Review
Abstract
Background: Large language models (LLMs) can support health care professionals in their daily work, for example, when writing and filing reports or communicating diagnoses. With the rise of LLMs, current research investigates how LLMs could be applied in medical practice and their benefits for physicians in clinical workflows. However, most studies neglect the importance of selecting suitable LLM architectures.
Objective: In this literature review, we aim to provide insights on the different LLM model architecture families (ie, Bidirectional Encoder Representations from Transformers [BERT]–based or generative pretrained transformer [GPT]–based models) used in previous research. We report on the suitability and benefits of different LLM model architecture families for various research foci.
Methods: To this end, we conduct a scoping review to identify which LLMs are used in health care. Our search included manuscripts from PubMed, arXiv, and medRxiv. We used open and selective coding to assess the 114 identified manuscripts regarding 11 dimensions related to usage and technical facets and the research focus of the manuscripts.
Results: We identified 4 research foci that emerged previously in manuscripts, with LLM performance being the main focus. We found that GPT-based models are used for communicative purposes such as examination preparation or patient interaction. In contrast, BERT-based models are used for medical tasks such as knowledge discovery and model improvements.
Conclusions: Our study suggests that GPT-based models are better suited for communicative purposes such as report generation or patient interaction. BERT-based models seem to be better suited for innovative applications such as classification or knowledge discovery. This could be due to the architectural differences where GPT processes language unidirectionally and BERT bidirectionally, allowing more in-depth understanding of the text. In addition, BERT-based models seem to allow more straightforward extensions of their models for domain-specific tasks that generally lead to better results. In summary, health care professionals should consider the benefits and differences of the LLM architecture families when selecting a suitable model for their intended purpose.
doi:10.2196/70315
Keywords
Introduction
Background
The last years brought an unmatched rise of large language models (LLMs) such as OpenAI’s ChatGPT [,], Google’s Bidirectional Encoder Representations from Transformers (BERT) [], and Meta’s Llama []. LLMs have dramatically extended the abilities of natural language processing through generating text by repeatedly adding the most likely following words []. Most LLMs are based on large amounts of general-purpose text data on which the models were trained [-]. Thereby, LLMs provide human-like responses to user-given prompts and can analyze [], explain [], and generate text []. These abilities offer several opportunities to facilitate and improve workflows in various domains, including health care.
To that end, LLMs have been increasingly applied and tested, for example, to assist physicians in diagnosis and treatment decisions [], researchers in identifying disease phenotypes [], or students in medical examination preparation []. Despite these opportunities, the use of LLMs in health care is subject to biases [], hallucinations (ie, presenting incorrect information in a factual correct form []), or a limited understanding of the complexity of current medical nomenclature []. Since health care–specific nomenclature is not fully covered in general-purpose LLMs, their applicability in health care is restricted. In addition, the model architectures of general-purpose LLMs differ. For example, BERT-based models use bidirectional encoding, meaning the input text is read in both directions, making them especially suitable for language understanding []. In contrast, generative pretrained transformer (GPT)–based models are built on unidirectional decoders, allowing them to exceed in text generation []. If practitioners and researchers in health care use general-purpose LLMs without being aware of their benefits and limits, the LLM’s unsuitability might risk patient health.
With the ever-evolving LLM landscape, it is crucial for health care researchers and practitioners to have a thorough overview and keep up with the current trends of using LLMs in health care. Having insights into models’ performances on specific tasks and the application of LLM architecture families to specific health care improvements could significantly ease the usage of LLMs in health care. However, prior literature reviews focus primarily on issues in applying LLMs in health care and lack a holistic overview of technical instantiations and research foci.
Prior Work
Due to the underrepresentation of medical nomenclature in the training of most general-purpose LLMs, models used in health care require unique adaptation. One prominent example is Med-PaLM, which can pass questions in the style of the United States Medical Licensing Examination (USMLE) []. Medical LLMs have by now been developed and adapted for different use cases such as documentation, communication, decision support, picture interpretation, or treatment plan creation [].
Previous review studies investigated the opportunities and limitations of LLMs in health care [-]. These reviews have started to aggregate the knowledge scattered across various medical specialties and focus on opportunities and risks (eg, hallucinations, legal and ethical concerns, or privacy issues) for using LLMs in biomedicine [] or on the development and applications of LLMs in health care (eg, information and knowledge discovery) []. These reviews have already provided a considerable description of the applications of LLMs and related them to the patient care journey [,]. For specific applications, one review also included an overview of selected LLM architectures in biomedical and health fields alongside their performance []. These reviews made a critical step in providing an overview of the emerging application areas and the performance of LLM architectures for specific biomedical datasets [,]. However, they focus on a selection of LLMs and do not include the breadth of available LLM architectures in health care research. Moreover, the application areas to which the LLM architectures are applied could help researchers and practitioners to choose the right LLM architecture for a specific use case. Since LLM architectures differ in how they process text [,], specific architectures can be more suitable for a particular medical task. Hence, a thorough investigation of current LLM architectures in health care is crucial as it can aid practitioners in selecting suitable LLM model architectures for their medical tasks.
Objectives
In this study, we investigate the use of LLM model architectures in current studies in health care. While our results regarding the target use are similar to other reviews [], we specifically provide an analysis of model architecture families for 4 research foci (ie, LLM Performance, LLM Societal Impact, LLM Comprehensibility, and LLM Innovation). For each research focus, we assess the used model architectures, the integration of LLMs in clinical practice, and the data types used in each manuscript. With that, we aim to provide insights into LLMs’ maturity, practical implementation, and innovation stages in health care. We conduct a scoping review [] to determine the breadth and boundaries of the literature on model architectures and the use of LLMs in the medical domain [].
Methods
In our study, we investigated PubMed as the primary database for identifying manuscripts on LLMs in health care. Since the topic is emerging and novel models are introduced rapidly, we also included the preprint databases arXiv and medRxiv in our search. We searched for TITLE-ABS-KEY(GPT OR LLM? OR Large Language Model? OR LLaMA OR Bard OR Med-PaLM) to focus on LLMs rather than the medical domain. We included specific model architectures in our search string to gain insights into the technical manifestations of widely recognized LLMs. Our initial search in January 2024 yielded 1842 hits, with arXiv having the most results (n=813). Before screening the manuscripts, we excluded 479 manuscripts published before 2018, which marked the introduction of GPT-1 [] as well as duplicates (n=117) and non-English publications (n=37).
For analyzing the manuscripts, we followed the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) guidelines that are summarized in []. During the screening of the title and the abstract of each manuscript, we investigated the topic of the manuscripts and excluded papers with no focus on health care (n=837). We aligned the decisions on borderline cases by having 2 authors (FL and RG) investigate the manuscript and reach a consensus. Excluded manuscripts were, for example, related to Llamas instead of the Llama architecture.
In addition, we excluded papers with no focus on LLMs (n=206) or gray literature (n=14). We also found 4 additional duplicates. During full-text analysis of the remaining 148 manuscripts, we further found 15 manuscripts with no focus on health care, 2 manuscripts with no LLM focus, 16 gray literature manuscripts such as blog posts, and excluded 1 publication with access restrictions. All included manuscripts can be seen in .
