Large Language Model Framework for Structured Biomarker Extraction and Triple-Negative Breast Cancer Identification...

Triple-negative breast cancer (TNBC) is an aggressive molecular subtype of breast cancer that needs to be identified accurately from pathology reports to inform decision-making. Large language models (LLMs) have been reported to show great potential in the automatic extraction of important oncology information from unstructured medical text.

We aimed to create and test an LLM-based approach for automatic biomarker extraction and TNBC subtype classification based on breast cancer pathology reports using prompt-engineering techniques.

In total, 532 breast cancer pathology reports from The Cancer Genome Atlas (TCGA) were used. Our approach extracted structured oncology features, such as estrogen receptor status, progesterone receptor status, HER2 status, tumor size, tumor grade, and tumor histology, which were used to identify TNBC. We tested 4 frontier models: GPT-4o mini (released July 18, 2024), GPT-4.1 nano (released April 14, 2025), GPT-5 nano (released August 7, 2025), and Gemini 2.5 Flash-Lite.

The designed framework was able to extract relevant biomarkers and clinicopathological attributes from unstructured pathology reports and allowed for the automated identification of patients with TNBC. Gemini 2.5 Flash-Lite showed superior performance in the extraction of clinicopathological attributes. For all models investigated, chain-of-thought (CoT) prompting was shown to outperform other strategies, such as zero-shot and few-shot prompting, for both biomarker extraction and TNBC classification. GPT-4o mini using CoT prompts reported the highest performance for TNBC classification.

Using prompt-engineering techniques based on LLMs proved to be an efficient method of translating unstructured breast cancer pathology reports into structured clinical information. Moreover, it can be used to accurately identify the TNBC subtype.