CAPTYN, a six-variable machine-learning model predicting clinical benefit of atezolizumab-bevacizumab in hepatocellular carcinoma: Development and external validation in IMbrave150

Background: Current prognostic models for hepatocellular carcinoma (HCC) treated with atezolizumab plus bevacizumab (AB) rely on limited variables and lack prospective validation. We aimed to develop and externally validate a machine-learning model integrating multiple clinical variables to predict clinical benefit to first-line AB in advanced HCC. Methods and Results: This multicenter study included 637 patients with unresectable HCC from three hospitals and one phase III prospective trial (IMbrave150), grouped into four AB cohorts. The training set comprised patients from CHA Bundang Medical Center (Korea, n=301) and the Medical University of Vienna (Austria, n=53), while external validation used IMbrave150 (n=99) and Severance Hospital (Korea, n=184) cohort. Clinical benefit (CB) was defined as CR, PR, or SD with PFS ≥6 months by RECIST v1.1; all other cases were classified as non-clinical benefit (NCB). Among 14 candidate variables, nine were identified by univariable Cox regression for OS and PFS. A recursive elimination procedure maximizing five-fold cross-validated AUC for NCB classification identified six optimal predictors-CRP, AFP, platelet, total bilirubin, lymphocyte, and neutrophil-which were used to train an XGBoost classifier, termed CAPTYN. In the training set, CAPTYN achieved an AUC of 0.93. SHAP-based interpretation showed that elevated CRP, AFP, and total bilirubin and reduced lymphocyte counts contributed to NCB, whereas platelet and neutrophil counts exhibited U-shaped associations. In external validation, CAPTYN achieved AUCs of 0.70 (95% CI, 0.59-0.81) in IMbrave150 cohort and 0.67 (0.59-0.75) in Severance cohort, outperforming CRAFITY, ALBI, and CRAPT-M (DeLong's test p<0.05). Calibration was acceptable (Brier score=0.22 and 0.24, respectively), and CAPTYN significantly stratified OS and PFS (IMbrave150 cohort: both p<0.001; Severance cohort: p=0.012 for OS, p=0.009 for PFS), whereas comparator models failed to discriminate PFS. Subgroup analyses across demographics and disease features in IMbrave150 consistently showed higher hazard ratios (>1.5) for OS and PFS in CAPTYN-predicted NCB patients. Conclusion: CAPTYN, a six-variable machine-learning model predicting CB to AB, was externally validated using a prospective trial and a real-world cohort, providing calibrated, interpretable probabilities that may inform individualized treatment decisions.