Design

Our study was cross-sectional. Our study used the CHART checklist, the Reporting Guideline for Chatbot Health Advice Studies (see S1 File).

For the first objective, we developed a conversational agent based on LLM and conducted a descriptive study of its technical aspects, system design, and functionality testing. For the second objective, we evaluated the safety and performance metrics of the conversational agent’s responses, which were assessed by the research team in a controlled setting. The performance of two LLMs was evaluated: GPT-4o, which was selected for being the most widely used OpenAI version at the time of writing the study protocol (May 2025), and Llama 3.1-8B, which was chosen for being an open-access, easy-to-implement, and free LLM. It should be noted that the LLMs were not subjected to fine-tuning or Retrieval-Augmented Generation (RAG). Both LLMs used the same prompt, which was designed to simulate a cognitive-behavioral therapist with a supportive and validating attitude, providing short responses. The full prompt is provided in S2 File. The research protocol was written and planned in May 2025, was approved by the ethics committee on September 19, 2025, and the evaluation and recruitment took place between September 20 and 21, 2025.

Setting

The Digital Health Research Center team designed a mental health care platform that integrates a conversational agent, mental health assessments, behavioral activation-based activities and an alert/emergency system. The platform was designed to be interoperable with the OpenMRS electronic medical record system. However, for the purposes of this study, we do not address OpenMRS interoperability and instead focus on the internal functioning of the platform. The platform is intended as an additional resource for both in-person and telehealth care in a private mental health clinic. It has been developed in both English and Spanish. This research is part of the Mental Health platform supported by the Artificial Intelligence study (MHAI Study). The target population consists of users aged 18–30 who own a smartphone, are digital natives, and are simultaneously undergoing psychotherapy or pharmacological treatment for depression or anxiety.

Participants

The participants in this study were four members of the research team. For the first objective, which focuses on the development of the conversational agent, we describe the technical aspects of its development, system design, and functionality testing. We also present the iterative improvement process of the conversational agent, based on a user experience approach.

For the second objective, the conversational agent’s responses were evaluated by research team members with expertise in psychotherapy and digital mental health. The evaluation of safety and performance metrics were based on simulated conversations between the conversational agent and potential users experiencing anxiety and depression. To ensure that researchers simulated comparable user cases, three user personas were designed as a basis for the simulated interactions (see S3 File). These user personas specified baseline characteristics that researchers were instructed to simulate, with each conversation consisting of a minimum of 17 interaction pairs (17 responses from the LLM and 17 from the researcher). Each interaction pair served as a unit of evaluation.

The number of interaction pairs was determined based on an independent-means difference analysis to evaluate potential differences in the performance of GPT-4o and Llama 3.1-8B (two groups). A minimum of 788 interaction pairs was estimated to be required for evaluation by the four raters, assuming a small effect size (d = 0.2), an alpha probability of 0.05, 80% power, a two-tailed distribution, and an allocation ratio of 1. Each researcher generated responses for both conversational agents in English and Spanish (x2), using both LLMs (x2), across the three user personas per LLM and language (x3), and a minimum of 17 interaction pairs in each case (x17), resulting in a total of 204 interaction pairs evaluated by each researcher. All research team members were adults (over 18 years old), fluent in both English and Spanish, held undergraduate degrees in health-related fields, and had at least three years of experience in mental health.

Measurements and procedures

Safety and conversational quality performance metrics.

A Likert scale from 1 (did not meet the criterion) to 5 (fully met the criterion) was used, and the evaluation criteria were defined in Table 1. The different criteria used had been applied in previous studies [7,17].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. The criteria used to evaluate LLMs’ responses.

https://doi.org/10.1371/journal.pone.0344939.t001

Analysis plan

For the main hypothesis (second objective), we evaluated which of the LLMs demonstrated better performance metrics and safety based on mixed-effects models adjusted for evaluator. Model adjusted by language (Spanish/English), user persona, and researcher. In addition, some of the responses were presented descriptively as examples of the style of responses for each model. We considered the hypothesis to be supported when the differences in performance metrics and safety were statistically significant (p < 0.05) and favored GPT-4o.

We conducted a one-way sensitivity analysis to contextualize the estimated token-based costs under realistic variations in session intensity. Using the mean prompt and completion token counts observed in our evaluation and the per-token prices. We computed the expected LLM execution cost for a standardized course of care comprising 10 sessions of approximately 45 minutes each. Assuming an interaction rate ranging from 1 to 3 model responses per minute (45–135 responses per session).

Ethical aspects

This study did not involve external human participants and therefore posed no ethical risk. The protocol was approved by the Institutional Research Ethics Committee of the Universidad Científica del Sur (N°1235-CIEI-CIENTÍFICA-2025). Because the study aimed to assess the safety of the conversational agent, no patients or potential end users were included, as exposure to potential hallucinations from the agent could pose an ethical risk to their safety or well-being. The researchers who evaluated the conversational agent’s responses completed and signed a written informed consent form and agreed to participate in the study. The informed consent form was sent by email, and the researchers returned it signed. The study did not include minors.

