This scoping review explored the use of LLMs in chronic disease management following the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews) [25]. A comprehensive search was conducted in PubMed, Scopus, IEEE Xplore, and Google Scholar, selected for their coverage of peer-reviewed medical, technical, and AI-related research. Google Scholar was included to capture a broad range of academic publications, including preprints and conference papers that may not be covered by traditional databases. Search terms included combinations of “Large language models,” “LLMs,” “ChatGPT,” “chronic diseases,” and “chronic disease management.” The search targeted both published and unpublished English-language articles from January 1, 2023, to January 15, 2025, ensuring a focus on recent advancements in LLM applications in health care.

Studies were included if they focused on applications of LLM in chronic disease management, provided full-text access, were published in English, and appeared between January 2023 and January 15, 2025. Exclusion criteria eliminated nonprimary research (reviews, editorials, viewpoints, and commentaries), abstracts without full text, non-English publications, and articles outside the date range. To capture emerging research and potentially studies, reputable non-peer–reviewed preprints from established repositories such as arXiv and medRxiv were included. Figure 1 illustrates the eligibility screening process with a decision tree.

The data from selected studies were extracted into a structured form that captured study objectives, the LLMs used, health care settings, study designs, disease management tasks, identified challenges, evaluation methods, and target users. These extracted data points were selected to ensure a structured and objective analysis aligned with the study’s aims. Study objectives provided insight into the intended applications of LLMs in chronic disease management, while health care settings contextualized their use across clinical and patient-centered environments. Study design and evaluation methods were included to assess methodological rigor and the validity of findings. In addition, data on LLM models, disease management tasks, and key challenges enabled a systematic evaluation of current applications, limitations, and areas for future research.

The Mixed Methods Appraisal Tool (MMAT, version 2018) [26] was used to perform a formal methodological quality assessment. The MMAT was selected due to its flexibility in evaluating diverse study designs included in the review. Two reviewers independently appraised each study using the 5 criteria relevant to its methodological design, with discrepancies resolved through discussion and reached consensus on final ratings. Consistent with scoping review methodology, the studies were not excluded based on quality, but results inform the interpretation of findings [25]. The findings from the included studies were synthesized and presented in alignment with the study objectives. A thematic analysis approach was used to categorize qualitative insights, grouping findings into patient-centered tasks, practitioner-centered tasks, and challenges. Discrepancies in data interpretation were resolved through consensus among the reviewers. Reference management and citation generation were conducted using Mendeley.

The PRISMA flow diagram (Figure 2) outlines each stage of the study selection process. The PRISMA-ScR framework was followed to ensure transparency and reproducibility, incorporating detailed search strategies and clearly defined exclusion criteria. A total of 446 records were identified from Google Scholar, 75 from Scopus, 56 from PubMed, and 28 from IEEE Xplore. After removing duplicates, 242 unique records underwent title and abstract screening, resulting in 61 articles for full-text review. Following the application of eligibility criteria, 29 articles were included in the final analysis.

Table 1 provides the characteristics of the 29 articles included in this scoping review. The articles were categorized based on publication type, with the majority of the studies being journal articles (n=18), followed by conference papers (n=8), and preprints (n=3). The studies used a variety of research designs, including experimental (n=10), qualitative research (n=3), comparative studies (n=4), cross-sectional study (n=2), case study (n=1), observational (n=3), retrospective cohort designs (n=3), mixed methods (n=1), pilot study (n=1), and prospective study (n=1). Table 1 provides more details on the characteristics of the reviewed studies.

As shown in Table 1, most studies were conducted in China (7/29, 24%) and the United States (4/29, 14%), with limited representation of low-resource settings. Experimental designs were predominant (10/29, 35%), and nearly half of the studies (14/29, 48%) focused on diabetes management. The studies primarily focused on adult populations (28/29, 96%), with only 1 study (1/29, 4%) specifically addressing pediatric applications. Human evaluation was the most common evaluation method (16/29, 55%), followed by automated evaluation metrics (10/29, 35%) used in prototype evaluation, with some studies using both approaches (3/29, 10%). Studies were carried out in diverse health care settings, with home care settings (10/29, 35%) and clinical settings (9/29, 31%) being the most common. This diversity in study characteristics reflects the broad application of LLMs across various health care contexts for chronic disease management.

Using the Mixed Methods Appraisal Tool [26], the studies were categorized as quantitative descriptive studies (19/29, 66%), comprizing observational, cross-sectional, case study, system development, prototype evaluation, and comparative analyses. Quantitative nonrandomized studies (3/29, 10%) included retrospective cohort and quasi-experimental designs, qualitative studies (5/29, 17%), and 2 studies with mixed methods design (2/29, 7%). Of the 29 studies, 18 studies (62.1%) were classified as high quality (meeting ≥4 of 5 criteria), 9 studies (31.1%) as moderate quality (meeting 2‐3 criteria), and 2 studies (6.9%) as low quality (meeting ≤1 criterion). The most common methodological limitations identified across studies included inadequate sampling strategy descriptions, limited participant demographic reporting, use of synthetic clinical data, and lack of external validation. Detailed quality appraisal results for each study are provided in Multimedia Appendix 1. Table 2 provides an overview of the LLM types, users, tasks, and challenges identified across the included studies.

