ACHBLF was defined as an acute hepatic insult manifesting as jaundice and coagulopathy, complicated within 4 weeks by ascites and/or encephalopathy in a patient with chronic HBV infection according to consensus recommendations of the Asian Pacific Association for the Study of the Liver in 2009. ACHBLF patients were assigned to a training cohort and a validation cohort randomly. One of the major limitations of ANN is over-training, which can lead to good performance on training sets but poor performance on relatively independent validation sets. To avoid over-training during building ANN, a part of ACHBLF patients were again randomly selected from the training group to train the network and the remaining were used for cross-validation.

ACHBLF was defined as an acute hepatic insult manifesting as jaundice and coagulopathy, complicated within 4 weeks by ascites and/or encephalopathy in a patient with chronic HBV infection according to consensus recommendations of the Asian Pacific Association for the Study of the Liver in 2009. To avoid over-training during building ANN, a part of ACHBLF patients were again randomly selected from the training group to train the network and the remaining were used for cross-validation.

The routine therapy of patients were same, including absolute bed rest, energy supplements and vitamins, intravenous drop infusion albumin, maintenance water, electrolyte and acid-base equilibrium, and prevention and treatment complications, etc. The start date of the follow-up was the date of the diagnosis of ACHBLF. In this study, patients receiving liver transplantation within 3 months were considered as death. All patients with ACHBLF were followed up for at least 3 months and the outcome (death or survival) of corresponding patient was recorded.

MELD score (R = 9.57 × ln (creatinine (mg/dL)) + 3.78×ln (bilirubin (mg/dL)) + 11.2×ln (INR) + 6.43) was used to measure the mortality risk in patients with end-stage liver disease. Given the lack of donors, MELD was used as organ allocation tool to increase graft success rate and patient survival rates, which was generally accepted. Recently, some adjustments were added to the original MELD formula to overcome limitations of MELD score. Published data suggested that MELD-Na (R = MELD + 1.59 × (135 - serum sodium (mmol/L))) might improve the prognostic accuracy [5]. Furthermore, several other scoring systems such as MELDNa (R = MELD - serum sodium (mmol/L) - (0.025 × MELD × (140 - serum sodium (mmol/L))) + 140), MESO (R = (MELD/serum sodium (mmol/L)) × 100), iMELD (R = MELD + (age(year) × 0.3) - (0.7 × serum sodium (mmol/L)) + 100)), etc had been described for predicting the mortality of end-stage liver disease accurately.

ANN can mimic a biological neural system both structurally and functionally. It consists of a set of highly complex, interconnected processing units (neurons) linked with weighted connections, and include an input layer, an output layer and one or more hidden layers. The input layer contains neurons which receive the data available for the analysis (e.g. various clinical, demographic or laboratory data) and the output layer contains neurons which export different predictive outcomes (e.g. clinical diagnosis or prognosis). The hidden layers are used to allow complex relations between the input and output neurons to evolve.In this study, we built ANN by using a graphical neural network development tool NeuroSolution V5.05 (Neurodimension, Florida, United State).

Statistical analysis was performed using SPSS 13.0 software and MedCalc 10.0 software. The Kolmogorov-Smirnov test was applied to determine whether sample data were likely to be derived from a normal distribution population. Continuous variables were expressed by mean ± standard deviation and compared using Wilcoxon signed rank test or Mann-Whitney U test when necessary. Categorical variables were described by proportions or count and compared using proportions Chi-square test or the Fisher's exact test when necessary.

Performances of the ANN prediction in the training cohort and in the validation cohort were tested using ROC analysis, in which AUROC was used to compare the performance of ANN and MELD-based scoring series using the Hanley and McNeil method. A value of P < 0.05 was considered statistically significant.