Extracting structured data from unstructured breast imaging reports with transformer-based models.

Area of Science:

• Natural Language Processing
• Medical Informatics
• Artificial Intelligence

Background:

• Clinical data is often unstructured free text, hindering research and decision-making.
• Structured clinical data is crucial for research and informed decision-making.
• This study addresses the challenge of converting unstructured breast imaging reports into structured data.

Purpose of the Study:

• To compare the performance of BERT-based and generative language models in structuring breast imaging reports.
• To evaluate models for converting unstructured text into tabular data for clinical and research use.
• To assess the efficacy of natural language processing in medical data extraction.

Main Methods:

• Evaluated five transformer-based models (BlueBERT, BioBERT, BioMedBERT, BioGPT, ClinicalT5) on 286 Spanish breast imaging reports translated to English.
• Employed classification for 19 categorical variables and extractive question answering for 4 entities.
• Tested various fine-tuning strategies and input configurations, using accuracy and macro F1 scores for evaluation.

Main Results:

• BioGPT achieved the highest performance in classification (96.10% accuracy, 90.30% F1 score), outperforming BERT-based models.
• BioGPT showed strong performance in extractive question answering (93.24% accuracy), comparable to other top models.
• BioGPT uniquely offered simultaneous classification and question-answering capabilities.

Conclusions:

• Generative models, especially BioGPT, provide a scalable solution for automating structured information extraction from breast imaging reports.
• BioGPT's superior performance and multi-task capability can significantly reduce manual data curation efforts.
• The findings support the efficient integration of imaging data into research and clinical workflows using advanced NLP.