Large Language Models in Mental Health: A Systematic Review of Applications, Innovations, and Ethical Challenges
A review of 92 studies shows LLMs are getting better at spotting depression and suicide risk online, but real clinical proof is still missing
This is a systematic review examining how large language models (LLMs) are used in mental health, covering social media analysis, therapy chatbots, and clinical decision-support tools. Following the PRISMA guidelines, the authors screened 304 candidate papers down to a final set of 92 high-quality studies. They found that while LLM-based detection of depression and suicidality has advanced rapidly, clinical validation, multimodal integration, and ethical safeguards remain limited.
METAL MEDIA explanatory visual
A review of 92 studies shows LLMs are getting better at spotting depression and suicide risk online, but real clinical proof is still missing
- 01Maps how LLMs use data from Twitter/Reddit posts, electronic medical records (EMRs), and counseling notes to enable early depression detection, suicide risk assessment, personalized therapy support, and psychoeducational content generation
- 02Searched seven academic databases (IEEE Xplore, ACM Digital Library, PubMed, Scopus, ScienceDirect, SpringerLink, Google Scholar) for 304 papers, narrowed to 115 after applying inclusion/exclusion criteria, then to a final 92 based on methodological transparency and reproducibility
- 03Compares five annotation strategies for labeling mental health data, ranging from expert clinician labeling and self-disclosure-based labeling to LLM-assisted annotation using zero-shot or few-shot prompting
- 04Documents how models like GPT-3.5, LLaMA2, and MentaLLaMA have expanded beyond depression detection into summarizing therapy sessions, supporting treatment planning, and engaging directly with patients in conversation
- 05Identifies explainability (showing the reasoning behind a model's output) and prompt engineering (designing instructions to guide a model's behavior) as key unresolved challenges for real-world clinical deployment
What they did
- Maps how LLMs use data from Twitter/Reddit posts, electronic medical records (EMRs), and counseling notes to enable early depression detection, suicide risk assessment, personalized therapy support, and psychoeducational content generation
- Searched seven academic databases (IEEE Xplore, ACM Digital Library, PubMed, Scopus, ScienceDirect, SpringerLink, Google Scholar) for 304 papers, narrowed to 115 after applying inclusion/exclusion criteria, then to a final 92 based on methodological transparency and reproducibility
- Compares five annotation strategies for labeling mental health data, ranging from expert clinician labeling and self-disclosure-based labeling to LLM-assisted annotation using zero-shot or few-shot prompting
- Documents how models like GPT-3.5, LLaMA2, and MentaLLaMA have expanded beyond depression detection into summarizing therapy sessions, supporting treatment planning, and engaging directly with patients in conversation
- Identifies explainability (showing the reasoning behind a model's output) and prompt engineering (designing instructions to guide a model's behavior) as key unresolved challenges for real-world clinical deployment
Why it matters
As mental health professional shortages and access barriers persist, this review clarifies which LLM applications are genuinely validated versus still experimental, helping developers and clinicians see where the technology can be trusted. It also surfaces unresolved ethical issues like privacy, bias, and accountability that anyone building these tools needs to address before deployment.
Terms in this paper
- PRISMA · An international standard protocol for conducting transparent, reproducible systematic literature reviews
- zero-shot/few-shot learning · A model's ability to handle new tasks with no task-specific examples (zero-shot) or only a handful of examples (few-shot)
- electronic medical records (EMR) · Digital hospital records of a patient's diagnoses, treatments, and outcomes over time
- explainability · An AI system's capacity to show understandable reasoning behind its conclusions
- PHQ-9 · A widely used 9-question clinical instrument for measuring depression severity
Original abstract (English)
We present a review on the applications of large language models (LLMs) in health, e.g., social media analysis, clinical conversational agents, therapy support tools, prompt engineering, multimodal learning, and ethical considerations. We integrate findings from interdisciplinary studies utilizing diverse data sources such as social media posts, electronic medical records, and multimodal inputs to enable early detection of depression, suicide risk assessment, personalized therapy support, and psychoeducational content generation. Our review highlights advancements in LLM models and annotation strategies that enhance interpretability and clinical relevance, while we also emphasize the critical role of prompt engineering for domain adaptation. We also discuss emerging multimodal fusion techniques integrating text, speech, and sensor data for improved mental health diagnosis and monitoring. Finally, we address ongoing ethical, sociotechnical, and regulatory challenges, and advocate frameworks to ensure safe, equitable, and accountable deployment of LLMs in real-world mental health care.
Read on arXivLatest papers
- SWE-bench Science: Can Coding Agents Resolve Engineering Tasks in Science?AI coding agents were tested on fixing real scientific software, and even the best one failed more than half the time
- FlashPrefill V2: Block-Sparse Prefill Attention for Long-Context LLM ServingMaking sparse attention fast enough and accurate enough for real LLM serving, not just papers
- PolicyGuide: From Guarding One Action to Guiding the Whole Workflow for Policy-Compliant LLM AgentsMaking customer-service AI agents follow the whole procedure, not just avoid one bad action
- EXIMO: VLM Guided Exploration of VLA PoliciesTeaching a robot new chores without human teleoperation, by letting a chatty AI supervise it
- EnvHarness: Awakening Static Worlds for Agent LearningInstead of building new training worlds from scratch, this work adds a plug-in layer that reshapes existing ones around each agent's actual weaknesses
- Bounded Sovereignty and the Control Tax: Pricing AI Oversight When the Deployer Does Not Own the ModelCompanies that rent AI instead of owning it can only do half of AI safety oversight
- PersonalBench: Measuring the Authorship Gap in LLM PersonalizationAI can be prompted to write 'like someone,' but its own voice never fully disappears
- Automated Summarization of Financial News Using Large Language Models and Retrieval-Augmented Generation: An Early Empirical Study (Fall 2023)Testing AI summaries of stock news, the simple approach beat the trendy retrieval-based one