RxBERT: Enhancing drug labeling text mining and analysis with AI language modeling.

Area of Science:

• Computational linguistics
• Regulatory science
• Artificial intelligence in medicine

Background:

• US drug labeling documents are vital for assessing drug efficacy and safety but pose challenges for traditional text mining due to their volume and free-text nature.
• Advances in artificial intelligence (AI) and natural language processing (NLP) offer new methods for extracting critical information from these complex regulatory documents.
• Existing NLP models may not be optimally tuned for the specific nuances and structure of FDA drug labeling data.

Purpose of the Study:

• To develop and evaluate RxBERT, a transformer-based NLP model specifically pre-trained on FDA human prescription drug labeling documents.
• To enhance the processing and utilization of drug labeling information for both research and regulatory review purposes.
• To demonstrate the potential of customized large language models (LLMs) for sensitive regulatory document analysis.

Main Methods:

• Developed RxBERT, a Bidirectional Encoder Representations from Transformers (BERT) model, by further pre-training BioBERT on FDA human prescription drug labeling documents.
• Evaluated RxBERT on multiple regulatory datasets, including the NIST TAC dataset, the FDA ADE Eval dataset, and a US Drug Labeling dataset for text classification.
• Compared RxBERT's performance against established NLP models like BERT and BioBERT.

Main Results:

• RxBERT achieved competitive or superior performance compared to other NLP approaches.
• Achieved 86.5 F1-scores on both the NIST TAC and FDA ADE Eval classification tasks.
• Reached a prediction accuracy of 87% on the US Drug Labeling dataset for classifying texts into labeling sections.

Conclusions:

• RxBERT, a transformer model tailored for drug labeling, demonstrates superior performance over the original BERT model.
• RxBERT shows significant potential to aid researchers and FDA reviewers in processing drug labeling information, thereby advancing drug effectiveness and public health safety.
• The study validates a pathway for creating customized LLMs for specialized, sensitive regulatory data.