IDDF2021-ABS-0141 Prediction of mortality among acute-on-chronic liver failure patients using artificial neural network analysis

BackgroundDespite the advances in understanding the pathophysiology and management of Acute-on-Chronic liver failure (ACLF), short-term mortality remains high. Though there have been numerous scoring systems to predict mortality, they lack predictive accuracy. We intend to study the accuracy of artificial neural network (ANN) analysis in predicting short mortality among such patients.MethodsPatients who were diagnosed as ACLF according to APASL criteria were included. Retrospective data of ACLF patients were collected, including the presence of decompensating events including ascites, encephalopathy, Gastrointestinal (GI) bleed, acute kidney injury (AKI) and infections. Basic blood investigations were recorded, including various prognostic scoring systems like CTP, MELD, CLIF-SOFA and AARC scores. A multi-layered perceptron ANN model with hidden nodes to make a prediction was constructed from 27 clinical and laboratory variables. The ANN was trained and validated internally using an adaptive moment estimation optimization algorithm. The primary endpoints were 30-day mortality.ResultsOne hundred patients were included with mean age of 46.04 ± 11.28 yrs. 88.4% were male. Most common acute precipitant of ACLF was alcohol followed by infections among 59.82% and 25% of patients respectively. DILI as acute cause was seen in 16.07%, while 7.1% had unknown cause. Commonest etiology of CLD was alcohol followed by viral among 77.27% and 7.95% of patients. 73.2% had overt HE. GI Bleed, AKI and infections were noted in 19.6%, 46.4% and 32.1% respectively. Mortality at the end of 30-days was 36.6%. Presence of AKI and GI bleed were significantly associated with mortality (p