Development and system design

A user-centered design approach was adopted, and the integration of different system components was analyzed. The system is structured into three main layers: frontend, backend, and database (see Fig 1 and S1 Table). Regarding deployment infrastructure, the application is hosted on Replit and configured to run on Cloud Run. The specific architecture detailing the connections between the various technologies used can be found in S1 Fig.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. APP architecture.

https://doi.org/10.1371/journal.pone.0344939.g001

APP Features: The app includes six main functionalities to reduce depressive and anxiety symptoms (see Fig 2): (1) User authentication system, which enables secure registration and access to the platform, ensuring privacy through secure session handling, tracking each user’s last login, and activity logging. (2) Psychological assessments, integrating the PHQ-9 and GAD-7 questionnaires for detecting and monitoring symptoms of depression and anxiety, ensuring data storage in the database. (3) Therapeutic chat with AI, utilizing the OpenAI GPT-4o model with a prompt designed to generate responses in a therapeutic style based on cognitive behavioral therapy (CBT) and incorporating CBT-based clinical manuals within the context window. (4) Therapeutic tasks, including mindfulness breathing exercises and scheduled physical activities, with weekly monitoring to assess user adherence to therapeutic recommendations. (5) An emergency alert system triggers an email alert for the clinical team to make immediate contact in critical situations. (6) Accessibility, security, and privacy features, including an encryption mechanism to protect sensitive data, a role-based access control system, and multilingual support in Spanish and English to enhance platform accessibility.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Platform features.

https://doi.org/10.1371/journal.pone.0344939.g002

The application’s workflow begins with user registration and login, ensuring secure authentication on the platform. Once logged in, users can access a variety of therapeutic features available on the dashboard, designed to address symptoms of anxiety and depression. S4 File provides a visual representation of the app’s user interface.

Safety and performance metrics

Our study evaluated the performance metrics based on quantitative response characteristics of two LLMs (GPT-4o vs. Llama 3.1-8B) and found that GPT-4o produced longer responses, with an average of 71.59 more characters than Llama 3.1-8B (95% CI: 51.23 to 91.94; p < 0.001), and exhibited greater lexical diversity, with a 7.77% increase compared to Llama 3.1-8B (95% CI: 6.86 to 8.69; p < 0.001). Additionally, GPT-4o used, on average, 3.16 more tokens for the input (95% CI: 1.37 to 4.95; p < 0.001) and 26.86 more tokens for the output (95% CI: 22.77 to 30.94; p < 0.001) than Llama 3.1-8B. This also translated into a higher cost per 100,000 tokens in USD for GPT-4o compared to Llama 3.1-8B (see Table 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. The difference between the large language model (GPT-4o vs Llama 3.1-8B) and mixed-effects models adjusted for evaluator (n = 816).

https://doi.org/10.1371/journal.pone.0344939.t002

In terms of conversational quality performance metrics, GPT-4o outperformed Llama 3.1-8B in several domains such like clarity of responses, the ability to handle ambiguous or incomplete questions while maintaining conversational coherence (robustness), comprehensive answers without including unsolicited information (completeness), the ability to restrict responses to mental health related content (boundaries), and the ability to respond in the same language in which the question was asked (target language) (see  Table 2). No significant differences were observed in tone, domain accuracy, or response safety.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Example of five pairs of consecutive interactions for the large language models (GPT-4o vs. Llama 3.1-8B).

https://doi.org/10.1371/journal.pone.0344939.t003

Consistent with our hypothesis, GPT-4o scored higher than Llama 3.1–8B on clarity, robustness, completeness, boundaries, and target language, while no differences were observed for tone, domain accuracy, or safety.

One-way sensitivity analysis of cost

The estimated total cost per patient by 10 sessions of 45 min each one, ranged from $0.22 to $0.66 for GPT-4o, compared with $0.0017 to $0.0053 for Llama 3.1-8B (see S2 Table). These results indicate that the relative cost difference is robust to plausible changes in message volume, although our estimates capture only variable token costs and exclude infrastructure and implementation overhead.

Examples of interaction pairs

We present examples of interaction pairs conducted by the authors while simulating the user personas (see Table 3). It should be noted that no interaction pairs that were clinically risky or unsafe received a score of 1, nor were there any particularly relevant cases that could be listed; therefore, no concrete examples of potentially unsafe interactions could be provided.