Our scoping review used a combination of predefined categories and open coding as shown in 
Table 1. We coded the identified manuscripts along 9 dimensions split between 3 usage, 5 technical dimensions, and an overarching research focus dimension. To align the understanding of all dimensions and generate codes inductively, we double-coded 20 manuscripts during full-text analysis (17.5%, 20/114). For these 20 manuscripts, we ensured that the author team understood the dimensions and their codes coherently. With the initial set of dimensions and codes, we proceeded with single coding for all dimensions. The coding process was monitored through regular meetings where we discussed coding of all dimensions. If the coding author was uncertain about a dimension, other members of the author team reinvestigated the manuscript and decided. We introduced new codes only if the entire author team agreed. In coding the identified manuscripts, we allowed multiple codes within each dimension to represent the variety of characteristics discussed within the manuscripts. Therefore, most coding dimensions excel 114, for example, since studies use multiple data modalities or LLM architectures.
| Coding dimension | Description | ||
| Usage | |||
| Medical specialty | Medical domain of the conducted study | ||
| Target audience | Stakeholders targeted by the LLMa application | ||
| Target use | Use case of the LLM application | ||
| Technical | |||
| Model integration | Focus of the LLM application | ||
| Model novelty | Maturity of model development | ||
| Model architecture | Applied LLM architecture | ||
| Data type | Used data type | ||
| Evaluation metrics | Used evaluation metrics | ||
| Research theme | Investigated improvement focus of the manuscript | ||
aLLM: large language model.
An overview of all dimensions and corresponding tables and figures can be found in . For usage dimensions, we investigated medical specialty (eg, radiology, pathology, or cardiology), target audience (eg, physicians, patients, or students), and target use (eg, diagnostic support or patient communication). These dimensions provided relevant information for the targeted application of LLMs in health care.
We further investigated the technical aspects of each manuscript. The coding dimension model integration differentiated between studies focusing on LLMs on a conceptual level, their usage, implementation, or evaluation. Relatedly, we captured the model novelty (ie, developing a new model, applying an existing model, extending an existing model, or not using a model at all). This categorization helped us understand the extent of current and new developments of LLMs in health care. Furthermore, we extracted the specific model architectures such as GPT-3.5 or BERT to highlight the variety of models in the manuscripts. We also noted the data type used in the studies such as clinical notes, images, or electronic health records (EHRs). We documented the evaluation metrics used to assess the models’ performance, such as accuracy, precision, recall, or F1-score.
Finally, we coded the targeted contributions of the manuscript as research themes. Similar to the other coding dimensions, we openly coded the research themes (eg, correctness and patient well-being). Since several contributions discussed across the identified research themes were within similar contexts (ie, fairness: n=1, bias: n=4, and societal impact: n=2), we combined and aggregated coding themes into research foci. After revising the research themes and double-coding by the involved authors, we combined the themes into 4 research foci.
Results
Our scoping review was carried out with the PRISMA-ScR guidelines (see 
Figure 1 for the PRISMA-ScR flowchart).
Usage Dimensions
Medical Specialty
In examining the medical specialty (Table 2), we find a frequent usage of LLMs in medical research (26/114, 22.81%), health education (17/114, 14.91%), and hospital medicine (17/114, 14.91%). LLMs in medical research are used for predictions [-], labeling [,,], or fine-tuning techniques [,] without focusing on a specific medical domain. Most studies using LLMs in health education (15/17, 88.23%) examine the performance of LLMs in medical domain-agnostic examinations such as the USMLE [], the Medical College Admission Test (MCAT) [], or the Chinese National Medical Licensing Examinations []. Other studies investigating LLMs for medical examinations conduct domain-specific examinations such as neurology board questions []. Studies that focus on hospital medicine investigate the usage of LLMs as a replacement for doctors’ medical advice [,-].
| Medical specialty | Values, n (%) |
| Anesthesiology | 3 (2.63) |
| Cardiology | 6 (5.26) |
| Clinical laboratory sciences | 6 (5.26) |
| Dermatology | 3 (2.63) |
| Emergency medicine | 2 (1.75) |
| Endocrinology | 3 (2.63) |
| Gastroenterology | 4 (3.51) |
| Health education | 17 (14.91) |
| Hospital medicine | 17 (14.91) |
| Intensive care medicine | 2 (1.75) |
| Internal medicine | 7 (6.14) |
| Medical research | 26 (22.81) |
| Nephrology | 2 (1.75) |
| Neurology | 6 (5.26) |
| Neurosurgery | 3 (2.63) |
| Oncology | 6 (5.26) |
| Ophthalmology | 5 (4.39) |
| Orthopedic surgery | 1 (0.88) |
| Otolaryngology | 2 (1.75) |
| Pathology | 5 (4.39) |
| Pediatrics | 2 (1.75) |
| Psychiatry | 4 (3.51) |
| Pulmonology | 2 (1.75) |
| Radiology | 7 (6.14) |
Target Use
In our sample, we identify 9 target uses of applying LLMs in health care. Within our samples of 114 studies, we identify 161 instances of LLM target use cases in total. The most prominent target use was decision support in 27.19% (31/114) of the studies, followed by information retrieval (24/114, 21.05%) and communication (24/114, 21.05%). Target use is related to our research foci, where the application of LLMs for communication and explanation targets at LLMs comprehensiveness. For example, Lim et al [] compare the perceived accuracy and comprehensiveness of GPT-3.5, GPT-4, and BARD on myopia-related questions.
Target Audience
We identify 8 different stakeholders in our sample. Most studies provide insights for physicians (69/114, 60.53%), researchers (32/114, 28.07%), patients (23/114, 20.18%), and students (17/114, 14.91%). Concerning the research foci, we find that physicians are more targeted in studies on the comprehensibility of LLMs (13/16, 81.25%), for example, to understand breast cancer survival []. Students are especially targeted for educative purposes to assess passing current medical licensing examinations with LLM support [,].
Technical Dimensions
Model Integration
In our sample, 2 studies (2/114, 1.75%) focus conceptually on LLM usage from an ethical or regulatory viewpoint [,]. Most manuscripts focus on using LLMs in clinical practice (65/114, 57.02%). Twenty-seven (23.68%) studies implement LLM models by extending existing LLM architectures or developing new models. However, only 2 manuscripts report on a full implementation already evaluated in a real-world setting [,]. Twenty papers (20/114, 17.54%) evaluate different model architectures, for example, to investigate racial presumptions in Claude, BERT, GPT-3.5, and GPT-4 [].
Model Novelty
Most papers apply LLMs (86/114, 75.44%) without focusing on technical changes, and only a fourth of the manuscripts extended existing models (29/114, 25.44%) by incorporating additional data, knowledge, or human assessment with 6 manuscripts doing both. We identify only 3 (2.63%) studies developing novel retrained model architectures in our sample. The 3 models include chain-of-thought-reasoning into the inference step of LLMs [], incorporate physician behavior to teach ChatGPT to reply in a doctor-specific manner [], or retrain GPT-3 on biomedical datasets containing clinical texts by question-answer challenges [].
Most model extensions are part of the LLM innovation research focus. While most studies use GPT-based architectures, studies in the innovation research name 64 model architectures, using on BERT-based and GPT-based models 23 times (23/64, 35.94%) each. The higher relative percentage indicates that BERT-based models are easier to extend for novel models than GPT-based models.
