From Information Extraction to Clinical Reasoning: A Systematic Scoping Review of Large Language Models in Cancer...

Area of Science:

• Oncology
• Medical Informatics
• Natural Language Processing

Background:

• Pathology reports are crucial for cancer diagnosis and staging but are difficult to translate into structured data.
• Existing clinical natural language processing (NLP) research has faced challenges in operationalizing pathology text at scale.
• Large language models (LLMs) present a promising avenue for improving the extraction and interpretation of pathology report data.

Purpose of the Study:

• To synthesize current research on LLM applications in cancer pathology.
• To evaluate LLM capabilities across a four-level framework of the pathology report data lifecycle.
• To identify challenges and future directions for LLM integration in oncology.

Main Methods:

• A systematic literature review following PRISMA-ScR guidelines was conducted.
• Searched four databases for studies on LLMs in cancer pathology up to September 2, 2025.
• Identified and analyzed 41 eligible studies based on a four-level capability framework (text preparation, information extraction, clinical reasoning, interpretive synthesis).

Main Results:

• Most studies focused on information extraction (level 2), with fewer studies addressing clinical reasoning (level 3) and interpretive synthesis (level 4).
• Encoder-based models (e.g., BioBERT) were common for extraction, while generative models (e.g., GPT, LLaMA) showed potential for staging and summarization.
• High performance on extraction tasks often did not translate to higher-level reasoning, particularly with inconsistent reports or missing clinical data.

Conclusions:

• LLM success in basic extraction does not guarantee proficiency in complex clinical reasoning or interpretation.
• Gaps between benchmark results and practical adoption stem from report inconsistencies and the need for inferred clinical assumptions.
• Future research must prioritize multi-site validation, clinically relevant error analysis, transparent evaluation, and privacy-preserving implementation for safe integration in oncology.