Revolutionizing Disease Diagnosis with Large Language Models: A Systematic Review

Purpose: The underlining success of large language models (LLMs) in general language understanding has revolutionized many domains, including healthcare. This systematic review aims to explore the application of LLM in disease diagnosis, acknowledging its success in reducing diagnosis delays in areas with limited health resources or when health costs are not affordable. Methods: We systematically reviewed the literature on LLMs used for disease diagnosis tasks, such as differential diagnosis, question-answering, and report generation, from January 2014 to September 2024. The included studies are categorized based on input data types (text, images, others), learning approaches (zero-shot, few-shot, retrieval-augmented generation, fine-tuning, etc.), and diagnostic contexts. Results: We observe that LLMs used for disease diagnosis primarily employ zero-shot learning, prompting, and retrieval-augmented generation techniques. While some LLMs are explicitly fine-tuned for this purpose, only a few have been trained from scratch. Most LLMs are trained using data from the Internet and research publications. Despite this, LLM models have successfully diagnosed many common chronic diseases and addressed some rare and uncommon diseases with reasonable accuracy. Despite these promising trends and successes, concerns about the potential for misdiagnosis persist, particularly with less common and rare diseases. Conclusion: The use of LLMs in disease diagnosis has a promising future in reducing diagnostic delay and complementing the work of health professionals in resource-constrained environments. The success of LLM in disease diagnosis calls for proper analysis of data from health domains and for encouraging the inclusion of information on otherwise uncommon diseases.