Using a large language model (ChatGPT) to assess risk of bias in randomized controlled trials of medical interventions: protocol for a pilot study of interrater agreement with human reviewers

Abstract

Abstract Background Risk of bias (RoB) assessment is an essential part of systematic reviews of treatment effect. RoB assessment requires reviewers to read and understand each eligible trial and depends on a sound understanding of trial methods and RoB tools. RoB assessment is a highly skilled task, subject to human error, and can be time-consuming and expensive. Machine learning-based tools have been developed to streamline the RoB process using relatively simple models trained on limited corpuses. ChatGPT is a conversational agent based on a large language model (LLM) that was trained on an internet-scale corpus and demonstrates human-like abilities in many areas, including healthcare. LLMs might be able to perform or support systematic reviewing tasks such as assessing RoB, which may reduce review costs, time to completion, and error. Objectives To assess interrater agreement in overall (cf. domain-level) RoB assessment between human reviewers and ChatGPT, in randomized controlled trials of interventions within medicine. Methods We will randomly select 100 individually- or cluster-randomized, parallel, two-arm trials of medical interventions from recent Cochrane systematic reviews that have been assessed using the RoB1 or RoB2 family of tools. We will exclude reviews and trials that were performed under emergency conditions (e.g., COVID-19) that may not exhibit typical RoB, as well as public health and welfare interventions. We will use 25 of the trials and human RoB assessments to engineer a ChatGPT prompt for assessing overall RoB, based on trial methods text. We will obtain ChatGPT assessments of RoB for the remaining 75 trials and human assessments. We will then estimate interrater agreement. Results The primary outcome for this study is overall human-ChatGPT interrater agreement. We will report observed agreement with an exact 95% confidence interval, expected agreement under random assessment, Cochrane’s 𝜅, and a p-value testing the null hypothesis of no difference in agreement. Several other analyses are also planned. Conclusions This study is likely to provide the first evidence on interrater agreement between human RoB assessments and those provided by LLMs and will inform subsequent research in this area.