The past decade has witnessed major advancements in the field of health care, particularly through the integration of artificial intelligence (AI). AI may be described as machines that mimic cognitive functions associated with the human mind, such as learning and problem-solving [1]. An area of progress involves the development of AI-enabled clinical decision support systems (AI-CDSSs) [2-4]. AI-CDSSs use machine learning algorithms to process vast amounts of data and provide case-specific advice to health care practitioners to aid clinical decision-making [5-7]. AI-CDSSs use clinical data both from structured (eg, laboratory test results) and unstructured (eg, clinician notes or imaging) sources. The interpretation of text-based data can be performed using natural language processing to transform text into usable data for clinical predictions [8]. In addition, deep learning models, including neural networks, can be used to generate recommendations based on image data, for example, in the detection of pneumonia from chest radiographs [9]. AI-CDSSs may improve the accuracy and efficiency of medical decision-making in several ways.

First, AI-CDSSs may offer structured rationales underpinning clinical decisions that can complement traditional care methods. This structured approach paves the way for clearer understanding, improved communication, and better tracking of the decision-making process in clinical settings [10,11]. Second, AI-CDSSs can integrate data from various sources to provide a comprehensive and personalized recommendation for every patient case [7]. Finally, AI-CDSSs promote the consistency of medical decisions. The use of AI algorithms may ensure that the same set of facts will consistently produce the same recommendations, thus minimizing harmful consequences due to human error [9].

Despite these advantages, the implementation of AI-CDSSs in clinical practice must still overcome numerous barriers. A major challenge in the deployment of AI-CDSSs is the variability in performance. This can occur when the data used to develop the AI models do not adequately represent the population for which the tool is intended. Another issue is when AI-CDSSs are not used as designed, which can be due to a range of factors, including user interface problems, lack of integration into clinical workflows, or insufficient training of health care professionals on how to use the system [7,12-14]. The resulting low performance casts doubt on the value of AI-CDSSs in assisting with clinical decision-making [12,15]. In addition, the lack of understanding of how AI recommendations are derived heightens clinicians’ reservations about using these systems [16-18]. There are also challenges related to the alignment of AI-CDSSs with existing workflows that can cause additional workload when new AI systems are incorporated into clinical procedures [7,19-21].

As the development of high-performing AI-CDSSs proceeds, understanding the factors that influence health care practitioners’ intention to use these systems becomes increasingly relevant. One of the most comprehensive theories to explain individual technology adoption is the Unified Theory of Acceptance and Use of Technology (UTAUT) [22]. The UTAUT proposes that a person’s intention to use a technology is determined by their beliefs and attitudes toward that technology, such as the perception of its performance or the perceived effort it would require to use it. The UTAUT’s comprehensive nature and its ability to account for various determinants of technology acceptance make it an appropriate model for examining the predictors of health care practitioners’ intention to use AI-CDSSs.

Research to identify predictors of the intention to use AI-CDSSs has accumulated over the past years [4,23-25]. However, the existing literature remains scattered and in need of systematic synthesis. Therefore, the overarching goal of this study was to quantitatively integrate existing studies on the predictors of health care practitioners’ intention to use AI-CDSSs. The proposed hypotheses were based on the UTAUT model and existing empirical evidence. With this meta-analysis, we make 4 major contributions to theory and practice. First, we used meta-analytic techniques to estimate the relationship between the predictors of the UTAUT and the intention to use AI-CDSSs, thus providing insights into the applicability of the UTAUT to the context of AI-CDSSs. Second, we identified additional predictors based on the existing literature and examined the relative contribution of the UTAUT and additional predictors in explaining the intention to use AI-CDSSs. With this approach, we contribute to a theoretical refinement and potential extension of the UTAUT model to the context of AI-CDSSs. Third, based on the UTAUT, we examined the role of contextual factors as moderators of the relationships between relevant predictors and use intention, thus shedding light on the conditions that influence the strength of these relationships. Finally, in line with the UTAUT model, this is the first meta-analysis that examines the role of mediators, thus allowing for a better understanding of the complex mechanisms through which use intention may be explained. The study protocol, including all hypotheses and research questions (RQs), has been preregistered through the Open Science Framework [26].

