Based on Medicine, The Now and Future of Large Language Models

Abstract

Objectives

This review explores the potential applications of large language models (LLMs) such as ChatGPT, GPT-3.5, and GPT-4 in the medical field, aiming to encourage their prudent use, provide professional support, and develop accessible medical AI tools that adhere to healthcare standards.

Methods

This paper examines the impact of technologies such as OpenAI's Generative Pre-trained Transformers (GPT) series, including GPT-3.5 and GPT-4, and other large language models (LLMs) in medical education, scientific research, clinical practice, and nursing. Specifically, it includes supporting curriculum design, acting as personalized learning assistants, creating standardized simulated patient scenarios in education; assisting with writing papers, data analysis, and optimizing experimental designs in scientific research; aiding in medical imaging analysis, decision-making, patient education, and communication in clinical practice; and reducing repetitive tasks, promoting personalized care and self-care, providing psychological support, and enhancing management efficiency in nursing.

Results

LLMs, including ChatGPT, have demonstrated significant potential and effectiveness in the aforementioned areas, yet their deployment in healthcare settings is fraught with ethical complexities, potential lack of empathy, and risks of biased responses.

Conclusion

Despite these challenges, significant medical advancements can be expected through the proper use of LLMs and appropriate policy guidance. Future research should focus on overcoming these barriers to ensure the effective and ethical application of LLMs in the medical field.

Introduction

Artificial Intelligence (AI) is an interdisciplinary domain aimed at emulating human intellect through computer systems that execute tasks traditionally performed by humans. Today, AI can multitask, learn, and generalize skills with minimal supervision [1]. Large Language Models (LLMs), a pivotal subset of AI, leverage algorithms to process extensive textual data, predicting subsequent content based on the context, thus producing human-like text [2]. A large language model developed by OpenAI called Generative Pre-trained Transformer (GPT) [3], with its first version (GPT-1) released in 2018 and GPT-4 released in 2023 [4, 5].In November 2022, OpenAI launched a conversational tool based on LLMs called ChatGPT (https://chat.openai.com). This tool gained widespread attention because of its user-friendly interface, human-like responses, and open access [6]. The current large language model behind ChatGPT is the 3.5th generation or 4th generation Generative Pre-trained Transformer (GPT-3.5 or GPT-4) [7]. According to relevant studies, performing GPT-4 is significantly superior to that of GPT-3.5 [8]. With advancing large language models, their applications in the field of medicine are expanding, especially in medical education [9], scientific research [10], clinical practice [11], and nursing [12], demonstrating vast potential (Fig. 1). Nonetheless, these advancements come with challenges, including ethical concerns, professional knowledge gaps, and the risk of generating misleading responses [13]. This review will primarily focus on the applications of large language models (particularly GPT and ChatGPT) in medicine and discuss potential difficulties and challenges, to promote the gradual development of LLMs into medical AI tools that provide accessible, professional support, while adhering to global healthcare standards, for the betterment of all individuals.

Fig. 1.

Large Language Models in Medicine: a simplified architecture behind ChatGPT, b potential applications of large language models in medicine

Application

In Education

In the past decade, artificial intelligence has been widely applied in various fields [14]. In education, the application of artificial intelligence is essential [15]. Large language models have been extensively trained and have accumulated rich medical knowledge. The United States Medical Licensing Examination (USMLE) is a three-step exam necessary for obtaining a medical license. The initial step assesses the candidate's proficiency in applying fundamental scientific concepts in medical practice, the subsequent step evaluates the candidate's comprehension of medical knowledge, skills, and clinical science foundations, and the final step examines the candidate's understanding of medical knowledge, biomedical sciences, and clinical practices [16]. Taking the United States Medical Licensing Examination (USMLE) as an example, even without specific training, ChatGPT can achieve or approach passing standards in various parts of the exam and provide logically coherent explanations based on its extensive medical knowledge. This performance not only highlights its rich reservoir of medical knowledge but also underscores the reliability of large language models (LLMs) in medical education, proving their value as learning and teaching aids [17–19]. Large language models like ChatGPT have demonstrated strong support in various educational settings [20]. They can assist medical educators in designing curricula that meet modern medical needs and enhance students' practical skills by simulating different clinical scenarios. Furthermore, LLMs have been applied to support personalized learning for medical students, significantly improving their ability to master and apply complex medical knowledge. The author will next introduce the specific applications of large language models in the field of education.

In terms of curriculum design, medicine is a complex and multi-layered field, and designing effective medical courses requires a comprehensive consideration of various factors such as the audience, educational objectives, methods, and content [21, 22]. In this process, large language models like ChatGPT show great potential [9]. ChatGPT can assist educators in drafting lesson plans [9]. By assessing needs, analyzing data, and offering insights, ChatGPT helps educators avoid teaching omissions and improve the comprehensibility of educational content [23]. Two educators in Australia used ChatGPT as a "virtual colleague" to assist in curriculum design. Through a Q&A approach, they first discussed the basic framework and educational objectives of the course with ChatGPT, and then gradually refined the specific teaching content and activities [24]. Similarly, an educator in China explored ChatGPT's performance in designing lesson plans for an anatomy course. He asked ChatGPT to provide lesson plans of about 300 words each for the digestive system and the endocrine system. In these lesson plans, ChatGPT first provided a summary introduction and then listed 16 teaching suggestions. These suggestions included not only the design of questions but also recommendations on teaching methods and ways to effectively obtain student feedback [25]. As a highly digitized AI tool, ChatGPT can design courses that meet students' essential digital skills and capabilities to effectively utilize and integrate advanced digital technologies in future medical practices [26]. It also promotes the development of personalized and interactive learning [27], aiding in understanding individual student situations in-depth and customizing courses according to each learner's unique needs to enhance personalized learning experiences. In addition, educators need to consider not only the framework and content of the curriculum but also stay updated with the latest developments in the medical field to ensure the content is cutting-edge and practical. In this regard, educators can directly ask ChatGPT to quickly obtain the latest research findings and medical developments, avoiding the cumbersome steps of traditionally relying on the internet to gather information themselves [13, 28]. This saves educators the time needed to collect information online, allowing them to delve deeper into designing course plans and teaching activities, ensuring that courses can timely reflect the latest trends and knowledge developments in the medical field. Although in practice, teachers have found ChatGPT particularly suitable for quickly constructing the initial framework of a course, saving valuable time for designing detailed, specific lesson plans, the final course content still needs to be manually refined and optimized by teachers according to specific teaching needs to ensure quality and depth.

