From diagnosis and treatment to prognosis: Clinical prospects of artificial intelligence in multimodal research of hepatocellular carcinoma.

PURPOSE: This review aims to critically evaluate the evolving role and clinical readiness of multimodal Artificial Intelligence (AI) in Hepatocellular Carcinoma (HCC), addressing the fundamental limitations of traditional single-modality approaches in characterizing tumor heterogeneity. PRINCIPAL RESULTS: Integrative analysis of heterogeneous data sources-specifically radiomics, pathomics, genomics, and clinical variables-demonstrates superior performance over uni-modal baselines in early diagnosis, microvascular invasion prediction, and immunotherapy response monitoring. However, the observed performance in published studies may present an "iceberg effect," where high internal validation metrics mask diminished generalizability to external cohorts, a discrepancy potentially attributable to publication biases and data drift. Comparative assessment indicates that while Late Fusion strategies currently provide greater robustness for clinical workflows, Early Fusion architectures hold promise for deeper biological insight. Furthermore, although Generative AI can alleviate data scarcity through synthetic augmentation, it introduces unaddressed risks of diagnostic hallucinations. SCIENTIFIC VALUE ADDED: Unlike descriptive surveys, this work highlights the critical necessity of shifting from passive data acquisition to active sensing and transitioning from correlation-based "black box" models to Causal AI and Chain-of-Thought reasoning to establish clinical trust. CONCLUSIONS: We conclude that bridging the translational gap requires immediate adherence to standardized reporting guidelines like TRIPOD-AI to ensure reproducibility. The future of HCC precision medicine lies in evolving toward pre-trained Foundation Models and "Digital Twins," transforming clinical management from empirical reliance to mechanism-driven computational intelligence over the next decade.