Human Versus Artificial Intelligence: Comparing Cochrane Authors' and ChatGPT's Risk of Bias Assessments

Systematic reviews and meta-analyses synthesize randomized trial data to guide clinical decisions but require significant time and resources. Artificial intelligence (AI) offers a promising solution to streamline evidence synthesis, aiding study selection, data extraction, and risk of bias assessment. This study aims to evaluate the performance of ChatGPT-4o in assessing the risk of bias in randomised controlled trials (RCTs) using the Risk of Bias 2 (RoB 2) tool, comparing its results with those conducted by human reviewers in Cochrane Reviews.

A sample of Cochrane Reviews utilizing the RoB 2 tool was identified through the Cochrane Database of Systematic Reviews (CDSR). Protocols, qualitative systematic reviews, and reviews employing alternative risk of bias assessment tools were excluded. The study utilized ChatGPT-4o to assess the risk of bias using a structured set of prompts corresponding to the RoB 2 domains. The agreement between ChatGPT-4o and consensus-based human reviewer assessments was evaluated using weighted kappa statistics. Additionally, accuracy, sensitivity, specificity, positive predictive value, and negative predictive value were calculated. All analyses were performed using R Studio (version 4.3.0).

A total of 42 Cochrane Reviews were screened, yielding a final sample of eight eligible reviews comprising 84 RCTs. The primary outcome of each included review was selected for risk of bias assessment. ChatGPT-4o demonstrated moderate agreement with human reviewers for the overall risk of bias judgments (weighted kappa = 0.51, 95% CI: 0.36-0.66). Agreement varied across domains, ranging from fair (κ = 0.20 for selection of the reported results) to moderate (κ = 0.59 for measurement of outcomes). ChatGPT-4o exhibited a sensitivity of 53% for identifying high-risk studies and a specificity of 99% for classifying low-risk studies.

This study shows that ChatGPT-4o can perform risk of bias assessments using RoB 2 with fair to moderate agreement with human reviewers. While AI-assisted risk of bias assessment remains imperfect, advancements in prompt engineering and model refinement may enhance performance. Future research should explore standardised prompts and investigate interrater reliability among human reviewers to provide a more robust comparison.