In addition to curriculum design, personalized learning is also an important component of medical education. AI-based feedback and assessment can help students identify weak areas and make timely adjustments [29]. Students can input information about their strengths, weaknesses, goals, and preferences to ChatGPT, which can generate personalized feedback and assessments for them. This level of personalization ensures consideration of each student's learning habits, enabling more effective assessment feedback [30]. Additionally, using data provided by students, ChatGPT may predict future challenges students may face and provide advanced strategies and resources. This predictive support can reduce student frustration and help them prepare before facing challenges. Through this approach, students can confidently grasp complex medical concepts and skills. Large language models also offer personalized tutoring services—as each output from the model is unique, it can provide customized answers and explanations based on student queries [31, 32]. Moreover, large language models like ChatGPT excel at summarizing materials, simplifying complex information, and aiding students in learning extensive medical knowledge [28, 33]. Although many browser extensions can summarize web content, ChatGPT can summarize key points faster and more accurately. Compared to search engines and textbooks, ChatGPT can investigate unknown questions more efficiently and comprehensively, saving students a significant amount of time [31]. In conclusion, personalized tutoring can help students better understand and apply their acquired knowledge.

In broader educational applications, ChatGPT also performs exceptionally well in standardized patient simulations [34]. In a study, the simulated medical records generated by ChatGPT received high praise from experienced physicians. Among the 10 generated medical records that were evaluated, two received a perfect score of 10, six received a score of 9, and two received a score of 8 [35]. This indicates the feasibility (albeit with some limitations) of using ChatGPT to simulate standardized patients, eliminating the need for preparing medical records or providing additional training to personnel. This saves a significant amount of time, labor, and resources. These differences in scores are mainly because when ChatGPT answers multiple questions at once, its responses can sometimes appear mechanical and rigid, which may not achieve the desired natural fluency in complex clinical dialogues. Nevertheless, using ChatGPT to simulate standardized patients still demonstrates its feasibility. It can eliminate the cumbersome steps of preparing medical records or providing additional training for relevant personnel, thereby saving a significant amount of time, labor, and resources. At the same time, this approach can help students practice communication with patients, improve clinical skills, and enhance their understanding of diseases and treatment principles, better preparing them for future clinical practice. In simulated healthcare communication scenarios, ChatGPT, a powerful large language model, can effectively portray a standardized patient. This allows students to practice patient communication and interaction skills, refine their clinical competencies, and gain a deeper understanding of disease management and treatment principles [2, 30, 36]. Additionally, ChatGPT can serve as a tool for evaluating students' clinical skills [28]. It offers an innovative approach to education, allowing students to gain real clinical experience in simulated environments and prepare for future clinical practice.

Although large language models (LLMs) like ChatGPT have shown their potential in education and various other fields, the biases they may exhibit cannot be overlooked. The output of these models can be biased due to the limitations of their training data or the unconscious biases of developers [37, 38], particularly when applied in sensitive areas such as education and healthcare. This bias is especially significant. Additionally, to prevent students from becoming overly reliant on these large language models, educators need to be vigilant about students potentially providing specific prompts to ChatGPT and copying and pasting the generated answers into their papers, which is a form of academic dishonesty. Educators can address this issue by requiring students to demonstrate their knowledge and skills in more interactive and engaging ways, such as through oral presentations or hands-on activities [39]. In summary, deeply understanding and actively addressing these issues is crucial to ensuring the fairness and effectiveness of LLMs in medical education applications (Table 1).

Table 1.

Different aspects of the application of LLMs in education and the Pros and Cons of each aspect

	Aspect	Pros	Cons
LLMs in Education	Educators	·Assist in curriculum and program design ·Provide professional knowledge insights ·Fill gaps in educational content ·Ensure comprehensive educational content ·Quickly understand the latest developments in the field	May lead educators to overly rely on technology, reducing the motivation for independent knowledge updates Bias in training data may affect the objectivity and comprehensiveness of educational content
	Student learning	·Personalized learning assistant ·AI-based feedback and assessment ·Customized answers and explanations ·Ability to predict and proactively address learning difficulties	Lack of emotional support and intuitive guidance from human teachers Personalized answers may lead to uneven mastery of knowledge Overreliance on LLMs in academia may lead to academic dishonesty Bias in training data may cause learning content to be biased
	Practical application	·Build standardized patients ·Enhance clinical skills through simulated healthcare communication ·Evaluate student clinical skills	May lack the complexity and unpredictability of real-life situations Simulation of patient dialogues may not fully replicate human emotions and subtle nonverbal communication cues Simulated environments and generated content may inherit biases from training data

In Scientific Research

The rapid development of artificial intelligence has ushered in the era of AI in the scientific research field [14]. This section will focus on the application of large language models (LLMs) in medical scientific research, demonstrating how these advanced technologies are driving the forefront of scientific development and outlining the real-world issues present in their application scenarios. Organ-on-a-chip technology can simulate the microenvironment and functionalities of human organs more realistically, providing more accurate models for drug development. Applying artificial intelligence on this basis can make drug evaluation more efficient [40]. Similarly, artificial intelligence can also be applied in organoid research, conducting rapid screening of construction strategies, concise analysis of multiple datasets, and precise preclinical evaluations, among others [41]. Many traditional Chinese medicine mechanisms are still unclear, and using artificial intelligence to assist in pharmacological research can facilitate precise treatment in traditional Chinese medicine [42].

