Extracting structured information from unstructured histopathology reports using generative pre-trained transformer 4...

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Digital pathology

Background:

• Deep learning for whole-slide histopathology images (WSIs) aids precision oncology but requires extensive labeled data.
• Unstructured pathology reports pose a challenge for data extraction, and structured reporting templates increase workload.
• Current methods for generating ground truth data for machine learning are time-consuming and expensive.

Purpose of the Study:

• To investigate the capability of large language models (LLMs) to extract structured data from unstructured histopathological reports.
• To evaluate the feasibility of using a zero-shot approach with LLMs for automated data extraction without model re-training.
• To assess the concordance of LLM-extracted data with human-generated structured data.

Main Methods:

• Utilized GPT-4, a large language model, to process unstructured plain language histopathological reports.
• Applied a zero-shot learning approach, requiring no model re-training.
• Focused on two extensive datasets of pathology reports for colorectal cancer and glioblastoma.

Main Results:

• GPT-4 successfully extracted structured data from unstructured histopathological reports.
• A high concordance was observed between the structured data generated by the LLM and data structured by human experts.
• The zero-shot approach demonstrated effectiveness in extracting relevant information.

Conclusions:

• Large language models can effectively extract structured ground truth data from unstructured pathology reports.
• LLMs offer a potential solution to overcome the data acquisition bottleneck in developing machine learning models for digital pathology.
• Routine use of LLMs could significantly enhance the development of AI tools for precision oncology and pathology.