The Applications of Large Language Models in Mental Health: Scoping Review

Introduction

Background

Mental health is a growing global public issue, impacting nearly a billion people worldwide, with an estimated 1 in 7 adolescents affected [1,2]. Nevertheless, >70% of individuals with mental health disorders lack access to essential support and services [3]. Furthermore, >720,000 people commit suicide annually, with nearly three-quarters of these suicides occurring in low- and middle-income countries [4]. Consequently, there is an urgent need in mental health to facilitate efficient detection from large-scale data; deliver effective treatments and interventions to large populations; and ensure private, affordable health care and increased access to specialized psychiatrists. To address inadequate access to effective and equitable mental health care, large-scale, innovative solutions are imperative.

Large language models (LLMs), emerging in 2019, are advanced natural language processing (NLP) models capable of analyzing vast textual data and generating humanlike language [5]. Notable LLMs such as GPT-3/4 [6], Pathways Language Model (PaLM) [7], and LLaMA [8], constitute a category of foundational models, each with billions of parameters, trained on extensive textual data [9]. Using the transformer architecture and self-supervised pretraining, LLMs are adept at tackling a variety of NLP tasks, including information extraction, interaction, content generation, and logical reasoning [10]. In comparison to prior NLP models [11,12], LLMs exhibit superior performance in computational efficiency, large-scale data analyses, interaction, and external validity and applicability [9]. Furthermore, LLMs can be fine-tuned to cater to specific domains, including mental health, thereby empowering them to engage in natural language interactions and accomplish mental health–related tasks. LLMs would help address insufficient mental health care system capacity and provide efficient or personalized treatments. Therefore, the application of LLMs in mental health is expanding across diverse domains [13-16].

Objective

Researchers have explored the applications of LLMs in mental health in various areas, encompassing screening or detecting mental disorders [17-19], supporting clinical treatments and interventions [20-22], and assisting in mental health counseling and education [17,20,23,24]. Nonetheless, few comprehensive reviews have yet synthesized these applications, assessed the performance of LLMs, or elucidated their advantages within the mental health domain [25,26]. Therefore, we conducted this scoping review to address 4 questions. First, we identified the challenges in mental health and compared the processes adopted by humans, nontransformer models, and LLMs. Second, we summarized the main areas of LLMs’ applications and presented specific processes of these applications. Third, we examined comparative performance studies between LLMs and humans, as well as among different LLMs. Finally, we presented the fine-tuning LLMs for mental health and compared their advantages and disadvantages, which could be directly used by researchers and psychiatrists. This review aimed to provide a foundational understanding of LLM applications in mental health by examining current trends, comparing performance, identifying challenges, and outlining a road map for future research and clinical practice.

Methods

Protocol Registration

We drafted the study protocol based on the relevant items from the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews; Multimedia Appendix 1). Compared with the original protocol, the previous query strings were capable of searching for articles related to the applications of LLMs in mental health. However, to enhance the reproducibility of the search process, we have added more detailed query strings. As for databases used, we excluded the Google Scholar database from our search strategy for two reasons: (1) the other 4 English language databases we selected already included the relevant articles and (2) it is not feasible to accurately count the specific number of articles retrieved from the Google Scholar database. Instead, we used official academic databases for the article search. These databases have strict selection criteria and typically include only peer-reviewed, high-quality publications from reputable publishers. In addition, we have updated the literature search period, revising the range from January 1, 2017, to August 31, 2024. This change is important for the following reasons. First, at the beginning of the study, we searched for articles published between January 1, 2017, and August 31, 2024. However, we found that all articles meeting the inclusion criteria were published after 2019. Second, previous studies [9,25] suggested that the term “LLM” was first proposed and widely used starting in October 2019. Therefore, we have revised our inclusion criteria to include only studies published from January 1, 2019, to August 31, 2024. The final protocol was registered prospectively in the Open Science Framework [27].

Search Strategy and Selection Criteria

A scoping review is a preliminary systematic review that aims to map the existing evidence on a specific topic or field of research [28]. It provides a broad overview of the literature by identifying the nature and extent of existing research, including types of studies, variables, and gaps in the evidence base [29]. This approach is particularly useful when the body of evidence is large or diverse, or when there is a need to understand the scope of a research area before conducting a more focused systematic review.

