LLM-based AI Dialogue Bot, Vignette Based Intervention
Conditions
Brief summary
The goal of this randomized controlled trial is to evaluate the role of large language models in enhancing laypeople's ability to self-diagnose and triage common diseases. The main questions it aims to answer are: * Does using an LLM help participants make more accurate self-diagnoses and care decisions for common illnesses, compared to their first guess without any help? * How much better is it when people work together with an LLM, compared to using a regular search engine, using the LLM alone, or how doctors would decide? Researchers will compare participants who were randomly assigned to either the LLM group (using DeepSeek) or the search engine group to see if the LLM-assisted approach leads to better clinical judgments. Participants will: * Read one of 48 short, realistic health vignettes; * Make an initial guess about what might be wrong by listing up to three possible causes, ranked from most to least likely, and choose a care level: seek immediate care, see a doctor within one day, see a doctor within one week, or manage at home without medical care. * Use their assigned tool (either DeepSeek or a standard search engine) to look up information and update their guess and care decision; * Submit their final diagnosis and care choice after using the tool. In addition, the study team evaluated the performance of four other AI models (GPT-4o, GPT-o1, DeepSeek-v3, and DeepSeek-r1) and 33 experienced general physicians on the same vignettes.
Interventions
BEHAVIORALAI-assisted health information seeking
Participants in this group used a large language model (DeepSeek) to search for medical information related to a clinical vignette after providing initial diagnostic and triage decisions. They were instructed to interact freely with the model to gather insights and then update their diagnoses and triage recommendations. The intervention simulates real-world use of AI tools for personal health decision-making
BEHAVIORALConventional internet search for health information
Participants in this group used mainstream internet search engines (e.g., Baidu, Google, Bing) to look up information about the clinical vignette after making initial diagnostic and triage decisions. They were allowed to search freely but were not permitted to use any named AI chatbot or large language model platform. This group represents typical self-directed online health information seeking behavior.
Sponsors
Huazhong University of Science and Technology
Study design
Allocation
RANDOMIZED
Intervention model
PARALLEL
Primary purpose
HEALTH_SERVICES_RESEARCH
Masking
SINGLE (Subject)
Eligibility
Sex/Gender
ALL
Age
18 Years to No maximum
Healthy volunteers
No
Inclusion criteria
* Age 18 years or older * Current resident of mainland China * History of high-quality participation in online surveys on Credamo platform (historical survey acceptance rate ≥ 80% and personal credit score ≥ 70)
Exclusion criteria
* Incomplete survey responses * Failure on embedded quality-check items * Implausibly short completion time (\<180 seconds for search engine group; \<360 seconds for LLM group) * Provision of non-diagnostic or irrelevant responses (e.g., "unknown", "don't know") * Consistent pattern of identical responses across all items
Design outcomes
Primary
| Measure | Time frame | Description |
|---|---|---|
| Top-3 Diagnostic Accuracy | Immediately after intervention (within the same survey session) | The primary diagnostic outcome was defined as the proportion of participants who included the correct diagnosis in their top three differential diagnoses after using the assigned tool (LLM or search engine). Accuracy was assessed for each of the 48 clinical vignettes and aggregated across all participants in each group. |
| Triage Accuracy (4-class exact match) | Immediately after intervention (within the same survey session) | Triage accuracy was defined as the proportion of participants who selected the correct triage level (emergent care, within one day, within one week, or self-care) that matched the reference standard. There were 12 vignettes per triage category. |
Secondary
| Measure | Time frame | Description |
|---|---|---|
| Top-1 Diagnostic Accuracy | Immediately after intervention (within the same survey session) | The proportion of participants who selected the correct diagnosis as their top (first) diagnosis after using the assigned tool. This measures the precision of laypeople's final diagnostic judgment. |
| Triage Accuracy (2-class binary match) | Immediately after intervention (within the same survey session) | — |
Countries
China
Contacts
PRINCIPAL_INVESTIGATORChenxi Liu
Huazhong University of Science and Technology
Outcome results
None listed