JMIR Publications

ABSTRACT

Background:

The integration of artificial intelligence (AI) in health care depends largely on acceptance of users. While prior research suggests that multiple factors drive AI adoption, particularly Effort Expectancy and Performance Expectancy, the literature remains inconclusive regarding the primary barriers, especially the relative roles of Perceived Risk and Trust. Furthermore, literature often rests on an unverified assumption that digital-native healthcare students are cognitively more receptive to AI than experienced physicians.

Objective:

This study aimed to quantitatively synthesize the structural mechanisms driving clinical AI adoption using an extended unified theory of acceptance and use of technology (UTAUT) model, and to evaluate structural invariance between healthcare students and physicians.

Methods:

Following PRISMA guidelines, a systematic search identified eligible cross-sectional studies. Moreover, a multivariate Meta-Analytic Structural Equation Modeling (MASEM) approach was executed with 2 stages. In Stage 1, a pooled correlation matrix was estimated using a random-effects restricted maximum likelihood (REML) model. In Stage 2, structural pathways and mediation effects were estimated using 95% likelihood-based confidence intervals (LBCI) to address multivariate non-normality. Additionally, One-Stage MASEM (OSMASEM) was utilized to test the generational hypothesis via multigroup global structural invariance.

Results:

Synthesizing data from 55 independent samples comprising 32,829 participants, Stage 1 analyses demonstrated substantial between-study heterogeneity across the pooled correlations (Q = 16842.50, P < .001; I² = 97.4%-99.2%). The Stage 2 structural model demonstrated excellent fit to the data (χ²₁ = 0.215, P = .643; CFI = 1.000; RMSEA = 0.000). Performance Expectancy (β = 0.312; 95% LBCI [0.185, 0.425]) and Effort Expectancy (β = 0.184; 95% LBCI [0.035, 0.312]) were significant positive predictors of Behavioral Intention. In contrast, Perceived Risk was not significantly associated with Behavioral Intention after accounting for the remaining structural pathways (β = –0.075; 95% LBCI [–0.260, 0.115]; P = .438). Trust significantly mediated the effects of both Performance Expectancy (βindirect = 0.100, 95% LBCI 0.042-0.182) and Effort Expectancy (βindirect = 0.119, 95% LBCI 0.048-0.218) on Behavioral Intention. The model explained 41.2% of the variance in Trust and 49.0% of the variance in Behavioral Intention. OSMASEM likelihood-ratio testing indicated no significant differences in structural pathways between healthcare students and practicing healthcare professionals (P = .777), supporting structural invariance across user groups.

Conclusions:

Perceived Risk is not an insurmountable barrier to healthcare AI utilization; its association with adoption intention appears to be attenuated when institutional and algorithmic trust are established. Furthermore, the psychological prerequisites for AI acceptance appear to be universal, thereby challenging the prevailing notion of a generational divide in medical informatics. Accordingly, future healthcare curricula and health policies should pivot from age-stratified technical training toward unified, explainable AI trust-calibration frameworks. Clinical Trial: PROSPERO CRD420261299478; https://www.crd.york.ac.uk/PROSPERO/view/CRD420261299478