Steatosis liver index: A validated machine learning model using clinical data distinguishes steatotic liver disease phenotypes.

INTRODUCTION AND OBJECTIVES: Steatotic liver disease (SLD) is a leading cause of liver disease worldwide. Itsstratified into metabolic dysfunction-associated steatotic liver disease (MASLD), metabolic dysfunction and alcohol-associated liver disease (MetALD), and alcohol-associated liver disease (ALD) based on self-reported alcohol use remains a challenge. We developed the Steatosis Liver Index (SLI), an ordinal machine-learning model that distinguishes these phenotypes using SLD clinical and laboratory variables without alcohol quantification data MATERIALS AND METHODS: National Health and Nutrition Examination Survey (NHANES) data cycles (1999 to 2006 and 2017 to March 2020) were analyzed. Participants meeting eligibility criteria (age ≥20 years and elevated alanine aminotransferase) were stratified into MASLD, MetALD, and ALD. Using an ordinal forest framework, SHAP-guided variable selection, and the Synthetic Minority Oversampling Technique (SMOTE) to address class imbalance, the model was derived from survey cycles 1999-2006 (n=3452) and validated on survey cycles 2017-2020 (n=1013) RESULTS: Of 57,034 participants, 4465 met eligibility criteria, 4145 MASLD, 245 MetALD, and 75 ALD. We derived a model including 15 variables (high-density lipoprotein, mean corpuscular volume, gamma glutamyl transferase, glycohemoglobin, height, mean diastolic blood pressure, ferritin, total cholesterol, monocyte percentage, globulin, iron, hemoglobin, mean systolic blood pressure, sex, and aspartate aminotransferase) with an accuracy of 0.855 in training and 0.848 in validation sets. In the validation dataset, c-statistics were 0.770 and 0.802 for distinguishing MASLD from MetALD and from ALD, respectively. CONCLUSION: This novel SLI may provide a clinical tool for both practice and research settings to stratify SLD phenotypes.