Model Architecture
We observe that authors frequently apply multiple LLM architectures. This resulted in 57 different LLM architectures used across all studies. Thirty-four of these architectures are named only once, highlighting the variety of LLMs developed for health care. In general, GPT-based (139/259, 53.66%) and BERT-based models (66/259, 25.46%) provide the vast majority of applied LLM architectures. Most of the studies apply OpenAI’s state-of-the-art models GPT-3 (n=12), GPT-3.5 (n=74), and GPT-4 (n=41), although BERT (n=17) and Llama (n=10) are frequently used as well. An overview of the different model architecture families and their use in different research foci is shown in Table 3.
| Large language model (LLM) architecture family | LLM performance | LLM societal impact | LLM Comprehensibility | LLM innovation |
| BARD | 6 (2.67) | 1 (2.13) | 1 (3.33) | 2 (3.13) |
| BERT | 62 (27.56) | 8 (17.02) | 6 (20.00) | 23 (35.94) |
| GLM | 3 (1.33) | 0 (0.00) | 0 (0.00) | 2 (3.13) |
| GPT | 119 (52.89) | 31 (65.96) | 17 (56.67) | 23 (35.94) |
| LaMDa | 2 (0.89) | 0 (0.00) | 0 (0.00) | 0 (0.00) |
| Llama | 9 (4.00) | 2 (4.26) | 4 (13.33) | 5 (7.81) |
| T5 | 5 (2.22) | 1 (2.13) | 0 (0.00) | 0 (0.00) |
| Vicuna | 6 (2.67) | 0 (0.00) | 0 (0.00) | 4 (6.25) |
| Other | 13 (5.78) | 4 (8.51) | 2 (6.67) | 5 (7.81) |
Data Types
Our sample accounts for multiple data modalities used as input or output dimensions of the LLMs. This results in 200 mentions of 11 data modalities within our sample as shown in Table 4. The most significant data type manifests as symptom descriptions (n=44), treatment options (n=31), and diagnosis (n=29). Since these modalities comprise the steps during diagnosis, they are also frequently used in combination, for example, to assess the treatment advice generated by ChatGPT if patients present themselves with symptom descriptions []. These data are also heavily used in manuscripts investigating the understandability of LLMs. The understandability of LLM is especially important when the models communicate their reasoning with patients. Interaction with patients, in turn, is required when treatment options, diagnosis, or symptom descriptions are provided.
| Data type | Values, n (%) |
| Adverse effects | 7 (6.14) |
| Clinical notes | 12 (10.53) |
| Diagnosis | 29 (25.44) |
| EHRa | 13 (11.4) |
| Examination questions | 28 (24.56) |
| Genomic data | 2 (1.75) |
| Images | 6 (5.26) |
| Patient communication | 7 (6.14) |
| Reports | 21 (18.42) |
| Symptom description | 44 (38.6) |
| Treatment options | 31 (27.19) |
aEHR: electronic health record.
In contrast, only very few manuscripts focus on traditional medical data modalities such as genomic data (n=2), imaging (n=6), or EHRs (n=13). These data modalities are usually used by physicians and, therefore, are used in innovation-driven papers, such as those by Gao et al [], who extend GPT-3.5 based on existing patient-doctor dialogues to generate a response for the diagnosis report given an ophthalmology image. The remaining papers revolve around traditionally text-related data such as patient communication or questions (n=8), clinical notes (n=12), or medical reports (n=21). Other frequently used data modalities include examination questions (n=28), for example, the performance of GPT-3.5 in the USMLE [], the MCAT [], or on Korean general surgery board examinations [].
Research Foci
In our analysis, we analyzed the targeted improvements of the manuscripts. We found 12 research themes that we aggregated into 4 research foci as shown in Table 5. Half of the 114 included manuscripts encompass multiple research foci (n=57).
| Focus and theme | Goal | Values, n (%) | |||
| LLMa performance | |||||
| Correctness | Ensuring predictive performance | 89 (78.07) | |||
| Performance improvement | Improving LLM performance | 14 (12.28) | |||
| Hallucinations | Analyzing and preventing hallucinations | 9 (7.89) | |||
| Consistency | Comparing LLM performance across contexts or over years | 8 (7.02) | |||
| LLM societal impact | |||||
| Ethics | Discuss ethical questions for LLM use in high-risk and sensitive contexts of health care | 12 (10.53) | |||
| Bias | Evaluate and mitigate biases such as stereotypes influencing health-related examinations in LLM | 6 (5.26) | |||
| Confidentiality | Evaluate risks and attack vectors in LLM use for health care data and discuss potential solutions | 4 (3.51) | |||
| Regulation | Discuss issues related to accountability and transparency requirements that need to be enforced by law to safely use LLMs in health care | 3 (2.63) | |||
| LLM comprehensibility | |||||
| Explainability | Enhance the plausibility and comprehensibility of health information provided by LLMs with additional explanations | 12 (10.53) | |||
| Interpretability | Retrace the inner mechanics of the reasoning processes of the LLM | 5 (4.39) | |||
| LLM innovations | |||||
| Novel applications | Develop novel LLM applications in various health care use cases and research | 14 (12.28) | |||
| Novel training techniques | Improve and develop novel training techniques for LLMs, for example, to improve the quality of LLM responses | 10 (8.77) | |||
aLLM: large language model.
Research Focus 1: LLM Performance
The research focus on LLM performance encompasses improvements in model performance, ensuring correctness (including hallucinations) and assessing prediction consistency. It underscores the importance of building models that improve performance by providing correct, hallucination-free, and consistent results. Overall, the studies show that domain-specific models improve performance and correctness.
Correctness
Ensuring correct predictions of LLMs constitutes the most prominent research theme. Correct predictions are especially crucial in health care as errors can lead to severe consequences for patients such as inappropriate treatment. Eighty-nine manuscripts [-] test the correctness of LLMs. However, the predictive performance of LLMs varies based on the source and structure of the input data [].
The potential of general-purpose LLMs for medical applications is highlighted in a manuscript that explores whether ChatGPT could generate helpful suggestions for improving a clinical decision support system. Clinicians rated the generated suggestions as highly relevant and clear, with low bias []. Other manuscripts evaluate the performance of general-purpose LLMs on medical and nursing examinations, for example, on the faculty of public health’s diplomat examination []. However, for health care–specific tasks, such as biological named entity recognition [], general-purpose LLMs perform insufficiently.
Therefore, another common theme is the development of specific-purpose LLMs and testing their correctness. For example, HuatuoGPT extends ChatGPT with real-world data from doctors to outperform other LLMs in terms of human evaluations and benchmarks, yielding highly accurate results in medical contexts []. Another study tests the application of LLMs in clinical trial enrollment. For that, GPT-3.5 is used to extract medical entity tags from patient information and classify these to identify suitable clinical trials. The application achieves a high accuracy for this task [].
Performance Improvement
Within the 14 studies in our sample, performance improvement is achieved by adapting LLMs to a specific domain or developing new, specialized LLMs with domain-specific data in LLM pretraining. GatorTronGPT is one example of a clinical LLM trained on clinical and general English text using a GPT-3 architecture and outperforms general-purpose LLMs in medical research []. Furthermore, domain adaptation of LLMs improves the predictive performance, for example, to classify nuclear medicine reports []. One manuscript compares models ranging from 220 million to 175 billion parameters and finds that smaller, domain-specific models often match or exceed the performance of larger general-purpose models [].
