Résumé : Early identification of patients with severe (discriminant function > or = 32) alcoholic hepatitis (AH) not responding to corticosteroids is crucial. We generated a specific prognostic model (Lille model) to identify candidates early on for alternative therapies. Three hundred twenty patients with AH prospectively treated by corticosteroids were included in the development cohort and 118 in its validation. Baseline data and a change in bilirubin at day 7 were tested. The model was generated by logistic regression. The model combining six reproducible variables (age, renal insufficiency, albumin, prothrombin time, bilirubin, and evolution of bilirubin at day 7) was highly predictive of death at 6 months (P < 0.000001). The area under the receiver operating characteristic (AUROC) curve of the Lille model was 0.89 +/- 0.02, higher than the Child-Pugh (0.62 +/- 0.04, P < 0.00001) or Maddrey scores (0.66 +/- 0.04, P < 0.00001). In the validation cohort, its AUROC was 0.85 +/- 0.04, still higher than the other models, including MELD (0.72 +/- 0.05, P = 0.01) and Glasgow scores (0.67 +/- 0.05, P = 0.0008). Patients above the ideal cutoff of 0.45 showed a marked decrease in 6-month survival as compared with others: 25% +/- 3.8% versus 85% +/- 2.5%, P < 0.0001. This cutoff was able to identify approximately 75% of the observed deaths.