Clinician responses to AI clinical decision support: survey data on barriers, trust and adoption intention
Description
Description This dataset contains the analysis file for a cross-sectional survey of 630 licensed clinicians directly involved in patient care, examining how perceived barriers to artificial-intelligence clinical decision support (AI-CDSS) relate to trust, resistance intention and intention to use, and how those relationships shift when the barrier is relevant to professional identity. Respondents were recruited purposively through professional gatekeepers and networks in 19 countries, with Malaysia (n = 317) and China (n = 193) predominating. The questionnaire was administered online in English and Chinese; the codebook reports the English form as fielded. The file holds only the variables entering the reported analysis: 630 rows and 47 columns. It contains an arbitrary respondent identifier; the indicators of eight measured constructs, namely attitude (3 items) and perceived clinical value (4), the lower-order dimensions of adoption favourability, clinical practice inertia (3), workflow barrier (4), perceived professional image and autonomy threat (4), trust (5), resistance intention (4), intention to use (4) and supervisory support (4); three marker-variable items for the common-method assessment; four controls entered in the intention-to-use equation alongside supervisory support, namely prior use of clinical AI tools, whether the workplace runs an AI-CDSS and two country indicators with Malaysia as the reference category; three demographic variables used to describe the sample and not entered in the model, namely age, gender and years of clinical experience; and a recruitment-group code used only for a sensitivity check. All focal items use seven-point scales; the marker items use five points. There are no missing values. The model was estimated by partial least squares structural equation modelling in SmartPLS 4.1.1. Adoption favourability is a Type I reflective-reflective higher-order construct estimated by the disjoint two-stage approach, so attitude and perceived clinical value enter the first stage as separate constructs and their latent variable scores serve as its indicators in the second stage. Inference uses 10,000 bootstrap subsamples with percentile intervals. No personal identifiers are included. The respondent identifier is arbitrary and cannot be traced to an individual; its values run from 1 to 673 with gaps because it was carried over from the full survey file before cleaning. The study was approved by the Universiti Malaya Research Ethics Committee (reference UM.TNC(P&I)/UMREC/2/6607, letter dated 26 June 2026) and all participants gave informed consent and took part anonymously. The data support the article "Which Barriers Matter When Professional Identity Is at Stake? Frontline Clinicians' Responses to AI Decision Support", submitted to the Journal of Service Management.
Files
Steps to reproduce
How the data were produced Forty-three enumerators in thirteen groups distributed an online questionnaire through professional networks. Six mandatory screening questions required a current licence to practise, clinical work, direct involvement in patient care, contribution to clinical decisions, completed supervised training, and English or Chinese; a negative answer ended the survey. This produced 673 complete questionnaires. Respondents then read a definition of an AI-CDSS as a tool that supports rather than replaces clinician judgement, then answered a comprehension check. To preserve the pre-adoption framing, 43 respondents were excluded because an AI-CDSS was already in routine use where they worked and they used AI tools occasionally or often, leaving the 630 cases deposited here. No further exclusions are needed. How to reproduce the reported analysis Import AI-CDSS-M3_analysis-data_n630.csv into SmartPLS 4.1.1. There are no missing values. Specify eight reflectively measured constructs from their indicator blocks: ATT1-3, PCV1-4, CPI1-3, WB1-4, PIAT1-4, TR1-5, RI1-4, ITU1-4 and SS1-4. Estimate adoption favourability as a Type I reflective-reflective higher-order construct by the disjoint two-stage approach. In stage one, ATT and PCV are separate constructs; their latent variable scores become its two reflective indicators in stage two. All other constructs keep their original indicators. In stage two, regress adoption favourability, trust and resistance intention on CPI, WB, PIAT and the interaction terms CPI x PIAT and WB x PIAT, formed from construct scores. Regress intention to use on the three appraisals. Enter the controls in the intention-to-use equation only: CTRL_PRIORUSE, CTRL_WORKPLACE_AI, CTRL_CHINA and CTRL_OTHER, with Malaysia as the reference category, plus supervisory support. CTRL_AGE, CTRL_MALE and CTRL_YEARS describe the sample and are not in the estimated model. Run the PLS algorithm with the path weighting scheme, then bootstrap with 10,000 subsamples. Decisions follow percentile intervals; the hypotheses are directional, so use one-tailed tests at 5 per cent with 90 per cent intervals. Compute each specific indirect association as the product of the barrier-to-appraisal and appraisal-to-intention coefficients. Test the six hypotheses with the index of moderated mediation, and report conditional indirect associations at the mean and at plus and minus one standard deviation of PIAT. Upsilon is the effect size for the indirect associations. For the common method assessment, form the marker from MV1-3, enter it as an antecedent of every endogenous construct, and compare coefficients and explained variance with and without it. For the recruitment sensitivity check, run a one-way analysis of variance of mean ITU across the thirteen groups in GRP_ENUM. For predictive assessment, run CVPAT with ten folds and ten repetitions against the indicator-average and linear-model benchmarks.
Institutions
- University of MalayaKuala Lumpur, Kuala Lumpur