The inevitable transformation of medicine and research by large language models: The possibilities and pitfalls

Large language models (LLMs) often refer to artificial intelligence models that consist of extensive parameters and have the ability to understand and generate human-like language. They are typically developed in a self-supervised learning manner and are trained on large quantities of unlabeled text to learn patterns in language. LLMs were initially used in natural language processing (NLP), but they have since been extended to a variety of tasks like processing biological sequences and combining text with other modalities of data. LLMs have the potential to revolutionize the way we approach scientific research and medicine. For example, by leveraging their ability to understand and interpret vast quantities of text data, LLMs can provide insights and make predictions that would otherwise be impossible.

In the medical domain, LLMs can be used to analyze immense electronic health records and improve communication between healthcare professionals and patients. For example, LLMs can be used to automate triage, medical coding, and clinical documentation, which can help to improve the accuracy and efficiency of these processes. They can also be used to improve NLP in medical chatbots and virtual assistants, allowing patients to interact with healthcare services more efficiently and effectively. They can also be used to process medical records and patient data, enabling better diagnoses and more personalized treatments. They can also be used to analyze clinical trial data and identify trends that could lead to better outcomes. Finally, LLMs can also be used to answer medical questions and provide guidance to healthcare professionals, which can help to improve the quality of care. In the accompanying Review, Zheng et al.¹ undertake a major effort to write a comprehensive review of this exciting and highly evolving field.

In research, LLMs can be used to search through diverse large datasets and identify patterns that would otherwise be difficult to detect. They can also be used to generate and test hypotheses and to summarize and analyze research papers. It is clear that LLMs will be transforming the way we communicate about medicine and research, and have the potential to revolutionize the field of healthcare.

The current state-of-the-art LLM is Generative Pre-trained Transformer 4 (GPT-4), developed by OpenAI, about which Technical details have not been made public yet.² Based on publicly available information, the number of parameters is comparable to its previous generation, GPT-3, which consists of 175 billion parameters. GPT-4 is a generative model, meaning it can generate human-like language and even create original content. Other notable LLMs include GPT-3, Bidirectional Encoder Representations from Transformers, and Text-to-Text Transfer Transformers, each with its unique strengths and capabilities. However, one example of an LLM developed specifically for the medical domain is GatorTron,³ which can process and interpret electronic health records. GatorTron was developed by a team of researchers from the University of Florida. The model is trained on >90 billion words of text, including >82 billion words of deidentified clinical text. GatorTron achieves good performance on five clinical NLP tasks, including clinical concept extraction, medical relation extraction, semantic textual similarity, natural language inference, and medical question answering. Besides, the results show that scaling up the number of parameters and the size of the training data can significantly improve the performance of these clinical NLP tasks. GatorTron's ability to accurately process unstructured clinical text can enhance medical AI systems and improve healthcare delivery. GatorTron is an example of the potential of LLMs to be tailored to specific domains or industries, allowing for more accurate and efficient language processing in specialized fields.

Despite the many potential benefits of LLMs in medicine and research, there are also risks and concerns. LLMs could be exploited for spreading false information or manipulating public opinion, such as during global health crises. LLMs are also fundamentally trained on all available information or data, including inaccuracies and biases. These inaccuracies and biases can be reflected in the output of hallucination, which refers to mistakes in the generated text that are semantically or syntactically plausible but are in fact incorrect or nonsensical. There are also privacy concerns with LLMs because they can potentially access and process sensitive personal data.

It is ultimately difficult to hold LLMs accountable for their outputs. Therefore, the accountability ultimately rests on the user. Human oversight and governance of LLM outputs, especially in medicine and research, is paramount. The implementation of LLMs in healthcare has to be subjected to the same rigor and standards as any other new interventions through clinical trials, to demonstrate that the application of LLMs is at least noninferior to current approaches. Ultimately, the use of LLMs in medicine and research requires a shared responsibility among all stakeholders, including researchers, technology companies, regulatory bodies, and society as a whole.

The power and potential of LLMs mean that it is here to stay, and its widespread implementation is inevitable. Recognition of its potential and ethical implementation is essential to ensure that they are used responsibly and for the benefit of all.

Yuanxu Gao, Daniel T. Baptista-Hon, and Kang Zhang wrote the manuscript. All authors have read and approved the final manuscript.

The authors declare no conflict of interest.