Large Language Models in Biomedicine and Health: A Holistic Evaluation of the Effectiveness, Reliability and Ethics using Altmetrics

The integration of Artificial Intelligence (AI), notably through Large Language Models (LLMs) such as ChatGPT and Bard, has led to a significant shift in the biomedical and healthcare landscape. These sophisticated models reshape clinical decision-making and patient interactions, demonstrating their increasing relevance across diverse applications. The range from diagnosing orthopedic diseases to enhancing disease detection from Electronic Health Records (EHRs) highlights their increasing relevance in healthcare.^[1] The influence of ChatGPT spans various sectors, influencing practices in finance, medical examinations, research authoring, student engagement with AI technologies, AI-assisted medical education, gender, cybersecurity, ethics in research, human resource management, sustainable development and higher education.^[2–12] The interdisciplinary nature of AI in medicine encourages progress in healthcare,^[13] precision medicine and public services.^[14] These models, derived from extensive data, merge diverse, intelligent tasks and medical expertise, suggesting the future of collaborative AI and medical practice.^[15] This necessitates strong ethical and regulatory frameworks for integrating ChatGPT into clinical workflows.^[16,17] The deployment of LLMs in biomedical contexts shows promise for enhancing clinical and research productivity despite challenges in ensuring accuracy^[18,19] and addressing safety concerns.^[20,21] Ethical considerations remain crucial, particularly in comprehensive AI implementation in healthcare.^[22,23] This calls for a systematic analysis of LLMs for biomedical applications, not just any aspect-specific review. The body of knowledge associated with literature accrued or rapidly piled by different fields of science and technology or subfields of such fields due to varying intensity and pace of publishing and patenting activities provides a plethora of opportunities for such an analysis. Despite being in its early stage, a large volume of literature is available on LLMs such as ChatGPT, especially because of its tremendous biomedical appeal. Thus, a systematic analysis of the body of literature related to major LLM tools and their role and impact on biomedicine and health can be useful for revealing various insightful implications that may benefit various stakeholders, including practitioners, researchers and policymakers. As there have been no such notable attempts thus far, we intend to address that gap in this work and the novelty of this research lies.

Scientometrics is a field of research that addresses the assessment of the progress of different fields as well as the impact assessment of various actors (or contributors) of science and technology via techniques and methods of science and technology. It offers several methods for mining and analysis of the body of scientific literature as well as for impact assessment of scholarly publications. However, there are certain perceived limitations to the scientometric approach for impact assessment. As the scientific impact of published works may take some time to manifest, traditional scientometric indicators such as citation counts are not suitable for early impact assessments of publications. Additionally, those indicators are not capable of capturing the societal impact of such works. In this era of expanding social media and online platform usage, there are means for assessing the attention garnered by published works on various social media platforms and other forums. Scientometricians are contemplating a fluid analogy between (i) the relationship between indicators such as citation to scientific impact and (ii) social media attention/visibility to societal impact, which led to the development of the field ‘Altmetrics’, which stands for alternative metrics and is often dubbed Scientometrics 2.0. Moreover, altmetrics, which is emerging as an innovative impact measurement tool alongside traditional citation counts, provides a detailed view of research impact, reflecting community engagement in scientific discourse.^[24–27] As both approaches can complement each other, a diligent combination of the traditional scientometric approach and the altmetric approach can be useful for the effective analysis of the body of literature related to different fields, subfields, etc., especially for cases such as the one we have to address.

This served as our motivation to employ the traditional scientometric approach and Altmetric approach to analyze the role and impact of LLMs in biomedicine and health. It assesses the engagement and influence of LLM-related research in these fields, using Altmetric data for insights into their effectiveness, reliability and ethical implications. Specifically, the pursuit of the following three key Research Questions (RQs) that can be addressed via the judicious combination of traditional scientometric approaches and Altmetric analysis that we adopt in this work will be beneficial for the abovementioned stakeholders and for providing a sense of direction for further research.

RQ1: Evaluating the influence of LLM research on public health and health services.

RQ2: Undertaking a Fields of Research (FoR) map and cluster analysis focusing on effectiveness, reliability and ethical considerations.

