Large language models show promise in predicting liver cancer treatment outcomes

A research team led by Prof. Li Hai from the Hefei Institutes of Physical Science of the Chinese Academy of Sciences has become the first to systematically explore how large language models (LLMs) can assist in predicting liver cancer treatment responses—offering a new path toward AI-powered precision medicine.

The findings were published in the Journal of Medical Systems.

Hepatocellular carcinoma (HCC) is one of the most common and deadly cancers worldwide. For patients with advanced HCC, combination therapies such as immune checkpoint inhibitors and targeted treatments offer some hope, but only about 30% of patients respond effectively. This makes accurate prediction of treatment response a critical unmet need in personalized oncology.

In this study, the researchers evaluated the performance of leading LLMs—GPT-4, GPT-4o, Google Gemini, and DeepSeek—in predicting treatment outcomes using zero-shot learning. This means the models were not specifically trained on liver cancer data beforehand. The dataset included clinical and imaging information from 186 inoperable HCC patients.

To enhance performance, the researchers tested various decision-making strategies, such as voting rules and logical combinations, and created a hybrid model named Gemini-GPT.

The Gemini-GPT model demonstrated predictive accuracy on par with senior doctors with more than 15 years of experience, while outperforming junior and midlevel clinicians in both speed and accuracy. It consistently produced stable results across various treatment types and disease stages, and proved especially reliable in identifying patients likely to benefit from therapy—often showing greater consistency than human doctors.

Applying simple logical strategies further improved its practical utility in clinical settings.

“This study shows how AI can help doctors make better decisions and offer more personalized treatment for cancer patients,” said Prof. Li Hai.

The work marks an important step toward trustworthy AI integration in real-world oncology, demonstrating that LLMs can do more than language—they can reason, predict, and support critical medical decisions.