Multimodal Modeling Distinguishes Treatment Response from Overall Survival in Hepatocellular Carcinoma Receiving...

While combined interventional therapies and targeted immunotherapy have improved outcomes for unresectable hepatocellular carcinoma (HCC), radiographic tumor shrinkage does not guarantee prolonged survival (the "responder paradox"). We hypothesized that these distinct endpoints may be associated with different clinical and biological factors and require separate predictive strategies. This study aimed to develop the predictive radiomics-integrated multimodal estimation (PRIME) system to independently predict and biologically characterize treatment response versus overall survival (OS).

In this multicenter study comprising 246 patients receiving combined interventional therapies and targeted immunotherapy, we integrated clinical data and dual-phase CT radiomics. We employed multimodal fusion algorithms to construct two models: PRIME-R (predicting 3-month objective response) and PRIME-S (predicting OS). Model performance was rigorously tested in an external validation cohort. Furthermore, we utilized SHapley Additive exPlanations and matched imaging-transcriptomic data to elucidate the specific clinical and molecular features associated with each outcome.

The PRIME models demonstrated superior accuracy compared to single-modality approaches. In external validation, PRIME-R achieved an AUC of 0.85, while PRIME-S achieved a C-index of 0.72. Significant risk stratification was confirmed (HR = 3.31, 95% CI: 2.06-5.31, P < 0.001). Crucially, feature analysis revealed a distinct divergence: PRIME-R was primarily driven by tumor morphology. In contrast, PRIME-S relied heavily on systemic inflammation and liver function reserve. Transcriptomic profiling supported this interpretation, showing that response was associated with acute immune activation, whereas survival was associated with metabolic adaptation and tissue homeostasis.

Short-term response and long-term survival in HCC are distinct clinical endpoints associated with different biological drivers. The PRIME system effectively distinguishes these outcomes, offering insights into the responder paradox and supporting dual-endpoint risk stratification pending further prospective validation.