Researchers present HCC-STAR, a large language model designed to read routine electronic medical record narratives for risk stratification and treatment guidance in hepatocellular carcinoma. The system jointly outputs risk score-based staging, ranked guideline-consistent treatments with evidence-based rationales, and individualized survival estimates.
- Trained on approximately 30,000 HCC cases from SEER expanded into EMR-style narratives via clinician-validated augmentation.
- Optimized using a step-verifiable composite reward to move beyond text-level memorization of clinical guidelines.
- Achieved state-of-the-art performance in a multi-center cohort of 6,668 patients across 12 hospitals in China.
- Surpassed GPT-5 and Gemini-2.5 Pro in treatment recommendation and risk stratification benchmarks.
- Hypothetical analysis showed median survival of 51 months under HCC-STAR recommendations versus 29 and 32 months for BCLC and CNLC guidelines.
- Blinded specialists rated the model's reasoning trustworthy, and it helped physicians make more accurate decisions faster as an assistant.
The findings support HCC-STAR as a reliable and verifiable decision-support system that enhances precision therapy and clinical outcomes for hepatocellular carcinoma patients.