This scoping review followed the five-stage framework: (1) identifying the research question; (2) identifying relevant studies; (3) study selection; (4) charting the data; and (5) collating, summarizing, and reporting the results. The search terms for mental health include the following: “psychiatr*,” “mental health,” “depress*,” “anxiety,” “posttraumatic stress disorder,” “PTSD,” “bipolar disorder,” “schizophrenia,” “obsessive-compulsive disorder,” “personality disorder,” “insomnia,” and “suicid*.” The keywords and search terms for LLMs in mental health include the following: “large language model,” “OpenAI language model,” “generative AI,” “generative artificial intelligence,” “BERT,” and “GPT.”

According to previous studies [9,25], the term “LLM” is used to distinguish language models based on their parameter scale (eg, containing tens or hundreds of billions of parameters). The term “LLM” has been proposed and widely used since October 2019. In addition, we planned to conduct this scoping review on August 31, 2024. Thus, we searched 4 English language databases (Web of Science, PubMed, IEEE Xplore, and Cochrane Library) and 3 Chinese language databases (Weipu, CNKI, and Wanfang) for peer-reviewed articles published between January 1, 2019, and August 31, 2024. We only included papers in Chinese published in high-quality journals. To find other possibly relevant studies and reports that were missed by the automated searches, the reference lists of the included articles and reports were examined.

The inclusion and exclusion criteria for studies are shown in Textbox 1.

The first round of screening was based on titles, keywords, and abstracts (4392/4859, 90.39%). These articles were divided into 6 subsets (each subset includes 732 papers). A total of 6 researchers (JL, PL, BW, YY, HZ, and CN) independently reviewed each part of the papers, including the titles, keywords, and abstracts. During the initial review, each article was categorized into one of the following groups: (1) fully met the inclusion criteria, (2) did not focus on mental health or mental disorders, (3) did not use LLMs, and (4) had unclear eligibility for inclusion. In the second round of screening, we conducted a cross-check process to ensure the classification of these articles (eg, CN reviewed the paper part handled by HZ, and HZ reviewed the paper part handled by JL). Articles from the fourth group were discussed by YJ and YW one by one for their eligibility. Several questions were discussed:

• NLP use uncertainty—some articles used NLP, but it was unclear whether LLMs were used. A full-text review was necessary to make a definitive determination. Thus, we reserved these articles for further review.
• Exclusion of cognitive disorders—it remains uncertain whether aphasia should be classified as part of the mental health domain. After the discussion, we excluded the articles about aphasia as they were not considered within the scope of mental health.
• Electronic medical records (EMRs) in the context—some articles mentioned EMRs, which could potentially be relevant to the mental health domain. However, these articles were not specifically focused on mental health. After the discussion, we decided to reserve these articles for further review.

The third round of review was based on the full texts to assess their eligibility (308/4392, 7.01%). These potentially eligible articles were divided into 6 parts (each part included 51 or 52 papers). A total of 6 researchers (JL, PL, BW, YY, HZ, and CN) independently reviewed each part of the papers, including the titles, abstracts, and full texts of each paper. To ensure the accuracy of paper screening, we conducted a double-check process. The researchers cross-checked each other’s work, reviewing the paper selection process (eg, JL reviewed the paper part handled by PL, and PL reviewed the paper part handled by BW). Disagreements were discussed with third reviewers (YJ and YW) until a consensus was reached. We excluded preprints, reviews, books, studies that did not use LLMs, and those not published in journals or conferences. Ultimately, 95 articles were retained for the final analysis. The search terms for English language databases and Chinese language databases are shown in Multimedia Appendices 2 and 3.

Data Extraction, Categorization, and Labeling

The final data collection form used for peer-reviewed articles is shown in Table 1 (N=95). The information of each study included categories, regions, application tasks, mental conditions, data sources, sample information, and applied models. These articles were divided into 6 parts (each part included 15 or 16 papers). A total of 6 researchers (JL, PL, BW, YY, HZ, and CN) independently extracted data from each part of the papers. To ensure the accuracy of data extraction, we conducted a double-check process. The researchers cross-checked each other’s work, reviewing the data extraction process. The excluded studies, along with the reasons for their exclusion (eg, “It’s a preprint”), are listed in Multimedia Appendix 4.