Strengths and limitations

This study has some limitations. First, we evaluated only two LLMs, GPT-4o and Llama 3.1-8B, without including other open-access models such as DeepSeek, larger Llama variants, or other state-of-the-art models such as Gemini 2.5-Pro, Grok-4, or GPT-o3. This restricts the generalizability of our findings and limits comparison with models that may present different performance–cost profiles. Although our results cannot be directly generalized to other LLMs, prior evidence indicates that state-of-the-art reasoning models generally outperform older or non-reasoning models [27]. Therefore, while our findings may not extend to earlier models, it is plausible that more recent state-of-the-art models would demonstrate similar or superior performance. Second, the evaluation was conducted using research team members simulating user personas rather than real patient interactions. While this approach limits ecological validity, we considered it appropriate for initial safety evaluation to avoid potential ethical risks associated with exposing vulnerable populations to unvalidated systems. Therefore, in later phases of the study, the platform should be validated in real-world settings. For example, this includes studies with real users (patients) in supervised environments, followed by randomized clinical trials. Third, although we employed a systematic method to evaluate the performance metrics based on Likert scale criteria, evaluations were conducted by expert evaluators from the research team, which may introduce subjective bias despite standardized training protocols. Fourth, because different evaluators rated different interactions, it was not possible to estimate inter-rater reliability coefficients (i.e., kappa or ICC). However, to mitigate evaluator-related bias, we used mixed-effects models adjusted for evaluator, thereby preventing the main analysis from being affected by rater severity or leniency, that is, by an evaluator being systematically more stringent or more lenient. Fifth, our study did not include fine-tuning or retrieval-augmented generation (RAG) at the model design stage. Although it has been shown that fine-tuned models exhibit performance comparable to general-purpose models for different health-related tasks [28], the use of RAG could represent an alternative to improve the model’s metric performance. Therefore, we estimate that subsequent stages of the study will include the use of RAG with transcripts of psychotherapy recordings, coding patient and therapist responses, and incorporating care manuals or clinical practice guidelines to improve the performance of the responses. Sixth, a cost-effectiveness analysis or clinical benchmarking were not possible because, although information on the cost per input and output token was available, there was no information on the clinical effect of the intervention. Therefore, future studies, such as randomized clinical trials, should include cost-effectiveness evaluations of this type of digital mental health intervention. Seventh, it was not possible to perform a sensitivity analysis of the findings for each LLM and user persona, as the available sample size was insufficient to replicate the main analysis stratified by these different conditions.

Conversely, this study demonstrates several important strengths that enhance the validity and applicability of our findings. First, we implemented a standardized evaluation protocol using validated metrics, enabling systematic and reproducible comparisons between models while maintaining consistency with established evaluation standards in health chatbot research [7]. Second, our study evaluates the potential of multilingual LLMs (English and Spanish), which is particularly relevant given the growing need for culturally and linguistically aligned mental health interventions [29]. Third, we make the dataset publicly available to support future analyses, in line with principles of transparency and open access.

Implications for future studies

Our study outlines a pathway for future research that addresses the limitations identified and ensures a substantial contribution to the field of digital mental health. First, we require evaluations with real users in controlled settings, that is, a preliminary usability assessment with patients and therapists to ensure that the platform is acceptable, usable, and can be adopted as a complement to psychotherapeutic processes. In addition, we should assess additional requirements that may be valuable to users and were not previously considered. Second, to ensure the ecological validity of the findings in real-world contexts, we must conduct a preliminary effectiveness evaluation in small groups, such as pilot clinical trials, or by using non-experimental methods, such as case series without a control group. This would allow us to determine whether the observed results are replicable in routine care settings and to assess preliminary clinical efficacy. Moreover, it would help identify the optimal platform prescription, such as the frequency of use and the number of sessions. Third, we should conduct robust clinical studies with control groups and adequate sample sizes. This would allow us to establish clinical efficacy under controlled conditions. Fourth, we should perform implementation studies in primary care and hospital settings to evaluate the real-world impact of MHAI in clinical contexts. At the same time, registration with the relevant regulatory authorities, such as the Food and Drug Administration (FDA) in the United States or the Dirección General de Medicamentos, Insumos y Drogas (DIGEMID) in Peru, will be required if the platform is to be considered a medical device for therapeutic use. Prior to this stage, the platform should be regarded as a tool that may improve overall emotional well-being, rather than as a formal therapy.

At the architectural level of MHAI, the platform must implement additional servers to support high traffic volumes through horizontal scalability, while maintaining compliance with data security and privacy standards such as HIPAA. Moreover, MHAI should not operate as an isolated platform; instead, it should be integrated into a real clinical care process and interoperable with electronic health record systems such as OpenMRS, Epic, or Cerner.

Regarding new features and functions, several developments are required. First, the emergency button must be tested under real-world conditions with real users, as only patients can determine whether it is helpful in practice. Although the current implementation is functional and operates correctly, it has not yet been tested in real-world contexts due to ethical constraints. Second, functions that allow users to obtain information from a licensed health professional, schedule a new appointment, or be redirected to a real clinical care pathway should be implemented. Third, additional large language models should be implemented to review conversations and extract clinically relevant information that can be validated by a health professional and subsequently entered into the patient’s electronic health record. For example, one model could identify suicide risk in conversations, while another could detect signs and symptoms that can later be validated and documented in the medical record.