A large language model-based named entity recognition framework for med-sig parsing

Medication Signatures (med-sigs) provide essential instructions for medication use, often documented with shorthand and abbreviations. While there is a widely accepted list of common abbreviations, these shortcuts can lead to medication errors, resulting in an estimated 44,000 to 98,000 hospital deaths annually in USA and costing between $37.6 to $50 billion in healthcare expenses, disability, and lost productivity. Standardising and translating medsigs across medical facilities is crucial. Natural language processing (NLP) and named entity recognition (NER) technologies are key in automating the interpretation of medical prescriptions, breaking down complex instructions into identifiable elements. This paper analyses state-of-the-art NER med-sig parsing models, evaluates their efficacy, and identifies gaps in their application. We propose adaptations and develop a pipeline using GPT-4 for NER on med-sigs. Analysing a dataset of 177 med-sigs, our pipeline outperformed nine existing parsing models, demonstrating its effectiveness.