DEEP LEARNING MODEL FOR PREDICTION OF PROGNOSIS IN PATIENTS WITH ACUTE-ON-CHRONIC LIVER FAILURE

초록

Background: Acute on Chronic Liver Failure (ACLF) is associated with high mortality due to multisystem organ failure. However, the accurate prediction of mortality is to help guide better dialysis therapies and wait longer for liver transplantation. We hypothesized that artificial intelligence models are more precise than standard models for predicting outcomes in ACLF. Methods: A novel deep-learning-based model was developed from data collected from patients with MELD ≥21 prospectively from July 2015 to August 2018 and retrospectively from January 2013 to December 2013 from the Korean Acute-on-Chronic Liver Failure (KACLiF) cohort. The prospective data was split into training and validation sets in a 7:3 ratio and used as a derivation cohort (n = 294), while the retrospective data was used as a validation cohort (n =177). We used a Decision Tree, Minimum Redundancy Maximum Relevance (MRMR), Elastic Net, and least absolute shrinkage and selection operator to refine the selection of important features. The selected features were evaluated using Random Forest to assess their ability to predict survival. Results: The mean MELD scores of the derivation cohort and the validation cohort were 29.6 (interquartile range [IQR] 25.5-33.4) and 28.2 (IQR, 24.7-33.6), respectively. The 30-day mortality rates of patients in the derivation and validation cohorts were 25.9% and 25.4%, respectively, and the 90-day mortality rates were 33.7% and 36.7%, respectively. For predicting 90-day mortality, the MRMR demonstrated the highest AUC in both derivation (0.745) and validation (0.617) cohorts. Baseline prothrombin time, previous acute deterioration events, circulatory failure, ascites, hepatic encephalopathy, GI bleeding, body temperature, respiratory rate, cause of liver disease, and albumin were the top features determining the 90- day outcomes. Although it demonstrated low predictive power, this model had the highest AUC among the previously reported models: MELD (0.522), MELD-Na (0.623), MELD 3.0 (0.543), CLIF-ACLF score (0.650), CLIF-C OFs (0.676), and CLIF-SOFA score (0.640). Conclusion: The deep-learning-based model performed better than the previous models for predicting mortality in patients with ACLF.