To qualify for inclusion, the following criteria had to be met. First, studies had to be published in English. Second, studies had to include AI-CDSSs. The second inclusion criterion was fulfilled if (1) one of the following terms—“artificial intelligence,” “AI,” “machine learning,” “deep learning,” or “deep neural networks”—was used to describe the technology [84] and (2) the technology was referred to as a clinical decision support system or it was described as providing recommendations regarding the diagnosis, treatment, or prognosis of health issues [6]. We included studies if AI-CDSSs were mentioned alongside other AI-enabled functionalities [85]. This led to the exclusion of studies that investigated the use intention of other health care technologies, such as telemedicine [86] or the Internet of Medical Things [87]. Notably, one study examined the intention to use explainable and nonexplainable AI-CDSSs in the same sample [38]. Because only one other study examined explainable AI [45], we included only the data for the nonexplainable AI-CDSSs. Third, studies had to include a measure of the intention to use AI-CDSSs as defined in the UTAUT [22], including self-developed scales based on the UTAUT scale. Fourth, studies had to be empirical. This led to the exclusion of nonempirical studies such as reviews or case studies [88]. Fifth, studies had to measure at least one predictor of the intention to use AI-CDSSs. Sixth, studies had to measure use intention among a sample of health care practitioners or medical students based on the list of health professionals by the World Health Organization [89]. Table S2 in Multimedia Appendix 1 shows a detailed overview of the inclusion criteria per included study.

This meta-analysis was conducted in accordance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines to ensure comprehensive and transparent reporting [90]. We used 5 steps to search for relevant data. First, relevant scientific articles, dissertations, and theses were searched using the electronic databases Embase, MEDLINE, ProQuest, PsycINFO, and Web of Science between October 15, 2022, and January 5, 2023. In total, 2 follow-up searches were conducted on May 2, 2023, and November 7, 2023. The search string was developed based on the participants, intervention, comparators, and outcome framework [91]. The framework was adapted to fit the research purpose, resulting in a 3-tiered search term including the population (health care professionals), technology (AI-CDSSs), and outcome (use intention) of interest. An overview of the search terms is presented in Table S3 in Multimedia Appendix 1. We used the search terms to search titles, abstracts, and keywords. We conducted follow-up searches in Google Scholar using the following search string: (“health care”), AND (“Artificial Intelligence”) AND (“UTAUT”). Second, we conducted forward searching of studies citing the seminal article by Venkatesh et al [22] via Google Scholar and backward searches of review articles [92-96]. Third, abstracts of relevant conference proceedings, including the Conference on Computer-Supported Cooperative Work and Social Computing, the Conference on Human Factors in Computing Systems, and the Institute of Electrical and Electronics Engineers, were searched. Fourth, we sent requests for unpublished articles and data using the mailing list of the German Psychology Association. Finally, authors of articles included in the meta-analysis were contacted and asked for unpublished data sets. No additional unpublished data were obtained.

We reached out to authors when critical information was needed to decide on the inclusion of a study or details essential for the meta-analytic synthesis, such as a correlation table, were missing. From the 24 authors contacted to procure missing information, we successfully obtained 6 data sets. These data sets were used to derive the missing information, for instance, to calculate missing correlations between variables of interest.

Figure S1 in Multimedia Appendix 1 shows the PRISMA diagram with the number of studies identified, included, and excluded, along with reasons for exclusion. The studies from the literature search were assessed following a 3-stage approach. First, titles were screened to identify relevant articles. Second, the abstracts of the remaining articles were reviewed. Third, full article texts were reviewed. As a result of a review of 107 full texts, 17 (15.9%) studies met the inclusion criteria (k=18 independent samples; N=3871).

