8% and 51.3%, respectively). Patients with ACLF by the APASL definition were significantly different from those defined by the EASL-CLIF definition in terms of patient characteristics, including higher bilirubin, lower creatinine, frequent ascites, less frequent encephalopathy, higher Child-Pugh score, and lower CLIF-SOFA score. Conclusions: The development of ACLF is associated with high short-term mortality. However, patient characteristics are significantly

different when ACLF is defined by the two different APASL vs. EASL-CLIF definitions. Thus refinement of selleck inhibitor the two ACLF definitions or a consensus definition is urgently needed. Disclosures: Dong Hyun Sinn – Speaking and Teaching: Gilead, Yuhan pharmacy The following people BIBW2992 have nothing to disclose: Tae Yeob Kim, Dong Joon Kim, Do Seon Song, Eileen L Yoon, Joo Hyun Sohn, Chang wook Kim, Young Kul Jung, Ki Tae Suk, Jin Mo Yang, Heon Ju Lee Background: Acute liver injury (ALI) is characterized by severe liver injury resulting in coagulopathy without encephalopathy in patients

without prior chronic liver diseases. While there are several prognosis models for acute liver failure (ALF) patients, there are none available to help identify ALI patients that are at the highest risk of progressing to ALF. We use the random forest (RF) statistical procedure to build a prediction model for ALI patients. A novel machine learning algorithm, RF can better distinguish important predictors of a given outcome than many traditional modeling procedures such as logistic regression. Aim: To derive a model to predict the risk that a patient with ALI will progress to ALF. Methods: 386 ALI subjects were pro-spectively enrolled in the ALF Study Group database between January 2008 and October 2013, defined as follows: INR ≥ 2.0 and ALT ≥ 10x ULN for acetaminophen (APAP) ALI, or INR ≥ 2.0, ALT ≥ 10x ULN, and bilirubin ≥ 3.0mg/dL for non-APAP ALI. 82 clinical variables from the database were entered into the RF for predicting ALI progression

to ALF. A selection procedure which minimizes the prediction error rate was implemented Thalidomide to determine variables for inclusion in the model. Analyses were carried out using R software. Results: Of the 82 variables entered into the model, etiology, INR, bilirubin, and jaundice days prior to hospital admission were the most predictive of progression to ALF. The resulting model yielded an overall prediction accuracy of 73%, sensitivity of 76%, specificity of 73%, and area under the receiver operating curve of 0.82. Etiologies with higher likelihood of progressing to ALF were autoimmune hepatitis, drug induced liver injury and indeterminate. APAP overdose, hepatitis B and shock/ ischemia were not as likely to progress to ALF. High values of INR, bilirubin, and jaundice days prior to hospital admission were associated with higher likelihood of progressing to ALF.