Large Language Models in Mental Health Care: a Scoping Review

1 Introduction

• Mental illness affected 22.8% of U.S. adults in 2021, while mental health disorders account for 30% of the global non-fatal disease burden and depression and anxiety cost $1 trillion yearly in lost productivity.The introduction presents these figures as evidence of individual, societal, and economic burdens.Source: §1 · p. 2
• Much mental health assessment and management occurs through language, including symptom evaluation and talk therapies, making recorded text and interactions a resource for research and practice.NLP has been used to identify conditions, support interventions, and build emotion-support chatbots using clinical and social-media data.Source: §1 · pp. 2–3
• LLMs can retrieve and summarize unstructured text and converse with users, creating potential for clinician insights, early intervention, tailored treatment, and accessible self-expression.Clinical therapeutic use is contingent on validation and integration into clinical workflows.Source: §1 · p. 3
• The review addresses a stated gap in reviews of mental-health LLMs, focusing on architectures developed since T5 and examining datasets, applications, challenges, and clinical practicality.The authors describe it as the first comprehensive review of LLM applications in mental health care.Source: §1 · p. 3

1.1 Background on Natural Language Processing and Large Language Models

• NLP progressed from rule-based and statistical approaches through neural and pretrained language models to the LLM era, with models described as larger and more versatile.The background names ELMo, BERT, T5, GPT, and Transformer architectures among these developments.Source: §1.1 · pp. 3–4
• Prompt-tuning instructs LLMs through natural-language prompts while keeping model weights fixed, unlike fine-tuning, which adjusts weights using additional data.The background also distinguishes open-source models that can be locally customized from hosted non-open-source models with different privacy and control considerations.Source: §1.1 · p. 4

1.2 Potential Applications in Mental Health Care

• Potential mental-health uses include predicting behavior, identifying stressors, emotional support, preliminary diagnostic assessment, disorder management, and therapeutic support such as treatment adherence.The passage conditions supplementary clinical roles on adequate regulatory, ethical, and privacy safeguards.Source: §1.2 · p. 4

2.1 Search Strategy and Selection Criteria

• Studies had to use at least one LLM published after T5 and directly address research questions about these models in mental-health-care settings.This scope deliberately excluded earlier pretrained models such as BERT and GPT-2.Source: §2.1 · pp. 4–5
• Figure 1 is identified as the screening prompt for ChatGPT-4.The review states that PRISMA guided its process and that its PRISMA diagram is shown in Figure 1.Source: §2.1 · p. 4, Figure 1 · p. 5
• The search covered original research published or disseminated from October 1, 2019, through December 2, 2023, without language restrictions and included peer-reviewed studies and preprints.Databases included arXiv, medRxiv, ACM Digital Library, PubMed, Web of Science, and Google Scholar.Source: §2.1 · p. 5

2.2 Screening with GPT-4

• After duplicate articles and records without abstracts were removed, 281 papers remained for initial screening, and GPT-4 was used as a secondary reviewer.The authors tested prompts before using GPT-4 for screening.Source: §2.2 · p. 6
• Human and GPT-4 title-and-abstract screening produced a Cohen’s Kappa of approximately 0.9024, with disagreements resolved through review-team discussion.GPT-4 tended to include more articles, while an uncertainty option supported comprehensive inclusion alongside precision.Source: §2.2 · p. 6

3 Findings

• 34 articles met the review criteria; research emerged in September 2022, with prompt-tuning and application studies predominating and only two dataset-and-benchmark papers identified.The review also found one study addressing ethics, privacy, and other challenges.Source: §3 · pp. 6–7

3.1.1 Application Areas

• LLM applications span conversational agents, resource enrichment, and classification, including support for individuals and recommendations or analyses for care providers.Sixteen studies aimed to improve conversational agents’ empathetic, context-aware responses; resource enrichment included educational content and synthetic data, while classification covered binary and multi-class tasks.Source: §3.1.1 · pp. 7–8

3.1.2 Mental Health Problems Involved

• 23 of the 34 reviewed articles addressed specific mental health problems; stress, suicide, and depression were the most frequently studied, with coverage limited by open-source dataset availability.The review also covered other conditions and circumstances, including anxiety, bipolar disorder, relationship problems, sleep problems, and bullying.Source: §3.1.2 · pp. 8–9

3.2.1 Distribution of Model Architectures and Prompting Algorithms

• Studies either prompted GPT-3.5 or GPT-4 without training or further-trained general LLMs; prompting variants included few-shot, chain-of-thought, chain-of-empathy, and diagnosis-of-thought approaches.The review notes that complete details of base-model training data remain undisclosed, including for models with some publicly available training-data information.Source: §3.2.1 · pp. 9–10

3.2.2 Fine-Tuning Techniques

• All reviewed studies involving fine-tuning used instruction fine-tuning, and studies fine-tuned existing models rather than training LLMs from scratch.Instruction fine-tuning uses instructions paired with tasks and can inject domain knowledge while improving instruction following.Source: §3.2.2 · p. 10

3.3 Dataset Characteristics

• The review identified 36 datasets across 34 studies, with 20 datasets (55.6%) collected from social media and substantial variation in dataset size, source, annotation, and licensing.For example, SMHD contains 4,399K Reddit posts, while D4 contains 1,339 expert-reviewed dialogues; 14 datasets did not specify annotator experience, and most datasets were publicly available.Source: §3.3 · p. 12

