Background: The prognosis in hepatitis B virus-associated acute-on-chronic liver failure (HBV-ACLF) is challenging due to heterogeneity. Radiomics may enable noninvasive outcome prediction.
Objective: This study aimed to evaluate ultrasound-based radiomics for predicting outcomes in HBV-ACLF.
Methods: We enrolled 264 HBV-ACLF patients, dividing them into a training cohort (n=184) and a validation cohort (n=80). From hepatic ultrasound images, 455 radiomic features were extracted. Radiomics-based phenotypes were identified through unsupervised hierarchical clustering. A radiomic signature was developed using a Cox-LASSO algorithm to predict 30-day mortality. Furthermore, we integrated the signature with independent clinical predictors via multivariate Cox regression to construct a combined clinical-radiomic nomogram (CCR-nomogram). Integrated discrimination improvement (IDI) and net reclassification improvement (NRI) assessed performance improvements achieved by adding radiomic features to clinical data.
Results: Both clustering and radiomic signature identified two distinct subgroups with significant differences in clinical characteristics and 30-day prognosis. In the training cohort, the signature achieved a C-index of 0.746, replicated in validation with a C-index of 0.747. The CCR-nomogram achieved C-indices of 0.834 and 0.819 for the training and validation cohorts. Incorporating radiomic features significantly improved the CCRnomogram over the signature and clinical-only models, evidenced by IDI of 0.108-0.264 and NRI of 0.292-0.540 in both cohorts (all p0.05).
Conclusion: Ultrasound-based radiomics offered prognostic information complementary to clinical data and demonstrated potential to enhance outcome prediction in HBV-ACLF.
Keywords: acute-on-chronic liver failure; hepatitis B virus; machine learning; nomogram; radiomics; ultrasound.
Copyright© Bentham Science Publishers; For any queries, please email at epub@benthamscience.net.