Liver Fat Fraction and Machine Learning Improve Steatohepatitis Diagnosis in Liver Transplant Patients

This study investigates the diagnostic accuracy of liver fat fraction (FF) and other biomarkers in differentiating metabolic dysfunction-associated steatohepatitis (MASH) from non-MASH conditions in a cohort of 127 liver transplant patients using H-1 MRS and machine learning techniques. Receiver operating characteristic analysis identified FF as the most significant predictor, achieving an area under the curve (AUC) > 0.96 for distinguishing MASH from non-steatosis and non-MASH metabolic dysfunction-associated steatotic liver disease (MASLD). Secondary biomarkers, including insulinemia and elastography, showed moderate discriminatory power (AUC = 0.7-0.8) and contributed to refining classification decisions within a decision tree model. The decision tree analysis, validated with 10-fold cross-validation and independent testing, demonstrated robust sensitivity and specificity, with FF contributing 60%-70% to decision-making. Secondary splits, such as insulinemia (similar to 16.21 mu IU/mL) and elastography (similar to 8 kPa), provided additional discriminatory power, particularly in cases with borderline FF values. Non-significant biomarkers, such as waist circumference and signals of diallylic protons resonating at 2.8 ppm, were excluded due to low discriminatory performance (AUC < 0.7). Compared to the general population (similar to 5.8% prevalence), MASH was significantly more common in liver transplant recipients (similar to 30%-50%). In patients with FF > 5.3%, the positive predictive value (PPV) for MASH ranged from 88% to 97%, more than twice the PPV observed in the general population (approximately 60%). These findings align with existing literature validating MRI-derived proton density fat fraction as a reliable biomarker for hepatic steatosis. However, liver fat percentage alone is insufficient for MASH diagnosis. Secondary biomarkers, particularly insulinemia and elastography, enhanced classification accuracy near the FF threshold of 5.3%. This multiparametric approach significantly improves ...