Validation of a Machine Learning Algorithm, EVendo, for Predicting Esophageal Varices in Hepatocellular Carcinoma

Background: Treatment with atezolizumab and bevacizumab has become standard of care for advanced unresectable hepatocellular carcinoma (HCC) but carries an increased gastrointestinal bleeding risk. Therefore, patients are often required to undergo esophagogastroduodenoscopy (EGD) to rule out esophageal varices (EV) prior to initiating therapy, which can delay care and lead to unnecessary procedural risks and health care costs. In 2019, the EVendo score was created and validated as a noninvasive tool to accurately screen out patients who were at low risk for having EV that required treatment. We sought to validate whether the EVendo score could be used to accurately predict the presence of EV and varices needing treatment (VNT) in patients with HCC.

Methods: This was a retrospective multicenter cohort study of patients with HCC from 9/2004 to 12/2021. We included patients who underwent EGDs within 1 year after their HCC diagnosis. We collected clinical parameters needed to calculate an EVendo score at the time of EGD and compared the EVendo model prediction to the gold standard endoscopic report in predicting presence of VNT.

Results: 112 with HCC were recruited to this study, with 117 qualifying EGDs. VNT occurred in 39 (33.3%) patients. The EVendo score had a sensitivity of 97.4% and a negative predictive value of 96.9%, supporting the validity in applying EVendo in predicting VNT in HCC.

Conclusion: In this study, we validated the use of the EVendo score in ruling out VNT in patients with HCC. The application of the EVendo score could safely defer about 30% of EGDs for EV screening in HCC patients. Although additional validation cohorts are needed, this suggests that EVendo score can potentially be applied in patients with HCC to avoid unnecessary EGDs, which can ultimately mitigate healthcare costs and delays in initiating HCC treatment with atezolizumab and bevacizumab.