3.4 Validation measures and metrics

• Evaluation combined human and automated methods in 19 articles, used human assessment alone in 5, and used automated methods alone in 10, with no unified framework widely adopted.F1 variants were predominant for mental-health applicability; language-generation studies used metrics such as BLEU and ROUGE, while human evaluations were frequent because available automated metrics do not fully capture context and nuance.Source: §3.4 · pp. 12–13

3.5 Benchmark studies

• Two benchmark studies were identified: Jin et al. evaluated models across six tasks, while Qi et al. focused on cognitive-distortion classification and suicidal-risk prediction using Chinese social-media data.The first compared GPT-4, GPT-3.5, Alpaca, Vicuna, and LLaMA-2; the second evaluated ChatGLM2-6B, GLM-130B, GPT-3.5, and GPT-4.Source: §3.5 · pp. 14–15

Discussion

• The review finds preliminary feasibility but notable gaps between current LLM research and actual clinical applicability.It synthesizes 34 studies and identifies three areas for improvement toward clinical practice.Source: §4 · p. 15

Improvements in data quality for LLM development and validation

• Prompt-tuning is prevalent for its simplicity and conversational style, but GPT-3.5 and GPT-4 may underperform in complex mental-health contexts and yield biased results.The review recommends investigating fine-tuning open-source LLMs because GPT-4 is not modifiable.Source: §4 · p. 15
• Social-media and dialogue datasets can support screening but are insufficient for formal diagnoses and may have unrepresentative samples and unvalidated, variable labels.Only one study distinguished clinical diagnoses from self-reported mental states; scarcity of robust clinical datasets likely sustains social-media reliance in 2024, raising bias and generalizability concerns.Source: §4 · pp. 15–16

Enhanced reasoning, empathy capabilities, and evaluation methods

• Mental-health dialogue tasks require advanced reasoning and empathy, yet reasoning may reflect memorization and empathy findings remain inconclusive.Memorization-based reasoning may be less robust across diverse user dialogues and language use.Source: §4 · p. 15
• No unified evaluation framework makes free-text outputs difficult to assess, while differing human-rating criteria produce potentially incomparable findings.Automated language metrics do not capture the nuances of mental-health dialogues; evaluation heterogeneity persisted in 2023 and likely into 2024.Source: §4 · p. 16

Addressing privacy, safety, and ethical/regulatory concerns

• Clinical use requires privacy, safety, and ethical safeguards, including data-protection compliance, bias and transparency standards, informed consent, and emergency protocols.The review calls for collaboration among developers, mental-health professionals, ethicists, and legal experts, with continuous monitoring and evaluation.Source: §4 · p. 16

Continuing Trends in 2024

• Interdisciplinary integration may advance context-sensitive models, but it also requires governance frameworks addressing ethical risks and equitable implementation.The review describes this collaboration as likely intensified in 2024.Source: §4 · p. 16

Conclusion

• LLMs show promise for mental-health care through their ability to process and generate large volumes of textual information.The review cautions that feasibility alone does not establish clinical applicability.Source: §5 · p. 17
• Lack of valid training sets, standardized evaluation metrics, and attention to privacy and ethics require consideration before clinical implementation.The review prioritizes addressing these challenges over further pilot studies that demonstrate feasibility.Source: §5 · p. 17
• The review calls for new collaborations to refine technology, standardize evaluations, and support ethical applications.These efforts are framed as ways to realize LLMs’ potential in mental-health care.Source: §5 · p. 17

Key References

• Xu et al. distinguishes clinical diagnoses from self-reported mental states, while Harrigian et al. examines social-media model generalizability.Both works inform the review’s focus on clinically relevant data and reliable application.Source: Key References · p. 17
• Lee et al. [47] addresses empathy capabilities, while Jin et al. proposes a dedicated mental-health LLM benchmark amid calls for standardized evaluation.The review notes that current empathy results remain inconclusive.Source: Key References · p. 17
• Yang et al. exemplifies open-model fine-tuning with social-media data, and Ma et al. [115] reports user perspectives on support-related implementation challenges.The cited challenges include data interpretation, ethics, safety, and potential over-reliance.Source: Key References · p. 17

A.2 Competing Interests

• JT reports research support from Otsuka and an adviser role at Precision Mental Wellness; all other authors report no conflict of interest.The disclosure names no other authors with conflicts.Source: §A.2 · p. 30

• Y.H. handled study design, data collection, and article screening.Source: §A.3 · p. 30
• Y.H., F.L., K.Y., and Z.L. conducted result analysis.Source: §A.3 · p. 30
• All authors reviewed the manuscript, and Y.H. held integral responsibility for the study.Source: §A.3 · p. 30

B.1 S.1. Screening prompt for ChatGPT-4

• The Figure B1 template was used to screen titles and abstracts.Source: §B.1 · p. 30, Figure B1 · p. 31
• The selected prompt scored 10 out of 10 on a randomly selected sample.Source: §B.1 · p. 30
• Explicitly requiring a focus on LLMs and mental health care helped avoid vaguely related articles.Source: §B.1 · p. 30

S.2.: Data Restriction and Licensing in Related Datasets

• Restrictions and licenses set the guidelines for dataset use, sharing, alteration, and dissemination.Source: §S.2 · p. 31
• The reviewed categories include public availability, research-only use, MIT and Apache 2.0 licenses, and CC-BY-4.0 and CC-BY-NC-4.0.Source: §S.2 · p. 31
• Datasets lacking a stated license or availability are marked N/P, and providers or creators may be contacted for usage information.Source: §S.2 · p. 32