CancerLLM: a large language model in cancer domain

Area of Science:

• Artificial Intelligence in Oncology
• Clinical Natural Language Processing (NLP)
• Computational Pathology

Background:

• Existing medical large language models (LLMs) lack specialization for cancer phenotyping and diagnosis.
• High parameter counts in current LLMs create significant computational challenges in healthcare.
• A need exists for efficient and accurate AI tools in oncology research and clinical practice.

Purpose of the Study:

• To develop and evaluate CancerLLM, a specialized, computationally efficient large language model for cancer applications.
• To assess CancerLLM's performance in cancer phenotype extraction and diagnosis generation.
• To compare CancerLLM's efficiency and robustness against existing LLMs.

Main Methods:

• Developed CancerLLM, a 7-billion-parameter Mistral-style model.
• Trained CancerLLM on 2.7 million clinical notes and 515,000 pathology reports across 17 cancer types.
• Fine-tuned CancerLLM for cancer phenotype extraction and diagnosis generation tasks.

Main Results:

• CancerLLM achieved high F1 scores: 91.78% for phenotyping extraction and 86.81% for diagnosis generation.
• Outperformed existing LLMs by an average F1 score improvement of 9.23%.
• Demonstrated superior efficiency in terms of time and GPU usage, alongside enhanced robustness.

Conclusions:

• CancerLLM offers a specialized and effective solution for cancer phenotyping and diagnosis.
• The model's efficiency and robustness make it suitable for real-world clinical settings.
• CancerLLM has the potential to significantly advance cancer research and clinical practice.