As shown in Table 2, GPT models were the most commonly used (14/29, 48%), followed by LLaMA variants (5/29, 17%), the Bard model (3/29, 10%), and BERT-based models (2/29, 7%). LLMs were primarily used for patient education and information provision (18/29, 62%), with most studies targeting patients (18/29, 62%) rather than health care providers (11/29, 38%). Inaccurate and inconsistencies in responses (18/29, 62 %) were the most frequently reported challenge across studies.

Our literature synthesis revealed that LLMs have significant potential to improve various chronic disease management tasks. The tasks identified have been broadly categorized into patient-centered tasks and practitioner-centered tasks.

The challenges identified and presented in Table 2 were categorized as follows.

This scoping review presents 3 significant findings from the 29 included studies (n=29): (1) LLMs are mostly used for both patient-centered (18/29, 62%) and practitioner-centered (11/29, 38%) tasks, with patient education and information provision emerging as the major application (18/29, 62%); (2) despite promising applications, significant challenges still exist, particularly regarding LLM response accuracy (18/29, 62% of studies), ethical concerns (10/29, 35%), and usability issues (9/29, 31%); and (3) methodological quality varies considerably across studies, with journal articles demonstrating higher quality (13/18, 72%) compared with conference papers (3/8, 38%) and preprints (1/3, 33%). These findings highlight both the considerable promise and significant limitations of current LLM applications in chronic disease management, which are examined in detail below.

Chronic disease management is an approach to managing chronic illnesses involving screenings, regular check-ups, monitoring, coordination of treatment, medication adherence, lifestyle modifications, and patient education [1]. The findings from this scoping review reveal an increasing interest in leveraging LLMs like ChatGPT to support both patient-centered and practitioner-centered tasks [6,16,22-24,30,31].

The majority of the studies (18/29, 62%) focused on patient-centered tasks, reflecting the emphasis on patient active engagement in chronic disease management [3,46]. Patients’ active engagement enables them to monitor their symptoms, disease progress, weight, and adverse drug effects, and adhere to medication and visits [3,46]. This review found that LLMs support various patient-centered applications, including patient education and information provision, disease monitoring and self-management, emotional support and therapeutic conversations, and diagnosis and treatment assistance.

Patient education and information provision emerged as the most prominent application (18/29, 62%), with LLMs providing health information about conditions, treatment plans, and medication schedules [16,17,21,23,24,27,28,32,33,37-41]. With the right health information, individuals with chronic diseases can easily self-manage their conditions. Diagnosis and treatment applications accounted for (6/29, 21%). Studies experimented using LLMs to suggest laboratory tests, for diagnosis, and for medication generation tailored to individual patient conditions [16,22,23,32,34,41]. However, using LLMs for diagnostic and treatment has been criticized due to concerns about hallucinations and misinterpretations of clinical guidelines [22,23], highlighting the need for continued research to ensure patient safety. Therefore, LLM usage should not replace health practitioners but instead serve as complementary tools.

Self-management and disease monitoring applications (8/29, 28%) demonstrated how LLMs can facilitate home-based monitoring of various physiological parameters [6,19-21,24,37,38,42]. Recent studies have highlighted that patient engagement with health monitoring technologies is crucial for improving health outcomes in chronic disease management [47]. Studies have also shown that wearable technologies integrated with LLMs provide real-time patient-centered health data that can better inform self-management decision-making [48]. However, ensuring consistent long-term engagement is still a challenge.

Emotional support and therapeutic conversations (4/29, 14%) represented an emerging application area [18,24,37,38]. Studies showed that LLMs can provide psychological support through tailored recommendations addressing specific stressors [18], identifying barriers to behavior change [24], and offering coping strategies for patients with diabetes [38]. Emotional support is increasingly recognized as essential in chronic disease management [49], which helps patients overcome psychological barriers to treatment adherence and lifestyle modifications.

Practitioner-centered tasks (11/29, 38%) mainly revolved around clinical decision support and medical predictions. Clinical decision support applications (8/29, 28%) provided health care practitioners with actionable information to enhance decision-making. Studies demonstrated that LLMs can generate personalized medical reports, generate treatment recommendations, and support diagnostic processes that assist health care specialists in making informed decisions [17,21]. However, while LLMs enhance diagnostic efficiency, concerns regarding inconsistent outputs pose barriers to clinical adoption. Medical prediction applications account for (7/29, 24%) of LLM use in chronic disease management, showing strong potential for early disease detection and risk stratification. By integrating structured and unstructured clinical data, such as lab results, clinical notes, and imaging, LLMs enable more comprehensive and accurate predictive models compared with traditional methods [17,29,31,34-36].

