Retrieval-Augmented Generation in Mental Health: A Systematic Review

Technical Summary: Retrieval-Augmented Generation in Mental Health: A Systematic Review

Problem Statement
Large Language Models (LLMs) are increasingly applied to mental health contexts, including counseling, screening, and psychoeducation. However, unconstrained generation poses significant risks regarding hallucinations and safety. While Retrieval-Augmented Generation (RAG) has been shown to ground LLM outputs in external knowledge and improve factual consistency in general biomedical settings, its specific application, architectural variations, and evaluation standards within mental health remain under-synthesized. Existing systematic reviews have either focused on general healthcare RAG, generative AI in mental health broadly, or specific multimodal streams (e.g., VR, wearables) without linking them to RAG architecture. There is a lack of a systematic review that explicitly applies RAG-specific inclusion criteria to synthesize primary mental health studies under PRISMA 2020 guidelines.

Methodology
This review adheres to PRISMA 2020 reporting guidelines. The search was conducted on July 4, 2026, querying PubMed (yielding 56 records) and equivalent terms in IEEE Xplore, ACM Digital Library, Scopus, and Web of Science (June 4–15, 2026). Forward and backward citation tracking of included studies and benchmark reviews identified an additional 38 records.

• Eligibility: Inclusion criteria required peer-reviewed articles, conference proceedings, or preprints reporting empirical RAG or retrieval-augmented LLM systems applied specifically to mental health, stress, anxiety, depression, psychological assessment, or digital therapeutics. Exclusions included reviews, pure LLM/fine-tuning without retrieval, general biomedical RAG without mental health application, and non-empirical reports.
• Selection Process: After deduplication (23 removed), 71 unique records underwent title/abstract screening, followed by full-text assessment of 22 reports. Fifteen studies met the final inclusion criteria (published 2022–2026).
• Data Extraction & Quality Control: A single reviewer (E.F.) performed screening and data extraction. Quality control relied on reproducible query verification, cross-checking against primary publications/DOIs, and internal consistency checks. Independent duplicate screening was not performed, which is noted as a limitation.
• Taxonomy: Systems were classified into three architectural categories based on primary publication descriptions:
  • Naïve RAG: Single-pass retrieval and generation.
  • Advanced RAG: Hybrid retrieval, reranking, query expansion, or adaptive retrieval.
  • Modular RAG: Multi-component pipelines with distinct retrieval, routing, or safety modules.
• Synthesis: Due to heterogeneous outcome metrics (e.g., accuracy, F1, Hit@k, ICC, empathy scores) and a lack of commensurable effect-size reporting, quantitative meta-analysis was not performed. Synthesis was qualitative and narrative.

Key Results

• Study Characteristics: Fifteen studies were included. The architectural distribution was: 5 Naïve, 7 Advanced, and 3 Modular. The most common conditions addressed were depression and general psychological distress.
• Performance: Six studies reported direct RAG-versus-baseline comparisons. In every instance, RAG or knowledge-augmented configurations outperformed the stated comparator on the study's primary reported outcome. Specific examples include:
  • Tural et al. (2025): Vector RAG achieved a Hit@5 of 78.9% compared to 50.4% for lexical search.
  • Kafi et al. (2026): A fine-tuned model with RAG achieved an empathy score of 3.72 versus 3.26 for a generalist model.
  • Xie et al. (2026): A Qwen3+RAG configuration achieved a depression F1 score of 0.9107, representing an approximate 0.15 gain over non-RAG baselines.
• Safety and Validation: Crisis-management protocols were documented in only two studies (Vakayil et al.; Campellone et al.). Source attribution and human-in-the-loop oversight were reported inconsistently. Only two studies (Guo et al.; Campellone et al.) reported exploratory randomized controlled trial (RCT) or within-subject clinical designs with validated instruments; the majority were technical evaluations.
• Retrieval Context: All 15 included studies utilized text-based retrieval only. No study indexed multimodal behavioral, VR, or wearable physiological stress signals as retrievable context, despite adjacent literature suggesting the viability of such integration.

Significance and Claims
The paper claims to be the first systematic review to explicitly apply RAG-specific inclusion criteria to synthesize primary mental health studies under PRISMA 2020. Its primary contribution is the identification and architectural mapping of 15 empirical studies, demonstrating a growing trend of RAG adoption in mental health applications.

The review highlights that while RAG configurations consistently outperform baselines in the reported studies, the field suffers from:

1. Heterogeneous evaluation methods that prevent quantitative pooling.
2. Sparse safety reporting regarding crisis management and human oversight.
3. A gap in multimodal integration, as no included study leveraged physiological or behavioral signals for retrieval, despite their relevance in adjacent stress-monitoring literature.

The authors conclude that while RAG is increasingly reported, standardized clinical evaluation and robust safety documentation remain limited. The review does not claim to prove clinical efficacy but rather maps the current technical landscape and identifies specific gaps in retrieval sources and safety governance.