For this scoping review, we developed a categorization framework based on the applications of LLMs in mental health: (1) screening or detection of mental disorders, (2) supporting clinical treatments and interventions, and (3) assisting in mental health counseling and education. At least 1 reviewer categorized each study manually by examining the title and abstract to assign categories. When the study categories could not be clearly determined, the methods and results sections were reviewed to assist with classification.

Statistical Analysis

Descriptive statistics were used to summarize the distribution of studies across different application areas. For each application task, we calculated frequencies and percentages. For the performance comparisons between humans and LLMs, between LLMs and nontransformer models, also between various LLMs, we collected the metrics results and plotted the results. Calculations and data charting were performed using R software (version 4.4.2) developed at Bell Laboratories by John Chambers and colleagues, and PyCharm software developed by JetBrains.

Results

Overview

The initial search yielded 4859 records, of which 467 duplicates were removed. Of the remaining 4392 records, 4084 records were removed due to the contents irrelevant to mental health or not using LLMs. After the full-text screening, 95 articles fulfilled the inclusion criteria (Figure 1). Table 1 demonstrates the basic information of each study, including categories, regions, application tasks, mental conditions, data sources, sample information, and applied models.

‎

Comparisons Between LLMs, Nontransformer Models, and Humans

The current challenges in the mental health field include difficulty in efficient mental disorders detection from large-scale data, effective treatment or intervention for large populations, private and low-cost health care, demand for professional psychiatrists, and so on. Compared with nontransformer models and humans, LLMs present higher capabilities in efficient parallel computing of massive data, textual generation, information capture, strong generalization ability, and fine-tuned mental health tasks. The process comparisons between LLMs, nontransformer models, and humans were presented with the framework (Figure 2, a higher resolution version of figure is also available in Multimedia Appendix 5). Therefore, LLMs have been applied to tasks such as detecting or predicting mental disorders, supporting clinical treatment, providing preliminary medical care, and generating educational materials—using datasets from sources such as social media platforms, EMRs, and counseling transcripts.

‎

Categorization of Studies Based on Mental Health Applications

We categorized studies in terms of the applications of LLMs in the mental health field. These applications were divided into 3 categories: screening or detection of mental disorders (67/95, 71%), supporting the clinical treatments and intervention (31/95, 33%), and assisting in mental health counseling and education (11/95, 12%; Multimedia Appendix 6). Each study was categorized with ≥1 applications; thus, the percentages sum to >100%. Most studies applied LLMs for depression detection and classification (33/95, 34.7%), supporting clinical treatments and interventions (14/95, 15%), and suicide risk prediction (12/95, 13%; Multimedia Appendix 6). These studies used data from social media platforms such as Reddit, Twitter, Facebook, and Weibo, or clinical datasets (Distress Analysis Interview Corpus-Wizard of Oz and Extended Distress Analysis Interview Corpus), as well as semistructured interviews from hospitals. When evaluating the performance of these LLMs, most studies measured the performance of LLMs with various metrics, such as F₁-score (54/95, 57%), precision (34/95, 36%), accuracy (45/95, 47%), and recall (32/95, 34%).

The number of studies mapped by country is presented in Figure 3A (a higher resolution version of figure is also available in Multimedia Appendix 5). The United States is the country that has explored the applications of LLMs in the field of mental health the most, followed by China and Canada. The number of included studies increased year on year, and this trend is shown in Figure 3B. We can find that since 2019, an increasing number of researchers have been exploring the applications of LLMs in the field of mental health. As for the application areas, the first area mainly focused on screening or detection of mental disorders, including depression detection, suicide risk prediction, sentiment analysis, and other mental disorders. The second area focused on supporting clinical treatments and interventions, including supporting clinical treatments, developing chatbots, and augmenting clinical data. The third area focused on assisting mental health counseling and education, including counseling assistance and mental health resource supplement (Figure 3C). These studies applied basic LLMs or fine-tuned LLMs (eg, MentaLLaMA [30], PsychBERT [31], and RoBERTa [68]) to detect or predict depression [30,32-41,49,61,69]; suicide risk [23,24,62]; and other mental disorders, such as anxiety [37,70], obsessive-compulsive disorder [71], and posttraumatic stress disorder [15]. The detailed process for these applications of LLMs is presented in Figure 3C.

