Evaluating the Accuracy of Large Language Models in Risk-of-Bias Assessment Using Version 2 of the Cochrane Risk-of-Bias Tool for Randomized Trials: Exploratory Feasibility Study

Abstract Background Large language models (LLMs) have the potential to improve the efficiency of evidence synthesis, but their reliability in performing complex tasks such as risk-of-bias (ROB) assessment in randomized controlled trials (RCTs) remains unclear. Objective This study aimed to evaluate whether LLMs can reliably assess ROB in RCTs using version 2 of the Cochrane ROB tool for randomized trials (ROB 2). Methods This study was conducted between December 28, 2024, and February 28, 2025, in adherence to American Association for Public Opinion Research reporting guidelines. Twenty-nine RCTs were selected from published Cochrane systematic reviews across diverse medical fields. We developed a structured prompt engineering framework that transformed ROB 2 decision trees into logical rules for the LLM. Each RCT was independently evaluated twice by ChatGPT, with Cochrane review authors’ assessments serving as the reference standard for comparison. The main outcomes were the accuracy and consistency of ROB 2 assessments at both the domain and trial levels, evaluated using accuracy, sensitivity, specificity, and F 1 -score. Consistency between the repeated assessments was quantified using the Cohen κ and prevalence-adjusted, bias-adjusted κ. Results The LLM demonstrated a moderate aggregate domain accuracy of 73.1% (95% CI 64.7%-81.5%) in the first assessment and 75.9% (95% CI 66.3%-85.4%) in the second assessment. Domain-averaged sensitivity decreased from 61.4% (95% CI 48.1%‐74.7%) to 53.4% (95% CI 41.7%‐65.0%), whereas domain-averaged specificity increased from 75.8% (95% CI 65.3%-86.3%) to 81.1% (95%CI 67.1%-95%), indicating a conservative tendency in identifying a high ROB. Domain-level accuracy ranged from 62.1% to 87.9%, with the lowest accuracy observed in domain 1 and the lowest F 1 -scores observed in domain 2. Consistency between repeated assessments was high, with a mean agreement of 89.0% (SD 7.5%), and Cohen κ values were 0.86, 0.39, 0.56, 0.84, and 0.85 in domains 1 to 5, respectively. Conclusions In this exploratory study, ChatGPT demonstrated moderate accuracy and high consistency in assessing ROB in RCTs using the ROB 2. However, its reliability diminished in complex scenarios requiring interpretation of implicit narratives or behavioral nuance. These findings suggest that LLMs may support methodological evaluations in systematic reviews by acting as automated screeners to reduce reviewer burden, but current implementation still requires expert oversight, particularly for trials involving subjective outcomes or nonstandard reporting.