Artificial intelligence-based multiclass classification of etiologies in patients with elevated liver enzymes: A prospective cohort study.

BACKGROUND AND AIM: Elevated liver enzymes are a common reason for hepatology referral, but etiological overlap complicates diagnosis. Most artificial intelligence applications address binary tasks, whereas routine practice requires multiclass discrimination. We developed and internally validated an explainable multiclass model to differentiate five major causes of persistent liver enzyme elevation. MATERIALS AND METHODS: This prospective proof-of-concept cohort included 280 consecutive adults with unexplained persistent aminotransferase elevation. Final diagnoses-metabolic dysfunction-associated steatotic liver disease, viral hepatitis, alcohol-related liver disease, drug-induced liver injury, or autoimmune hepatitis-were independently adjudicated by three hepatologists. Data were split into training (70%) and independent test (30%) sets, and four supervised algorithms were evaluated. To assess incorporation bias, a basic model using routine clinical and biochemical variables was compared with an extended model including serological markers. Interpretability was assessed using SHapley Additive exPlanations. RESULTS: eXtreme gradient boosting achieved the best test-set performance, with an overall accuracy of 0.86 (95% CI 0.78-0.93) and balanced accuracy of 0.87 (95% CI 0.82-0.92). Class-specific AUROC values ranged from 0.92 to 0.95. Serological markers improved discrimination for autoimmune hepatitis (AUROC increase, 0.15; p<0.03) and viral hepatitis (increase, 0.11; p<0.02). Interpretability analysis identified clinically plausible predictors. Misclassifications occurred in 12 cases (14.3%), mainly in overlapping drug-induced liver injury-autoimmune hepatitis presentations. CONCLUSION: An explainable multiclass model differentiated five major etiologies of persistent liver enzyme elevation. Serological markers added value for autoimmune and viral etiologies. External validation is required before clinical implementation.