In the second application area, most studies explored the capability of LLMs in supporting clinical treatments and interventions [17,21,23,62,63,86], developing chatbots, and augmenting unbalanced clinical data [95,104-106]. These studies applied LLMs to provide treatment advice, assist diagnostic services, and assess prognosis through a question-answering approach. The performance of LLMs was evaluated by professional clinicians and compared with the related performance of humans. A total of 3 studies also applied LLMs to elicit emotion [20-22]. Furthermore, to address the imbalance of clinical data and enhance diagnosis and treatment, LLMs could augment clinical data and targeted dialogues in safe and stable ways.

Moreover, LLMs have been applied for counseling assistance [14,104-106] and educational resource supplements [19,111]. These results showed that the introduction of any interaction (video or chatbot) improved intent to practice and overall experience compared to the baseline [107]. Furthermore, LLMs showed the potential to generate educational materials for training.

‎

Performance Comparisons Between LLMs and Humans and Between Various LLMs

Several studies compared the performance between humans and LLMs. The metrics’ results of performance were displayed in Multimedia Appendix 7. Figure 4A (a higher resolution version of figure is also available in Multimedia Appendix 5) presents the results of 3 studies on obsessive-compulsive disorder identification [71], chatbots efficiency [22], and treatment support [21]. According to these results, most LLMs showed efficient and promising performance for mental health tasks. Several LLMs, such as GPT-4, Claude, and Bard, aligned closely with mental health professionals’ perspectives.

Figures 4B and 4C show the comparisons of model performance between different LLMs in mental disorder diagnosis [72], clinical treatment assistance [86], and depression detection [32,40,42-46]. These results found that the latest LLMs (eg, ChatGPT) perform better than traditional and previous models. The complete results are presented in Multimedia Appendix 8.

‎

Existing Fine-Tuned LLMs for Mental Health

Table 2 provides the fine-tuned LLMs for mental health, including availability, base models, the number of parameters, training strategy, and published year. These fine-tuned LLMs could be applied specifically to mental health tasks.

Common Advantages and Disadvantages

Figure 5 summarizes the common advantages and disadvantages of LLMs in mental health. These LLMs could be divided into the BERT series, GPT series, LLaMA series, and others. For example, the shared strengths of BERT-based models make them well-suited for fine-tuning on specific mental health issues. However, the BERT series models require large computational resources. The ChatGPT series models can conduct multiround dialogue, even based on small-sample learning. Nevertheless, the accuracy of ChatGPT models should be improved. Furthermore, GPT-4 could receive multimodal data and show more powerful performance in comprehension and generation. As for the LLaMA series models, they are open source for the public and beneficial for interactive applications, although their complex task performance is inferior to large-scale proprietary models.

‎

Discussion

Principal Findings

This scoping review explored the applications of LLMs in the mental health field and summarized trends, application areas, performance comparisons, challenges, and prospective future directions. The applications of LLMs were categorized into 3 key areas: screening or detection of mental disorders, supporting clinical treatments and interventions, and assisting in mental health counseling and education. Most studies used LLMs for depression detection and classification (33/95, 35%), clinical treatment support and intervention (14/95, 15%), and suicide risk prediction (12/95, 13%). Compared with nontransformer models and humans, LLMs demonstrate higher capabilities in information acquisition and analysis and efficiently generating natural language responses. Furthermore, we summarized the fine-tuning LLMs for mental health and compared their advantages and disadvantages, offering insights for future researchers and psychiatrists.

Advantages of LLMs’ Applications in Mental Health

Compared to nontransformer models and humans, LLMs demonstrate higher capabilities in information acquisition, analysis, and generating professional responses. These enhanced capabilities posit LLMs as potential tools for the detection and prediction of mental disorders through the analysis of extensive datasets, including social media content [30,32,33], EMRs [21,86], and counseling notes [87,88]. Their applications in treatment and intervention are noteworthy. LLMs could assimilate patient clinical records, summarize treatment sessions, and support diagnoses for mental disorders. This potential streamlines the workflow for patients and mental health care systems efficiently. On the basis of the information interaction and generation ability, LLMs can be instrumental in the development of chatbots designed for initial medical consultation and emotional support. Such applications help to offer discreet and affordable health care solutions to individuals who, due to stigma or financial constraints, are reluctant to seek assistance for mental health issues.

