Extracting lung cancer staging descriptors from pathology reports: A generative language model approach

Area of Science:

• Oncology
• Medical Informatics
• Natural Language Processing

Background:

• Electronic health records contain vital oncology data but are challenging to process.
• Natural Language Processing (NLP) and large language models (LLMs) offer solutions for cancer research.
• Extracting pathological staging information from reports can update cancer staging.

Purpose of the Study:

• Evaluate fine-tuned generative language models (GLMs) for extracting lung cancer pathological staging information.
• Assess the feasibility of using smaller GLMs in resource-constrained environments.

Main Methods:

• Collected lung cancer surgical pathology reports from a tertiary hospital.
• Defined 42 key descriptors for tumor-node (TN) classification, creating a gold standard.
• Trained and evaluated GLMs using prompt-response pairs derived from reports and the gold standard.

Main Results:

• Six GLMs were trained and evaluated for information extraction and TN classification.
• The Deductive Mistral-7B model achieved the highest performance.
• Achieved 92.24% exact match ratio for information extraction and 0.9876 accuracy for concurrent T and N classification.

Conclusions:

• Training GLMs with deductive datasets enhances information extraction performance.
• Small GLMs (approx. 7 billion parameters) can achieve high performance.
• The GLM-based method supports clinical decision-making, lung cancer staging, and research.