RQ3: Investigating the role of social media in shaping public perception and academic engagement.

As our 1^st objective (RQ1) was to extract insights into the effectiveness, reliability, ethical considerations and overarching impact of LLMs in healthcare at a very early stage, Altmetric data analysis is needed. Thus, RQ1 is addressed via the identification of journals with top Altmetric scores (sum of Altmetric scores of individual articles). The identification of top notch journals (those that managed to receive substantial early attention), the identification of dominant areas in LLM healthcare research, the analysis of the top articles that gained the most attention (those with the most AAS garnering articles) and the identification of fields of research associated with those publications are the keys for identifying works that might effectively discuss the abovementioned aspects that can be major determinants of LLM usage in medicine and healthcare. The science mapping approach (scientometric approach) can address RQ2. In this approach, a FoR map associated with the body of literature related to LLMs needs to be created. This is augmented by a cluster analysis delineating the interconnections between various research fields, thereby partitioning the FoR map into discernible clusters. Subsequently, a content analysis of the leading publications within each thematic cluster was conducted to extract more specific insights that can inform various key stakeholders. In addition, it is important to analyze the dominant sources (such as geographical regions) that constitute the social media activities that shaped the early pattern of social media attention in LLM research (i.e., RQ3). This issue is addressed by analyzing the geographical patterns of Twitter (a platform that dominantly contributes to Altmetric score computation) users. The pursuit of the three RQs is achieved after the selection of a suitable database that indexes published works that form the body of literature related to major LLM usage in biomedical and healthcare.

Our major data source for collecting publication data related to ChatGPT and Bard is the Dimensions database.^[28] We used the search terms (chatgpt OR “chat GPT” OR chat-GPT OR gpt3* OR gpt4*) OR (“bard”)) in the title and abstract of the publications, as ChatGPT product versions and Bard are the most promising LLMs found thus far. All publication types were included except Preprints. The data were collected on November 14, 2023. Publications in 2022 and 2023 up to that day were covered.

Of the 4968 publications retrieved, we focused on 942 publications belonging to two major Fields of Research (FoRs), Biomedical and Clinical Sciences and Health Sciences, using the ANZSRC 2020 code.^[28] We collected the altmetric details of the 942 publications using the DOI from Altmetric.com. The Altmetric Attention Score (AAS) is a composite score of mentions (that represent attention) in social media and other platform mentions. We also identified the most frequently mentioned source titles for these publications. The thematic clusters within FoRs are identified using cocitation maps via VOSviewer software, which enables the visualization and analysis of scientific landscapes, revealing patterns and relationships within research fields.^[29]

The results obtained by applying our framework to the selected data (942 publications) related to LLM for biomedicine and health for addressing three specified RQs are discussed next.

As mentioned above, to address RQ1, we determined the journals that received the highest Altmetric Attention Score (sum of the AASs of individual publications) from the dataset of 942 selected publications. The top 10 journals with the highest AAS are shown in Table 1. Upon analysis of the top AAS-receiving publications and FoR mapping, the profound linkage of LLMs to diverse fields of research in biomedical and clinical sciences and health sciences and their potential applications are revealed. The impact of LLM research on public health policies and health services is evident in the diverse range of journals featured in this research. Various areas/aspects of LLM influence in biomedical and healthcare research and practice, as evident from publications with top AASs in the top 10 journals with the highest AAS sums, are summarized below:

Validation and Transparency: The paper in JAMA Internal Medicine (i.e., the 1^st paper in Table 1 and the only one in our dataset), “Comparing Physician and Artificial Intelligence Chatbot Responses to Patient Questions Posted to a Public Social Media Forum”,^[30] which falls under Health Sciences (specifically Health Services and Systems), underscores the importance of transparency and the rigorous validation of AI tools in healthcare, reflecting the medical community’s focus on the accuracy and ethical implications of AI interactions with patients.

Second, Radiology features the paper “Appropriateness of Breast Cancer Prevention and Screening Recommendations Provided by ChatGPT”,^[31] which belongs to Biomedical and Clinical Sciences. This paper focuses on the role of AI in diagnostic processes and preventive medicine, highlighting the potential impact of AI on patient care in radiology.

