記住我
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 []. An area of progress involves the development of AI-enabled clinical decision support systems (AI-CDSSs) [-]. 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 [-]. 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 []. 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 []. 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 [,]. Second, AI-CDSSs can integrate data from various sources to provide a comprehensive and personalized recommendation for every patient case []. 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 [].
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 [,-]. The resulting low performance casts doubt on the value of AI-CDSSs in assisting with clinical decision-making [,]. In addition, the lack of understanding of how AI recommendations are derived heightens clinicians’ reservations about using these systems [-]. 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 [,-].
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) []. 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 [,-]. 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 [].
Theory and Hypothesis DevelopmentThe UTAUT and the Intention to Use AI-CDSSsThe UTAUT integrates 8 former technology use theories and has become one of the most prominent technology use models [,]. The UTAUT has been applied to investigate factors influencing the acceptance and use of technology in different contexts, including health care [-]. The primary outcome measure considered in the UTAUT, alongside actual use, is the intention to use a technology [,,]. Intentions are indicators of motivation and reflect the level of determination that individuals have to actually perform a certain behavior []. The successful deployment of any technology depends largely on the user’s intention to use it []. Accordingly, understanding the predictors of the intention to use AI-CDSSs may help overcome individual-level impediments thwarting the adoption of AI-CDSSs in health care.
The UTAUT consists of 4 core predictors of individual use intention: performance expectancy, effort expectancy, social influence, and facilitating conditions []. The relationships between these variables and use intention are proposed to be moderated by gender, age, experience, and voluntariness of use []. The UTAUT model is shown in . All relationships included in the UTAUT were proposed as hypotheses, whereas all additional relationships and moderators that were derived based on empirical findings and other theories were proposed as RQs.
Figure 1. The Unified Theory of Acceptance and Use of Technology model. Predictors of the Intention to Use AI-CDSSs Based on the UTAUTPerformance expectancy refers to the extent to which individuals believe that using a technology will improve their job performance. AI-CDSSs have the potential to enhance job performance by aiding clinicians in deriving diagnoses or making treatment decisions []. If clinicians perceive their decisions to be improved by using AI-CDSSs, then performance expectancy will be high [,]. Hypothesis 1 is that performance expectancy is positively related to the intention to use AI-CDSSs.
Effort expectancy concerns the perceived ease of use of a technology. It is suggested that a system that is perceived to be easy to use is more likely to be accepted than one that is perceived to be complicated to use []. If, for example, the perceived effort of using an AI-CDSS in one’s existing clinical workflows is perceived to be high, health care practitioners may be less willing to use it [,,]. Hypothesis 2 is that effort expectancy is positively related to the intention to use AI-CDSSs.
Social influence refers to the impact of social factors, such as the expectations and influence of peers, on an individual’s intention to use a technology. The positive relationship between social influence and the intention to use AI-CDSSs has consistently been supported in empirical studies [,]. For example, it has been found that medical professionals holding the belief that their colleagues, top management, and professional bodies endorse the use of AI-CDSSs in clinical settings are more willing to adopt them []. Hypothesis 3 is that social influence is positively related to the intention to use AI-CDSSs.
Facilitating conditions represent the organizational and technical infrastructure necessary for technology adoption []. It has been argued that, if users believe that the resources and support are in place to facilitate the use of AI-CDSSs, they are more likely to intend to use them [,,]. Hypothesis 4.1 is that facilitating conditions are positively related to the intention to use AI-CDSSs. In addition, according to the UTAUT, there is a direct relationship between facilitating conditions and actual technology use []. Facilitating conditions refer to the resources and support available to use a technology, including the access to the necessary tools and knowledge. This practical aspect makes the influence of facilitating conditions on use immediate as users are more likely to use technology when they perceive a supportive environment and available resources. Unlike other predictors in the UTAUT, facilitating conditions are proposed as direct antecedents of actual use []. Hypothesis 4.2 is that facilitating conditions are positively related to actual use of AI-CDSSs.
Additional Predictors of the Intention to Use AI-CDSSsThe UTAUT has been modified, and additional predictors have been added over time to account for various settings and technologies [,-]. However, a meta‐analytic review is limited to the relationships that have been studied in the literature. Following previous research and methodological best practices, we included additional predictors beyond the UTAUT in the meta-analysis that were examined in at least 3 independent samples [,]. Following this criterion, we identified attitude, trust, perceived risk, AI anxiety, and personal innovativeness as additional predictors of the intention to use AI-CDSSs.
