The integration of artificial intelligence (AI) is instrumental in driving innovation and enhancing operational efficiencies across diverse sectors, ranging from finance and manufacturing to transportation and entertainment [1,2]. In health care, AI has emerged as a game changer, offering unprecedented opportunities to transform the delivery of medical services and improve patient outcomes [2,3]. The integration of AI in health care is expected to expand into various domains, including drug discovery, personalized medicine, and predictive analytics [4,5]. An important aspect of AI integration into health care is the readiness of health care practitioners to accept and use AI technology in their practice [6]. Various studies have attempted to understand the factors affecting the integration of AI into health care. While some studies focused on the perspectives of students [7-11], the focus of the current literature review is on the perspectives of health care professionals as the main players in the integration of AI in health care.

The United Arab Emirates presents an interesting case study in AI adoption across health care contexts. AI adoption in the United Arab Emirates is a topic of growing interest across various sectors. Research has explored AI’s impact on different fields in the United Arab Emirates, such as air quality monitoring [31], journalism practices [32], e-innovative projects in the public sector [33], and the legal framework surrounding AI decisions and explanations [34]. These studies highlight the importance of understanding AI mechanisms, addressing biases, and establishing ethical guidelines to ensure responsible AI use.

In the dental domain, a cross-sectional exploratory study involving 134 participants (including 72 undergraduate dental students, 19 academic staff, and 44 practicing dentists) investigated perceptions, knowledge, and organizational readiness for AI implementation [35]. Using structured questionnaires, the study found that the majority of participants had medium to high knowledge of AI, particularly among female students. Significant associations were observed between demographics and AI perceptions, underscoring the role of background characteristics in shaping adoption readiness. In the radiology domain, a cross-sectional study of 153 radiology professionals (radiologists and radiographers) in UAE hospitals explored knowledge, perceptions, readiness, and challenges regarding AI integration [36]. The findings revealed a lack of AI awareness, particularly among radiographers. While there was some interest in AI for postprocessing and dose management, most participants were unaware of AI’s broader potential in radiology. The study concluded that AI literacy remains low, with wide disagreement about AI’s importance in clinical radiology.

Using partial least squares structural equation modeling (SEM), a study with 53 participants (from IT and health departments) identified the critical success factors for AI project implementation in the UAE public health sector [37,38]. These included managerial, operational, organizational, IT infrastructure, and strategic factors. All were found to be significantly associated with both the perceived usefulness and ease of use of AI technologies. The study emphasizes that structural and institutional support are key to ensuring AI adoption success, making these variables crucial candidates for inclusion in SEM. Another large-scale study of 553 nurses across the United Arab Emirates assessed knowledge, perception, and organizational readiness toward AI in nursing [39]. Using descriptive statistics and chi-square tests, the study revealed that perceptions of AI varied significantly by age, educational background, and work experience. Notably, older and more experienced nurses reported feeling more threatened by AI, suggesting an age-related barrier that could influence trust and perceived risk in SEM models. Additionally, an analytic hierarchy process–based study conducted with 27 health care executives in the United Arab Emirates evaluated factors affecting AI adoption, including concerns over accuracy, data security, and privacy [40]. While this study is not open access, it contributes to the understanding that managerial perceptions often prioritize ethical and infrastructural safeguards, which are essential for implementation readiness.

This study aims to develop and validate a comprehensive model of factors associated with health care professionals’ acceptance and readiness to adopt AI in clinical practice within the UAE health care system. By spanning diverse professional groups, using advanced analytical techniques, and integrating multiple theoretical frameworks, this study aims to provide actionable insights for health care leaders, policymakers, technology developers, and educators seeking to facilitate successful AI implementation in health care settings.

This study makes a significant contribution to the literature by specifically examining AI adoption within the UAE health care context. The United Arab Emirates presents a unique research setting characterized by rapid technological advancement, substantial investment in health care infrastructure, and a highly diverse health care workforce representing various cultural and educational backgrounds. Understanding AI adoption patterns in this context has particular relevance given the “We the UAE 2031” vision [41] and the National Strategy for Artificial Intelligence 2031 [42] and its emphasis on developing world-class health care infrastructure supported by advanced technologies. The findings from this research will directly inform implementation strategies aligned with these national priorities.