The 4^th publication in Table 1 titled “ChatGPT and Generating a Differential Diagnosis Early in an Emergency Department Presentation”^[32] in the Annals of Emergency Medicine, also in the Biomedical and Clinical Sciences field, reflects the urgency of integrating AI in emergency medicine and a strong commitment toward collaborative AI applications that can potentially enhance diagnostic accuracy in time-sensitive medical settings.

In the interdisciplinary journal PLOS Digital Health, the paper “Performance of ChatGPT on USMLE: Potential for AI-assisted medical education using large language models”^[33] (i.e., the 5^th one in Table 1), which spans information and computing sciences as well as biomedical and clinical sciences, demonstrates interest in utilizing AI for educational purposes, particularly in medical training and assessment.

The 6^th article in Table 1, featured in the Journal of Medical internet features “Assessing the Utility of ChatGPT Throughout the Entire Clinical Workflow: Development and Usability Study”.^[17] which falls under Health Sciences (Health Services and Systems), assesses LLM’s ability to support clinical decision-making and highlights limitations such as possible model hallucinations.

The 9^th work in Table 1 is published in the American Journal of Gastroenterology, “Chat Generative Pretrained Transformer Fails the Multiple-Choice American College of Gastroenterology Self-Assessment Test”,^[34] which is categorized under Biomedical and Clinical Sciences and enumerates the challenges that AI faces in specialized medical fields and the need for domain-specific AI training.

The 8^th paper in Table 1 is from npj Digital Medicine, “The imperative for regulatory oversight of large language models (or generative AI) in healthcare”,^[35] which falls under Health Sciences (Health Services and Systems) and calls for attention to the need for governance in the use of AI in healthcare, resonating with the community’s concern for responsible implementation.

The 3^rd paper in Table 1 is an interdisciplinary paper titled “Abstracts written by ChatGPT Fool Scientists”,^[36] which appeared in Nature and cuts across Biological Sciences, Language, Communication and Culture and reflects on concerns about the reliability of scientific communication in the age of AI and the importance of discerning AI-generated content.

JAMA Network Open features the 7^th work in Table 1 titled “Evaluating Artificial Intelligence Responses to Public Health Questions”,^[37] which is about health sciences and public health, indicating a significant interest in the accuracy and reliability of AI in disseminating public health information.

Finally, the paper “Chatting with ChatGPT to learn about the safety of COVID-19 vaccines-A perspective” in human vaccines and immunotherapeutics addresses the biomedical and clinical sciences fields such as immunology, medical biotechnology and medical microbiology and highlights AI’s potential role in public health dialogs, especially concerning vaccine safety and misinformation.^[38]

Thus, the top 10 articles according to altmetric attention scores are found to shed light on the influence of LLMs on public health and health services and the imperative factors that contribute to these influences, such as effectiveness/performance, transparency, reliability, ethical and regulatory considerations and risks and safety issues. Therefore, RQ1 is successfully addressed via the use of an altmetric approach and content analysis.

The Fields of Research map (Figure 1), which is a coaffiliation map of FoRs (coupled via publications in our dataset), represents the interlinkage between various FoRs associated with the body of literature analyzed. In the map, the size of the vertices (FoRs) indicates high linkage to other vertices (FoRs). Upon clustering, the map delineates two principal clusters: Biomedical and Clinical Sciences (green cluster) and Health Sciences (red cluster). Analysis of both these clusters and the top 10 publications from the list of publications that are responsible for the formation of green clusters and red clusters addresses RQ2. These issues are discussed next.

Figure 1:
Fields of the research map showing two major clusters.

As already mentioned, a heavy dominance of the FoR ‘Biomedical and Clinical Sciences’ can be seen in the green cluster. Other FoRs, such as ‘Biological Sciences’ and ‘Engineering’, are also found to indicate the ever-increasing interconnectedness between life sciences and engineering sciences disciplines. The major sub-FoRs that are deeply influenced by LLM research seem to include ‘Clinical Sciences’, ‘Immunology’, ‘Ophthalmology and optometry’ and ‘Dentistry’. Some sub-FoRs of the FoR ‘Health Sciences’, such as ‘Allied Health and Rehabilitation Science’ and ‘Sports Science and Exercise’, are also found in the green cluster (while FoR Health Sciences dominates the red cluster). This indicates the deep connectedness between ‘Biomedical and Clinical Sciences’ and ‘Health Sciences’ and the ability of LLMs and AI to influence these major FoRs individually and simultaneously.