In this scoping review, most studies applied LLMs in detecting depression and suicidal risk [34-41,49,61,69]. These studies have demonstrated the potential of LLMs such as MentaLLaMA [30], PsychBERT [31], RoBERTa [68], and GPT-4 [89], in detecting and identifying mental disorders from social media platforms and clinical datasets. These models have been trained to recognize depression and recommend evidence-based treatments, with some, such as GPT-4 and Claude, closely aligning with the perspectives of mental health professionals [20]. Furthermore, the integration of linguistic features and the appending of mental disorders’ background information have been shown to enhance classification performance and calibration of these LLMs [17,23,62,63]. LLMs perform comparably to experienced clinicians in identifying the risk of suicide ideation, with the addition of suicide attempt history enhancing sensitivity. GPT-4 has demonstrated superior performance over GPT-3.5 in recognizing SI, despite a tendency to underestimate resilience [24]. Other studies showed that the use of LLMs produced effective strategies for predicting suicidal risk with sparsely labeled data [23,62]. This efficiency in labeling and analysis of large datasets is a better advancement over previous methods, which were often hindered by the requirement of manual annotation and consequently limited by small sample sizes or finite datasets.

Due to their efficient ability to acquire information and generate humanlike language, LLMs show potential in preliminary care [14], providing treatment guidelines, augmenting unbalanced clinical datasets [95,104-106], and assisting in training professionals. LLMs also have shown potential to assist with cognitive behavioral therapy tasks, such as reframing unhelpful thoughts and identifying cognitive biases [89]. Several studies also investigated the potential of LLMs in addressing health care regional disparities and enhancing diagnostic and therapeutic services, although these LLMs have only shown initial results so far [20-22]. Furthermore, in the diagnosis of posttraumatic stress disorder [15], LLM-augmented datasets have shown improved performance over original datasets, with both zero-shot and few-shot approaches outperforming the original dataset, highlighting the effectiveness of LLMs in enhancing diagnostic accuracy with minimal additional data. According to the results of performance comparisons, LLMs present similar performance as professionals and may even surpass physicians [22]. These results demonstrate the potential of LLMs in clinical assistance and support. However, when evaluating the quality of LLM-generated responses to queries and educational health materials, several studies have indicated that, despite GPT-4’s strong face validity, it remains insufficiently reliable for direct-to-consumer use [111]. Therefore, the outputs by LLMs for educational health require cautious human editing and oversight. These findings underscore the growing importance of LLMs in mental health research and practice, offering new avenues for early detection, risk assessment, and intervention strategies.

Prompt Engineering Techniques of LLMs

Prompt engineering techniques are becoming increasingly essential across various applications and mental health conditions [109]. By optimizing prompt strategies in a targeted manner, models can more effectively perform tasks such as the preliminary detection of mental disorders, intervention recommendations, and emotional assessments, all while adhering to professional and ethical standards. Previous studies have reported that the use of diverse prompts when applying LLMs can lead to substantial variations in the stability and accuracy of the models’ outputs [112,113]. Notably, without modifying the underlying architecture of the LLMs, adjusting the input prompts alone can significantly influence the quality and relevance of the outputs. This underscores the critical role of prompt design in optimizing model performance. Thus, it is vital to standardize prompts and share them openly in academic and clinical contexts to enhance the robustness and reproducibility of research findings. In mental health dialogue scenarios, prompts must balance professional standards with ethical considerations to prevent the generation of misleading or inappropriate content [114]. Clinicians should be aware that poorly designed prompts may pose potential risks, such as inaccurate treatment suggestions or unsuitable discussion topics. The ongoing refinement and standardization of prompt engineering not only enhance the performance and explainability of LLMs but also enable health care professionals and researchers to provide more efficient and safer preliminary support and services to patients [115].

