Large Language Models for Traumatic Dental Injuries Across Web-Based and Mobile-Based Interfaces: Assessing Accuracy,...

Area of Science:

• Artificial Intelligence in Dentistry
• Clinical Decision Support Systems
• Information Retrieval for Traumatic Dental Injuries

Background:

• Traumatic dental injuries (TDIs) necessitate rapid, guideline-based clinical decisions.
• Accessing accurate information on TDIs can be challenging for practitioners.
• Large language models (LLMs) are emerging as quick online information resources.

Purpose of the Study:

• To evaluate the accuracy and consistency of multiple LLMs in answering TDI-related questions.
• To assess the influence of different user interfaces (web vs. mobile) on LLM performance.
• To compare the performance of ChatGPT-4o, DeepSeek-V3, Gemini 2.0 Flash, and Qwen2.5-Max.

Main Methods:

• Twenty TDI questions based on 2020 IADT guidelines were formulated (10 open-ended, 10 yes-no).
• Four LLMs were queried via web and mobile interfaces over five days, generating 800 responses.
• Open-ended answers were scored using GQS and mDISCERN; yes-no answers were compared to a key.

Main Results:

• Qwen2.5-Max achieved higher Global Quality Scores (GQS) and modified DISCERN (mDISCERN) scores.
• Yes-no question accuracy ranged from 86% to 91% with no significant model differences.
• ChatGPT-4o performed better on the web, while Qwen2.5-Max excelled on mobile; Qwen2.5-Max showed higher temporal consistency.

Conclusions:

• LLMs can be supplementary tools for guideline-based TDI information, particularly for closed-ended questions.
• LLM performance varies by model, interface, and question type.
• All LLM-generated information for TDIs must be cautiously interpreted and verified by dental professionals.