The top 10 publications in the biomedical and clinical sciences clustered according to the Altmetrics Attention Score, along with their further relevant details, are presented in Table 2. In Table 2, the FoR column predominantly shows works in clinical sciences, highlighting the significant focus on the practical applications and implications of LLMs in various medical fields. This interdisciplinary paper combines information and computing sciences with clinical sciences, emphasizing the growing intersection of AI technology and health care. The range of disciplines extends to Immunology, Medical Biotechnology, Medical Microbiology, Ophthalmology and Optometry and Oncology and Carcinogenesis, showing that LLMs have varied impacts across different biomedical and clinical areas. They illustrate how LLMs are being explored for their potential in medical education, diagnostic processes and specific areas such as vaccine safety and cancer research.

According to the 1^st publication in Table 2, the ChatGPT achieved a passing score or nearly moderate performance in the USMLE,^[33] indicating that LLMs will soon mature enough to impact clinical medicine and that such a scenario demands open science research infrastructure in medical education.

The 2^nd publication in Table 2 underscored that, in some cases, ChatGPT generated incorrect answers or explanations,^[32] indicating that inconsistency in outputs underscores the unpredictability of LLMs and highlighting the need to restrict the role of LLMs as an aid rather than replace physicians’ judgment.

The 3^rd work suggested that although some improvement is needed, the LLM holds great potential for automating patient educational information about breast cancer prevention and screening^[27] and recommended physician supervision throughout the process.

Through a comparative analysis, the 4^th work revealed that GPT-4 significantly outperformed GPT-3.5 and Bard, with a lower rate of hallucinations observed for GPT-4 patients,^[39] indicating the potential of LLMs in neurosurgical education and clinical decision-making.

An experiment reported in the 5^th work revealed that ChatGPT provided a confident, human-like answer^[38] and suggested that the use of technology, especially LLMs, can heavily impact medical writing, particularly in terms of speed and accuracy.

The 6^th work^[34] involved a comparative analysis of two versions of the ChatGPT on gastroenterology-related examination and found that both versions failed the test (as the pass score required was 70%), suggesting the need for substantial improvement of LLMs before implementation in healthcare.

The 7^th work revealed the overall ability of ChatGPT to detect counterfeit questions related to vaccines and vaccination^[40] and highlighted that the current language used is optimal (fewer technical terms without compromising scientific details), suggesting that the potential LLM is reliable medical information.

According to the 8^th paper, LLMs such as the ChatGPT cannot substantially assist in training for board certification in ophthalmology,^[24] as the performance of the ChatGPT was the best in general medicine and the poorest in the retina and vitreous.

The 9^th paper revealed that the ChatGPT performed well in terms of basic knowledge, lifestyle and treatment,^[41] but a lack of knowledge about regional guideline variations (such as HCC screening criteria) is an area that should be properly worked upon and recommended to restrict its usage as an adjunct informational tool.

The 10^th paper also identified the potential of ChatGPT for producing accurate results, but as responses were not actionable and were found to be at the college reading level,^[42] it recommended its use as a supplementary tool (not as a primary source of medical information).

In the red cluster, FoRs such as Information and Computing, Psychology, Education, etc., have the highest linkages to other FoRs or sub-FoRs. This indicates the profound importance and possible impact that LLMs can have on FoRs related to medical research and practice. The presence of FoR Information and Computing Sciences highlights how the future of medical research and practice is dependent on the advancement of information, computing and learning (AI) technologies. Nursing, epidemiology, etc., are the dominant sub-FoRs within the FoR Health Sciences that are influenced by LLM advancement. In FoR psychology, sub-FoR ‘biological psychology’ is the subfield in which the LLM has the greatest impact at this early stage. The sub-FoR ‘Neurosciences’ (a subfield of Clinical Sciences) is also found in this cluster, showing its profound linkage to FoRs such as ‘Health Sciences’ and ‘Information and Computing Sciences’.

