Abstract

Background:

The IMbrave 050 trial recently demonstrated positive results of the adjuvant atezolizumab plus bevacizumab therapy for hepatocellular carcinoma (HCC) patients at high-risk of recurrence after curative treatment. However, the IMbrave 050 criteria to define high-risk of recurrence might be suboptimal. The aim of this study was to compare the performance of current prediction models and IMbrave 050 criteria in predicting recurrence after curative resection for HCC.

Methods:

Consecutive 1444 HCC patients receiving curative resection were retrospectively enrolled, including 984 (68.1%) patients fulfilling IMbrave 050 high risk criteria. In the high risk group, patients were classified as group 1 (up to three tumors, with largest tumor >5, n=458), 2 (four or more tumors, with largest tumor ≤5 cm, n=15) and 3 (up to three tumors, with largest tumor ≤5 cm with vascular invasion and/or poor tumor differentiation, n=511) by IMbrave 050 definition. The performance of ERASL-post and an artificial intelligence (AI)-derived clinical-radiomic GARSL postoperative model were compared with IMbrave 050 criteria in predicting recurrence after resection.

Results:

Besides tumor size, tumor number and microvascular invasion, AFP level, ALBI grade and FIB-4 score were also independent predictors of early recurrence by multivariate analysis. The median RFS in patients with IMbrave low risk, and high risk groups 1, 2 and 3 were 80.0, 22.0, 16.2 and 49.6 months, respectively (p<0.001). The median RFS in patients with ERASL low-, intermediate-, high-risk groups, and GARSL low-, high-risk groups were 69.1, 21.5, 5.0, 96.6 and 10 months, respectively. The area under the receiver operating characteristic curves (AUCs) of IMbrave 050 criteria, ERASL and GARSL models for predicting early recurrence of HCC within 2 years were 0.614, 0.674 and 0.857, respectively. In the IMbrave high risk group, 46.9% and 56.2% of patients were classified as low risk by ERASL and GARSL models, respectively. In the IMbrave low risk group, 0% and 25.3% of patients were classified as high risk by ERASL and GARSL models, respectively.

Conclusion:

The high risk patients defined by the IMbrave 050 criteria were heterogeneous and outcomes varied widely. Prediction models using more comprehensive prognostic factors, especially the AI-derived GARSL model, performed better to select high-risk candidates for adjuvant immunotherapy.