Related Experiment Video
Updated: Aug 28, 2026
Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
Published on: December 6, 2024
Quality and reliability of YouTube videos on PET/CT radiation safety: a comparative analysis between human experts
1Independent Researcher, Taoyuan, Taiwan. petershih5@gmail.com.
Purpose:
To evaluate the quality and accuracy of YouTube videos regarding PET/CT radiation safety and to assess the feasibility of using a Large Language Model (LLM) as an automated tool for content moderation compared to a human expert.
Methods:
A systematic search was conducted on YouTube using keywords related to "PET/CT radiation safety." A total of 42 videos were included and categorized by uploader source (Professional vs. Non-Professional). Video quality was assessed using the modified DISCERN (mDISCERN) tool and the Global Quality Scale (GQS). Usefulness and engagement metrics (views, likes) were analyzed. Additionally, inter-rater reliability between a human physician with clinical work experience in a nuclear medicine department and an AI model (Gemini-3) was evaluated using Cohen's Kappa (κ).
Results:
The majority of content (90.5%) originated from professional sources. Professional videos demonstrated significantly higher information quality compared to non-professional videos (mean mDISCERN: 3.71 vs. 2.50, P = 0.032; mean GQS: 4.18 vs. 2.50, P = 0.005). However, video popularity was not an indicator of quality, as no significant correlation was found between view counts and mDISCERN scores (ρ = 0.124, P = 0.432). Notably, the agreement between the human expert and the AI model was only slight (κ = 0.191). Discrepancy analysis revealed that the AI model systematically overestimated the quality of content containing commercial bias while penalizing highly technical academic videos for their presentation style.
Conclusion:
YouTube contains high-quality information on PET/CT safety, predominantly from professional sources; however, the platform's algorithm does not prioritize clinical accuracy, often favoring lower-quality content. Furthermore, current LLMs lack the contextual judgment required to reliably detect commercial bias and misinformation, underscoring the irreplaceable role of physician oversight in digital patient education.