JMIR Publications

ABSTRACT

Background:

Artificial intelligence early warning systems are increasingly embedded in inpatient nursing workflows, yet the psychometric performance of generic trust-in-automation scales in this setting remains uncertain.

Objective:

This study examined whether the classic Jian automation trust scale could function as a meaningful measurement tool in nurses using workflow-embedded artificial intelligence early warning systems under mandatory deployment conditions.

Methods:

We conducted a multicenter cross-sectional survey among 712 nurses from 8 tertiary hospitals in Shaanxi Province, China. Reporting was guided primarily by CROSS, with supplementary reference to STROBE and DECIDE-AI; TRIPOD+AI was used only to clarify reporting boundaries. Trust_Score and Distrust_Score were evaluated using internal consistency indices, item distributions, and multigroup confirmatory factor analysis diagnostics. Latent profile analysis based on 4 interaction indicators was followed by BCH 3-step distal comparisons with 1000 bootstrap draws. Separate ordinary least squares models with HC3 robust standard errors were then fitted using a single-item global trust rating as the exploratory outcome, and 10-fold cross-validated R² was used to evaluate out-of-sample performance.

Results:

Trust (α=-0.100; ω=0.120) and Distrust (α=0.082; ω=0.108) showed extremely poor internal consistency, indicating that the instrument did not form an interpretable latent construct in this setting. Trust items were heavily compressed at the high end of the response scale. Multigroup confirmatory factor analysis for the AI-training grouping did not support measurement invariance (configural CFI=0.682; metric CFI=0.285). Latent profile analysis identified 4 interaction profiles, but BCH comparisons showed no between-profile differences in Trust_Score or Distrust_Score after Holm correction. Regression models had very limited explanatory value (adjusted R²=0.011 and 0.016), and all cross-validated R² values were negative (-0.065 to -0.019), indicating worse performance than a naive mean-prediction baseline.

Conclusions:

A generic trust-in-automation scale showed marked measurement misfit in a strongly workflow-embedded clinical artificial intelligence early warning system context. The present cross-sectional self-report data do not support strong claims about institutional decoupling or ceremonial adoption. The inverse association between better perceived experience and lower single-item global trust in high-risk departments should be treated as an exploratory signal requiring confirmation with system logs and system-level performance data.