Table 3 shows that the top publications in health sciences predominantly feature papers classified under ‘Health Services and Systems. This focus reflects an academic and practical interest in understanding how AI tools can integrate into and enhance healthcare services, from clinical workflows to regulatory oversight. Notably, several studies have compared AI-generated responses with those of healthcare professionals, assessing the accuracy, reliability and utility of the ChatGPT in clinical settings. This highlights the growing curiosity about the potential role of AI in augmenting or assisting medical decision-making processes. Papers such as “Evaluating Artificial Intelligence Responses to Public Health Questions” also extend the exploration to public health, indicating a broader scope of LLM applications beyond individual patient care to community health issues.

Furthermore, studies on the role of AI in generating clinical vignettes, scientific abstracts and even the creation of discharge summaries point to the diverse and innovative ways LLMs are utilized. These studies suggest that AI has transformative potential for reshaping various facets of healthcare service delivery and medical communications, opening doors to new methods of patient engagement, medical education and healthcare management. This trend underscores the importance of continued research and regulatory considerations as AI becomes increasingly embedded in health services systems.

The 1^st paper in Table 3 shows that the LLM is competent in generating quality, empathetic and more detailed (than physicians) responses to queries in online forums,^[30] indicating its ability to significantly reduce the burnout of physicians and suggesting further exploration of the use of LLMs in clinical settings.

The 2^nd study reported that the ChatGPT performed well in terms of general medical knowledge,^[17] but its performance is inferior for differential diagnosis and clinical management and highlights the ability to improve accuracy for more clinical information as the major strength of LLMs.

An investigation in the 3^rd work revealed that the ChatGPT consistently provided evidence-based answers^[43] and recommended encouragement of public-private partnerships for the use of AI and LLMs in healthcare wherein public health agencies should develop and maintain database resources for AI companies to use.

A comparative analysis of the ChatGPT with physicians carried out in the 4^th work revealed that the ChatGPT showed no gender or socioeconomic biases in response to the recommendations^[44] and confirmed the congruence of the ChatGPTs 3.5 and 4 with the accepted guidelines for managing mild and severe depression.

The 5^th work reported the results of an analysis of LLM’s performance in medical writing^[45] and as LLMs demonstrated versatility in generating results customizable to the literacy levels of patients, their ability to perform medical writing and editing is indicated, though not without supervision, due to caveats.

The 6^th work recommended the development of a specific and distinct regulatory framework for LLMs (rather than using a common framework for LLMs and non-LLM AI technologies) and such a framework should target the regulation of companies developing LLMs rather than the regulation of every iteration of LLMs.^[46]

The 7^th work revealed the ability of LLMs to generate accurate and comprehensive information^[47] and enumerated limitations such as hallucination, inability to seek clarifications for ambiguous queries and tendency to make inaccurate citations (that questions transparency).

The 8^th work reported the results of an examination of LLM-generated abstracts and original abstracts by human reviewers^[48] and confirmed the ability of the ChatGPT to write seemingly believable scientific abstracts and highlighted the need for careful discussions to determine ethical considerations.

According to the 9^th work, ChatGPT produced fraudulent papers that may convince researchers, even experienced ones,^[21] though there were inaccurate and nonexistent references, duplicate citations, etc. and emphasized the importance of effective and responsible harnessing of LLMs.

The 10^th work recommended that ChatGPT be upgraded to autoscrap medical information from patients’ digital records by sufficiently addressing data privacy concerns^[49] after (i) sufficient pilot testing, (ii) gathering stakeholders’ perspectives and (iii) careful considerations of technology failure strategies.

Thus, FoR map analysis revealed important FoRs and interlinkages between FoRs belonging to two major clusters. Additionally, the presence of sub-FoRs of FoRs belonging to the red cluster in the green cluster and vice versa indicates the interdependence of the two major clusters and the possible ability of LLMs to impact both clusters together. In-depth analysis of both clusters selected for 10npublications (according to the AAS) also reinforced this fact and such an exercise helped to garner useful insights related to almost all the imperative aspects related to LLMs, such as accuracy and performance, diverse applicability, reliability, transparency, other limitations and ethical considerations. Thus, as envisaged, a combination of a scientometric approach and an altmetric approach helped to successfully address RQ2. The insights gained from the in-depth analysis and possible recommendations based on such insights are discussed in the ‘Future directions’ section.

