Effective Natural Language Processing Algorithms for Gout Flare Early Alert from Chief Complaints

In this study, we extend the exploration of gout flare detection initiated by Osborne, J. D. 16et al, through the utilization of their dataset of Emergency Department (ED) triage nurse chief com- 17plaint notes. Addressing the challenge of identifying gout flares prospectively during an ED visit, 18where documentation is typically minimal, our research focuses on employing alternative Natural 19Language Processing (NLP) techniques to enhance the detection accuracy. This study investigates 20the application of medical domain-specific Large Language Models (LLMs), distinguishing between 21generative and discriminative models. Models such as BioGPT, RoBERTa-large-PubMed-M3, and 22BioElectra were implemented to compare their efficacy with the original implementation by Os- 23borne, J. D. et al. The best model was Roberta-large-PM-M3 with a 0.8 F1 Score on the Gout-CC-2019 24dataset followed by BioElectra with 0.76 F1 Score. We concluded that discriminative LLMs per- 25formed better for this classification task compared to generative LLMs. However, a combination of 26using generative models as feature extractors and employing SVM for the classification of embed- 27dings yielded promising results comparable to those obtained with discriminative models. Never- 28theless, all our implementations surpassed the results obtained in the original publication.