ChatGPT for assessing risk of bias of randomized trials using the RoB 2.0 tool: A methods study

Abstract

BackgroundThe assessment of risk of bias is a critical component of systematic review methods. Assessing risk of bias, however, can be time- and resource-intensive. AI-based solutions may increase efficiency and reduce burden.ObjectiveTo evaluate the reliability of ChatGPT for performing risk of bias assessments of randomized trials.MethodsWe sampled recently published Cochrane systematic reviews of medical interventions (up to October 2023) that included randomized controlled trials and assessed risk of bias using the Cochrane-endorsed revised risk of bias tool for randomized trials (RoB 2.0). From each eligible review, we collected data on the risk of bias assessments for the first three reported outcomes. Using ChatGPT-4, we assessed the risk of bias for the same outcomes using three different prompts: a minimal prompt including limited instructions, a maximal prompt with extensive instructions, and an optimized prompt that was designed to yield the best risk of bias judgments. The agreement between ChatGPT’s assessments and those of the systematic reviewers was quantified using weighted kappa statistics.ResultsWe included 34 systematic reviews with 157 unique trials. We found the agreement between ChatGPT and systematic review authors for assessment of overall risk of bias to be 0.16 (95% CI: 0.01 to 0.3) for the maximal ChatGPT prompt, 0.17 (95% CI: 0.02 to 0.32) for the optimized prompt, and 0.11 (95% CI: -0.04 to 0.27) for the minimal prompt. For the optimized prompt, agreement ranged between 0.11 (95% CI: -0.11 to 0.33) to 0.29 (95% CI: 0.14 to 0.44) across risk of bias domains, with the lowest agreement for the deviations from the intended intervention domain and the highest agreement for the missing outcome data domain.ConclusionOur results suggest that ChatGPT and systematic reviewers only have “slight” to “fair” agreement in risk of bias judgments for randomized trials. ChatGPT is currently unable to reliably assess risk of bias of randomized trials. We recommend systematic reviewers avoid using ChatGPT to perform risk of bias assessments.

Exploring the potential of Claude 2 for risk of bias assessment: Using a large language model to assess randomized controlled trials with RoB 2

“…Just recently, other authors have used LLM to conduct RoB assessment, with mixed results. Pitre et al (34) found comparably low agreement between ChatGPT-4 and Cochrane authors when assessing RoB of 157 RCTs from 34 Cochrane Reviews using RoB 2 (Cohen's κ of 0.16 for the overall assessment). Testing the use of ChatGPT (GPT-4) for RoB assessment of non-randomized studies of intervention using ROBINS-I (65), Hasan et al (35) also obtained only slight agreement (Cohen's κ of 0.13 for the overall assessment).…”

“…Currently, there are very limited methods to support RoB assessment using ML (5). However, also using ChatGPT alone for RoB assessments seems not recommendable, neither for RCTs (33,34) nor for non-randomized studies of interventions (35), due to limited agreement in RoB judgements between ChatGPT and humans.…”

GPT's Performance in Identifying Outcome Changes on ClinicalTrials.gov

“…33 Researchers are exploring the capabilities of GPT-4 in various scholarly tasks, 34 including reviewing scientific papers, [35][36][37] implementing edits based on reviewer comments, 38 and systematic review tasks, such as article screening, data extraction, 39,40 and assessing the risk of bias in included studies. 41 While GPT-4 has shown satisfactory performance in some of these tasks, the results in others have been less than optimal.…”

Potential Roles of Large Language Models in the Production of Systematic Reviews and Meta-Analyses

“…with an estimated review time of 10-15 minutes per trial. However, automated tools such as RobotReviewer can streamline the extraction and evaluation process in batches [51][52][53], improving efficiency-though manual verification is still necessary. Additionally, chatbots based on LLMs can aid in risk of bias assessment (see Figure S8), and studies indicate that their accuracy is comparable to human evaluations [23].…”

“…For randomized controlled trials, tools such as Risk of Bias (RoB) [ 62 ] or its updated version RoB 2 [ 63 ] are typically used, with an estimated review time of 10-15 minutes per trial. However, automated tools such as RobotReviewer can streamline the extraction and evaluation process in batches [ 51 - 53 ], thereby improving efficiency, although manual verification is still necessary. Additionally, chatbots based on LLMs can aid in risk of bias assessment (see Multimedia Appendix 8 ), and their accuracy appears to be comparable to that of human evaluations [ 23 ].…”