Notably, real-time risk assessment tools, like the ambulatory device developed for sickle cell anemia management [20], demonstrate how LLMs can predict complications before symptoms appear. However, challenges relating to medical accuracy still limit their seamless integration into clinical workflows [6,16,17,21-23,27]. A significant emerging trend is the development of retrieval-augmented generation (RAG) frameworks that enhance prediction accuracy by dynamically incorporating up-to-date medical knowledge [22,35,50-53]. Future advancements should focus on seamless integration of LLMs with existing medical systems, wearable health technologies, and mobile health applications to improve interoperability and trustworthiness.

The methodological quality assessment revealed notable trends that should be considered while interpreting the results of this scoping review. High-quality studies (18/29, 62%) were not uniformly distributed across LLM application tasks, studies examining medical prediction applications [29,31,35,36] and patient education and information provision [28,32,38,39] had significantly stronger methodological rigor. Studies investigating emotional support showed mixed quality, with some high-quality qualitative research [18,24], a quantitative descriptive study [37] alongside a moderate-quality mixed study [38]. Notably, self-management and disease monitoring studies showed significant methodological heterogeneity, with quality ratings ranging from low to high.

Journal articles demonstrated a substantially higher proportion of high-quality studies (13/18, 72%) compared with conference papers (3/8, 38%) and preprints (1/3, 33%), suggesting that peer-review processes enhance methodological quality. The identified common limitations, including inadequate sampling strategies, limited participant demographic reporting, and insufficient methodological transparency, were more prevalent in lower-tier publication sources, including conference papers and preprints. Quantitative descriptive studies, especially those focused on system design and prototype testing (10/29, 34.5%), showed mixed quality ratings ranging from low to high, with a common limitation being the use of synthetic data, lack of clinical validation, as they commonly prioritized technical utility. These methodological patterns significantly influence the reliability of the findings.

This section discusses the key challenges identified in the review and presents corresponding recommendations to mitigate these issues. Each challenge is followed by practical solutions to enhance the applicability of LLMs in chronic disease management.

Inaccurate and inconsistent responses emerged as a significant challenge (18/29, 62%), primarily due to poor data quality, inherent biases in training datasets, and the limited scope of knowledge constrained by the model’s training cutoff [6,22,23,31]. Given that LLM performance is intrinsically linked to data quality, flawed datasets inevitably propagate errors in outputs, a manifestation of the “garbage in, garbage out” principle [54-56]. The issue of biases in AI models has gathered significant attention in recent literature [54-56], prompting possible migration strategies [57-59]. These limitations carry critical implications for health care, as erroneous LLM outputs may lead to incorrect clinical decisions, posing significant risks to patient safety [8,12]. Hence, researchers are exploring technical solutions such as advancing domain-specific fine-tuning techniques [60,61], leveraging retrieval-augmented generation (RAG) frameworks [22,50-52], and refining outputs through reinforcement learning (RL) and prompt engineering techniques [62-64]. In addition, implementing expert validation protocols has emerged as a crucial safeguard to ensure adherence to evidence-based practice.

The scarcity of high-quality datasets for chronic diseases accounted for (6/29, 21%). Studies highlighted limitations and narrow coverage of publicly accessible clinical training datasets due to data privacy and institutional restrictions [16,19,21-23,39]. Given that experimental studies often require substantial model training datasets, the absence of adequate data poses a significant challenge to the effectiveness and success of these studies. Therefore, to address this gap, synthetic datasets and data augmentation techniques have been explored by studies [15,65]. However, these methods risk amplifying pre-existing biases in source data [54-56]. Therefore, dataset validation is essential to ensure quality, collaborative partnerships with health care institutions to access real-world clinical datasets and knowledge distillation techniques, where smaller models can be trained on outputs from larger, clinically validated models, reducing dependency on raw data volume, can be explored [66].

The computational demands of training LLMs for health care applications remain a significant challenge (6/29, 21%). Consequently, low computing resources approaches, such as quantization, parameter-efficient fine-tuning (PEFT) techniques like Low-Rank Adaptation (LORA), Quantized LoRA (QLORA), Weight-Decomposed Low-Rank Adaptation, and REFT [67-73], are evolving and are popularly used in fine-tuning LLMs in low-resource computing environments. In addition, adapter-based tuning methods provide lightweight alternatives by injecting trainable adapter layers into frozen pretrained models, enabling task-specific fine-tuning without updating the entire model parameters [74,75]. Building on these advances, the development of lightweight LLMs optimized for mobile devices presents a promising direction for extending AI-based chronic disease management to resource-constrained settings [76].

