Ahead, a Novel Validated Machine Learning Model for Clinical Diagnosis of Alcohol Associated Hepatitis

Background: Current criteria from the National Institute on Alcohol Abuse and Alcoholism (NIAAA) for clinical diagnosis are not based on evidence. We developed a machine-learning model for the clinical diagnosis of AH. Methods and Results: 467 subjects with alcohol associated liver disease (ALD) across seven centers were randomly assigned to training or testing cohorts. Three machine learning algorithms: Random Forest, Gradient Boosting Machine, and XGBoost were derived on the training cohort to identify clinical characteristics associated with histologic AH. Combined Alcohol Hepatitis Ensemble Algorithm Development (AHEAD) model integrating AST, ALT, total serum bilirubin was superior to NIAAA criteria to predict histologic AH in test cohort AUC of 0.695 (0.608–0.782) vs. 0.590 (95% CI 0.494–0.685) (Bootstrap p = 0.049). Using 50% probability cutoff, AHEAD model was 61.5% specific (107 of 174 non-cases correctly identified) and 63.3% sensitive (19 of 30 cases correctly identified) for AH diagnosis. NIAAA model was 54.6% specific (95 of 174 non-cases correctly identified) and 63.3% sensitive (19 of 30 cases correctly identified). Conclusion: The AHEAD model (https://aihepatology.shinyapps.io/AHEAD/) is more specific than the widely utilized NIAAA criteria for clinical diagnosis of AH. It may be useful in clinical trials by reducing enrolment of subjects without AH.