Benchmarking Human-AI Collaboration for Common Evidence Appraisal Tools

Abstract Background It is unknown whether large language models (LLMs) may facilitate time- and resource-intensive text-related processes in evidence appraisal. Objectives To quantify the agreement of LLMs with human consensus in appraisal of scientific reporting (PRISMA) and methodological rigor (AMSTAR) of systematic reviews and design of clinical trials (PRECIS-2). To identify areas, where human-AI collaboration would outperform the traditional consensus process of human raters in efficiency. Design Five LLMs (Claude-3-Opus, Claude-2, GPT-4, GPT-3.5, Mixtral-8x22B) assessed 112 systematic reviews applying the PRISMA and AMSTAR criteria, and 56 randomized controlled trials applying PRECIS-2. We quantified agreement between human consensus and (1) individual human raters; (2) individual LLMs; (3) combined LLMs approach; (4) human-AI collaboration. Ratings were marked as deferred (undecided) in case of inconsistency between combined LLMs or between the human rater and the LLM. Results Individual human rater accuracy was 89% for PRISMA and AMSTAR, and 75% for PRECIS-2. Individual LLM accuracy was ranging from 63% (GPT-3.5) to 70% (Claude-3-Opus) for PRISMA, 53% (GPT-3.5) to 74% (Claude-3-Opus) for AMSTAR, and 38% (GPT-4) to 55% (GPT-3.5) for PRECIS-2. Combined LLM ratings led to accuracies of 75-88% for PRISMA (4-74% deferred), 74-89% for AMSTAR (6-84% deferred), and 64-79% for PRECIS-2 (18-88% deferred). Human-AI collaboration resulted in the best accuracies from 89-96% for PRISMA (25/35% deferred), 91-95% for AMSTAR (27/30% deferred), and 80-86% for PRECIS-2 (76/71% deferred). Conclusions Current LLMs alone appraised evidence worse than humans. Human-AI collaboration may reduce workload for the second human rater for the assessment of reporting (PRISMA) and methodological rigor (AMSTAR) but not for complex tasks such as PRECIS-2.