Following the approach of previous meta-analyses, we only included relationships that were identified in a minimum of 3 separate samples [43,97]. We grouped overlapping variables into construct categories (see Table S4 in Multimedia Appendix 1 for definitions of superordinate constructs and subconstructs. Studies from both the primary and the follow-up literature search were coded by 2 researchers each (AK and SG for the primary search and JD and AK for the follow-up search). Any conflicts in the coding were resolved in weekly consensus meetings. In addition, in line with approaches to ensure accuracy in coding established in previous meta-analyses [98], a random sample of 56% (10/18) of the independent samples was recoded by JC and AS. We included agreement on correlations, reliabilities, and moderator categories into the assessment of interrater agreement. Overall interrater agreement was high (94.7%). Notably, no disagreements were observed regarding correlations. Some mistakes in the coding of reliabilities occurred during recoding due to referencing an incorrect line from the source document. The final code sheet used for the analyses is available on request from the Corresponding Author.

All analyses were conducted using RStudio (Posit Software, PBC) [99] using the R packages psychmeta [100] and metaSEM [101].

The overall mean age of the participants was 36.2 (SD 13.5; range 21-53) years, and 48.7% were female. A total of 41% (7/17) of the studies focused on diagnostic AI-CDSSs, 12% (2/17) focused on treatment AI-CDSSs, 24% (4/17) focused on treatment and diagnostic AI-CDSSs, and 24% (4/17) focused on unspecific AI-CDSSs. In total, 65% (11/17) of the studies were conducted in Asia (6/11, 55% in China), 18% (3/17) were conducted in Europe, 6% (1/17) were conducted in the United States, and 12% (2/17) were conducted worldwide in English-speaking countries.

In the following sections, we report sample size–weighted and reliability‐corrected correlations (r_c) for the relationships between relevant antecedent variables and AI-CDSS use intention. In line with Cohen [113], we classified our reported effects as weak (r_c=0.1), moderate (r_c=0.3), and strong (r_c=0.5).

To assess the robustness of the meta-analytic findings, we used cumulative meta-analysis. This approach involves conducting a sequence of iterative meta-analyses, with each analysis adding an effect size for a specific relationship. Effect sizes are added in order of decreasing precision, meaning that the initial effect sizes added represent the most accurate population effect size estimates. If less precise studies tend to skew the meta-analytic estimates, this will be observable as a shift in cumulative results when these studies are included [106]. The results of the cumulative meta-analyses are shown in Figure S3 in Multimedia Appendix 1. A total of 5 “drifts” were identified, and all relationships drifted toward stronger effects as less precise studies were added, indicating an overestimation of the true effect. However, meaningful differences were not observed for any of the relationships after half the studies were added compared to after all the studies were added (Table S6 in Multimedia Appendix 1). Accordingly, we conclude that none of the drifts influenced the meta-analytic conclusions.

The primary goal of the meta-analysis was to gain a better understanding of the predictors of intention to use AI-CDSSs among health care practitioners based on the UTAUT and its extensions. The results of the meta-analysis provide empirical support for the applicability of the UTAUT to the context of AI-CDSSs. As predicted, performance expectancy, effort expectancy, social influence, and facilitating conditions were positively related to the intention to use AI-CDSSs. These findings are largely in line with the findings of UTAUT meta-analyses in other fields [27,32,40,61]. We provide a summary of the main findings for our hypotheses and RQs in Tables 7 and 8.