Figure 2 shows the notable presence of LLM research discussions on Twitter, underlining the importance of such platforms in spreading and deliberating scientific findings. The tweet analysis points to a conversation dominated by the Global North, particularly the United States, underscoring its leading technological role. This dominance is somewhat less evident but still noticeable in countries such as the United Kingdom, Japan, Canada and various European nations. Conversely, the Global South, represented by nations such as Brazil, India and South Africa, appears less engaged in these AI dialogs, suggesting a disparity in digital and developmental engagement. This situation underscores the imperative for the inclusive development of technologies with far-reaching effects.

Figure 2:
Leading countries based on Tweets.

Thus, the analysis of the geographical distribution of tweets (major contributors to AAS) provided insights into the existence of global disparities (between the Global North and South) related to the evolution of LLMs, emphasizing the dominance of certain countries or regions in shaping public opinion and perceptions related to technologies, especially those with game-changing potential. This also underlines the importance of altmetric sources and platforms in the progress of LLM research and applications in the biomedical and healthcare fields and points to the need for some nations/regions to consider proactive social media engagement to popularize their research outcomes and to remain updated about advancements. Thus, RQ3 is also effectively addressed by almetric analysis, as expected.

This framework, which is based on a combination of scientometric and altmetric data-driven analyses of ChatGPT and Bard’s applications in biomedicine and health, offers a comprehensive view of their potential applications, challenges and interdisciplinary impact.

The study highlights active engagement with LLM research in biomedicine and health, particularly in publications such as JAMA Internal Medicine and Radiology. This study emphasizes the accuracy and performance of LLMs and AI in patient care and diagnostic processes. Science mapping and cluster analysis revealed key FoRs and sub-FoRs that are mostly influenced by LLM research and advancement. Analysis of the top publications within two major clusters encompassing ChatGPT’s performance on the USMLE^[33] and its applications in emergency medicine^[32] hints at the potential of LLMs across different functional tiers in these sectors. Scholarly works^[50,51] have shown improvements in the accuracy and thoroughness of LLMs, indicating a progression in their adaptation to medical contexts.

Thus, our study revealed the growing influence of LLMs, from emergency medicine to education and regulation; LLMs and AI transform medical training, diagnostics, public health and scientific communication, indicating their ability to reshape healthcare service delivery, education and management. LLMs in clinical sciences and health services and systems are noted for their practical applications and potential to transform healthcare, with research spanning diverse fields, from immunology to oncology.

An updated altmetric analysis revealed significant academic and public interest in LLMs, driven by high attention scores and social media engagement. This widespread interest revolves around LLM efficacy and reliability and their role in bridging the gap between language, AI and healthcare. This integration shapes current research trends, highlighting the potential of LLMs to advance healthcare research and inform future policies.

This study also highlights ethical implications identified from various important works identified through altmetric analysis of the whole literature and literature related to two important clusters. Some works have stressed the importance of transparency and the need for the validation of LLMs for healthcare. Some acknowledged its limitations in specialized tests and vulnerability to misinformation. The reliability of the LLMs was also challenged in some works because they are prone to hallucinations and because of their tendency to add inaccurate and/or nonexistent references. Additionally, as LLMs and AI integrate more deeply into health systems, robust research and regulations are crucial. The need to separate LLMs from non-LLM AI applications in medicine and healthcare is also identified as a crucial challenge, but if it is addressed properly, it can be converted into a great plethora of opportunities.

From a critical perspective, these data showcase the promise and the perils of integrating generative AI tools such as ChatGPT and Bard into healthcare. The high attention scores reflect the public and professional community’s vested interest in these developments. However, they also call for cautious and considered integration of these technologies, emphasizing validation, transparency and ethical usage to ensure that they truly benefit patient care and outcomes. Given the rapid evolution of AI technologies, continuous monitoring, evaluation and adaptation of these systems are required to ensure that they align with healthcare goals and ethical standards. This integration shapes current healthcare research trends, highlighting the contributions of LLMs to advancing healthcare research and informing future policy-making. The ethical challenges of specific domains, such as dermatology, emphasize the need for varied datasets and conscientious AI deployment in healthcare.^[11]

Our findings suggest several future research possibilities for applying Large Language Models (LLMs) in diverse healthcare contexts. These prospects are informed by the latest data, observed trends and identified limitations of LLM tools, as revealed in widely referenced studies and various exploratory endeavors. However, as the key limitations contribute immensely to further exploration and development, these limitations are summarized again as follows for quick perusal.