Firstly, large language models have demonstrated their exceptional capabilities in biomedical natural language processing and medical text evaluation. LLMs can effectively enhance the efficiency of medical data processing and analysis, thereby accelerating the research process. For example, the GatorTronGPT model, based on the GPT-3 architecture, is specifically designed for biomedical natural language processing, clinical text generation, and medical text evaluation. This model surpasses many existing models in extracting relationships such as drug-drug interactions, chemical-disease relationships, and drug-target interactions [43]. However, large language models are highly dependent on data, and variations in data quality can lead to deviations in the model's output. Additionally, due to the diversity and complexity of data in the biomedical field, existing models may not fully cover all variables and anomalies, which can affect their universality and robustness in specific applications [44].

The natural language processing (NLP) capabilities of LLMs can also be applied to the processing of genetic and protein structure data. This ability stems from the fact that genetic and protein information is often expressed in textual form, allowing large language models to extract valuable information by understanding and analyzing the text. Specific applications include predicting protein structures based on amino acid sequences, designing protein sequences with specific biological functions, and identifying promoter regions in bacterial DNA [7]. However, these applications rely on the model's deep understanding and accurate parsing of biological texts. At the current stage, LLMs still face challenges such as high computational complexity and significant resource consumption when processing high-dimensional biological data.

In terms of experimental design and execution, LLMs also offer significant advantages. LLMs can assist not only in data analysis but also play an important role in experimental automation. For example, the GPT-4-driven Coscientist system can automatically design, plan, and execute complex experiments, including tasks such as chemical reaction optimization, hardware control, and data analysis [45]. The Suzuki and Sonogashira reactions, discovered in the 1970s, utilize palladium metal as a catalyst to facilitate the formation of chemical bonds between carbon atoms in organic molecules [46, 47]. These reactions play a critical role in developing novel drugs, particularly for treating inflammation, asthma, and other diseases. In the final testing phase, Coscientist was tasked with executing the Suzuki and Sonogashira reactions. In less than four minutes, Coscientist designed a precise program based on the team's provided chemicals to achieve the target reaction. The Coscientist system demonstrated its powerful capabilities in automation, error detection, and correction while executing Suzuki and Sonogashira reactions. When the Coscientist system attempted to use a robot in the real world to execute its program, an error occurred in the code controlling the equipment for heating and vibrating liquid samples. Coscientist identified the issue without human intervention, consulted the technical manual of the equipment, corrected the code, and retried, ultimately successfully completing the reaction. This study highlights the immense potential of large language models in accelerating research progress, self-checking and correction, and enhancing the efficiency and diversity of experimental design. Researchers can use large language models to (semi-)automate the planning of complex experimental steps, optimize reaction conditions, and explore new research pathways with unprecedented efficiency and scale [45]. Drug repurposing, which involves applying existing drugs to new disease treatments, is an effective drug discovery method. Compared to traditional drug development, drug repurposing is valued for its low cost, short cycle, and high safety. Its standard process includes drug screening, target confirmation, laboratory testing, preclinical research, clinical trials, and regulatory approval, but this process often faces challenges such as time consumption, high costs, and heavy data processing. A recent study explored the application of ChatGPT in the field of drug repurposing, demonstrating its effective support in drug screening, mechanism analysis, and clinical trial design [48]. Through interactions with ChatGPT, researchers selected 20 potential drugs for treating Alzheimer's disease (AD) and tested them in subsequent experiments, ultimately finding that three of these drugs were associated with reducing AD risk [48] This effectively accelerates the research speed of drug repurposing. Although these innovations have provided unprecedented momentum for the interdisciplinary development of the medical and chemical fields, the case of the Coscientist system also reveals the potential challenges that large language models may face in real-world environments, particularly in terms of their limitations in self-identifying and correcting errors. This underscores the importance of enhancing the self-correction capabilities of large language models, pointing to future development directions for applying these models in the medical field.

