Correctness, Harmfulness, and Diversity of Large Language Models for Colonoscopy Preparation Assistance: Comparative...

Area of Science:

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

Background:

• Colorectal cancer is a major health concern, with colonoscopy being crucial for early detection.
• Inadequate bowel preparation frequently leads to postponed colonoscopies due to patient difficulties with instructions.
• Existing digital tools offer limited support for patient-specific preparation questions.

Purpose of the Study:

• To evaluate the correctness, safety, and diversity of AI-generated dialogues for colonoscopy preparation.
• To compare the performance of leading large language models (LLMs) in simulating patient-AI Coach interactions.
• To assess the effectiveness of a safety filter in mitigating errors during AI-driven patient support.

Main Methods:

• Five leading LLMs generated 250 simulated patient-AI Coach dialogues each.
• Dialogues covered diet, medications, and preparation instructions, using a multiprompt approach for question diversity.
• Human experts and an LLM-as-a-judge system evaluated dialogue correctness, error types, and potential harmfulness.

Main Results:

• Leading LLMs showed potential but did not achieve adequate performance for colonoscopy preparation.
• Closed-weight models (GPT-5.1, GPT-4.1, o3) outperformed open-weight models (Llama, Mistral) in response correctness.
• All models generated harmful errors; a safety filter reduced but did not eliminate these issues.

Conclusions:

• LLMs offer potential for colonoscopy preparation support but require further development for safe deployment.
• Addressing persistent harmful errors and improving safety mechanisms are critical for patient-facing applications.
• Validation with real patient queries is essential before widespread use of LLM-based preparation tools.