Ensuring that LLMs consistently and accurately respond to medical inquiries is paramount.

Unpredictability and hallucination: Reducing the generation of incorrect or irrelevant information is crucial for the safe use of LLMs in medical education and clinical practice.

Regional guideline awareness: LLMs must be trained on regional medical guidelines and regulations.

AI output detectors: AI output detectors are critical for identifying and eliminating fictitious, AI-generated scientific content.

Social acceptance: Addressing data privacy and security concerns is essential for building public trust in LLMs in healthcare.

Building upon the findings of our study on the impact of Large Language Models (LLMs), such as ChatGPT and Bard, on biomedicine and health, we propose the following recommendations. These findings are closely aligned with our research insights and aim to enhance the efficacy, ethical application and sustainable integration of LLMs in healthcare environments:

Encouraging public-private partnerships and investment in research infrastructure: Given the study’s revelations on the interdisciplinary and significant influence of LLMs and the need for sophisticated research infrastructure, we advocate nurturing Public-Private Partnerships (PPPs) and augmenting research infrastructure. Such PPPs should be structured and operationalized within a robust and comprehensive legal framework that, in turn, is set up on top of a standard ethical framework. Special emphasis should be placed on the sharing and misuse of clinical data, other sensitive information and the structure and nature of the engagement of various entities within PPPs to ensure the safety, security and dignity of people.

Technological enhancement of AI LLMs: The analysis of LLMs’ performance underscores the necessity for improvements in accuracy and reliability, which would facilitate their evolution as a source of medical knowledge, an informational tool and potentially as an assistant in diagnosis and decision-making for physicians and health researchers. Special care should be taken to avoid the known, forewarned and unknown caveats of technological overreliance and overuse in biomedical and health research and practice, as these FoRs are directly associated with the life and health of patients. This can be addressed via constant explorations and discussions on aspects wherein technology should be used and not used and fostering technological development and advancement in that direction so that room can be always reserved for human supervision and intervention wherever and whenever necessary. Human-centric technological development is thus desirable for AI and LLMs.

The findings on substantial engagement with LLM research on social media platforms underscore the importance of strategically using these channels. The varied applications of LLMs highlight the need for cooperative approaches across distinct disciplines. Such collaboration, knowledge exchange and pooling can improve overall LLM performance by reducing limitations such as hallucination tendencies. Such interdisciplinary as well as multinational collaborations might be useful for determining global regulations and standards and at the same time determining locally acceptable exceptions in regional guidelines.

In line with our findings on high Altmetric scores and engagement levels, we recommend the implementation of stringent validation protocols for LLMs to ensure their dependability and effectiveness, particularly in specialized medical fields. This aligns with the observed interest and engagement in LLMs, emphasizing the necessity for precision and reliability. As proposed in some of the works with high Altmetric Attention Scores, effective validation frameworks can serve as benchmarks and be further refined. The role of interdisciplinary and multinational collaboration can also be vital in this respect and validation exercises based on multiple relevant criteria (as applicable) should be considered.

Table 1, which shows journals with high Altmetric Attention Scores, highlights the need for diverse policies to govern LLM integration in healthcare, reflecting the interdisciplinary applications of these policies. Policies should be adaptable; evolving in tandem with LLM tools and include systematic determinations for the equitable use of AI in medical and academic publishing. Regulatory frameworks in strict congruence with ethical and legal frameworks should be established to monitor and enforce preventive and corrective action wherever applicable in the case of materialized or potential ethical breaches as well as legal violations.

We would like to express our immense gratitude to our beloved Chancellor, Mata Amritanandamayi Devi (AMMA), for providing the motivation and inspiration for this research.