Detection of esophageal varices and prediction of hepatic decompensation in unresectable hepatocellular carcinoma using AI

BACKGROUND & AIMS: In hepatocellular carcinoma (HCC) with cirrhosis, portal hypertension worsens outcomes. Esophagogastroduodenoscopy (EGD), the current screening method for esophageal varices (EVs), is invasive and may delay therapy. We aimed to develop and externally validate non-invasive models to detect EVs and predict hepatic decompensation (bleeding, ascites or hepatic encephalopathy), a major cause of mortality in patients with HCC, using routine contrast-enhanced CT and clinical data. METHODS: This multicenter retrospective study included 489 patients with unresectable HCC treated with atezolizumab-bevacizumab (AtezoBev) from five French centers, divided into a development cohort (n = 279) and an external validation cohort (n = 210). Arterial-phase contrast-enhanced CTs were processed through a Deep Learning pipeline using a foundation model (HepatoSageCT). Logistic and Cox models generated clinical models and combined models integrating the HepatoSageCT scores with key clinical variables for EVs and hepatic decompensation. Performance was assessed using AUROC, sensitivity, specificity, C-index and cause-specific hazard ratios. RESULTS: Portosystemic shunts (PSS) at imaging identified EVs with an AUROC of 0.78, increasing to 0.84 when combined with HepatoSageCT. A decision algorithm incorporating PSS and HepatoSageCT missed 4.2% of varices needing treatment, compared to 8.4% when using only PSS, while missing 0% of large EVs. HepatoSageCT predicted hepatic decompensation in the validation cohort (C-index: 0.73, hazard ratio: 3.17) with significant stratification (p