What question did this study set out to answer?

This study aims to evaluate the performance of large language models in answering thyroid eye disease-related questions and assess customization strategies.

April 20, 2026Open Access

Evaluating and enhancing the performance of large language models in thyroid eye disease through customization and Chain-of-Thought strategies

CLChaoyu LeiShanghai Jiao Tong University CZChen ZhaoShanghai Ninth People's Hospital KJKaiyuan JiShanghai Jiao Tong University

Key Points

This study aims to evaluate the performance of large language models in answering thyroid eye disease-related questions and assess customization strategies.
Evaluated several large language models on multiple-choice questions related to thyroid eye disease.
Customized top-performing models (GPT-4 and Claude 3.5) into TED-GPT and TED-Claude using chain-of-thought strategies.
Assessed accuracy and performance across multiple-choice, short-answer, and case questions using the QUEST framework.
The best accuracy for multiple-choice questions was observed in TED-Claude at 89.1% and TED-GPT at 86.1%.
Customization and chain-of-thought methods significantly improved performance in answering questions compared to original models.
TED-Claude outperformed others in accuracy, readability, reasoning, and likelihood of harm.

Abstract

The paucity of public awareness regarding thyroid eye disease (TED) usually leads to delayed medical care. While large language models (LLMs) hold great potential for augmenting patient education, their ability in answering TED-related questions has yet to be comprehensively evaluated. This study aims to assess the capability of LLMs to address TED-related questions and explore the practicability of customizing LLMs for disease-specific domains. Considering the diverse LLM candidates, we deployed a cascade pipeline to search for the best model for TED. We first evaluated performances of several prevailing LLMs on multiple-choice questions. The best-performing models, GPT-4 and Claude 3.5, were selected and customized to create TED-GPT and TED-Claude. Chain-of-Thought (CoT) was then utilized, resulting in CoT-GPT and CoT-Claude. We also evaluated newer LLMs with native CoT capabilities (GPT-4-o1, GPT-4-o3, Gemini-2.0-Flash, Gemini-2.5-Pro, Claude 3.7). These models, along with their original versions, were then assessed and compared on multiple-choice questions. The better-performing TED-GPT and TED-Claude were evaluated on short-answer and case questions, with comparisons made to their original ones using the QUEST framework (Quality, Understanding/Reasoning, Expression, Safety/Harm, Trust). For multiple-choice questions, GPT-4 and Claude 3.5 arrived competitive accuracies (76.2% and 83.2%, respectively). The addition of CoT, along with customization into GPT-4 and Claude 3.5, led to improvement in the accuracy (CoT-GPT 86.1%, CoT-Claude 87.1%, TED-GPT 86.1%, TED-Claude 89.1%), outperforming all other newer LLMs. For case and short-answer questions, the customized TED-GPT and TED-Claude also performed better than their original versions. TED-Claude showed the best performance in accuracy, readability, comprehensiveness, likelihood of harm and reasoning. Thus, LLMs, particularly TED-Claude, achieved relatively satisfactory performances in answering TED-related questions. Besides, using LLMs’ customization modules, along with CoT, effectively enhanced model performances. This indicates that clinicians can use these simple and universal methods to construct LLMs suitable for specific medical domains.

Ask AI

Helpful

Bookmark

View Full Paper