Hallucinations
In our sample, we found 9 studies investigating hallucinations of LLMs. Hallucinations occur especially in summaries of textual information [] or in 12.5% of GPT-based cancer treatment recommendations []. Another study finds that using LLMs for drug-drug interactions also produced many hallucinations [].
Consistency
The consistency of LLMs evaluates performance across different contexts or over time. One manuscript evaluates ChatGPT’s performance on the Japanese National Nursing Examination across multiple years []. In another case, researchers assess the consistency across GPT-3.5 and GPT-4 on questions related to heart failure []. Overall, the importance of reproducibility in complex decision-making tasks is highlighted [].
Research Focus 2: LLM Societal Impact
The second research focus pertains to the societal impact of LLMs. The societal considerations surrounding LLMs in health care highlight complex challenges related to ethics, bias, confidentiality, and regulation. While LLMs have the potential to improve health care provision, their deployment must be carefully managed to mitigate bias, ensure patient data confidentiality, and ensure ethical appropriateness. Moreover, the absence of clear regulations and accountability frameworks presents a critical barrier to their widespread and safe use.
Ethics
Twelve studies investigate ethical concerns such as appropriate empathy of LLMs. For instance, 1 manuscript finds that LLM responses to patient questions express appropriate empathy []. On the contrary, another manuscript suggests that LLMs fail to resolve high-risk scenarios, such as suicide prevention []. In their manuscript, only a few models provide essential crisis resources such as suicide hotlines, highlighting the need for model refinement to comply with ethical standards [].
Bias
Six studies show that LLMs can replicate and amplify harmful stereotypes in medical content. For example, LLMs propagate race-based misconceptions when responding to historically involved race-based medicine []. Another study shows location-based biases in cardiovascular disease risk assessment by ChatGPT []. Bias extends beyond location, for example, to gender as ChatGPT changes its predictions more frequently when race or gender descriptors are added to medical texts []. While these findings highlight the importance of addressing bias in LLMs to prevent exacerbating health care inequalities, research also discusses potential solutions. For example, LLMs are shown to help identify bias within a corpus of text [].
Confidentiality
Ensuring data confidentiality is still an issue for LLMs in health care []. Studies highlight how adversarial attacks could be used to generate misinformation or leak private patient data []. Research tries to tackle these issues with a privacy-aware data augmentation approach for LLM-based patient-trial matching [] or using GPT-3.5 to identify confidential content within clinical notes [].
Regulation
Three studies emphasize the need for better regulation to ensure safe usage of LLMs in health care [,,]. This lack of regulation poses questions about who is held accountable for errors in diagnosis or treatment recommendations [,] and the transparency of LLM functionality [].
Research Focus 3: LLM Comprehensibility
The third research focus deals with the comprehensibility of LLMs, particularly in terms of explainability and interpretability. On the one hand, explainability provides insights into LLM predictions and recommendations. Interpretability, on the other hand, offers explanations for the internal processes of LLMs. Both explainability and interpretability are crucial for health care, as health care professionals need to rely on clear, accurate information to make decisions.
Explainability
In our sample, 12 studies focus on improving the explainability of LLMs either by enhancing LLMs with additional information or using self-explanations. For example, LLMs incorporate additional information in the form of in-text citations and references [] or include causal graphs to explore the effect of different genetic perturbations on the survival rates of patients with breast cancer [].
LLMs self-explaining their reasoning helps students understand complex medical conditions and generate personalized explanations of MCAT-related questions []. Similarly, 1 manuscript focuses on using LLMs to explain its reasoning for diagnosing certain mental health conditions [].
Interpretability
Five studies focus on the inner reasoning processes of LLMs. For example, 1 manuscript assesses the reasoning of ChatGPT behind the responses for 119 public health examination questions []. In another study, model interpretability is enhanced by using the Alpaca architecture to map radiologist findings to radiological images easing comprehensibility for physicians []. Another approach uses GPT-3 to generate interpretability by grouping patients based on the similarity of their embeddings for medication prescriptions [].
Research Focus 4: LLM Innovation
The fourth research focuses on LLM innovation through novel applications and training techniques. This focus explores the adaptation and extension of LLMs to accelerate, improve, and automate health care tasks. At the same time, novel training techniques, such as multimodal learning and context-enriched approaches, are extending the capabilities of LLMs to more complex medical contexts.
Novel Applications
Fourteen studies present novel applications of LLMs in specific areas. One manuscript reports that LLMs could reduce the dimensionality of complex health care data such as ICD-10 (International Statistical Classification of Diseases, Tenth Revision) codes while maintaining the original embeddings []. Another application uses LLMs to create patient clusters based on pharmacy records for more personalized treatment recommendations []. In biomedical research, LLMs trained on additional medical literature identify and validate disease-relevant targets []. Studies also show that LLMs can be used for theme-driven analyses (eg, of opioid use disorder) uncovering both frequent and clinically undocumented patterns [].
In health care education, LLMs are used to create examination questions to accelerate examination generation, standardize assessments, and reduce bias in questions []. When preparing for examinations, LLMs could respond to individual learning styles and needs in various medical domains [,].
Novel Training Techniques
Ten studies in our sample develop novel training techniques to adapt LLMs to health care. Some of these approaches [,,] show that medical data can be directly used to pretrain LLMs. Adapting LLMs or prompts to medical domains improves their performance [,,,] and could simulate expert thought processes [].
In a more general setting, the research explores the integration of external knowledge sources to improve the accuracy of Llama in diagnosing disease symptoms [] or the reliability of ChatGPT responses []. Furthermore, LLMs are used to clean and annotate medical data to fine-tune additional LLMs [].
Other studies develop a novel multimodal model fine-tuned on images, diagnostic reports, and LLM-based dialogues to enhance condition diagnosis and response generation [,]. These multimodal training approaches provide LLMs with a broader range of inputs to increase diagnosing versatility.
Discussion
Principal Results
With the rise of LLMs in the past years, research using LLMs in health care has significantly increased. We found that the manuscripts in our literature sample targeted a broad variety of stakeholders in the health care sector such as physicians (69/114, 60.53%), researchers (32/114, 28.07%), patients (23/114, 20.18%), or medical students (17/114, 14.91%).
We identified 4 primary research foci for LLMs in health care: LLM performance, LLM societal impact, LLM comprehensibility, and LLM innovation. With several manuscripts tackling 2 research foci simultaneously, most of our manuscripts focused on the performance of LLMs (87/114, 85.09%).
A large number of studies focusing on LLM performance show the importance of accurate, reliable LLM results in paving the way for the application of LLMs in health care. This research focus supports the idea that general-purpose LLMs are not yet ready to be applied to health care tasks and instead need domain adaptation. As the performance of general-purpose LLMs progresses, we found that the predictive performance seems to be better in BERT-based models when investigating a limited sample of 13 manuscripts that used BERT-based and GPT-based models. While the performance difference was often marginal, GPT-based models seem to have a higher recall but lower precision [,] or are used to augment training data for BERT-based models [,]. This could result from the model architectures. While both model architectures use transformer techniques, BERT-based models use a bidirectional encoder [] and GPT-based models use a unidirectional decoder []. Therefore, BERT can consider words in both reading directions when understanding text, while GPT processes text unidirectionally (usually from left to right) and is optimized for generating texts. This shows an interesting conflict of LLM architectures and highlights the benefits of BERT-based models for medical tasks such as knowledge discovery or report generation [], while GPT-based models are better suited for language generation and communicative purposes (eg, examination preparation [] or dialogues with patients []).
