Dynamic Early Survival Prediction Model for Hepatocellular Carcinoma Patients Treated With Atezolizumab and Bevacizumab: A Longitudinal Deep Learning Analysis.

Abstract

Background: Atezolizumab combined with bevacizumab has become the standard first-line systemic therapy for unresectable hepatocellular carcinoma (uHCC). Although this regimen offers statistically significant and clinically meaningful benefits, accurately predicting overall survival (OS) remains a challenge. This study aims to identify potential biomarkers to improve early OS prediction in patients with uHCC treated with atezolizumab and bevacizumab.

Methods: A longitudinal survival analysis was conducted using data from the GO30140 and IMbrave150 trials. Multiple deep learning architectures for dynamic survival prediction in HCC (DynSurv-HCC) were evaluated to assess their prognostic performance.

Results: Of 415 patients with unresectable HCC, 291 and 124 were randomly assigned to training and validation sets in a 7:3 ratio. The DynSurv-HCC model with the random survival forest (RSF) method outperformed other deep learning approaches. In the training set, the DynSurv-HCC model achieved AUCs of 0.93 (95% CI: 0.89-0.97), 0.91 (95% CI: 0.87-0.94), and 0.91 (95% CI: 0.84-0.96) at 6, 12, and 24 months, respectively. In the validation set, the model achieved an AUC of 0.90 (95% CI: 0.82-0.98) at 6 months. Importantly, the DynSurv-HCC model demonstrated robust and consistent predictive accuracy across varying etiologies and baseline α-fetoprotein (AFP) levels.

Conclusions: The DynSurv-HCC model with RSF demonstrated promising early OS prediction in patients with HCC receiving atezolizumab and bevacizumab, regardless of etiology or baseline AFP levels. Our findings underscore its clinical potential in guiding personalized treatment strategies and enhancing prognostic assessments for patients with uHCC.