Despite rapid national AI initiatives, little is known about front-line practitioners’ acceptance and readiness in the United Arab Emirates. This study addresses that gap using a model based on the Technology Acceptance [43], the Unified Theory of Acceptance and Use of Technology (UTAUT) [44], and the Theory of Trust and Acceptance of Artificial Intelligence Technology (TrAAIT) [45] models.

This study adhered to the Checklist for Reporting Results of Internet E-Surveys (CHERRIES) for web-based survey research [46] (Checklist 1). A cross-sectional online survey was designed based on the Technology Acceptance [43], UTAUT [44], and TrAAIT [45] models. Ethical approval was obtained, and informed consent was secured electronically. The instrument was pretested and refined before full deployment, using validated constructs. Participants signed an electronic informed consent through a tick box, and the survey was deployed on a web-based platform accessible via desktop and mobile. Since the survey was shared through the professional network of the authors, the response rate could not be precisely calculated. To ensure data integrity, duplicate entries were prevented using cookies and IP filtering, and the survey form was configured to accept only fully completed responses. For analysis, confirmatory factor analysis (CFA) and SEM were performed using the SAS (SAS Institute) software [47]. The study acknowledges limitations related to potential self-selection and response bias. The methodology used in this study consists of sequential steps, as demonstrated in Figure 1.

This is an initial step before starting the analysis, as it tests the reliability and validity of the collected data. The internal consistency of each dimension was measured using Cronbach α and composite reliability (CR). In addition, the sample size adequacy for SEM was conducted using the Kaiser-Meyer-Olkin test [67]. A Kaiser-Meyer-Olkin value between 0.7 and 1 indicates that the sample is adequate. Any variables that reduced the value of α will be removed from subsequent analysis. The average variance extracted (AVE) quantifies the amount of variance in indicators that is explained by the underlying latent construct. A higher AVE value indicates that the indicators are strongly related to the construct and explain a substantial portion of their variance, suggesting good convergent validity. A value of 0.5 or higher is deemed an acceptable level of AVE. Discriminant validity was evaluated using the Heterotrait-Monotrait criterion, using <0.85 as a conservative cutoff or <0.90 as a more permissive cutoff. Because this study used a cross-sectional, self-reported questionnaire to measure both predictors and outcomes, we evaluated the potential for common method bias. We applied the Harman single-factor test and used the conventional decision rule that common method bias is unlikely to be a dominant concern when the first factor explains <50% of the total variance.

A CFA was conducted to validate the measurement model. This step confirmed that the observed survey items were statistically aligned with their corresponding latent constructs, including Perception, Trust, Risk, Benefit, Acceptance, and Readiness. The CFA demonstrated strong factor loadings and good model fit, supporting construct validity. Following this, SEM was applied to test the relationships among these validated constructs. The model evaluated how Trust, Benefit, Risk, and Perception are linked to Acceptance, and how Acceptance, in turn, is associated with Readiness to adopt AI technologies among health care professionals.

This study received ethics approvals from the American University of Sharjah Institutional Review Board (protocol number 24‐077) and from Dubai Health/MBRU (MBRU IRB-2024‐491). Participants reviewed an online information sheet and provided electronic informed consent before accessing the survey. No compensation or incentives were offered. No direct identifiers were collected; analytic data were deidentified before analysis and stored on access-controlled systems. No individual is identifiable in this paper.

This section presents the empirical findings from the analysis of survey responses collected from health care practitioners regarding their attitudes and readiness toward AI integration.

The study included a total of 182 health care practitioners from different specialties and practice settings across the United Arab Emirates. The demographics of the studied sample are shown in Table 3. These demographics reflect a predominantly female and relatively young group with diverse professional roles and varying levels of experience. Purposive, network-based recruitment may oversample digitally engaged professionals and certain Emirates. We report the full profile and restrict inference to the cohort. The majority of respondents lack AI training, and a substantial proportion are uncertain about recognizing AI in software, highlighting areas for potential educational interventions and professional development. Nurses comprised around 39% (71/182) of respondents, potentially skewing attitudes toward nursing workflows.

