Development of a machine learning model for automatic data extraction from breast cancer pathology reports

Area of Science:

• Medical Informatics
• Computational Biology
• Oncology Research

Background:

• Manual data extraction from medical records is time-consuming and prone to errors.
• Natural Language Processing (NLP) and Machine Learning (ML) offer potential for automating clinical data extraction.
• Accurate extraction of clinicopathologic features from breast cancer pathology reports is vital for research and patient outcomes.

Purpose of the Study:

• To develop and evaluate an NLP pipeline for automated data extraction from breast cancer pathology reports.
• To compare the performance of different NLP models, including ClinicalBERT, PubMedBERT, BioMedRoBERTa, and Mistral-Nemo LLM.
• To assess the impact of pre-training on the SQuAD question-answering dataset for improved model accuracy.

Main Methods:

• An NLP pipeline was constructed by selecting the best performing model from four candidates.
• Models were evaluated on 1,795 breast cancer pathology reports.
• BERT-based models underwent further pre-training using the SQuAD dataset.
• Model performance was quantified by comparing extracted data against human annotations.

Main Results:

• PubMedBERT, pre-trained on SQuAD, achieved the highest overall accuracy of 97.4%.
• 30 out of 32 Fields of Interest (FOIs) demonstrated accuracy exceeding 95.0%.
• The developed NLP model outperformed a previous rule-based algorithm (95.6% accuracy).

Conclusions:

• A high-performing question-answering NLP pipeline can automate the extraction of clinicopathologic features from breast cancer pathology reports.
• This automated approach offers a scalable solution for high-fidelity data extraction, enhancing research efficiency.
• Improved data extraction can contribute to better clinical outcomes in breast cancer care.