ChatGPT combines cutting-edge technologies such as natural language processing and machine learning to generate text that closely resembles human style. This makes it exceptionally difficult to discern whether an article was written by a human or generated by ChatGPT [49]. Generative AI tools like ChatGPT also demonstrate great potential in scientific writing. ChatGPT can quickly understand and integrate input literature information to draw conclusions, making it more efficient than traditional methods of literature reading and information integration [50]. ChatGPT not only helps researchers organize and analyze articles but also provides writing suggestions to improve the quality and efficiency of papers [51]. Furthermore, ChatGPT can assist researchers in maintaining the coherence of logical structure when writing papers, ensuring clear and strong arguments, and providing customized suggestions based on the user's research field and writing style, such as optimizing language expression [50, 52], which greatly helps authors who struggle with writing [53]. In the field of medical research, English has become the mainstream language for international communication [54], For non-native English researchers, artificial intelligence has significantly improved the accuracy of machine translation [55]. ChatGPT and GPT-4 perform excellently in handling complex translation tasks, being able to identify ambiguous words and clarify their meanings based on context [56]. GPT-4 surpasses many machine translation systems in translating entire documents, ensuring that the translated text is not only accurate but also reads naturally and fluently, maintaining the original mood and style of the text [57]. This ability is particularly important as it involves not only direct vocabulary and grammar translation but also an accurate grasp of the deep meaning and context of the text. In the final stage of writing a paper, ChatGPT can also assist in proofreading, helping to detect grammar errors, spelling mistakes [52], and ensuring that the submitted paper meets academic publishing standards. Although ChatGPT combines cutting-edge technologies such as natural language processing and machine learning to generate text in a style similar to human writing, this makes it exceptionally difficult to discern whether an article is written by a human or generated by ChatGPT [49]. Therefore, rigorous review and proofreading by researchers are still necessary.

Additionally, ChatGPT can be used for data processing in simulated research. Its application is broad, capable of participating in various stages from data analysis to the writing of research reports. For example, in a study, researchers created a simulated database containing 100,000 healthcare workers with varying ages, body mass index (BMI), and risk characteristics. Some virtual characters in the database were vaccinated with an imaginary vaccine aimed at reducing their chances of hospitalization after infection. Researchers provided this database and information about the fictional vaccine to ChatGPT, which then processed the simulated data, generated R language code to analyze the data results, validate the vaccine's effectiveness, and write related research papers [58]. This study demonstrates ChatGPT's ability to handle data, write, and correct R language code and also highlights its role in writing scientific research articles. In another study, researchers used ChatGPT to extract and summarize key functional information of lncRNAs in the updated EVLncRNAs database, showcasing its utility in literature summarization [59]. In the research on the use of generative AI in medical 3D printing, ChatGPT excels in introducing the history and basic knowledge of 3D printing, generating PubMed search terms, and providing research recommendations for medical 3D printing [60]. This will provide substantial help to researchers in their scientific work. ChatGPT can also provide suggestions on experimental design and guide researchers on how to conduct research and adhere to research ethics [61], further expanding its potential applications in the field of scientific research. Drug repurposing, the process of applying existing drugs to new therapeutic indications, is an effective approach to drug discovery. Compared to traditional drug development, drug repurposing has gained significant attention because of its low cost, shorter timelines, and higher safety profile. The standard process includes drug screening, target validation, laboratory testing, preclinical studies, clinical trials, and regulatory approvals, but this process often faces challenges such as lengthy timelines, high costs, and cumbersome data processing. A recent study explored the application of ChatGPT in the field of drug repurposing, demonstrating its effective support in drug screening, mechanistic analysis, and clinical trial design, which could significantly accelerate the pace of drug repurposing research [48]. Through interactive engagement with ChatGPT, researchers identified 20 potential drugs for the treatment of Alzheimer's disease (AD) and evaluated 10 of these drugs for their potential to reduce the risk of AD in individuals aged 65 and older using two large clinical databases – the Vanderbilt University Medical Center and the All of Us Research Program. The results revealed that Metformin, Simvastatin, and Losartan were associated with a reduced risk of AD [48].

In summary, LLMs have tremendous potential for application in medical science research. They can not only optimize existing research processes but also open up new research directions, injecting new vitality into scientific exploration. However, the practical application of these technologies must take into account the aforementioned limitations and potential issues to ensure their reliability and effectiveness in the field of medical research.

In Clinic

Large language models have demonstrated tremendous potential for application in clinical settings [2, 32, 62]. Outpatient letters serve as a significant mode of communication between most hospital specialists and General Practitioners (GPs), holding considerable importance in the field of medicine. These letters typically serve as the primary means of recording consultations for outpatient departments and hospitals, as well as the main mode of contact and communication between hospital staff and GPs. Recently, researchers have begun to utilize ChatGPT for drafting these outpatient letters [63]. By leveraging ChatGPT, doctors can enhance the efficiency of the letter-writing process. This not only improves communication effectiveness but also ensures the accuracy and consistency of information. Moreover, some existing large language models possess adequate clinical knowledge, approaching the level of clinical experts [11]. This indicates that large language models have sufficient foundational conditions for application in clinical settings. As a chatbot, ChatGPT can effectively engage in communication with patients [2]. Patients can ask ChatGPT questions related to diseases, treatments, lifestyle, and more. It can provide easily understandable and implementable advice, assisting patients in better managing their health, enhancing adherence to medical prescriptions, reducing misunderstandings between patients and doctors, and thereby facilitating the establishment of positive relationships between patients and healthcare providers [64].