The results of relative weight analyses showed that all 4 UTAUT predictors together explained 50% of the variance in use intention among health care practitioners, reaffirming the relevance of the UTAUT predictors in the context of AI-CDSSs. Among the UTAUT predictors, performance expectancy emerged as the most relevant, accounting for 31% of the total explained variance, followed by social influence (28%), facilitating conditions (26%), and effort expectancy (15%). In most UTAUT research, performance expectancy is more relevant than effort expectancy, possibly because performance expectancy is inherently connected to the primary motives behind technology use [32,40]. That is, it directly relates to the perceived benefits that users expect to gain from using a technology [32,40]. Effort expectancy refers to the expected ease of using a technology [22]. While important, the ease of use may become a secondary consideration if the technology does not meet the primary performance-related objectives. In other words, users might be willing to overcome a steeper learning curve if they believe the payoff in performance is worthwhile [63]. This could explain why performance expectancy accounts for a higher percentage of the variance in technology acceptance and use intentions compared to effort expectancy. Overall, the findings of this meta-analysis reflect a common finding in technology acceptance research where the anticipated improvement in performance is often found to be a stronger driver of user acceptance than the anticipated effort to learn and use the technology [32,40,61,63].

Among the UTAUT predictors, effort expectancy and facilitating conditions had the widest CRs (0.56 and 0.66, respectively), suggesting the presence of moderating influences [102,114]. For example, previous research suggests that radiologists, accustomed to complex machines and heavy workloads, may be willing to invest effort into learning how to use new technology if it reduces their workload, indicating a moderating influence of occupation on the relationship between effort expectancy and use intention [4,37]. In addition, the strength of the relationship between effort expectancy and use intention has been shown to differ between AI-CDSSs for feedback versus decision support [24].

In addition to the core UTAUT variables, we identified attitude, trust, perceived risk, AI anxiety, and personal innovativeness as predictors of the intention to use AI-CDSSs. Although all the included studies (17/17, 100%) reported a negative relationship between AI anxiety and use intention, the CI for AI anxiety included 0. This lack of an observed relationship may be due to the low sample sizes (the total sample size was 391) and the resulting high uncertainty in the true effect. Interestingly, in the relative weight analyses, trust proved to be a more relevant factor than performance expectancy in explaining variance in the intention to use AI-CDSSs. The relevance of trust may be explained by the lack of transparency in how AI recommendations are generated coupled with the high stakes associated with clinical decision-making [115]. Indeed, research has suggested that even highly efficient AI-CDSSs may face resistance in clinical applications if health care practitioners do not trust the system’s safety [116-118]. The findings of this meta-analysis align with those of research advocating for the inclusion of trust in the UTAUT model [116].

Furthermore, this meta-analysis emphasizes the need to consider both drivers and inhibitors of the intention to use AI-CDSSs for a more comprehensive understanding of the adoption process [119]. The relative weight analyses demonstrate that AI anxiety explained approximately 10% of the variance in the intention to use AI-CDSSs after trust (29%), facilitating conditions (20%), performance expectancy (19%), and social influence (14%) and before effort expectancy (7%). The relevance of perceived risk as a predictor of use intention was small (approximately 2% after all other predictors). Risk perception is a cognitive assessment of the potential losses and gains from using AI-CDSSs, which is based on logical evaluation and can be mitigated by providing relevant information [120]. In contrast, AI anxiety is an emotional response that encompasses fears and insecurities about AI technology [121]. Accordingly, AI anxiety is less rational and more difficult to alleviate because it can be deeply rooted in concerns about AI’s impact on job security, professional autonomy, and the quality of patient care [24,122,123].

Even for relationships assessed in a substantial number of independent samples, such as performance expectancy, effort expectancy, attitude, trust, and perceived risk, the CRs were wide (>0.34), suggesting the presence of moderators [102,114]. This observation is supported by the modest amount of variance accounted for by statistical artifacts, indicating that there may be other reasons for substantial variance between individual studies [102,114]. While we considered multiple moderators suggested by the UTAUT and additional contextual and methodological moderators, we only found 2 moderation effects.

