Measuring healthcare professionals' benefit-risk perceptions of medical AI: Scale development and validation.
TL;DR
The AI-BRPS has demonstrated satisfactory initial psychometric properties and shows promise for assessing healthcare professionals' benefit expectations and risk concerns regarding medical AI in similar settings, with potential applications in pre-implementation assessment and future evaluation of intervention outcomes.
Abstract
Objective
To develop and validate the Artificial Intelligence Benefit-Risk Perception Scale (AI-BRPS) for healthcare professionals and explore the associative structure and central nodes of perceived AI benefits and risks.
Methods
This methodological study generated an initial item pool through literature review, qualitative interviews, expert consultation, and pilot survey. From December 2025 to January 2026, 794 healthcare professionals were recruited using convenience sampling from 36 tertiary hospitals in China to evaluate the reliability and validity of the preliminary scale. Network analysis was conducted at the dimension and item levels as a complementary approach to explore associative patterns and identify central and bridge nodes.
Results
The final AI-BRPS comprised the 13-item Perceived AI Benefit Subscale (PABS) and the 16-item Perceived AI Risk Subscale (PARS), covering eight dimensions rated on a five-point Likert scale. Exploratory factor analysis (EFA) identified four-factor structures for both subscales, explaining 67.531 % and 61.657 % of the variance, respectively. Confirmatory factor analysis (CFA) showed good model fit. The Cronbach's α coefficients of the PABS and PARS were 0.924 and 0.934, and the McDonald's ω coefficients were 0.925 and 0.934, indicating satisfactory reliability. Network analysis indicated that social benefits and professional risks occupied relatively influential positions, while cost risks occupied a potential bridging position between benefit and risk perceptions in the observed associative network.
Conclusion
The AI-BRPS has demonstrated satisfactory initial psychometric properties and shows promise for assessing healthcare professionals' benefit expectations and risk concerns regarding medical AI in similar settings, with potential applications in pre-implementation assessment and future evaluation of intervention outcomes. Further evaluation of test-retest reliability, measurement invariance, and cross-cultural applicability is warranted. The findings may also highlight clinically relevant considerations for AI implementation, including clinical reliability, clear responsibility boundaries, decision-making authority, and manageable learning and use costs.