Challenges of LLMs’ Applications in Mental Health

While LLMs have demonstrated high performance on certain benchmarks, their practical applications in clinical settings are still limited. Recent studies have shown that even advanced LLMs struggle with complex clinical tasks, such as interpreting medical records and providing accurate treatment recommendations [30,32,33]. The efficiency gains from LLMs in clinical settings must be balanced against their limitations and potential risks [108]. As these models continue to evolve, their role in mental health support is likely to expand, with a focus on enhancing data privacy, biases, and ethical considerations in clinical implementation. LLMs are trained on vast amounts of data, which may include sensitive personal and health information. Ensuring compliance with privacy regulations such as the general data protection regulation is essential to protect patient confidentiality [116,117]. The inadvertent inclusion of personally identifiable information in pretraining datasets can compromise patient privacy, and LLMs can make privacy-invading inferences from seemingly innocuous data. Implementing measures such as data anonymization and safe data storage procedures is crucial to address these issues. Biases in LLMs are another critical issue that must be considered. These models would perpetuate and magnify biases in their training data, leading to differential treatment and outcomes, particularly in populations considered vulnerable [118]. For instance, biases in gender, race, and socioeconomic status can result in inaccurate or misleading information, which may exacerbate existing health disparities [119]. To mitigate these risks, it is important to develop techniques for identifying, alleviating, and preventing biases in LLMs. Ethical concerns are paramount when using LLMs in mental health. The potential for LLMs to generate false or misleading information is a particular concern in health care. Ensuring that LLMs are rigorously tested and monitored is crucial to maintaining reliability in patient care. In addition, transparency is vital for health care professionals to trust the results of LLMs. These LLMs should provide clear, understandable insights for operation with sufficient explanation.

Prospective Future Directions of LLMs in Mental Health

In the future, the application trend of LLMs in mental health is expected to continue to rise, and the aspects of their applications will be broader. Initially, most studies based on the LLMs with textual data and multimodal data, such as pictures, videos, and sounds, could be integrated with multimodal LLMs. Various data types might further improve the performance of mental disorders’ detection and identification. Several studies have explored multimodal LLMs in mental health research [72,108]. Moreover, existing studies mainly focus on depression and suicidality; more mental disorders should be investigated with LLMs, especially for rare mental disorders, such as borderline personality disorder and bipolar disorder. The applications of LLMs in treatments, interventions, and preliminary care would be beneficial for these patients. Furthermore, although several studies have developed chatbots for early detection or intervention in mental disorders, further research is needed to enhance the accuracy and robustness of LLMs. In addition, it is important to provide open-resource and fine-tuned LLMs for mental health, especially for low- and middle-income countries. Although LLMs show great performance in various applications in mental health, several areas of LLMs should be improved. First, there is a need for more high-quality, diverse, and representative datasets to train LLMs for mental health applications, ensuring that the models can understand and respond to a wide range of mental health–related queries and scenarios. Second, while LLMs can generate coherent and contextually appropriate responses, they still lag behind human performance in terms of empathy and emotional intelligence, which are crucial in mental health support. Third, LLMs need to improve their ability to reason and understand the context and nuances of mental health dialogues, which often require a deep understanding of human emotions and psychological states.

Finally, it is important to establish standards for privacy, safety, and ethical considerations when LLMs process sensitive personal and health information [120,121]. These standards are essential to mitigate the potential risks and ensure the responsible use of LLMs in health care settings. On the one hand, the model training process for LLMs inevitably involves the collection of stigmatizing or biased data [118], which can generate hallucinatory content that may mislead or harm patients. For instance, biased data can result in inaccurate or inappropriate medical advice, which could have serious consequences for individuals seeking health information [105]. Several studies underscore the risk of sensitive data exposure and emphasize the prevention of harmful content generation [71,73]. On the other hand, the convenience and low cost of LLMs may lead teenagers to become overly dependent on them for mental support. This excessive reliance could result in several negative outcomes. It may lead to addiction to the internet-based world, negatively impacting their daily lives and social interactions. Moreover, it could delay the optimal time for teenagers to seek professional help, potentially exacerbating their mental health issues. While LLMs can provide initial support and guidance, they cannot replace the nuanced understanding and empathy that human professionals can offer. Furthermore, LLMs cannot address crises effectively [122]. Although they can identify extreme emotions during conversations, they typically only suggest seeking professional assistance without providing direct and effective measures. LLMs lack the clinical judgment required to handle emergencies, which means they cannot offer the immediate support that may be necessary in critical situations. Therefore, in the practical application of LLMs, professional intervention is essential [123]. Experts should develop a dedicated system of ethical standards to guide the use of LLMs in health care. This system should include regular supervision and evaluations to ensure that LLMs are used responsibly and ethically. LLMs should be used as supplementary tools rather than complete replacements for human roles. They can provide initial support and guidance, but ultimately, the responsibility for clinical judgment and patient care should remain with trained health care professionals. Future advancements depend on collaborative efforts to refine technology, develop standardized evaluations, and ensure ethical applications, aiming to realize LLMs’ full potential in supporting mental health care.