First, age moderated the relationship between social influence and use intention, with older health care practitioners exhibiting a weaker relationship between social influence and use intention. This finding does not align with the UTAUT proposing that older individuals place more relevance on the opinion of relevant others when intending to use a new technology [22]. An explanation for this discrepancy may be that practitioners value their professional independence increasingly more with age, thus relying less on the opinion of others regarding technology use as they get older. While differences in professional values and behaviors have been shown to exist between younger and older health care practitioners [124], a systematic examination of the effect of age on the relationship between social influence and the intention to use AI-CDSSs is lacking. It has to be noted that the observed moderation effect was based on only 22% (4/18) of independent samples, underscoring the need to systematically study the influence of age on the relationship between social influence and use intention.

Second, the relationship between effort expectancy and use intention was stronger for diagnostic AI-CDSSs than for devices that combined diagnostic and treatment recommendations. When clinicians assess a tool solely for diagnostic purposes, they may find it easier to anticipate the required effort to use it as the task is less complex and the outcome is more direct. This clear understanding may strengthen the relationship between effort expectancy and use intention. Indeed, the perceived risk associated with smart devices has been found to negatively influence the relationship between effort expectancy and use intention [125]. The multifaceted nature of combined tools may make it more challenging for clinicians to evaluate the effort needed to understand and use them. This uncertainty possibly leads to a weaker relationship between effort expectancy and use intention as clinicians may not be able to adequately assess the effort required, thus not being able to use it as a source of information when it comes to indicating their intention to use it. Future research is needed to investigate the moderating influence of device type.

The results of the mediation analyses indicate that the relationship between facilitating conditions and use intention may be explained through effort expectancy and performance expectancy. This finding aligns with the UTAUT proposing that, when performance and effort expectancy are considered, facilitating conditions lose their importance in predicting use intention [22]. An explanation for the relevance of effort expectancy as a mediator may be that issues related to the support infrastructure, a critical aspect of facilitating conditions, are also conceptually addressed by effort expectancy [22]. That is, if health care organizations establish the appropriate support infrastructure, the effort required to use AI-CDSSs becomes lower [22,54]. Similarly, if a user perceives that the technology is supported by adequate facilitating conditions, they may be more likely to believe in the performance benefits when using the system, explaining the mediating role of performance expectancy.

Performance expectancy and trust emerged as the 2 most relevant predictors of AI-CDSS use intention, suggesting that measures targeted toward health care practitioners’ beliefs in the performance and trustworthiness of AI-CDSSs may be effective in enhancing their intention to use them. However, the consistently positive link between performance expectancy and use intention also suggests that health care institutions need to take measures to deter the perception of low-performing systems as high performing, which could potentially cause more harm than benefit [126]. Health care practitioners require transparent communication regarding the performance and limitations of AI-CDSSs alongside adequate training to ensure their correct use. In addition, regulatory bodies such as the Food and Drug Administration need to ensure that available AI-CDSSs meet certain safety and performance standards [84,127,128]. Adequate policies and oversight in these contexts may ensure a balance between the adoption and safe application of AI-CDSSs in health care decision-making.

Trust in technology is a multifaceted construct including users’ perceptions of a system’s functionality, helpfulness or benevolence, and integrity [48,49]. Consequently, actions taken to enhance performance expectancy may not be sufficient for building trust in a system [116]. If organizations aim to improve health care practitioners’ trust in AI-CDSSs, they need to address the various facets relevant to trust in technology. This includes dealing with ethical issues related to data privacy and the potential misuse of AI-CDSSs as well as addressing the lack of transparency and explainability in AI-generated recommendations [129,130]. For example, trust has been associated with the system’s capability to explain its decision-making process, emphasizing the role of explainable AI as a path to building trust in AI-CDSSs [116,131]. In addition, regulatory strategies should be designed to promote and maintain trust in AI-CDSSs along with safe patient outcomes. This might include the use of postmarket surveillance systems to monitor the performance of deployed AI-CDSSs over time, which has been suggested as a method for identifying and mitigating issues of utility and safety in real-world clinical settings [132,133]. Developers can integrate user-centered design principles to tailor AI-CDSSs to the needs and workflows of specific clinical specialties and roles. The early inclusion of user feedback may facilitate the development of user-friendly AI-CDSSs and increase trust in these systems [134,135]. Health care administrators may foster practitioners’ trust by providing training programs to increase familiarity with the technology and by designing evaluation metrics that can monitor system performance and user satisfaction [136].

