Machine learning outperforms large language models for survival prediction in advanced hepatocellular carcinoma: a...

Area of Science:

• Oncology
• Artificial Intelligence
• Medical Informatics

Background:

• Accurate prognostic prediction is crucial for advanced hepatocellular carcinoma (HCC).
• Machine learning (ML) models show promise in outcome prediction.
• The utility of large language models (LLMs) for structured clinical prognostication in HCC is not well-defined.

Purpose of the Study:

• To compare the prognostic performance of ML algorithms against LLMs in advanced HCC.
• To evaluate the ability of ML and LLMs to stratify patients based on survival risk.

Main Methods:

• A multicenter retrospective study of 1031 advanced HCC patients undergoing interventional therapy and targeted treatment.
• Development and comparison of six ML algorithms against two LLMs (ChatGPT-4o, DeepSeek-v3).
• Identification of seven key predictors: age, comorbidities, albumin-bilirubin grade, tumor burden, portal vein tumor thrombus, and alpha-fetoprotein level.

Main Results:

• Support Vector Machine (SVM) achieved the highest AUC (0.658), followed by XGBoost (0.654) in the test cohort.
• LLMs demonstrated limited discriminative ability (AUC = 0.590-0.591), significantly lower than ML models (P < 0.05).
• ML models effectively stratified patients into distinct 1-year survival risk groups, unlike LLMs.

Conclusions:

• ML models significantly outperform current LLMs for structured prognostic prediction in advanced HCC.
• ML models offer more reliable support for risk stratification and clinical decision-making in advanced HCC.
• Further research may explore optimizing LLMs for clinical prognostic tasks.