Artificial intelligence can play a pivotal role in medical imaging, such as combining AI and radiomics in magnetic resonance imaging to predict the response to treatment of rectal cancer [65]. A study has shown that ChatGPT has the potential to provide patients with accurate information on breast cancer prevention and screening [66]. Since 2018, the traditional approach to addressing specific natural language processing tasks has involved fine-tuning large language models and additional task-specific training [67]. In the field of radiology, models have been developed specifically for well-defined tasks, such as extracting measurements of pancreatic cystic lesions from CT and MRI reports [68], identifying sites of metastatic disease [69], or rapidly categorizing tumor response categories(TRC) from free-text oncology reports(FTOR) [70]. However, the new generation of language-based base models (such as GPT-3 and its subsequent versions like GPT-3.5, GPT-4, etc.) can dynamically address tasks by using prompts without requiring additional training [71]. GPT-4 can interpret images [5], so ChatGPT based on GPT-4 can analyze and interpret images. It makes a more efficient and accurate auxiliary method for diagnosing medical images [72]. ChatGPT can even assist radiologists in identifying the smallest anomalies in medical images that may be overlooked and can reduce variability and errors in image interpretation (i.e., different interpretations of the same image by different radiologists) [73], effectively ensuring the accuracy of patient diagnostic results. Thanks to its ability to generate coherent, grammatically correct text, ChatGPT is helpful for automating the writing of radiology reports [74, 75]. Compared to GPT-3.5, GPT-4 generated structured radiology reports are more extensive and in-depth, while GPT-3.5 is more concise [76]. Furthermore, in a survey investigating the effectiveness of large language models in analyzing CT reports of 424 lung cancer patients, researchers found that GPT-4 had a higher success rate in extracting lesion diameters (98.6%) and was more accurate in identifying metastatic disease (98.1%), as well as performing better in describing or assessing patients' disease progression in medical reports [77].

In the clinical application of internal medicine, large language models (LLMs) are demonstrating revolutionary potential, particularly in diabetes management and the development of personalized treatment strategies. By conducting an in-depth analysis of patients' medical records, laboratory test results, and other relevant information, LLMs can provide strong decision support for healthcare providers. In diabetes treatment, LLMs can generate personalized treatment recommendations based on a wealth of clinical data and the latest research findings, including medication regimens and lifestyle adjustment suggestions. This is particularly crucial for healthcare providers in offering medication advice, helping them optimize treatment plans to enhance therapeutic outcomes [78, 79]. In the field of gastroenterology, ChatGPT shows high accuracy in addressing common gastrointestinal diseases (such as irritable bowel syndrome (IBS) and inflammatory bowel disease (IBD)), especially in providing information and education to patients [80]. This indicates that ChatGPT has the potential to serve as a medical assistant tool for gastroenterologists and help bridge the information gap between patients and their knowledge of gastroenterology.

In the clinical application of surgery, large language models are also widely utilized. In the field of neurosurgery, a study evaluated the performance of GPT-3.5 and GPT-4 by comparing their responses to 50 questions related to head injuries, vascular malformations, tumors, and neurological disorders with the responses of three neurosurgeons of varying expertise levels (high, medium, low). The results indicated that the ability of GPT-3.5 was comparable to that of a low-expertise neurosurgeon, while the ability of GPT-4 was equivalent to that of a high-expertise neurosurgeon [81]. Glioblastoma is the most common primary malignant brain tumor [82]. ChatGPT based on GPT-3.5 can provide reasonable treatment recommendations for glioblastoma patients based on their clinical conditions, surgical outcomes, imaging text information, and immunopathology results. Although its ability to consider unique patient circumstances (such as tumor subtypes and patient functional status) still needs improvement, it can serve as an auxiliary tool to collaborate with experts and enhance the efficiency and quality of treatment decisions for brain tumors [83]. In the field of spinal surgery, ChatGPT based on GPT-3.5 or GPT-4 can simplify data collection and analysis, assist in surgical planning, and provide real-time support during endoscopic spine surgery to aid spine surgeons [84], who often need to make critical decisions in a short amount of time. In this regard, ChatGPT can effectively reduce errors in judgment during spine surgery [85]. Additionally, ChatGPT has similar applications in urology, plastic surgery, oral and maxillofacial surgery, and podiatric surgery [86–89].

The specific terminology used in pathological reports is necessary for providing optimal recommendations for final diagnoses and treatment plans, but these terms may be difficult for patients unfamiliar with certain diseases to comprehend, and may even be misleading. To assess ChatGPT's ability to interpret pathological reports, a study conducted experiments using ChatGPT based on GPT-4. Researchers uploaded a hematopathology report analyzing bone marrow, showing signs of plasma cell neoplasms along with reactive plasmacytosis, which can obscure or interfere with some features of plasma cell neoplasms, making the situation complex. Researchers asked ChatGPT to explain the report to the patient and found that ChatGPT was able to interpret the report's complex terms and related data in an easily understandable way for the patient, providing the necessary information [90]. This not only makes it more convenient for patients but also saves doctors' time. Another research team built a specialized knowledge base and employed zero-shot learning methods to enable LLMs to accurately diagnose heart diseases and sleep apnea in electrocardiograms (ECGs) without direct training data. The success of this method demonstrates the potential of LLMs in handling highly specialized medical data, particularly in their ability to rapidly and accurately provide diagnostic support [91] Skin, as the first line of defense against external threats such as bacteria, viruses, and other harmful microorganisms, is susceptible to common skin diseases such as acne, bacterial skin infections, pressure sores, fungal skin diseases, itching, and psoriasis. Traditional methods for diagnosing skin diseases rely on dermatologists' observations and experience. As the first interactive diagnostic tool based on advanced vision-language models, SkinGPT-4, through pre-training (analyzing 52,929 images of skin diseases and doctors' records and clinical concepts using MiniGPT-4), offers a new diagnostic approach for patients. Patients can upload their skin photos to SkinGPT-4 for diagnosis, where it autonomously analyzes the images, identifies skin characteristics, and provides interactive treatment recommendations. Tested with 150 real cases, SkinGPT-4 has demonstrated its ability to accurately diagnose skin diseases. This will greatly alleviate the shortage of dermatologists in remote areas. While SkinGPT-4 cannot replace dermatologists, it can expedite the diagnosis of skin diseases effectively and has the potential to promote people-centered healthcare and health equity in underdeveloped regions [92]. Additionally, the Medical Multimodal Large Language Model (Med-MLLM) developed by Liu F. et al. has many advantages in clinical applications, particularly its rapid response to newly emerging or rare diseases. Med-MLLM can efficiently and accurately diagnose diseases with limited data. This feature makes Med-MLLM very useful for disease diagnosis and prognosis assessment, as well as providing support for establishing standard treatment protocols and plans. The model demonstrates its unique advantages in handling specific epidemics (such as COVID-19) by utilizing deep learning and self-supervised learning methods to learn from a large amount of unlabeled medical images and texts, capturing complex medical knowledge and clinical phenotypes. By fine-tuning (cross-entropy optimization) Med-MLLM, tasks related to COVID-19 can be optimized, such as generating COVID-19 reports (medical report generation), disease classification (determining if it is COVID-19, and if so, whether it is the Omicron or Delta variant), and survival prediction (COVID-19 prognosis). Research also indicates that after analyzing patients' chest images, Med-MLLM can generate image reports in different languages (English, Spanish, Chinese) that are rich in information and accurate [93].

