A clinical AI model for liver cancer staging just outperformed both specialist physicians and frontier models in a large multi-center trial.
Researchers built HCC-STAR, a large language model designed to read routine electronic medical records and produce risk-stratified staging, ranked treatment recommendations with evidence-based rationales, and individualized survival estimates. The team trained it on roughly 30,000 hepatocellular carcinoma cases sourced from the SEER database, expanded into EMR-style narratives through a clinician-validated augmentation workflow. Evaluated across 6,668 patients from 12 hospitals in China, HCC-STAR outperformed GPT-5, Gemini-2.5 Pro, and both resident and attending physicians on treatment accuracy. Blinded hepatobiliary specialists rated its reasoning as trustworthy.
The survival numbers are the sharpest part of the claim: hypothetical analysis put median overall survival at 51 months when following HCC-STAR recommendations, versus 29 and 32 months under the two dominant clinical staging systems, BCLC and CNLC. That gap — if it holds in prospective trials — would be meaningful for a cancer where most patients are diagnosed at advanced stages. The model also helped physicians make more accurate decisions faster when used as an assistant, which is the more defensible near-term use case than full autonomy.
The usual caveats apply: this is a retrospective study, the survival figures are hypothetical, and the training data came from a single national registry. Whether the approach generalizes outside China's hospital network is the next question worth asking.