Evaluating the Potential of Large Language Models for Respiratory Disease Consultations
- Conditions
- Acute Upper Respiratory InfectionLung CancerPneumoniaAcute BronchitisBronchiectasisAsthmaPulmonary EmbolismTuberculosisHay FeverPulmonary Fibrosis
- Interventions
- Diagnostic Test: Diagnosis by three human doctorsDiagnostic Test: Diagnosis by ChatGPT-3.5 (with search capabilities)Diagnostic Test: Diagnosis by ChatGPT-3.5 (without search capabilities)Diagnostic Test: Diagnosis by ChatGPT-4.0 (with search capabilities)Diagnostic Test: Diagnosis by ChatGPT-4.0 (without search capabilities)Diagnostic Test: Diagnosis by Claude instant (with search capabilities)Diagnostic Test: Diagnosis by Claude instant (without search capabilities)Diagnostic Test: Diagnosis by Claude 2 (with search capabilities)Diagnostic Test: Diagnosis by Claude 2 (without search capabilities)Diagnostic Test: Diagnosis by Gemini Pro (with search capabilities)Diagnostic Test: Diagnosis by Gemini Pro (without search capabilities)
- Registration Number
- NCT06457269
- Lead Sponsor
- North Sichuan Medical College
- Brief Summary
The clinical trial aimes to evaluate multiple large language models in respiratory disease consultations by comparing their performance to that of human doctors across three major medical consultation scenarios.
The main question aims to answer are:
* How do large language models perform in comparison to human doctors in diagnosing and consulting on respiratory diseases across various clinical scenarios?
In three clinical scenarios including the online query section, the disease diagnosis section and the medical explanation section, research assistants or volunteers will be asked to cross-question all LLMs or real doctors using predefined online questions and their own issues. After each questioning session, a short washout period is implemented to eliminate potential biases.
- Detailed Description
Not available
Recruitment & Eligibility
- Status
- COMPLETED
- Sex
- All
- Target Recruitment
- 703
- Self-reported symptoms of common respiratory diseases, such as cough, chest tightness, fever, and wheezing
- Ability to engage in LLM dialog operations independently or with minimal peer training
- A health status deemed suitable for study participation by the pulmonology experts
- Excessively poor health status
Study & Design
- Study Type
- INTERVENTIONAL
- Study Design
- CROSSOVER
- Arm && Interventions
Group Intervention Description Cross-comparison group(the disease diagnosis section) Diagnosis by Claude 2 (with search capabilities) Cross-comparison group (including human doctor controls and all LLMs) Cross-comparison group(the disease diagnosis section) Diagnosis by Gemini Pro (with search capabilities) Cross-comparison group (including human doctor controls and all LLMs) Cross-comparison group(the disease diagnosis section) Diagnosis by Gemini Pro (without search capabilities) Cross-comparison group (including human doctor controls and all LLMs) Cross-comparison group(the medical explanation section) Diagnosis by ChatGPT-4.0 (with search capabilities) Cross-comparison group (including human doctor controls and all LLMs) Cross-comparison group(the disease diagnosis section) Diagnosis by three human doctors Cross-comparison group (including human doctor controls and all LLMs) Cross-comparison group(the disease diagnosis section) Diagnosis by ChatGPT-3.5 (with search capabilities) Cross-comparison group (including human doctor controls and all LLMs) Cross-comparison group(the medical explanation section) Diagnosis by three human doctors Cross-comparison group (including human doctor controls and all LLMs) Cross-comparison group(the medical explanation section) Diagnosis by ChatGPT-3.5 (with search capabilities) Cross-comparison group (including human doctor controls and all LLMs) Cross-comparison group(the medical explanation section) Diagnosis by Claude 2 (with search capabilities) Cross-comparison group (including human doctor controls and all LLMs) Cross-comparison group(the medical explanation section) Diagnosis by Claude 2 (without search capabilities) Cross-comparison group (including human doctor controls and all LLMs) Cross-comparison group(the medical explanation section) Diagnosis by Gemini Pro (with search capabilities) Cross-comparison group (including human doctor controls and all LLMs) Cross-comparison group(the disease diagnosis section) Diagnosis by ChatGPT-3.5 (without search capabilities) Cross-comparison group (including human doctor controls and all LLMs) Cross-comparison group(the disease diagnosis section) Diagnosis by ChatGPT-4.0 (with search capabilities) Cross-comparison group (including human doctor controls and all LLMs) Cross-comparison group(the disease diagnosis section) Diagnosis by Claude instant (with search capabilities) Cross-comparison group (including human doctor controls and all LLMs) Cross-comparison group(the disease diagnosis section) Diagnosis by Claude instant (without search capabilities) Cross-comparison group (including human doctor controls and all LLMs) Cross-comparison group(the disease diagnosis section) Diagnosis by Claude 2 (without search capabilities) Cross-comparison group (including human doctor controls and all LLMs) Cross-comparison group(the medical explanation section) Diagnosis by Gemini Pro (without search capabilities) Cross-comparison group (including human doctor controls and all LLMs) Cross-comparison group(the disease diagnosis section) Diagnosis by ChatGPT-4.0 (without search capabilities) Cross-comparison group (including human doctor controls and all LLMs) Cross-comparison group(the medical explanation section) Diagnosis by Claude instant (with search capabilities) Cross-comparison group (including human doctor controls and all LLMs) Cross-comparison group(the medical explanation section) Diagnosis by ChatGPT-3.5 (without search capabilities) Cross-comparison group (including human doctor controls and all LLMs) Cross-comparison group(the medical explanation section) Diagnosis by ChatGPT-4.0 (without search capabilities) Cross-comparison group (including human doctor controls and all LLMs) Cross-comparison group(the medical explanation section) Diagnosis by Claude instant (without search capabilities) Cross-comparison group (including human doctor controls and all LLMs)
- Primary Outcome Measures
Name Time Method Expert indicators-Accuracy For each participant, starting from the day of random conversation, a maximum participation time of one week will be given. As for subjective expert indicators, the evaluation will be conducted within two months. Based on the doctors' responses to patients' issues, a 5-point scale will be used for scoring by an expert panel: 5- The responses are completely accurate, addressing all of the patient's questions or diagnosing by identifying the key points of the patient's complaint. 4- The responses are mostly accurate, generally addressing the patient's questions or diagnosing by identifying the key points of the patient's complaint. 3- The responses are moderately accurate, addressing the patient's questions or diagnosing by identifying the key points of the patient's complaint. 2- The responses are rarely accurate, barely addressing the patient's questions or diagnosing by identifying the key points of the patient's complaint. 1- The responses are very inaccurate, not addressing the patient's questions or diagnosing by identifying the key points of the patient's complaint at all.