After the initial demographic inquiries, participants engaged in questions aimed at assessing their knowledge about AI (Table 4). Despite 49% (roughly half; 89/182) of participants indicating a high level of confidence (agree n=58/182 and strongly agree 31/182) in their AI knowledge, the objective measurement questions unveiled a contrasting result. While approximately half (n=99/182) of the respondents answered the first question correctly, out of 182, only 60 (32%), 31 (17%), and 20 (11%) answered the second, third, and fourth questions correctly, respectively.

We assessed internal consistency, composite reliability, and convergent validity via AVE. Thresholds follow standard psychometrics or SEM texts. For the sample in this paper, the Kaiser-Meyer-Olkin value was 0.883, which implies that the sample was adequate. It is worth mentioning that the variable Benefit 1 reduced the value of α and was removed from the analysis. Cronbach α and CR values for most factors are approximately between 0.67 and 0.92, which are consistent with the acceptable level of 0.7. Although the Perception construct shows a marginal Cronbach α value of 0.67, it has been retained because all standardized indicator loadings exceed the acceptable level of 0.5. However, the Knowledge factor showed low reliability with α and CR equal to 0.14 and 0.42, respectively. Therefore, the Knowledge factor was excluded from the subsequent analysis.

Although perception (AVE=0.41) and risk (AVE=0.42) fall below the 0.50 benchmark, both factors show adequate internal consistency and composite reliability (Perception: α=.67, CR=0.67; Risk: α=.81, CR=0.81) with most loadings≥0.50. According to Fornell and Larcker [68], AVE is a conservative index of convergent validity, and a construct with AVE<0.5 may be retained when it has an acceptable level of reliability. Accordingly, these two factors were kept in the final model. The reliability and validity measures for the collected data are demonstrated in Table 5 (with the CFA results).

Discriminant validity was assessed using the heterotrait-monotrait ratio (HTMT). All interconstruct HTMT values were below the conservative threshold of 0.85, supporting discriminant validity across the measurement model. HTMT matrix is reported in Table 6. Overall, the HTMT results suggest that the latent variables represent distinct constructs rather than redundant measurements of a single underlying factor. The Knowledge construct is not included in the table, as it is excluded from subsequent analysis (see below).

Harman single-factor test indicated that the first unrotated factor had an eigenvalue of 10.9480 and accounted for 36.49% of the total variance, which is below the 50% benchmark. This suggests that a single common method factor is unlikely to be the primary driver of the observed associations among constructs in this dataset.

In this step, CFA was conducted on the covariance matrix in order to validate the measurement model. The CFA results are shown in Table 5. All indicators were significant (P <.05) with acceptable loading sizes (standardized parameter estimates). Across constructs, internal consistency ranged from excellent to borderline, with the knowledge construct failing. Trust, Acceptance, Readiness, and Benefit demonstrated strong reliability. By contrast, Knowledge exhibited inadequate reliability and was excluded from the SEM. The goodness-of-fit indices (GFIs) for the CFA model are as follows: standardized root-mean-square residual (SRMR)=0.0680, root-mean-square error of approximation (RMSEA)=0.0913, GFI=0.8022, adjusted goodness-of-fit index (AGFI)=0.7629, and Bentler comparative fit index (CFI)=0.9058. The SRMR value falls within the acceptable threshold of less than 0.08. The RMSEA value exceeds the acceptable upper limit of 0.05. The CFI falls within the acceptable level of greater than 0.9. Both GFI and AGFI are less than the acceptable level of 0.85. Although the SRMR and CFI met commonly cited thresholds, the RMSEA and GFI fall outside conventional cutoffs for good model fit, indicating moderate to mediocre overall fit of the proposed model to the observed data [69].

The theoretical model that reflects the hypotheses 1 to 8 is depicted in Figure 2. PROC (procedure) covariance analysis of linear structural equations in SAS software [47] is used to fit the model to the data. The results of the structural model are depicted in Table 7. Based on the results, all Hypotheses 1 to 8 are significant. First, the health care practitioners’ perception and perceived benefit are positively associated with trust (Hypotheses 1 and 2). While perceived risk is negatively associated with trust (Hypothesis 3). Also, practitioners’ acceptance of AI is positively associated with trust, perception, and perceived risk (Hypotheses 4 to 6) but negatively associated with perceived risk (Hypothesis 7). The health care practitioners’ readiness to adopt AI in their practice is positively associated with acceptance (Hypothesis 8). The goodness-of-fit tests for the SEM are as follows: SRMR=0.0680, RMSEA=0.0913, GFI=0.8022, AGFI=0.7629, and Bentler CFI=0.9058. The SRMR value falls within the recommended level of less than 0.08. CFI is marginally above the acceptable level of 0.90. Although the GFI and AGFI values are below the recommended level of 0.85, and RMSEA is higher than the recommended level of 0.05, indicating moderate to mediocre overall fit of the proposed model to the observed data. The final SEM structure is shown in Figure 2.

