AI in Qualitative Health Research Appraisal: Comparative Study.

Abstract

Background: Qualitative research appraisal is crucial for ensuring credible findings but faces challenges due to human variability. Artificial intelligence (AI) models have the potential to enhance the efficiency and consistency of qualitative research assessments.

Objective: This study aims to evaluate the performance of 5 AI models (GPT-3.5, Claude 3.5, Sonar Huge, GPT-4, and Claude 3 Opus) in assessing the quality of qualitative research using 3 standardized tools: Critical Appraisal Skills Programme (CASP), Joanna Briggs Institute (JBI) checklist, and Evaluative Tools for Qualitative Studies (ETQS).

Methods: AI-generated assessments of 3 peer-reviewed qualitative papers in health and physical activity-related research were analyzed. The study examined systematic affirmation bias, interrater reliability, and tool-dependent disagreements across the AI models. Sensitivity analysis was conducted to evaluate the impact of excluding specific models on agreement levels.

Results: Results revealed a systematic affirmation bias across all AI models, with "Yes" rates ranging from 75.9% (145/191; Claude 3 Opus) to 85.4% (164/192; Claude 3.5). GPT-4 diverged significantly, showing lower agreement ("Yes": 115/192, 59.9%) and higher uncertainty ("Cannot tell": 69/192, 35.9%). Proprietary models (GPT-3.5 and Claude 3.5) demonstrated near-perfect alignment (Cramer V=0.891; P<.001), while open-source models showed greater variability. Interrater reliability varied by assessment tool, with CASP achieving the highest baseline consensus (Krippendorff α=0.653), followed by JBI (α=0.477), and ETQS scoring lowest (α=0.376). Sensitivity analysis revealed that excluding GPT-4 increased CASP agreement by 20% (α=0.784), while removing Sonar Huge improved JBI agreement by 18% (α=0.561). ETQS showed marginal improvements when excluding GPT-4 or Claude 3 Opus (+9%, α=0.409). Tool-dependent disagreements were evident, particularly in ETQS criteria, highlighting AI's current limitations in contextual interpretation.

Conclusions: The findings demonstrate that AI models exhibit both promise and limitations as evaluators of qualitative research quality. While they enhance efficiency, AI models struggle with reaching consensus in areas requiring nuanced interpretation, particularly for contextual criteria. The study underscores the importance of hybrid frameworks that integrate AI scalability with human oversight, especially for contextual judgment. Future research should prioritize developing AI training protocols that emphasize qualitative epistemology, benchmarking AI performance against expert panels to validate accuracy thresholds, and establishing ethical guidelines for disclosing AI's role in systematic reviews. As qualitative methodologies evolve alongside AI capabilities, the path forward lies in collaborative human-AI workflows that leverage AI's efficiency while preserving human expertise for interpretive tasks.