At the end of this chapter, we need to emphasize that, based on existing research, we cannot guarantee that current large language models can accurately perform autonomous diagnoses under all circumstances. Additionally, these models exhibit high sensitivity to the order and amount of information when processing clinical data, which can lead to inconsistent diagnostic accuracy [94]. In practical applications, this means that clinicians must strictly supervise and control these models to ensure that patient care is not affected. Furthermore, when using large language models (LLMs) for clinical decision support, clinicians must pay close attention to ethical considerations. Particularly when using these AI tools to provide diagnostic and treatment recommendations, it is essential to ensure that patients are fully informed and consent to the use of these technologies [95]. Simultaneously, protecting patients' data security and privacy rights is indispensable [95], and all necessary measures must be taken to prevent the leakage of sensitive medical information to ensure the ethicality and compliance of medical services. To this end, developing a "private version" of large language models might be an effective solution, ensuring that the input images and text information are 100% secure from leakage. Overall, thoroughly exploring and addressing these challenges in the process of utilizing LLMs to support clinical decision-making is crucial for safeguarding patient rights and enhancing the quality of healthcare services.

In Nursing

In nursing, with the advancement and increased availability of large language models, nursing practices should adapt to and embrace technological developments [96]. ChatGPT can reduce repetitive tasks for nurses. It provides automated support in tasks such as emergency triage, data entry, scheduling, appointment reminders, and discharge record generation, thereby improving work efficiency [97, 98]. When nurses communicate with patients who speak different languages, ChatGPT can enhance communication between patients and nurses, aid in the patients' understanding of the communication content, and ultimately help generate a care plan that aligns with their preferences [99]. In specific patient care applications, the performance of GPT-4 surpasses that of GPT-3.5; for example, while GPT-3.5 may provide incorrect advice when advising on the care of patients with mechanical intestinal obstruction, GPT-4 can offer more accurate and comprehensive steps. This improvement is reflected not only in the accuracy of grammar and language expression but also in the correction of content-level errors, thereby enhancing the reliability of nursing recommendations [2].

Electronic Health Records (EHR) are a critical part of nursing work. They are used to record patient health information, including treatment history, diagnoses, treatment plans, test results, and medication information. Nurses need to use EHRs to review patient information, document care activities and observations, and share this information with other nurses or doctors [100]. LLMs can improve the accuracy and efficiency of clinical documents in EHR systems. They can automatically review and analyze documents, identify missing or incomplete information, and provide appropriate supplementation or clarification suggestions. For example, LLMs can detect inconsistencies between diagnoses and treatment plans, ensuring that documents accurately reflect the actual clinical situation [101].

Additionally, by analyzing historical patient data, Large Language Models have the ability to identify potential health risks for patients. For instance, LLMs can predict the likelihood of a patient being readmitted, identify risks for the development of certain diseases, or evaluate the likelihood of adverse events based on patient characteristics [101]. A specific example is the "Foresight" model developed by Kraljevic, Zeljko et al. based on GPT, which can predict a patient's future health status and even anticipate potential scenarios the patient may encounter at any point in the future. One notable advantage of this model is its scalability, as it can be easily applied to a broader patient population and its performance improves with more data being obtained. The exceptional performance of the Foresight model has been confirmed through testing at King's College Hospital and South London and Maudsley Hospital in the UK, as well as on the MIMIC-III database in the United States. This model has been publicly shared by researchers online (https://foresight.sites.er.kcl.ac.uk/) [102]. These capabilities of LLMs enable nurses to intervene proactively and provide personalized care. When nurses communicate with patients in different languages, ChatGPT can facilitate communication between patients and nurses in various languages, helping patients understand communication content and ultimately generate care plans in line with their preferences [99]. Furthermore, ChatGPT can reduce repetitive tasks for nurses. For instance, it can perform tasks such as emergency triage and automate repetitive tasks including data input, scheduling, appointment reminders, and generating discharge records [97, 98].

Cancer is one of the leading causes of death worldwide, with over 10 million people dying from cancer each year [103]. In cancer care, ChatGPT can enhance the level of cancer management services and address the physical, psychological, and behavioral health needs of cancer patients [104]. Cancer patients, especially those with hepatocellular carcinoma, require extensive and personalized care to improve outcomes [105]. ChatGPT assesses patients, assists caregivers in creating-driven personalized care plans tailored to the patients' needs and preferences, optimizes treatment effectiveness, and minimizes the risk of complications [106, 107]. Research shows that ChatGPT achieves an accuracy rate of up to 96.9% when answering questions on the "Common Cancer Myths and Misconceptions" web page [108]. ChatGPT can analyze cancer patients' medical records, predict health risks such as the likelihood of postoperative complications, interpret next-generation sequencing reports, and provide a list of potential clinical trial options, promoting continuous management and intervention over the years [109]. Overall, in cancer management and care, ChatGPT can help cancer patients achieve better treatment outcomes and maximize their quality of life.