This exploratory, cross-sectional survey examined how Trust, Perception, perceived benefits, and perceived risks relate to Acceptance and Readiness for AI among United Arab Emirates–based health care practitioners. Trust reflects foundational beliefs about an AI system’s functionality, reliability, and helpfulness, and in our instrument spans propensity to trust technologies, understandability, technical competence, reliability, helpfulness, personal attachment, user autonomy, faith, and institutional credibility; conceptually, it precedes more specific evaluations. Perception summarizes a practitioner’s overall stance toward AI in practice, operationalized as perceived professional impact and preparedness for AI. Perceived risk and perceived benefit index respondents’ appraisal of potential harms versus advantages from AI use, treated as distinct, opposing evaluations. Acceptance aggregates intention-related and attitudinal components (eg, perceived ease of use or usefulness, attitude, social influence, perceived behavioral control, performance and effort expectancy, enjoyment, perceived fee, technicality, perceived value, behavioral, and purchase intentions), distinguishing it from the upstream belief constructs. Readiness differs from acceptance by focusing on actionable preparedness, knowledge, skills, and attitudes to apply AI in clinical care and is captured through cognition, ability, vision, and ethics. Finally, we recorded knowledge as both subjective awareness and objective (tested) knowledge for context and calibration, although the knowledge factor was not retained in the SEM.

One of the most significant findings is the central role of trust in AI technology. Trust’s positive link to perceived benefit and perception, coupled with its negative link to perceived risk, suggests that building institutional and technical trust in AI systems could accelerate its adoption among UAE health care professionals. The findings of this study align with prior research by Gefen et al [57], Karaca et al [51], and Mcknight et al [56], which positioned trust as an essential precursor in technology acceptance frameworks. Perception and benefit were positively and significantly associated with acceptance of AI, in line with existing literature indicating that when health care professionals perceive AI to be useful, accurate, and aligned with ethical standards, their willingness to accept its implementation increases [45]. Furthermore, risk showed a significant negative relationship with acceptance, confirming that concerns related to privacy, safety, and misuse can hinder adoption.

Standardized path coefficients from trust to perception are stronger than the paths from trust to risk and benefit, suggesting that fine-grained risk-benefit campaigns alone will yield limited gains unless accompanied by measures that raise global attitudes and institutional trust. In addition, trust holds both a sizable direct link to acceptance and sizable indirect links through perception, benefit, or risk. This pattern of relationships is coherent with early adoption contexts. Limited hands-on exposure to production AI systems likely dampens risk-benefit appraisals and elevates reliance on assurance cues (eg, reliability and accountability), which are captured in Trust. Programs that raise AI literacy alone may be insufficient unless they also build institutional or technical trust (eg, governance, transparency, or audits).

Another key insight is the significant relationship between acceptance and readiness, indicating that attitudinal intention precedes practical preparedness [29]. In our measurement, Acceptance (TAM) reflects willingness and intention shaped by usefulness or ease, norms, and control, whereas readiness (Medical Artificial Intelligence Readiness Scale) indexes whether clinicians possess the cognition, ability, vision, and ethics to act. The magnitude of this path suggests that when clinicians mentally commit to using AI, they are far more likely to report the skills, confidence, and situational preparedness needed to implement it. Practically, this means interventions that raise Acceptance (eg, governance transparency, clear accountability, and visible clinical benefit) can unlock and amplify the returns from skills training and workflow enablement. At the same time, acceptance is necessary but not sufficient: high intention without competency risks superficial adoption. These results, therefore, prioritize two-step programs; first, build acceptance through trust building and perceived value, and second, immediately pair that shift with competency-based training and supervised practice to consolidate readiness.