In addition, the consistency of LLMs was rarely investigated in our sample. Systematic evaluation over time is missing to allow reliable decisions on the application of LLMs. Some studies have started to compare domain-adapted models against domain-agnostic models [,]. Our findings suggest that domain-adapted BERT-based models provide the best predictive performance. Further research should strengthen these findings by conducting computational benchmark evaluations between models. Relatedly, hallucinations were evaluated only for general-purpose LLMs in the retrieved manuscripts. Moreover, many hallucinations were present in the target use case of text summarization, although the textual information was directly given to the LLM. For other use cases (eg, knowledge discovery), hallucinations have not been examined in our sample. Several mitigation strategies for hallucinations were developed, such as enabling users to detect hallucinations themselves [] or reinforcing model inference by incorporating human feedback []. As a second research focus, the societal impact of LLMs was investigated (25/114, 21.93%). Our results hint that LLMs in health care come along with race-based misconceptions that reproduce harmful stereotypes [] and provide insufficient support in high-risk scenarios such as suicide prevention []. Similar to research on artificial intelligence, societal and ethical impact should be considered when using LLMs in health care []. This indicates that further research is needed to determine how bias in health care can be prevented. This is in line with existing reviews [-] that also note various challenges of applying LLMs in health care (eg, hallucination, fairness, bias, privacy, and legal concerns) underlying the relevance of this research focus. In our sample, societal impact is predominantly evaluated with GPT-based models instead of other model architectures such as BERT. When accounting for the societal impact of LLMs, an investigation of multiple model architectures could provide new insights. Further manuscripts highlighted confidentiality and privacy issues when using and adapting LLM with sensitive health care data. Our results indicate that future research still needs to determine privacy-preserving strategies when using LLMs in health care. A possible avenue for future research could be to introduce privacy-preserving training approaches such as federated learning that have been shown to work in health care settings []. The third research focus on model comprehensibility included a limited number of 16 studies (16/114, 14.04%). In this selection, a minimal tendency can be made toward extending state-of-the-art LLM models instead of simply applying them. In contrast, despite this limited number of manuscripts, the focus suggests that LLM comprehensibility is tailored toward physicians in 13 of 16 (81.25%) studies. In addition, the frequent use of LLMs for decision support (9/16, 56.25%) and classification tasks (4/16, 25.00%) further emphasizes explanations targeted toward medical personnel. While artificial intelligence research has already highlighted the need for explanations on model reasoning to improve model reliance [], the explanations should not only adhere the physician’s perspective but also include other stakeholders. Patients’ perspectives should mainly be investigated when assessing model comprehensibility. To that end, our sample suggests that the comprehensibility studies instead comprised patient-specific data types such as symptom descriptions (8/16, 50.00%), treatment options (7/16, 43.75%), EHR (3/16, 18.75%), or images (3/16, 18.75%). As LLMs become increasingly integrated into health care, improving explainability and regulations will be crucial to ensure the safe and responsible use of LLMs and create an environment where LLMs can be used effectively in health care settings [].
As a fourth research focus, we identified studies focusing on model innovation. In our literature review, we found that 75.44% of the studies applied existing models and 25.44% of the studies extended models (with 6 studies doing both). The studies in our sample rarely implemented new models (2.63%). The studies mostly used state-of-the-art general-purpose models such as GPT-3.5, GPT-4, or BERT. While this holds true for the entire dataset, innovation (ie, model innovation or novel applications) was comparatively more performed on BERT-based models (35.94%) than on GPT-based models (35.94%). This includes health care–specific extensions of LLMs such as BioBERT, ClinicalBERT, or MedBERT that enhance the performance of these general-purpose models for the health care domain by including external knowledge sources. This again builds on the aforementioned differences in model architectures where BERT-based model architectures are better suited for knowledge discovery tasks and GPT-based model architectures rather than solving communicative purposes.
These insights help practitioners to select appropriate LLM model architectures. If the focus of the usage of the LLM is to ease communication, GPT-based models seem more appropriate. When LLMs are used for innovative purposes (such as knowledge discovery), our sample suggests the superiority of BERT-based models. However, most general-purpose models still require domain adaptation and could benefit from the incorporation of domain knowledge [,].
Limitations
This study is not without limitations. First, all codings and assessments of the dimensions were subjectively assessed by the authors of this manuscript. While we tried to minimize the subjectiveness through double-coding and regular meetings to discuss and ensure a similar understanding of all dimensions and manifestations, the final assessment is still subject to subjective assessments. Second, our comparison of GPT-based and BERT-based models focuses on a limited sample of 13 manuscripts that report on the performance of both. Future benchmark evaluations of these manuscripts are needed to support these tendencies. Third, we conducted the assessment based on a literature search in January 2024. While we included early works from prepublication databases such as arXiv or medRxiv to reduce the impact of later publication, maintaining a current and up-to-date review in this fast-changing field is challenging. For example, the introduction of GPT Playground might have facilitated the implementation and adaptation of GPT-based models. Fourth, in our search term, we focused on LLMs and did not include further related concepts (eg, transformer-based models). Although we included established LLMs in our search query, this might have led to the exclusion of relevant studies. Finally, we assume that manuscripts retrieved from prepublication databases have sufficient scientific rigor. While we excluded only 2 publications that we deemed of insufficient quality, manuscripts included in our sample might not hold up to highest scientific standards.
Conclusions
In this literature review, we identified 4 research foci for LLMs in health care and highlighted specifics to each focus. Most manuscripts in our review focused on the performance of LLMs. Our sample shows a tendency that GPT-based models provide higher recall and are more heavily involved in societal questions that require interaction and communication. In contrast, BERT-based models show higher precision and are more intensely used for innovation such as classification or knowledge discovery. This is in line with the architectural design of the models where the unidirectional design of GPT is better suited for generative purposes. In contrast, the bidirectional design of BERT is better suited for question-answering. We also find that more manuscripts use diagnosis-related data such as treatment recommendations or symptom descriptions than manuscripts using traditional medical data such as images or EHR. This, again highlights the novel ways to communicate with LLMs. While these findings seem to hold through the manuscripts comparing different LLMs, future research should conduct thorough comparative benchmark evaluations.
With this review, we aid health care professionals in understanding the benefits of specific LLM technologies. GPT-based models are better suited for communicative purposes due to their enhanced ability to generate text. However, in the current state, general-purpose LLMs might be able to pass medical examinations but are unable to explain or reason their predictions. When incorporating medical-specific knowledge, the performance of LLMs was further increased. Since we investigate only published literature, future research should extend the comparison of performance and bias of domain-adapted and general-purpose LLM models. Without these benchmarks, we should teach physicians to use LLMs as supporting tools while raising awareness to potential model insufficiencies.
Acknowledgments
The authors would like to acknowledge the assistance of their internal friendly reviewers who helped them revise this manuscript before submission. The authors used the generative artificial intelligence tools ChatGPT by OpenAI and Grammarly to revise the writing of the manuscript.
Conflicts of Interest
None declared.
Multimedia Appendix 1
PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) checklist.
DOCX File , 110 KBMultimedia Appendix 2
Overview of all included publications along with our coding dimensions and our analysis of each dimension for each research focus.