According to the Centers for Disease Control and Prevention (CDC) definition, chronic diseases broadly refer to illnesses lasting for a year or longer, requiring continuous medical attention, limiting daily activities, or both [110]. The World Health Organization (WHO) states that Non-Communicable Diseases (NCDs), also known as chronic diseases, cause 41 million deaths annually, accounting for 74% of all global deaths [111]. Currently, ChatGPT is applied in personalized care management and remote healthcare for cardiovascular and cerebrovascular diseases, effectively improving the prognosis care outcomes for patients [112]. In the management of Chronic Obstructive Pulmonary Disease (COPD), ChatGPT can enhance patients' self-care management, help patients understand symptoms, causes, and treatment options, change lifestyles, select appropriate medications, and provide emotional support [113]. Additionally, ChatGPT has been utilized to assist in managing diabetes and HIV/AIDS [114, 115]. It offers personalized management options, promotes HIV testing and prevention, and provides useful information support for diabetes self-care. Furthermore, ChatGPT is well-versed in diabetes [116], which benefits diabetic patients by providing useful information support before self-care. With the increasing prevalence of mental chronic illnesses, ChatGPT, as a chatbot, can explain the impact of negative behaviors that may lead to self-harm or suicide, and offer solutions, playing a critical role in providing mental health support and interventions for young people [117–119]. In conclusion, ChatGPT can provide positive feedback in the management of chronic diseases, offer companionship in the daily lives of elderly patients with chronic conditions, and provide initial treatment advice before professional therapy and care.

Although LLMs have demonstrated many advantages in nursing practice, there are also some limitations. Firstly, in complex clinical situations, LLMs may make incorrect judgments, necessitating the professional judgment of nursing staff for supplementation and verification [94]. Secondly, subtle differences in language translation by LLMs may lead to misunderstandings, so nurses must exercise caution when using them. Additionally, for highly specialized medical information, ChatGPT cannot replace the judgment of professional healthcare providers, and reliance on technology might lead patients to overlook their own responsibility for health [95]. Lastly, when using LLMs to assist in nursing practice, ethical issues must be addressed to ensure that AI tools complement rather than replace traditional nursing services, maintaining humanistic care and emotional support, and considering cultural sensitivity and the diverse needs of patients to promote an inclusive care environment [120]. A comprehensive focus on these drawbacks is key to ensuring the effective and responsible application of LLMs in the field of nursing.

Limits and Challenges

Large language models like ChatGPT are widely used in the medical field, prompting healthcare workers to pay attention to the ethical issues brought by these models [95, 121], particularly their ability to mimic human thought processes and creativity. As Alan Turing eloquently posed in 1950, "Can machines think?" This question has become a key philosophical issue under the backdrop of large language models technology [122]. The potential ethical issues are numerous and complex, primarily involving three aspects-humanities ethics, algorithm ethics, and legal ethics. 1. Humanities Ethics: When using ChatGPT in healthcare, protecting privacy becomes a crucial issue [121]. The collection, storage, and processing of sensitive patient information raise significant privacy concerns, with one major concern being the risk of unauthorized access or data breaches. As ChatGPT interacts with patients and healthcare providers, it may collect and store personal health information, including medical history, test results, diagnoses, and other sensitive data. Another privacy issue involves the risk of data re-identification, even if the data collected by ChatGPT undergoes de-identification, it is still possible to re-identify specific individuals by combining it with other available data sources [123]. Transparency in data usage is crucial [124]. Patients should be informed of how their data will be used by ChatGPT and have the right to consent to or refuse such usage. The natural language processing and machine learning functionalities of ChatGPT may pose privacy risks [125, 126], potentially inadvertently disclosing sensitive information or providing inaccurate responses that could harm patient privacy and health. Moreover, despite significant advancements in natural language processing, ChatGPT remains a machine learning model lacking genuine human emotion and emotional understanding [10, 12, 127, 128], which may result in responses appearing cold, mechanical, and devoid of human warmth. In medical scenarios, patients often seek warm and empathetic communication and interaction with healthcare providers, especially when dealing with health issues and illnesses [129, 130]. Patients may expect ChatGPT to provide more warm and personalized responses to meet their needs. 2. Algorithm Ethics: A key issue at the algorithm level with ChatGPT is the lack of transparency and interpretability behind the large language model algorithm [131, 132]. Transparency refers to the openness and clarity of algorithms and artificial intelligence systems in their operations, decisions, and generation of outputs [133]. Transparent AI systems allow users and relevant stakeholders to understand the internal workings to comprehend factors influencing their output. Interpretability means that artificial intelligence systems can provide understandable explanations for their decisions and recommendations [133]. Interpretability is crucial for users, such as healthcare professionals and patients, as it can help build trust in artificial intelligence systems and understand the reasons behind the decisions. Complex deep learning models often operate as black boxes [131], making it challenging to understand the specific algorithmic processes guiding them in assisting healthcare providers in making clinical decisions, potentially neglecting global factors while only focusing on specific aspects. This leads to another issue in algorithm ethics-algorithm biases or biases in training data. Data bias refers to biases present in the training data used to develop artificial intelligence models, which may not represent the existing biases in the data or could contain systematic biases [134]. Algorithm bias refers to biases arising from the design, implementation, or decision-making process of the algorithm itself [38]. In healthcare assistance, such as formulating treatment plans, if training data contains bias or is not representative, artificial intelligence tools like ChatGPT may generate inaccurate or discriminatory results [132]. When algorithmic feature selection, biased model design, or biased decision rules exist, even when trained on unbiased data, the outputs may still produce erroneous results due to these sources of algorithmic bias [38, 135]. The presence of algorithmic bias can exacerbate and intensify social, cultural, or historical biases already existing in the healthcare environment, leading to unfair treatment or discrimination against certain individuals or groups [136, 137]. Additionally, the output results of ChatGPT will be used for future iterative training, which means any biases could persist without human intervention [138]. 3. Legal Ethics: The reasons causing legal ethical issues are related to both human ethics and algorithm ethics. On one hand, in humanities ethics, the training data used by ChatGPT may contain sensitive personal health information. Unauthorized use of such information would violate data protection laws in some countries, such as the General Data Protection Regulation (GDPR) in Europe or the Health Insurance Portability and Accountability Act (HIPAA) in the United States [139, 140]. On the other hand, the potential of ChatGPT in real cases providing inappropriate medical advice raises significant legal issues [95], leading to complexities in assigning responsibility when adverse outcomes result from the medical advice provided by ChatGPT. OpenAI explicitly states in its terms of use that it assumes no responsibility for the content generated by GPT [141]. Hence, it appears that the burden of any errors entirely falls on the users. This poses a question: if inaccurate or inappropriate advice leads to harm, who should be held accountable [142]? The complexity of medical decision-making increases, and AI’s involvement may blur the lines of responsibility. The primary reason for the difficulty in determining responsibility lies in algorithm ethics, where the algorithms of such AI tools lack transparency, making it challenging to accurately understand their thought processes and decision-making procedures [132, 142]. Moreover, tools like ChatGPT could impact students' critical thinking skills, potentially negatively affecting their ability to discern valuable information from errors and irrelevant information [2, 143] Concerns arise about students or scientists using ChatGPT's generated texts as their own to complete assignments or publish articles [144], which is considered deceptive. As to whether ChatGPT can be considered an author of a paper, most scientists take a negative stance [145, 146]. ChatGPT may at times provide incorrect answers, which could be particularly dangerous in medical situations where errors or falsehoods may be subtle and often presented convincingly, leading users to believe in their authenticity [32]. A study showed that in answering questions related to nephrology, ChatGPT based on GPT-3.5 or GPT-4 had accuracy rates lower than the passing threshold and average scores of nephrology exam takers, demonstrating limitations in its accuracy and consistency aspects [147]. When ChatGPT exhibits flaws in clinical settings, there is a significant likelihood of reducing patients' trust in using ChatGPT.

