Machine Learning-Based Biomarker Identification for Early Diagnosis of Metabolic Dysfunction-Associated Steatotic Liver Disease

Metabolic dysfunction-associated steatotic liver disease (MASLD) is an umbrella term for simple hepatic steatosis and the more severe metabolic dysfunction-associated steatohepatitis. The current reliance on liver biopsy for diagnosis and a lack of validated biomarkers are major factors contributing to the overall burden of MASLD. This study investigates the association between biomarkers and hepatic steatosis and stiffness measurements, measured by FibroScan®. Data from the National Health and Nutritional Examination Survey (2017-2020) was collected for 15,560 patients. Propensity score matching balanced the data with a 1:1 case-to-control for age and sex allowing for preliminary trend assessment. Random Forest machine learning determined variable importance for the incorporation of key biomarkers (age, sex, race, BMI, HbA1C, PFG, insulin, total cholesterol, LDL-cholesterol, HDL-cholesterol, triglycerides, ALT, AST, ALP, albumin, GGT, LDH, iron, total bilirubin, total protein, uric acid, BUN, and hs-CRP) into logistic regression models predicting steatosis (MASLD indicated by a controlled attenuation parameter™ score of >238 dB/m) and stiffness (hepatic fibrosis indicated by a median liver stiffness >7 kPa). Sensitivity analysis using XGBoost and Recursive Feature Elimination was performed. The Random Forest models (the most accurate) predicted MASLD with 79.59% accuracy (p<0.001) and specificity of 84.65% and, hepatic fibrosis with 86.07% accuracy (p<0.001) and sensitivity of 98.01%. Both the steatosis and stiffness models identified statistically significant biomarkers, with age, BMI, and insulin appearing significant to both. These findings indicate that assessing a variety of biomarkers, across demographic, metabolic, lipid, and standard biochemistry categories, may provide valuable initial insights for diagnosing patients for MASLD and hepatic fibrosis.