Expert indicators-Comprehensiveness For each participant, starting from the day of random conversation, a maximum participation time of one week will be given. As for subjective expert indicators, the evaluation will be conducted within two months. Based on the doctors' responses to patients' issues, a 5-point scale will be used for scoring by an expert panel: 5-The responses are highly comprehensive, addressing various aspects of potential diseases corresponding to the patient's symptoms, providing detailed advice, and offering its own extended interpretations. 4-The responses are mostly comprehensive, covering most aspects of potential common diseases related to the patient's symptoms, and providing fairly detailed advice. 3-The responses are moderately comprehensive, addressing some aspects of potential common diseases related to the patient's symptoms, and offering basic advice. 2-The responses are rarely comprehensive, failing to consider various aspects of potential common diseases related to the patient's symptoms, and providing very limited advice. 1-The responses are not comprehensive at all, overlooking most potential diseases related to the patient's symptoms, and failing to provide any advice.
Empathy indicators For each participant, starting from the day of random conversation, a maximum participation time of one week will be given. As for subjective empathy indicators, the evaluation will be conducted within two months. Results from CARE scales concerning the doctor-patient relationship, which were completed by patients following each diagnostic session. Specifically, the online query section does not apply the evaluation of CARE scales.
Expert indicators-Ethical compliance For each participant, starting from the day of random conversation, a maximum participation time of one week will be given. As for subjective expert indicators, the evaluation will be conducted within two months. Based on the doctor's response to the patient's question, an expert panel will review each item in accordance with the Declaration of Helsinki and the International Code of Medical Ethics which aims to determine whether there are any responses or suggestions that could potentially harm the patient or violate ethical guidelines. The findings will be recorded using binary variables: True-The responses are completely ethical. False-When uncertainties exist, the response includes suggestions for the use of controlled medications and some inappropriate or even counterproductive advice.
Expert indicators-Correctness For each participant, starting from the day of random conversation, a maximum participation time of one week will be given. As for subjective expert indicators, the evaluation will be conducted within two months. Based on the doctors' responses to patients' issues, a 5-point scale will be used for scoring by an expert panel: 5- The responses are completely correct, with no inappropriate or ambiguous statements. 4- The responses are mostly correct, with most statements being appropriate and unambiguous. 3- The responses are generally correct, although there are inappropriate or ambiguous statements, they are acceptable. 2- The responses are partially correct, with few statements being appropriate or unambiguous. 1- The responses are completely incorrect, with nearly all statements being inappropriate and full of ambiguities.
- Secondary Outcome Measures
Name Time Method Regular indicators-Total number of conversations For each participant, starting from the day of random conversation, a maximum participation time of one week will be given. After the completion of the dialogues, the system will automatically summarize all objective indicators and dialogue information. The total number of dialogs in a complete conversation between a user and LLMs or a real doctor, where each dialog consists of one question and one answer.
Regular indicators-Total conversation cost ($) For each participant, starting from the day of random conversation, a maximum participation time of one week will be given. After the completion of the dialogues, the system will automatically summarize all objective indicators and dialogue information. The total cost in dollars for completing the entire conversation.
Regular indicators-Total conversation time (min) For each participant, starting from the day of random conversation, a maximum participation time of one week will be given. After the completion of the dialogues, the system will automatically summarize all objective indicators and dialogue information. Timing starts from the user's input and stops when the LLMs or real doctors completes the output of the last sentence.
Regular indicators-Follow-up words For each participant, starting from the day of random conversation, a maximum participation time of one week will be given. After the completion of the dialogues, the system will automatically summarize all objective indicators and dialogue information. The number of words in follow-up questions asked by the LLM or real doctor to the patient after providing basic answers in a complete conversation.
Regular indicators-Number of output statements For each participant, starting from the day of random conversation, a maximum participation time of one week will be given. After the completion of the dialogues, the system will automatically summarize all objective indicators and dialogue information. The total number of words output by the LLMs or real doctors.
Regular indicators-Total number of questions For each participant, starting from the day of random conversation, a maximum participation time of one week will be given. After the completion of the dialogues, the system will automatically summarize all objective indicators and dialogue information. The number of follow-up questions asked by the LLM or real doctor to the patient after providing basic answers in a complete conversation.
Regular indicators-Number of input statements For each participant, starting from the day of random conversation, a maximum participation time of one week will be given. After the completion of the dialogues, the system will automatically summarize all objective indicators and dialogue information. The sum of the number of characters entered by the user.
Trial Locations
- Locations (1)
The Affiliated Hospital of North Sichuan Medical College
🇨🇳Nanchong, Sichuan, China