagström, Jérôme Boursier, José Luis Calleja, George Boon-Bee Goh, Wah-Kheong Chan, Rocio Gallego-Durán, Arun J. Sanyal, Victor de Lédinghen, Philip N. Newsome, Jian-Gao Fan, Laurent Castéra, Michelle Lai, Céline Fournier-Poizat, Grace Lai-Hung Wong, Angelo Armandi, Atsushi Nakajima, Wen-Yue Liu, Ying Shang, Marc de Saint-Loup, Elba Llop, Kevin Kim Jun Teh, Carmen Lara-Romero, Amon Asgharpour, Mandy Sau-Wai Chan, Manuel Romero-Gomez, Lin Hh, Seung Up Kim, Terry Cheuk-Fung Yip, Vincenza Calvaruso, Mirko Zoncapè, Jimmy Che‐To Lai, Boyu Yang, Hye Won Lee, Gabriele Di Maria, Marco Enea, Salvatore Contino, Vincent Wai-Sun Wong, Giansalvo Cirrincione, Calogero Cammà · 发表于：Clinical Gastroenterology and Hepatology · 年份：2026 · DOI：10.1016/j.cgh.2026.07.006 · 研究领域：Liver Disease Diagnosis and Treatment、Artificial Intelligence in Healthcare、Hepatitis C virus research

BACKGROUND & AIMS: Accurate staging of liver fibrosis is crucial for risk stratification in patients with metabolic dysfunction-associated steatotic liver disease. We aimed to develop and validate artificial intelligence-based models capable of distinguishing fibrosis stages. METHODS: We developed and validated machine learning models to predict fibrosis stages in more than 3600 biopsy-confirmed patients with metabolic dysfunction-associated steatotic liver disease using 22 clinical features, liver stiffness measurement, and controlled attenuation parameter. Models included Feature Tokenizer Transformer, TabNet variants with distance-aware losses, ordinal Multilayer Perceptron with the COnditional RAnk Logits framework. Three centers were prespecified for geographic external validation. In the remaining centers, we used a stratified 75/25 split into a training pool and an internal random test set, tuned hyperparameters by stratified 10-fold cross-validation on the training pool, and trained 1 model per imputed dataset (M = 5) with an internal 80/20 train-validation split for early stopping and operating-point selection. Performance was pooled across imputations using Rubin's rules and reported for the internal random test set and the held-out centers. RESULTS: Feature Tokenizer Transformer and TabNet achieved the highest performance in binary tasks: area under the receiver operating characteristic curve = 0.860 and 0.855 (F ≥3 vs 0-F2) and area under the receiver operating ch...