Usability and accessibility concerns accounted for (9/29, 31%), including issues with text-only interfaces for some LLMs, cultural misalignments, and outputs ill-suited for pediatric or elderly populations [18,28,40]. While studies highlighted text-only interfaces as a critical limitation of LLMs in health care [28,40], recent advances in multimodal architectures have addressed this gap. These models now integrate and interpret diverse data modalities, including medical images, audio, and structured documents, while generating composite textual and visual outputs [31,77-79]. In addition, having dynamic and simplified interfaces to accommodate low-digital-literacy users, pediatric users, and for different cultural and language settings could further improve LLM usability and adaptation.

Legal, ethical, and regulatory concerns (10/29, 35%) remain a key issue in LLM adoption in health care. Studies identified data privacy, biases, misinformation, responsibility, and accountability for LLM-generated content as key concerns. Although several AI frameworks have been proposed, the absence of standardized guidelines and regulatory approvals creates significant gaps [27,80-82]. This regulatory vacuum risks inconsistent model development and validation practices, unaddressed ethical dilemmas regarding accountability for AI-generated recommendations, and potential mismatches between rapidly evolving LLM capabilities and static health care regulations [27,83]. Addressing these ethical concerns requires standard guidelines and rules to ensure responsible use in health care settings[84]. Therefore, future efforts must prioritize the development of a comprehensive regulatory framework.

Evaluation gaps accounted for (5/29, 17%), reflecting critical shortcomings in current methodologies for assessing LLM performance in clinical contexts. The existing automated evaluation metrics mainly focus on predictive performance using metrics such as accuracy and F₁-scores that lack medical and treatment knowledge [22,85]. These metrics may produce misleading conclusions if not appropriately selected [85]. Furthermore, human evaluation, although valuable, introduces subjectivity and interrater variability, yet the absence of a standardized LLM evaluation framework makes attaining consensus among human raters challenging [30,85]. A promising direction involves a hybrid evaluation approach that integrates human expert reviews with automated metrics. Future efforts should prioritize the development of standardized LLM evaluation frameworks tailored to health care settings. Table 3 summarizes the challenges associated with applying LLMs in chronic disease management with proposed recommendations.

Although quality assessment was conducted using MMAT, studies were not excluded based on their methodological rigor. As a result, including moderate and low-quality studies may have influenced the reliability and consistency of the reviewer’s findings. The varied methodological designs across studies may have affected the interpretation of results and conclusions drawn. Furthermore, the review was limited to only English-language publications, which may have introduced language bias and limited the generalizability of our findings, particularly in contexts where LLMs are being adapted to local languages or integrated into culturally specific health care practices. The exclusion of databases such as Embase and Web of Science may have limited the comprehensiveness of the search. Future research can consider broader database coverage and non-English sources to enhance diversity and scope.

This scoping review reveals several critical implications for the integration of LLMs in chronic disease management. First, it is essential to address the accuracy issues identified in 18/29, 62% of studies. This calls for both technical solutions (domain-specific fine-tuning, reinforcement learning (RL), and retrieval-augmented generation (RAG) frameworks [61,64]) and nontechnical solutions (expert validation and collaborative partnerships with health care institutions to access and use real-world clinical data). Second, enhancing accessibility across diverse patient populations requires developing culturally adapted LLM interfaces, implementing age-appropriate interaction modes [40], and integrating with low-resource platforms such as SMS-based systems and lightweight mobile apps for various populations [76]. Third, robust governance frameworks must be established to address the ethical, legal, and privacy concerns noted in 10/29, 35% of studies to ensure regulatory compliance.

Finally, future research should focus on multimodal LLMs that can synthesize diverse data inputs from wearable devices and mobile health applications for holistic patient monitoring [19,31,35] and develop resource-efficient deployment techniques. The predominance of diabetes-focused studies (14/29, 48%) highlights a potential research gap in addressing other prevalent chronic conditions. Similarly, there is a research gap in age-inclusive LLM design, given the overwhelming focus on adult populations (28/29, 96%) and the lack of pediatric studies. Addressing these gaps would enhance the clinical relevance and equitable application of LLMs across the full spectrum of chronic disease management.

This scoping review highlights the growing potential of LLMs in supporting chronic disease management through patient education, diagnosis and treatment, emotional support, self-management support, decision support, and prediction tasks. While LLMs offer promising capabilities, their effective integration into health care still requires addressing key challenges related to accuracy, accessibility, usability, and ethical and privacy concerns. Future research should focus on integrating LLMs with mobile and wearable technologies, creating culturally and age-appropriate interfaces, and exploring integration with low-resource platforms. Addressing these research gaps will ensure equitable and safe use of LLMs across diverse health care settings.