XLSX File (Microsoft Excel File), 367 KBMultimedia Appendix 3
Additional tables and figures for the analysis of each dimension.
DOCX File , 156 KBReferences
- OpenAI. Introducing ChatGPT. Nov 30, 2022. URL: https://openai.com/blog/chatgpt [accessed 2025-03-12]
- Hu K. ChatGPT sets record for fastest-growing user base—analyst note. Feb 01, 2023. URL: https://www.reuters.com/technology/chatgpt-sets-record-fastest-growing-user-base-analyst-note-2023-02-01/ [accessed 2025-05-09]
- Devlin J, Chang MW, Lee K, Toutanova K. BERT: pre-training of deep bidirectional transformers for language understanding. May 24, 2019. URL: https://arxiv.org/abs/1810.04805 [accessed 2025-04-18]
- Touvron H, Lavril T, Izacard G, Martinet X, Lachaux MA, Lacroix T, et al. LLaMA: open and efficient foundation language models. ArXiv. posted online on 2023. 2023:1-27. [FREE Full text]
- Kasneci E, Sessler K, Küchemann S, Bannert M, Dementieva D, Fischer F, et al. ChatGPT for good? On opportunities and challenges of large language models for education. Learn Indiv Differences. 2023;103(April 2023):1-9. [FREE Full text] [CrossRef]
- Carlini N, Tramèr F, Wallace E, Jagielski M, Herbert-Voss A, Lee K, et al. Extracting training data from large language models. 2021. Presented at: 30th USENIX Security Symposium (USENIX Security 21) USENIX Association; 2023 August 9:2633-2650; Anaheim CA USA. URL: https://www.usenix.org/conference/usenixsecurity21/presentation/carlini-extracting
- Zhao WX, Zhou K, Li J, Tang T, Wang X, Hou Y, et al. A survey of large language models. ArXiv. :1-144. Preprint posted online 2023. [FREE Full text]
- Contreras Kallens P, Kristensen-McLachlan RD, Christiansen MH. Large language models demonstrate the potential of statistical learning in language. Cognitive Science. Feb 25, 2023;47(3):1-6. [CrossRef] [Medline]
- Chew R, Bollenbacher J, Wenger M, Speer J, Kim A. LLM-assisted content analysis: Using large language models to support deductive coding. ArXiv. :1-27. Preprint posted online in 2023
- He X, Bresson X, Laurent T, Perold A, LeCun Y, Hooi B. Harnessing explanations: LLm-to-LM interpreter for enhanced text-attributed graph representation learning. ArXiv. :1-22. Preprint posted online in 2023
- Zhao Z, Song S, Duah B, Macbeth J, Carter S, Van MP, et al. More human than human: LLM-generated narratives outperform human-LLM interleaved narratives. 2023. Presented at: C&C '23: Proceedings of the 15th Conference on Creativity and Cognition; June 19-21, 2023:368-370; Virtual Event USA. [CrossRef]
- Liu Z, Zhong A, Li Y, Yang L, Ju C, Wu Z, et al. Radiology-GPT: a large language model for radiology. ArXiv. :1-16. Preprint posted online in 2024. [FREE Full text] [CrossRef]
- Shyr C, Hu Y, Bastarache L, Cheng A, Hamid R, Harris P, et al. Identifying and extracting rare diseases and their phenotypes with large language models. Journal of Healthcare Informatics Research. 2024;8(2):438-461. [FREE Full text] [CrossRef] [Medline]
- Han Z, Battaglia F, Udaiyar A, Fooks A, Terlecky SR. An explorative assessment of ChatGPT as an aid in medical education: use it with caution. Medical Teacher. 2024;46(5):657-664. [CrossRef] [Medline]
- Meskó B, Topol EJ. The imperative for regulatory oversight of large language models (or generative AI) in healthcare. npj Digital Medicine. 2023;6(1):120. [CrossRef]
- Jin Q, Yang Y, Chen Q, Lu Z. GeneGPT: augmenting large language models with domain tools for improved access to biomedical information. Bioinformatics. 2024;40(2):btae075. [FREE Full text] [CrossRef] [Medline]
- Singhal K, Azizi S, Tu T, Mahdavi SS, Wei J, Chung HW, et al. Large language models encode clinical knowledge. Nature. 2023;620(7972):172-180. [FREE Full text] [CrossRef] [Medline]
- Li J, Dada A, Puladi B, Kleesiek J, Egger J. ChatGPT in healthcare: a taxonomy and systematic review. Computer Methods and Programs in Biomedicine. 2024;245:108013. [FREE Full text] [CrossRef] [Medline]
- Thirunavukarasu AJ, Ting DSJ, Elangovan K, Gutierrez L, Tan TF, Ting DSW. Large language models in medicine. Nature Medicine. 2023;29(8):1930-1940. [CrossRef] [Medline]
- Yang R, Tan TF, Lu W, Thirunavukarasu AJ, Ting DSW, Liu N. Large language models in health care: development, applications, and challenges. Health Care Science. 2023;2(4):255-263. [FREE Full text] [CrossRef] [Medline]
- Tian S, Jin Q, Yeganova L, Lai PT, Zhu Q, Chen X, et al. Opportunities and challenges for ChatGPT and large language models in biomedicine and health. Briefings in Bioinformatics. 2023;25(1):1-13. [FREE Full text] [CrossRef] [Medline]
- Arksey H, O'Malley L. Scoping studies: towards a methodological framework. International Journal of Social Research Methodology. 2005;8(1):19-32. [CrossRef]
- Paré G, Trudel MC, Jaana M, Kitsiou S. Synthesizing information systems knowledge: A typology of literature reviews. Information & Management. 2015;52(2):183-199. [CrossRef]
- Radford A, Narasimhan K, Salimans T, Sutskever I. Improving language understanding by generative pre-training. 2018. URL: https://cdn.openai.com/research-covers/language-unsupervised/language_understanding_paper.pdf [accessed 2025-04-18]
- Tricco AC, Lillie E, Zarin W, O'Brien KK, Colquhoun H, Levac D, et al. PRISMA Extension for scoping reviews (PRISMA-ScR): checklist and explanation. Annals of Internal Medicine. 2018;169(7):467-473. [FREE Full text] [CrossRef] [Medline]
- Zagirova D, Pushkov S, Leung GHD, Liu BHM, Urban A, Sidorenko D, et al. Biomedical generative pre-trained based transformer language model for age-related disease target discovery. Aging. 2023;15(18):9293-9309. [FREE Full text] [CrossRef] [Medline]
- Romano W, Sharif O, Basak M, Gatto J, Preum SM. Theme-driven keyphrase extraction to analyze social media discourse. 2024. Presented at: Proceedings of the International AAAI Conference on Web and Social Media; 2024 May 28:1315-1327; New York, NY. [CrossRef]
- Moradi M, Blagec K, Haberl F, Samwald M. Gpt-3 models are poor few-shot learners in the biomedical domain. ArXiv. :1-9. Preprint posted online in 2021
- Hu M, Alkhairy S, Lee I, Pillich RT, Fong D, Smith K, et al. Evaluation of large language models for discovery of gene set function. Nature Methods. 2025;22(1):82-91. [CrossRef] [Medline]
- Susnjak T. Applying BERT and ChatGPT for sentiment analysis of Lyme disease in scientific literature. In: Gilbert L, editor. Borrelia burgdorferi. New York, NY. Springer US; 2024:173-183.
