Related Experiment Video
Updated: Jan 16, 2026
Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
Published on: December 6, 2024
Using Large Language Models to Assess the Consistency of Randomized Controlled Trials on AI Interventions With
Xufei Luo1,2,3,4,5, Zeming Li6, Zhenhua Yang7
1Evidence-Based Medicine Center, School of Basic Medical Sciences, Lanzhou University, 199 Donggang West Road, Chengguan District, Lanzhou, 730000, China, 86 13893104140.
Background:
Chatbots based on large language models (LLMs) have shown promise in evaluating the consistency of research. Previously, researchers used LLM to assess if randomized controlled trial (RCT) abstracts adhered to the CONSORT-Abstract guidelines. However, the consistency of artificial intelligence (AI) interventional RCTs aligning with the CONSORT-AI (Consolidated Standards of Reporting Trials-Artificial Intelligence) standards by LLMs remains unclear.
Objective:
The aim of this study is to identify the consistency of RCTs on AI interventions with CONSORT-AI using chatbots based on LLMs.
Methods:
This cross-sectional study employed 6 LLM models to assess the consistency of RCTs on AI interventions. The sample selection is based on articles published in JAMA Network Open, which included a total of 41 RCTs. All queries were submitted to LLMs through an application programming interface with a temperature setting of 0 to ensure deterministic responses. One researcher posed the questions to each model, while another independently verified the responses for validity before recording the results. The Overall Consistency Score (OCS), recall, inter-rater reliability, and consistency of contents were analyzed.
Results:
We found gpt-4-0125-preview has the best average OCS on the basis of the results obtained by JAMA Network Open authors and by us (86.5%, 95% CI 82.5%-90.5% and 81.6%, 95% CI 77.6%-85.6%, respectively), followed by gpt-4-1106-preview (80.3%, 95% CI 76.3%-84.3% and 78.0%, 95% CI 74.0%-82.0%, respectively). The model with the worst average OCS is gpt-3.5-turbo-0125 on the basis of the results obtained by JAMA Network Open authors and by us (61.9%, 95% CI 57.9%-65.9% and 63.0%, 95% CI 59.0%-67.0%, respectively). Among the 11 unique items of CONSORT-AI, Item 2 ("State the inclusion and exclusion criteria at the level of the input data") received the poorest overall evaluation across the 6 models, with an average OCS of 48.8%. For other items, those with an average OCS greater than 80% across the 6 models included Items 1, 5, 8, and 9.
Conclusions:
GPT-4 variants demonstrate strong performance in assessing the consistency of RCTs with CONSORT-AI. Nonetheless, refining the prompts could enhance the precision and consistency of the outcomes. While AI tools like GPT-4 variants are valuable, they are not yet fully autonomous in addressing complex and nuanced tasks such as adherence to CONSORT-AI standards. Therefore, integrating AI with higher levels of human supervision and expertise will be crucial to ensuring more reliable and efficient evaluations, ultimately advancing the quality of medical research.
