JMIR Publications

ABSTRACT

Background:

Large Language Models (LLMs) are increasingly powering conversational agents in digital mental health interventions (DMHI). Despite their growing use, there remains a lack of clarity regarding the models’ development, evaluation, and deployment processes, as well as their alignment with ethical and clinical standards.

Objective:

This systematic review aims to examine the design, implementation, and evaluation of LLM-based mental health chatbots and agentic AI systems, focusing on their underlying model architectures, development methodologies, evaluation strategies, and deployment approaches.

Methods:

We conducted a systematic search of peer-reviewed publications and preprints through databases including PubMed, IEEE Xplore, ACL Anthology, and arXiv. Twenty studies were selected based on predefined eligibility criteria. Data extraction covered LLM types, training approaches, system architecture, evaluation metrics, and deployment context. Studies were assessed for methodological rigor, including whether external validation or clinical trial registration was conducted.

Results:

Among the 20 included studies 45% (n=9) employed GPT-based models (GPT-2, GPT-3, GPT-4), while only 55% (n=11) used fine-tuned or domain-specific variants (e.g., ClinicalT5, LLaMA, ChatGLM, Qwen). Chatbot deployment types included standalone applications (65%, n=13), virtual agents (25%, n=5), and embedded platforms (15%, n=3). Evaluation strategies were predominantly qualitative (65%, n=13), including thematic analysis and rubric-based scoring, while 90% (n=19) also used quantitative metrics such as BLEU, ROUGE, and perplexity. Only 10% (n=2) conducted any form of external validation, and none reported psychometric validation or standardized clinical outcome measurement. No included study reported trial registration or randomized controlled trial data.

Conclusions:

LLM-based mental health systems show potential to enhance user engagement and personalization in DMHI, particularly through adaptive, multi-modal agent structures. However, the current literature reflects limited methodological rigor, with gaps in external validation, standardization, and ethical compliance. To ensure safe and effective deployment, future research should prioritize clinical validation, robust evaluation frameworks, and transparent governance of AI behaviors.