The “Knowledge” construct exhibited poor reliability (α=.14) and convergent validity (AVE=0.20) and was excluded from the SEM. These numbers are probably due to the absence of a prevalidated measure for AI literacy and the reliance on simple true or false statements to test the participants on AI facts. Therefore, the knowledge domain was assessed narratively based on descriptive statistics alone. Interestingly, although many participants expressed confidence in their knowledge of AI, objective assessments revealed notable knowledge gaps. This discrepancy between perceived and actual knowledge suggests the need for structured AI training programs that not only elevate baseline understanding but also address specific misconceptions. In addition, this gap suggests that self-ratings are insufficient proxies for deployable literacy. For training and policy, competency-based assessment (objective scoring and feedback) should accompany awareness sessions. Tailored professional development courses and the inclusion of AI modules in medical education could significantly enhance knowledge and confidence levels. Moreover, demographic analysis offers valuable nuance. The participant pool was diverse in terms of roles and years of experience, yet uniformly reflected limited prior AI training and inconsistent ability to recognize AI applications. These findings indicate that AI knowledge is not merely a function of professional role but is associated with broader systemic factors such as institutional support and exposure to technology in the workplace.

The clear path from trust to acceptance to readiness offers a practical roadmap for health care leaders and policymakers aiming to align AI adoption with national strategies. Aligned with the UAE’s AI Strategy 2031, we translate the trust-centered pathway into four linked levers. First, mandated, role-specific professional development, with tiered AI modules (foundation to specialty) and objective assessments. Second, governance checkpoints in procurement and deployment, audit trails, bias or risk communication, accountability mapping, and publicly available model factsheets to strengthen institutional trust. Third, usability and workflow fit through co-designed clinical pathways, pilot sites, supervised practice, and real-world feedback loops so that benefit and ease are evident at the point of care. And finally, readiness activation that converts acceptance into capability via hands-on credentialing (simulations and supervised cases) and change-management supports (local champions and help-desk service-level agreements).

Across studies from Saudi Arabia, the United Arab Emirates, Oman, and Egypt, health care professionals report moderate-to-high awareness of AI but limited practical experience and training, producing a persistent Knowledge-Readiness gap. For example, Almalki et al [70] found 76.6% high computer proficiency, alongside 62.1% low AI-specific knowledge, and only 20.6% of radiology professionals in Hamd et al [71] had received AI training. Attitudes are cautiously positive when AI is seen as useful and easy to use, for instance, 81% in Alhashmi et al [72], anticipated performance gains; nurses in Baraka et al [73] were receptive despite little prior use. Yet enthusiasm coexists with worries about job displacement, data confidentiality, reliability, and liability [74,75]. Barriers cluster into five repeatable categories: technical, professional, ethical, legal or regulatory, and resource, with technical and resource issues most frequently cited [71,73,76]. Facilitators mirror this profile with targeted education and curricular integration, organizational and IT infrastructure support, clear policy frameworks, and user-centered design being consistently linked to higher acceptance and readiness [72,73,77]. Acceptance also varies by demographic and professional factors; males, informatics-adjacent roles, technicians, and younger professionals tend to be more accepting, whereas senior physicians report lower familiarity and greater legal concerns [70,75,77,78].

This study’s central pattern aligns with findings across the Arab region studies that attitudes improve when AI is perceived as useful or easy and when organizational scaffolding is present. Our structural model clarifies a mechanism that prior surveys often imply but rarely quantify. We also replicate the well-noted Awareness-Readiness gap, that is, high subjective confidence coexists with uneven objective performance and limited formal training. Whereas most regional papers are cross-sectional and descriptive, our work advances the field by (1) reporting standardized path coefficients for all hypothesized relationships, (2) positioning trust as an upstream driver that reshapes perception and benefit and risk appraisals before linking to acceptance, and (3) explicitly quantifying the Acceptance-Readiness linkage rather than treating readiness as a descriptive end point. This moves the conversation from lists of correlates to a mechanism-oriented account of adoption. Moreover, our discussion echoes the reviewed literature corpus on the importance of readiness enablers that target the perceived usefulness and ease of AI use.

Set against prior work, our findings support a trust-centered adoption pathway; governance and organizational signals elevate trust, which shapes global perceptions more than a narrow risk-benefit tally; acceptance then gates implementation readiness. Regional studies document the same ingredients, but our modeling clarifies how these pieces combine and where interventions should focus: trust-building governance paired with role-aligned, competency-based training to convert positive attitudes into deployable practice.

