Background: Several prognostic models have emerged in alcoholic hepatitis (AH), but lack of external validation precludes their universal use.

Aim: To validate the Maddrey Discriminant Function (DF); Glasgow Alcoholic Hepatitis Score (GAHS); Mayo End-stage Liver Disease (MELD); Age, Bilirubin, INR, Creatinine (ABIC); MELD-Na, UK End-stage Liver Disease (UKELD), and three scores of corticosteroid response at 1 week: an Early Change in Bilirubin Levels (ECBL), a 25% fall in bilirubin, and the Lille score.

Methods: Seventy-one consecutive patients with biopsy-proven AH, admitted between November 2007-September 2011, were evaluated. The clinical and biochemical parameters were analysed to assess prognostic models with respect to 30- and 90-day mortality.

Results: There were no significant differences in the areas under the receiver operating characteristics curve (AUROCs) relative to 30-day/90-day mortality: MELD 0.79/0.84, DF 0.71/0.74, GAHS 0.75/0.78, ABIC 0.71/0.78, MELD-Na 0.68/0.76, UKELD 0.56/0.68. One-week rescoring yielded a trend towards improved predictive accuracies (30-day/90-day AUROCs: 0.69–0.84/0.77–0.86). In patients with admission DF ≥ 32 (n = 31), response to corticosteroids according to ECBL, 25% fall in bilirubin and the Lille model yielded AUROCs of 0.73/0.73, 0.78/0.72 and 0.81/0.82 for a 30-day/90-day outcome respectively. All models showed excellent negative predictive values (NPVs; range: 86–100%), while the positive ones were low (range: 17–50%).

Conclusions: MELD, DF, GAHS, ABIC and scores of corticosteroid response proved to be valid in an independent cohort of biopsy-proven alcoholic hepatitis. MELD modifications incorporating sodium did not confer any prognostic advantage over classical MELD. Based on excellent NPVs, the models are best to identify patients at low risk of death.