Medical Misinformation in AI-Assisted Self-Diagnosis: Development of a Method (EvalPrompt) for Analyzing Large...

Area of Science:

• Artificial Intelligence in Healthcare
• Medical Informatics
• Natural Language Processing

Background:

• Large language models (LLMs) are rapidly integrating into healthcare, prompting discussions on their potential to improve quality and accessibility.
• While LLMs can pass medical exams, their use for self-diagnosis and potential to spread misinformation requires evaluation.

Purpose of the Study:

• To assess the effectiveness of LLMs, specifically ChatGPT, for individual self-diagnosis.
• To evaluate the clarity, correctness, and robustness of LLM responses in simulated self-diagnosis scenarios.

Main Methods:

• Developed the EvalPrompt methodology using medical licensing examination questions.
• Experiment 1: Prompted ChatGPT with open-ended questions simulating self-diagnosis.
• Experiment 2: Assessed robustness by using sentence dropout in responses to mimic missing information.

Main Results:

• ChatGPT-4.0 responses were deemed correct for only 31% of questions in Experiment 1.
• In Experiment 2, 61% of responses remained correct, indicating robustness but also potential for misinformation.
• ChatGPT-4.0 fell below the 60% passing threshold for clarity and correctness.

Conclusions:

• LLM capabilities for self-diagnosis are currently modest, with responses often unclear and inaccurate.
• Caution is advised when using LLM-generated medical advice due to significant misinformation risks.
• Development of a comprehensive self-diagnosis dataset is needed to improve LLM reliability in healthcare.