Discussion and Prospects

The primary advantage of large language models like ChatGPT lies in their ability as medical assistant tools to support various aspects of medical education, research, clinical practice, and nursing by processing and analyzing large volumes of data. However, their main limitation is that they are only auxiliary tools and cannot fully replace professional healthcare workers, educators, or researchers. Despite GPT-4 demonstrating higher intelligence and robust capabilities compared to GPT-3.5, it still cannot independently take on the roles of teachers, researchers, or healthcare providers. This limitation underscores the irreplaceability of human professional knowledge and experience while also highlighting the importance of integrating artificial intelligence technology with human expertise in practical applications. Despite existing limitations and challenges, ChatGPT and similar large language models undoubtedly serve as powerful auxiliary tools that effectively help individuals complete tasks, and improve work efficiency and quality. However, proper use of these tools is crucial. On one hand, maintaining the right mindset, avoiding over-reliance on artificial intelligence, holding a critical attitude toward ChatGPT-generated content, and carefully discerning information is essential. On the other hand, understanding how to pose appropriate questions to ChatGPT is crucial, as asking the right questions can significantly enhance efficiency. Research on question-asking techniques for ChatGPT by Heston, Thomas, and Khun, Charya provides significant assistance in this regard [31]. When facing ethical issues, developers must strictly adhere to global ethical standards and regulations in training large language models. Healthcare professionals also need to safeguard patient privacy and continually learn and adapt in the face of rapid technological advancements to keep their professional knowledge updated and stay abreast of developments in the field of medical AI. If hospitals adopt ChatGPT, it is essential to assign individuals to delve deeply into the differences between training data and patients, and be vigilant about whether any doctors overlook crucial advice due to adherence to old practices or mistrust of ChatGPT. Ensuring high visibility of large language model tools (i.e., their warnings and guidelines are easily noticeable) and ensuring that doctors have a thorough understanding of these tools (i.e., high operational transparency, enabling doctors to assess the reliability of its recommendations) is critical. In the future, leveraging the powerful capabilities of state-of-the-art large language models, we aim to develop an advanced medical assistance tool designed to serve hospitals globally. This tool will not only provide professional clinical support in line with medical qualifications but will also include voice functions, ensuring ease of use for all users, including visually impaired individuals and other groups requiring accessible services. Through this approach, we hope to break the boundaries of traditional medical services, enabling people from different regions and backgrounds worldwide to benefit from the medical convenience offered by AI technology, achieving true medical equality and inclusivity.

Author Contributions

Xingliang Dai and Ziqing Su conceived and designed the manuscript. Ziqing Su, Guozhang Tang, Rui Huang, Yang Qiao, Zheng Zhang, Xingliang Dai discussed, wrote and revised the manuscript. All authors discussed the results and commented on the manuscript.

Declarations

Competing of interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.