- Alrdahi H, Han L, Suvalov H, Nenadic G. Medmine: Examining pre-trained language models on medication mining. ArXiv. :1-7. Preprint posted online in 2023
- Deng Z, Gao W, Chen C, Niu Z, Gong Z, Zhang R, et al. OphGLM: An ophthalmology large language-and-vision assistant. Artificial Intelligence in Medicine. Nov 2024;157:103001. [FREE Full text] [Medline]
- Li Y, Li Z, Zhang K, Dan R, Jiang S, Zhang Y. ChatDoctor: a medical chat model fine-tuned on a large language model meta-AI (LLaMA) using medical domain knowledge. Cureus. 2023;15(6):e40895. [FREE Full text] [CrossRef] [Medline]
- Kung TH, Cheatham M, Medenilla A, Sillos C, De Leon L, Elepaño C, et al. Performance of ChatGPT on USMLE: potential for AI-assisted medical education using large language models. PLOS Digital Health Public Library of Science (PLoS). 2023;2(2):e0000198. [FREE Full text] [CrossRef] [Medline]
- Bommineni VL, Bhagwagar S, Balcarcel D, Davatzikos C, Boyer D. Performance of ChatGPT on the MCAT: the road to personalized and equitable premedical learning. medRxiv, posted in 2023. 2023:1-19. [CrossRef]
- Zong H, Li J, Wu E, Wu R, Lu J, Shen B. Performance of ChatGPT on Chinese national medical licensing examinations: a five-year examination evaluation study for physicians, pharmacists and nurses. BMC Medical Education. 2024;24(1):143. [FREE Full text] [CrossRef] [Medline]
- Schubert MC, Wick W, Venkataramani V. Performance of large language models on a neurology board-style examination. JAMA Network Open. 2023;6(12):e2346721. [FREE Full text] [CrossRef] [Medline]
- Zhang H, Chen J, Jiang F, Yu F, Chen Z, Li J, et al. HuatuoGPT, towards taming language model to be a doctor. ArXiv. :1-21. Preprint posted online in 2023. [FREE Full text] [CrossRef]
- Liu S, McCoy AB, Wright AP, Carew B, Genkins JZ, Huang SS, et al. Leveraging large language models for generating responses to patient messages-a subjective analysis. J Am Med Inform Assoc. May 20, 2024;31(6):1367-1379. [FREE Full text] [CrossRef] [Medline]
- Nastasi AJ, Courtright KR, Halpern SD, Weissman GE. A vignette-based evaluation of ChatGPT's ability to provide appropriate and equitable medical advice across care contexts. Scientific Reports. 2023;13(1):17885. [FREE Full text] [CrossRef] [Medline]
- Lim ZW, Pushpanathan K, Yew SME, Lai Y, Sun CH, Lam JSH, et al. Benchmarking large language models’ performances for myopia care: a comparative analysis of ChatGPT-3.5, ChatGPT-4.0, and Google Bard. eBioMedicine. 2023;95:104770. [CrossRef]
- Farooq M, Hardan S, Zhumbhayeva A, Zheng Y, Nakov P, Zhang K. Understanding breast cancer survival: using causality and language models on multi-omics data. 2023. URL: http://arxiv.org/abs/2305.18410 [accessed 2024-10-12]
- Fleming SL, Morse K, Kumar A, Chiang CC, Patel B, Brunskill E, et al. Assessing the potential of USMLE-like exam questions generated by GPT-4. medRxiv. :1-6. Preprint posted online in 2024. [FREE Full text]
- Rahimzadeh V, Kostick-Quenet K, Blumenthal Barby J, McGuire AL. Ethics education for healthcare professionals in the era of ChatGPT and other large language models: do we still need It? The American Journal of Bioethics. 2023;23(10):17-27. [FREE Full text] [CrossRef] [Medline]
- Leiser F, Rank S, Schmidt-Kraepelin M, Thiebes S, Sunyaev A. Medical informed machine learning: a scoping review and future research directions. Artif Intell Med. 2023;145:102676. [CrossRef] [Medline]
- Omiye JA, Lester JC, Spichak S, Rotemberg V, Daneshjou R. Large language models propagate race-based medicine. npj Digital Medicine. 2023;6(1):195. [FREE Full text] [CrossRef] [Medline]
- Wu CK, Chen WL, Chen HH. Large language models perform diagnostic reasoning. ArXiv. :1-10. Preprint posted online in 2023. [FREE Full text]
- Peng C, Yang X, Chen A, Smith KE, PourNejatian N, Costa AB, et al. A study of generative large language model for medical research and healthcare. npj Digital Medicine. 2023;6(1):210. [FREE Full text] [CrossRef] [Medline]
- Oh N, Choi GS, Lee WY. ChatGPT goes to the operating room: evaluating GPT-4 performance and its potential in surgical education and training in the era of large language models. Annals of Surgial Treatment and Research. 2023;104(5):269-273. [FREE Full text] [CrossRef] [Medline]
- Lehman E, Hernandez E, Mahajan D, Wulff J, Smith MJ, Ziegler Z, et al. Do we still need clinical language models? ArXiv. :1-23. Preprint posted online in 2023. [FREE Full text]
- Chen Z, Balan MM, Brown K. Language models are few-shot learners for prognostic prediction. ArXiv. :1-9. Preprint posted online in 2023
- Bumgardner VKC, Mullen A, Armstrong SE, Hickey C, Marek V, Talbert J. Local Large Language Models for Complex Structured Tasks. AMIA Jt Summits Transl Sci Proc. 2024;2024:105-114. [FREE Full text] [Medline]
- Zheng H, Zhu Y, Jiang LY, Cho K, Oermann EK. Making the most out of the limited context length: predictive power varies with clinical note type and note section. In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop). Association for Computational Linguistics; 2023. Presented at: 61st Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop); July 10-12, 2023:104-108; Toronto, Canada. URL: http://arxiv.org/abs/2307.07051 [CrossRef]
- Liu S, Wright AP, Patterson BL, Wanderer JP, Turer RW, Nelson SD, et al. Assessing the value of ChatGPT for clinical decision support optimization. medRxiv. :1-20. Preprint posted online in 2023. [FREE Full text] [CrossRef] [Medline]
- Davies NP, Wilson R, Winder MS, Tunster SJ, McVicar K, Thakrar S, et al. ChatGPT sits the DFPH exam: large language model performance and potential to support public health learning. BMC Medical Education. 2024;24(1):57. [FREE Full text] [CrossRef] [Medline]
- Tang R, Han X, Jiang X, Hu X. Does synthetic data generation of LLMs help clinical text mining? ArXiv. :1-10. Preprint posted online in 2023. [FREE Full text]
- Peikos G, Symeonidis S, Kasela P, Pasi G. Utilizing ChatGPT to enhance clinical trial enrollment. ArXiv. :1-39. Preprint posted online in 2023. [FREE Full text] [CrossRef]
- Huemann Z, Lee C, Hu J, Cho SY, Bradshaw TJ. Domain-adapted large language models for classifying nuclear medicine reports. Radiology Artificial Intelligence. 2023;5(6):e220281. [FREE Full text] [CrossRef] [Medline]
- Hulman A, Dollerup OL, Mortensen JF, Fenech ME, Norman K, Støvring H, et al. ChatGPT—versus human-generated answers to frequently asked questions about diabetes: a turing test-inspired survey among employees of a danish diabetes center. PLoS One. 2023;18(8):e0290773. [FREE Full text] [CrossRef] [Medline]
