Truace
Subscribe

Large Language Model Framework for Structured Biomarker Extraction and Triple-Negative Breast Cancer Identification From Pathology Reports: Development and Evaluation Study

Background: 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. Objective: We aimed to create and test an LLM-based approach for automatic biomarker extraction and TNBC subtype classification based on breast cancer pathology…

By the Truvace desk2 min read1 sourcePeer-reviewedTRV-2026-2228· cite & verify
Large Language Model Framework for Structured Biomarker Extraction and Triple-Negative Breast Cancer Identification From Pathology Reports: Development and Evaluation Study
"Capillaritis-Acute hemorrhage - Milk allergy Case 107" by Pulmonary Pathology is licensed under CC BY-SA 2.0. To view a copy of this license, visit https://creativecommons.org/licenses/by-sa/2.0/.

In brief

Researchers built a prompting framework that reads free-text breast cancer pathology reports and pulls out structured details such as hormone receptor status, HER2 status, tumor size, grade, and histology. Those fields were then used to pick out patients with the triple-negative subtype. The work used 532 reports from The Cancer Genome Atlas and tested four commercial language models.

The approach matters because triple-negative breast cancer is aggressive and has to be identified accurately from pathology reports to inform decisions, and these reports are usually unstructured. The abstract reports that the method worked, but it gives no numeric performance figures, and it does not show use in routine clinical care.

Main points

  1. The study used 532 breast cancer pathology reports from The Cancer Genome Atlas.
  2. Four models were compared: GPT-4o mini, GPT-4.1 nano, GPT-5 nano, and Gemini 2.5 Flash-Lite.
  3. Chain-of-thought prompting beat zero-shot and few-shot prompting for both biomarker extraction and TNBC classification across all models.
  4. Gemini 2.5 Flash-Lite was best at extracting clinicopathological attributes, while GPT-4o mini with chain-of-thought prompts scored highest for TNBC classification.

The gain

Prompt-engineered language models converted unstructured pathology reports into structured clinical features and identified triple-negative breast cancer patients automatically.

The rundown

The framework extracted estrogen receptor, progesterone receptor, and HER2 status along with tumor size, grade, and histology from unstructured pathology text, then used those structured features to classify TNBC.

Model results differed by task. Gemini 2.5 Flash-Lite led on extracting clinicopathological attributes, but GPT-4o mini using chain-of-thought prompts reported the highest performance for TNBC classification.

Prompting strategy mattered consistently: for every model tested, chain-of-thought prompting outperformed zero-shot and few-shot approaches. The authors conclude that prompt engineering is an efficient route from unstructured reports to structured clinical information.

Sources

  1. Peer-reviewedJMIR Formative Research2026-10-23

The debate