Social influence has been demonstrated to be a relevant predictor of health care practitioners’ intentions to use AI-CDSSs, particularly among younger professionals. Institutions aiming to adopt AI-CDSSs can leverage the important role of social influence by establishing a culture that values technological advancements and by engaging key opinion leaders to advocate and exemplify the use of these systems. In addition, trainings can be structured not only to educate but also to establish a shared understanding and a community of practice that positively reinforces the application of AI-CDSSs [137,138]. By addressing the social aspects of technology acceptance, health care institutions can ensure that their investment in AI is met with a user base that is both competent and willing to integrate these tools into their daily practice.

The importance of facilitating conditions underscores the need for health care organizations to provide a supportive infrastructure that simplifies the integration of AI-CDSSs into existing workflows. For instance, the provision of training programs, allowing health care practitioners to gain firsthand experience, and setting up accessible support teams ready to address system-related issues can considerably boost health care practitioners’ intention to use such systems [23,139].

AI anxiety emerged as a barrier to the intention to use AI-CDSSs in the relative weight analysis. Therefore, hospitals and other health care institutions should consider measures to counteract any irrational negative emotional reactions to AI before and during the integration of AI-CDSSs into clinical workflows. A potential method to mitigate AI anxiety involves increasing medical staff involvement in the development process [123] or providing more training opportunities to increase their exposure to AI-enabled devices, thus reducing irrational fears [137].

This meta-analysis is not without limitations. First, this study offers insights into the predictors of use intention as the key determining factor of actual use. However, some health care practitioners may express intention to use AI-CDSSs but be hesitant when it comes to their actual implementation. Few studies included in the meta-analysis (3/17, 18%) examined the predictors of actual use, underscoring the need for additional research on predictors of the actual use of AI-CDSSs [4,48].

Second, we were unable to explain the considerable variation in some of the effects based on moderator analyses. We could not evaluate 3 UTAUT moderators—experience with AI-CDSSs, voluntariness of use, and occupation—owing to insufficient samples incorporating these variables. In addition, although all studies including AI anxiety (3/17, 18%) reported negative relationships with use intention, the CI of the meta-analytic estimate included 0 due to the low sample size and the associated high uncertainty in the estimate. More studies on the relationship between AI anxiety and intention to use AI-CDSSs are needed. The large CRs and the low correlations between estimates and statistical artifacts suggest the existence of moderating factors not included in the meta-analysis [102,114]. Future research should explore moderating effects such as differences in the observed relationships among health care practitioners working in different fields or roles to better understand the boundary conditions that affect the relationships between predictors and the intention to use AI-CDSSs.

Third, the 9 relevant predictors could not be examined in a single relative weight analysis. The use of multiple models with subsets of predictors is a pragmatic approach to addressing data sparsity. However, the selected approach hinders definitive conclusions regarding the importance of all considered predictors. Furthermore, innovativeness could not be assessed in the relative weight analysis due to a lack of available samples assessing this predictor. The compromises that had to be made in the relative weight analyses highlight the need for an updated meta-analysis that includes complete predictor sets.

Fourth, the insights derived from the meta-analysis are primarily confined to unspecific AI-CDSSs. Given that AI-CDSS adoption is still limited, only a handful of studies have delved into exploring predictors of the use of specific AI-CDSSs with distinctive features [4,24,38]. The results of these studies show that the attitude toward AI-CDSSs may vary depending on use cases and system features. Future research should examine the adoption of individual systems and variations in effects across different types of AI-CDSSs.