First, like the broader literature, our sampling is constrained. Recruitment used purposive, network-based sampling, which precluded calculation of a response rate and prevented formal assessment of nonresponse bias; thus, selection effects may have favored more digitally engaged practitioners.

Second, although the sample size (N=182) is acceptable for exploratory SEM, it is modest relative to model complexity and limits power for subgroup comparisons and measurement-invariance testing.

Third, the generalizability of our findings is limited by the composition of the study sample. Participants were predominantly recruited from Dubai-based health care settings (103/182, 57%), and nurses constituted the largest professional group (71/182, 39%), which may bias observed relationships toward nursing workflows and the Dubai Health regulatory environment. Because health care governance, digital maturity, and implementation practices vary across Emirates and between federal and local systems, the results should not be interpreted as nationally representative of the entire UAE health care workforce. Accordingly, we interpret the SEM as an exploratory model validated within this cohort and recommend replication using larger, more representative samples (eg, stratified sampling across professions, care settings, and Emirates) before extending conclusions to the broader UAE health care workforce.

Fourth, convergent validity in some cases was marginally less than the acceptable threshold (AVE <0.50 for Perception and Risk); nevertheless, the factors were retained because their internal consistency and composite reliability were within the acceptable limits and their exclusion worsened overall SEM fit. However, the marginal psychometric strength of these constructs warrants cautious interpretation of pathways involving this construct and motivates replication with refined measurement in larger samples. Although the Acceptance-Readiness HTMT (0.848) is near the conservative cutoff, the value remains below the discriminant-validity threshold, indicating that Acceptance and Readiness are closely coupled but not empirically redundant, and future work should confirm this separation in larger samples and through additional validity checks.

Fifth, model fit could be improved with an increased sample size. Fit indices like the RMSEA, GFI, and AGFI all fall outside the acceptable range, indicating mediocre, suboptimal overall fit of the proposed model to the observed data. The fit indices were reported transparently to allow future work to build on this study. Our exploratory model should be validated, refined, and better established with improved fit.

Sixth, the Knowledge construct failed reliability or validity and was excluded from SEM, constraining conclusions about AI literacy. Any recommendations made regarding the knowledge gap came from the observed descriptive discrepancy as many respondents self-rated as informed, while objective item correctness was low for several items.

Future work should prioritize upgrading the knowledge instrument. A validated, scenario-based scale with graded difficulty and objective scoring would better capture AI literacy. Reporting composite reliability and, where feasible, item-response metrics could establish performance benchmarks suitable for continuing professional development. Confirmatory testing with broader sampling is needed to strengthen external validity. Stratified or probability-based samples across Emirates, roles, and care settings, powered for multigroup and measurement-invariance tests (eg, nurses vs physicians and junior vs senior), should be paired with prespecified fit targets. Evaluation should move beyond attitudes to implementation outcomes. Adoption, fidelity, workflow time and interruptions, help-desk use, and safety signals (overrides and near-misses) ought to be tracked and linked to unit-level readiness. Human-centered workflow integration, co-designed pilots, electronic health record–embedded pathways, and iterative usability testing could define success via task completion time, error rates, and perceived workload. Finally, equity-focused targeting (eg, for professions or seniority groups with lower readiness) and transparent reporting of study instruments would facilitate replication and cumulative science.

In the Arab-region literature, acceptance and readiness to implement AI depend on five interlocking domains: (1) knowledge or experience, (2) attitudes or perceived usefulness, (3) barriers (technical, professional, ethical, legal, and resource), (4) facilitators (education, organizational or IT support, policy frameworks, and usable design), and (5) demographic or professional influences. Current evidence suggests that positive attitudes need to be complemented with capability building and organizational or policy scaffolding; without them, the observed awareness does not translate into practice.

In this study cohort, trust and overall perception had stronger relationships with acceptance than specific benefit or risk appraisals, and acceptance was tightly linked to readiness. These results provide a mechanism-focused account of AI adoption in a UAE clinical cohort and identify governance- and training-related levers that can be operationalized in future, evaluative work. Replication with validated measures, broader sampling, and longitudinal designs is warranted.