Prediction Of Acute-on-chronic Liver Failure In Patients With Acutely Decompensated Cirrhosis By Deep Learning Based On Two-dimensional Shear Wave Elastography

Purpose:Accurate predictions of acute-on-chronic liver failure(ACLF)development in patients with acute decompensation(AD)of cirrhosis are helpful in prognosis assessment and individualized treatment.This study aimed to predict ACLF develop-ment through deep learning of shear wave elastography(SWE).Methods:This study we prospectively enrolled 288 acutely decompensated patients,follw-up 1-year,of which 202 patients were the primary cohort and 86 patients were the test cohort.Through transfer learning by Resnet-50 to analyze two-dimensional SWE images,a deep learning signature(DLS)was constructed for 1-year ACLF development prediction.A nomogram was established integrating deep learning SWE information and laboratory data based on a Cox regression analysis.Time-dependent receiver operating characteristic(ROC)curve,Harrell’s concordance index(C-index)and Kaplan–Meier survival analysis were assessed predictive performance.Furthermore,the risk of ACLF in subgroups with different alanine aminotransferase,body mass index and cirrhosis etiology statuses was evaluated with the nomogram.Results:Areas under curve(AUCs)of DLS for 28-day,3-month and 1-year ACLF development were 0.911,0.837 and 0.860,and the C-index was 0.821 in the test cohort.Prediction of the combined nomogram was significantly better than that of prognostic scores(all p<0.05).AUCs for 28-day,3-month and 1-year were 0.938,0.925 and 0.914,and the C-index was 0.876 in the test cohort.The nomogram stratified patients in the cohorts and subgroups into high-and low-risk groups of ACLF development(all p<0.05).Conclusions:Deep learning based on 2D-SWE images has the potential to predict the development of ACLF and can accurately stratify the risk of AD patients,contributing to clinical treatment.