Trajectory-Based Integration of Biomarker Dynamics and Clinical Features for Prognostic Prediction in Hepatocellular Carcinoma: A Retrospective Cohort Study.

Purpose: Dynamic changes in circulating biomarkers may reflect tumor biology and host responses more accurately than static measurements. This study aimed to develop and internally validate a dynamic prognostic model for progression-free survival (PFS) and overall survival (OS) in patients with hepatocellular carcinoma (HCC) by integrating longitudinal biomarker trajectories with clinical variables. Patients and Methods: We retrospectively analyzed 379 patients with HCC treated at a single center in China. Latent class mixed models were used to identify circulating biomarker trajectories, which were integrated with clinical variables to construct the Integrated Multi-Biomarker Trajectory (IMBT) model. The model was compared with BCLC stage, a baseline biomarker model, and an AFP trajectory model. Discrimination, calibration, and clinical utility were assessed using C-indices and time-dependent AUCs, calibration plots, calibration slopes and Brier scores, and decision curve analysis (DCA), respectively. Internal subgroup validation was performed in a treatment-defined subset of 327 patients from the same center and source cohort who received PD-(L)1 inhibitor plus molecular targeted therapy. A predefined AFP-negative subgroup analysis was performed using baseline/cycle-0 AFP <20 ng/mL, which was used only to define the subgroup. Results: AFP, D-dimer, and absolute lymphocyte count (ALC) trajectories independently predicted PFS and OS. Sharp-falling AFP trajectories were associated with favorable outcomes, whereas persistently high or rising AFP and D-dimer trajectories and lower ALC trajectories indicated poorer prognosis. The IMBT model achieved C-indices of 0.741 for PFS and 0.728 for OS. For PFS, the C-index exceeded those of the BCLC, baseline biomarker, and AFP trajectory models by 0.108, 0.124, and 0.039, respectively. The corresponding C-indices were 0.730 and 0.724 in the internal PD-(L)1 plus MTT subgroup. Calibration slopes at 12 and 24 months were 1.176 and 1.509 for PFS and 0.898 and 1.014 for OS, respectively, while DCA demonstrated positive net benefit across clinically relevant thresholds. In AFP-negative patients, the PFS/OS C-indices were 0.647/0.634 in the full cohort and 0.692/0.683 in the PD-(L)1 plus MTT subgroup. Conclusion: Longitudinal trajectories of AFP, D-dimer, and ALC provide independent and complementary prognostic information. By integrating dynamic biomarker patterns with clinical features, the IMBT model improved risk stratification compared with conventional staging, baseline biomarker, and AFP-only trajectory models. These findings support further prospective external validation of the model for dynamic prognostic assessment and risk-adapted monitoring in HCC.