Fifth, the existing body of research on AI-CDSS adoption primarily relies on cross-sectional observational studies, with questionnaires as the main method of data collection. These studies inherently limit the establishment of causal relationships, thus underscoring the need for future research to include longitudinal or experimental designs for more robust evidence of causality. Longitudinal studies may also be used to shed light on the development of use intention and the relevance of relevant predictors over time. For example, it is possible that initial trust plays a crucial role during the implementation phase but becomes less relevant once a system has been successfully implemented.

Sixth, we selected the UTAUT as a general theoretical framework to examine the predictors of intention to use AI-CDSSs. However, there has been some criticism of the UTAUT [64,140]. For example, the UTAUT may not answer questions related to the determinants and processes involved in value-adding technology use [64,141,142]. We found support for the prediction that beliefs about the performance and ease of use of AI-CDSSs lead to a higher intention to use these systems. However, based on the UTAUT, whether these beliefs are well founded (ie, whether positive expectations actually lead to beneficial use because the system is indeed high performing and easily implementable) may not be resolved. Another criticism pertains to the UTAUT’s narrow viewpoint on individual use. Other models such as the nonadoption, abandonment, scale-up, spread, and sustainability framework [143] adopt a system perspective. This approach enables the examination of predictors at micro (individual technology users), meso (organizational processes and systems), and macro (national policy and wider context) levels, thereby more accurately representing the complex processes involved in technology adoption [143,144]. In addition, the UTAUT focuses on an individual’s intention to use a technology and does not fundamentally consider how well the technology fits the task it is being used for. Theories such as the task-technology fit model examine the interconnectedness between task and technological characteristics. The model delves into how features of both the assigned task and the technology at hand shape the task-technology fit, thereby influencing the overall performance and use intentions [145].

Finally, some of the predictors included in this meta-analysis, particularly the additional predictors beyond the UTAUT, may not be adequately represented using standard measurement instruments. For instance, AI anxiety has a multitude of dimensions, such as privacy violation anxiety, bias behavior anxiety, job replacement anxiety, learning anxiety, or ethics violation anxiety [146]. This meta-analysis did not distinguish between these different aspects of AI anxiety as separate predictors of AI-CDSS use intention. Similarly, trust in AI is a multifaceted construct that includes perceptions of the system’s benevolence, competence, and integrity [48,49]. Moreover, trust may refer to different aspects of an AI-CDSS, such as trust in the reliability of its predictions when being applied to different contexts, trust in legal protection if harm to patients occurs from using the AI-CDSS, and trust in data privacy [4,47,147]. More research is needed that explores the relevance of different elements of trust (ie, benevolence, competence, and integrity) and elements of the AI-CDSSs that may be trusted to different degrees (eg, reliability, legal and liability issues, and privacy concerns) for the intention to use AI-CDSSs in clinical decision-making [147].

This meta-analysis underscores the relevance of the UTAUT to examine the predictors of intention to use AI-CDSSs in health care. The results indicate that performance expectancy, effort expectancy, social influence, and facilitating conditions are positively related to the intention to use AI-CDSSs among health care practitioners. The analyses further revealed the relevance of the additional predictors attitude, trust, personal innovativeness, AI anxiety, and perceived risk. The results of mediation analyses show that effort expectancy and performance expectancy explain the relationship between facilitating conditions and use intention. Despite identifying age and AI-CDSS type as moderating influences, there is scope for future research to investigate other possible moderators to explain the variability in the observed effects. While the UTAUT model provides a theoretical framework for studying health care practitioners’ intention to use AI-CDSSs, it remains relatively silent on the predictors of value-adding use of AI-CDSSs. Future research could investigate the conditions that encourage value-adding use by applying comprehensive frameworks that consider both individual and broader organizational processes (eg, clinic systems and administrative hurdles). Finally, the findings of this meta-analysis provide starting points for the development and integration of AI-CDSSs that are likely to be adopted by health care practitioners as end users.