Comparison With Previous Studies

Several reviews have explored the applications of LLMs in mental health care from various perspectives. A viewpoint article and a preprint summarized the opportunities and risks associated with using LLMs in mental health [115,124]. The viewpoint article focuses on application scenarios, including education, assessment, and intervention, while the preprint clarifies the potential opportunities and risks of LLMs from an ecological conceptualization perspective, encompassing the care seeker, caregiver, institution, and society [124]. The analysis methods and application categories in these reviews differ from those in this scoping review. Another preprint scoping review also identified diverse applications of LLMs in mental health care, including preprint articles [25]. On the basis of 34 articles screened from 313 articles published between October 1, 2019, and December 2, 2023, the researchers categorized the application areas into 3 domains: the deployment of conversational agents, resource enrichment, and detailed diagnosis classification. However, this scoping review delved deeper into LLMs and training techniques, dataset characteristics, validation measures and metrics, and challenges. In contrast, our study offers a more targeted and comprehensive reference guide for LLM applications in mental health. Specifically, we compare the application effects of LLMs with human evaluations and among different models, aiming to provide a detailed and structured overview of LLMs’ performance and potential in mental health scenarios. Another systematic review with 40 articles assessed the clinical applications of LLMs in mental health [26], focusing on their potential and inherent risks. Although this review summarized the results of each article in tabular form, it did not extract key information. Our scoping review searched 4 English language databases and 3 Chinese language databases with more specific search terms and a wider time range. We categorized these applications into 3 key areas: screening or detection of mental disorders, supporting clinical treatments and interventions, and assisting in mental health counseling and education based on 95 articles screened from 4859 articles. We also extracted key information about LLMs’ applications and developed frameworks for each application area. These frameworks are designed to help researchers and clinicians understand the use of LLMs in mental health, even without a background in artificial intelligence. Furthermore, we provided the fine-tuned LLMs for mental health and compared the advantages and disadvantages of LLMs. This comparative approach offers a more nuanced understanding of how different models and human expertise can be leveraged in mental health applications, aiding researchers and clinicians in selecting suitable LLMs for their specific mental health tasks.

Although this scoping review explored the applications of LLMs in mental health, several limitations should be considered. First, there is an absence of assessment of the risk of bias due to the unique nature of these included studies. Moreover, we cannot perform a meta-analysis due to the diversity of methods and tasks in the included studies. Second, with the rapid development of LLMs, the results of comparative studies would be cautious. The performance of these LLMs may have significantly improved. Third, the preprint studies (eg, from arXiv and medRxiv platforms) were not included in this review due to the lack of peer review, though they were published recently. Finally, this review is limited to studies published in English and Chinese, which may compromise the comprehensiveness and representativeness of the findings. This approach may overlook the applications and contributions of LLMs in other regions that speak different languages, thereby failing to provide a comprehensive global overview of LLM applications in mental health. In addition, different languages present unique challenges and opportunities for LLM applications in mental health. Although this scoping review identified a limited number of studies applying LLMs for depression detection in Malay dialects [41] and developing therapeutic dialogue agents in the Polish language [106], the performance of LLMs across various languages can vary based on linguistic characteristics, data availability, and cultural contexts. Limiting the review to English and Chinese studies may underestimate these differences. Moreover, low- and middle-income regions, which often rely on technologies such as LLMs to enhance mental health service accessibility and quality, may use languages other than English or Chinese. Limiting the review to English and Chinese studies could neglect the actual needs and potential contributions of these regions, thereby affecting the development and application of LLMs in those areas. To provide a more holistic view of LLM applications in mental health, future scoping reviews should consider a broader range of languages to provide a more comprehensive understanding.

Conclusions

This scoping review summarized the applications of LLMs in the mental health field, including trends, application areas, performance comparisons, challenges, and future directions. The applications mainly focused on 3 areas: screening or detection of mental disorders, supporting clinical treatments and interventions, and assisting in mental health counseling and education. Compared with nontransformer models and human experts, LLMs demonstrate higher capabilities in information acquisition and analysis, generating natural language responses, and addressing complex reasoning problems. As these models continue to evolve, their role in mental health support is likely to expand, with a focus on enhancing accuracy, sensitivity, and ethical considerations in clinical implementation.