- Tang L, Sun Z, Idnay B, Nestor JG, Soroush A, Elias PA, et al. Evaluating large language models on medical evidence summarization. npj Digital Medicine. 2023;6(1):158. [FREE Full text] [CrossRef] [Medline]
- Perlis RH. Research Letter: Application of GPT-4 to select next-step antidepressant treatment in major depression. medRxiv. :1-9. Preprint posted online in 2023. [CrossRef]
- Taira K, Itaya T, Hanada A. Performance of the large language model ChatGPT on the national nurse examinations in Japan: evaluation study. JMIR Nursing. 2023;6:1-8. [FREE Full text] [CrossRef] [Medline]
- King R, Samaan JS, Yeo YH, Mody B, Lombardo DM, Ghashghaei R. Appropriateness of ChatGPT in answering heart failure related questions. Heart Lung and Circulation. 2024;33(9):1314-1318. [FREE Full text] [CrossRef] [Medline]
- Heston TF. Evaluating risk progression in mental health chatbots using escalating prompts. medRxiv. :1-16. Preprint posted online in 2023. [CrossRef]
- Marafino BJ, Liu VX. Performance of a large language model (ChatGPT-3.5) for pooled cohort equation estimation of atherosclerotic cardiovascular disease risk. medRxiv. :1-13. Preprint posted online in 2023. [CrossRef]
- Guevara M, Chen S, Thomas S, Chaunzwa TL, Franco I, Kann BH, et al. Large language models to identify social determinants of health in electronic health records. npj Digital Medicine. 2024;7(1):6. [FREE Full text] [CrossRef] [Medline]
- Zakka C, Shad R, Chaurasia A, Dalal AR, Kim JL, Moor M, et al. Almanac — Retrieval-Augmented Language Models for Clinical Medicine. The New England Journal of Medicine AI. Jan 25, 2024;1(2):1. [FREE Full text] [CrossRef]
- Yuan J, Tang R, Jiang X, Hu X. Large Language Models for Healthcare Data Augmentation: An Example on Patient-Trial Matching. AMIA Annual Symposium Proceedings. 2023;2023:1324-1333. [FREE Full text] [Medline]
- Rabbani N, Brown C, Bedgood M, Goldstein RL, Carlson JL, Pageler NM, et al. Evaluation of a large language model to identify confidential content in adolescent encounter notes. JAMA Pediatrics. 2024;178(3):308-310. [FREE Full text] [CrossRef] [Medline]
- Guo E, Gupta M, Sinha S, Rössler K, Tatagiba M, Akagami R, et al. neuroGPT-X: toward a clinic-ready large language model. Journal of Neurosurgery. Apr 01, 2024;140(4):1041-1053. [CrossRef] [Medline]
- Yang K, Ji S, Zhang T, Xie Q, Kuang Z, Ananiadou S. Towards interpretable mental health analysis with large language models. 2023. Presented at: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing; 2023 Dec 1:6056-6077; Singapore. [CrossRef]
- Liu Z, Wu Z, Hu M, Zhao B, Zhao L, Zhang T, et al. PharmacyGPT: The AI Pharmacist. ArXiv. :1-23. Preprint posted online in 2023. [FREE Full text]
- Kane MJ, King C, Esserman D, Latham NK, Greene EJ, Ganz DA. A compressed large language model embedding dataset of ICD 10 CM descriptions. BMC Bioinformatics. 2023;24(1):482. [FREE Full text] [CrossRef] [Medline]
- Xie Y, Seth I, Hunter-Smith DJ, Rozen WM, Seifman MA. Investigating the impact of innovative AI chatbot on post-pandemic medical education and clinical assistance: a comprehensive analysis. ANZ Journal of Surgery. 2024;94(1-2):68-77. [CrossRef] [Medline]
- Sarker S, Qian L, Dong X. Medical data augmentation via ChatGPT: a case study on medication identification and medication event classification. ArXiv. :1-5. Preprint posted online in 2023
- Ma L, Han J, Wang Z, Zhang D. CephGPT-4: an interactive multimodal cephalometric measurement and diagnostic system with visual large language model. ArXiv. :1-10. Preprint posted online in 2023. [FREE Full text]
- Lu Y, Srinivasan G, Preum S, Pettus J, Davis M, Greenburg J, et al. Assessing the impact of pretraining domain relevance on large language models across various pathology reporting tasks. medRxiv. :1-63. Preprint posted online in 2023. [FREE Full text] [CrossRef]
- Xu L, Sanders L, Li K, Chow JCL. Chatbot for health care and oncology applications using artificial intelligence and machine learning: systematic review. JMIR Cancer. 2021;7(4):e27850. [FREE Full text] [CrossRef] [Medline]
- Chen S, Li Y, Lu S, Van H, Aerts HJWL, Savova GK, et al. Evaluating the ChatGPT family of models for biomedical reasoning and classification. Journal of the American Medical Informatics Association. 2024;31(4):940-948. [CrossRef] [Medline]
- Kirk R, Mediratta I, Nalmpantis C, Luketina J, Hambro E, Grefenstette E, et al. Understanding the effects of RLHF on LLM generalisation and diversity. ArXiv. :1-34. Preprint posted online in 2023. [CrossRef]
- Morley J, Machado CCV, Burr C, Cowls J, Joshi I, Taddeo M, et al. The ethics of AI in health care: a mapping review. Social Science & Medicine. 2020;260:113172. [CrossRef] [Medline]
- Rieke N, Hancox J, Li W, Milletarì F, Roth HR, Albarqouni S, et al. The future of digital health with federated learning. npj Digital Medicine. 2020;3(1):119. [FREE Full text] [CrossRef] [Medline]
- Buçinca Z, Malaya MB, Gajos KZ. To trust or to think: cognitive forcing functions can reduce overreliance on AI in AI-assisted decision-making. 2021. Presented at: Proceedings of the ACM Human-Computer Interaction; 2021 Apr 13; New York, NY, USA. [CrossRef]
- von Rueden L, Mayer S, Beckh K, Georgiev B, Giesselbach S, Heese R, et al. Informed machine learning—a taxonomy and survey of integrating prior knowledge into learning systems. IEEE Trans Knowl Data Eng. 2021;35(1):614-633. [CrossRef]
Abbreviations
| BERT: Bidirectional Encoder Representations from Transformer |
| EHR: electronic health record |
| GPT: generative pretrained transformer |
| ICD-10: International Statistical Classification of Diseases, Tenth Revision |
| LLM: large language model |
| MCAT: Medical College Admission Test |
| PRISMA-ScR: Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews |
| USMLE: United States Medical Licensing Examination |
Edited by J Sarvestan; submitted 19.12.24; peer-reviewed by D Hu, S Kim, P Wang; comments to author 07.01.25; revised version received 28.01.25; accepted 04.04.25; published 19.06.25.
Copyright©Florian Leiser, Richard Guse, Ali Sunyaev. Originally published in the Journal of Medical Internet Research (https://www.jmir.org), 19.06.2025.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in the Journal of Medical Internet Research (ISSN 1438-8871), is properly cited. The complete bibliographic information, a link to the original publication on https://www.jmir.org/, as well as this copyright and license information must be included.