Individual behavior is driven by intention, which is, in turn, a function of an individual’s attitude toward the behavior and subjective norms [,]. Indeed, a positive attitude toward AI-CDSSs has been identified as a relevant predictor of the intention to use AI-CDSSs [-]. Because the relationship between positive attitude and use intention is not included in the UTAUT, we propose the following RQ (RQ 1) to explore the relationship between positive attitude and the intention to use AI-CDSSs []: is there a positive relationship between a positive attitude toward AI-CDSSs and the intention to use AI-CDSSs?
Trust becomes relevant if the outcome of a situation is uncertain or the possibility of undesirable outcomes exists []. Trust has been argued to be a particularly relevant predictor of the intention to use AI-CDSSs due to a lack of transparency of how recommendations are derived and the high stakes of erroneous decisions in health care [,]. Generally, we may differentiate between initial trust as the judgment of the truster before being exposed to the trustee and knowledge-based trust that may be established after the truster has interacted with the trustee []. In the context of AI-CDSSs, some studies refer to initial trust in terms of beliefs in the reliability and safety of AI-CDSSs before the user has been exposed to or actively used the system [,,]. An example item for initial trust is “I believe AI could provide completely accurate diagnosis assistant service” []. Another aspect of trust that has been explored in empirical studies is trust in different attributes of the technology, namely, its functionality (being able to do a required task), its helpfulness or benevolence (being able to provide effective help when needed or act in the best interest of the user), and its integrity (operating reliably or consistently without failing) [,,]. An example item for trust in the system’s integrity regarding data security is “I trust that recommendations from the AI-powered care pathway are reliable” []. Because trust is not included in the UTAUT model, we propose an RQ (RQ 2) to explore whether there is a positive relationship between trust and the intention to use AI-CDSSs []: is there a positive relationship between trust and the intention to use AI-CDSSs?
Perceived risk is determined by the unpredictability and perceived intensity of outcomes []. In the context of AI-CDSSs, perceived risk refers to the perceived potential negative consequences associated with their use, including performance failure and data insecurity []. An example item for perceived risk of a performance failure is “There is a possibility of malfunction and performance failure, so the system might fail to deliver accurate contouring areas and could mislead my work with inaccurate contouring” []. Health care professionals may be reluctant to engage with new services fearing that their perceived risk may result in negative user experience or even harm to them or their patients []. Different forms of perceived risk have been found to be negatively associated with the intention to use AI-CDSSs [,,,]. For example, it has been found that performance and legal risk associated with AI-CDSSs are negatively related to the intention to use AI-CDSSs []. Because perceived risk is not included in the UTAUT model, we propose the following RQ (RQ 3) to investigate whether perceived risk is negatively associated with the intention to use AI-CDSSs []: is there a negative relationship between perceived risk and the intention to use AI-CDSSs?
AI anxiety encompasses general fears and insecurities regarding AI technology. It represents an intuitive, negative affective reaction to AI technologies, for example, based on the fear of making mistakes [,]. AI anxiety is often measured using the AI anxiety scale []. An example item is “I feel apprehensive about using the system.” If health care professionals experience anxiety in using AI-CDSSs, their intention to use them is presumably low. Indeed, AI anxiety has been identified as a negative predictor of the intention to use AI in health care []. However, because AI anxiety is not included as a predictor of use intention in the UTAUT, we propose the following RQ (RQ 4) to explore whether AI anxiety is negatively associated with the intention to use AI-CDSSs []: is there a negative relationship between AI anxiety and the intention to use AI-CDSSs?
Personal innovativeness describes an individual’s readiness to experiment with and embrace a new technology []. Those demonstrating a high degree of personal innovativeness have greater capabilities and, therefore, demonstrate greater readiness to use a new technology [,]. Indeed, there is empirical evidence for a positive link between personal innovativeness and the intention to use AI-CDSSs [,]. RQ 5 is as follows: is there a positive relationship between personal innovativeness and the intention to use AI-CDSSs?
The Relationship Between AI-CDSS Use Intention and Actual UseThe UTAUT proposes that an individual’s intention to use a technology is the main predictor of its actual use []. However, this relationship has not yet been extensively researched in the context of AI-CDSSs. The limited investigation of actual use may be attributed to the restricted number of AI-CDSSs implemented in clinical practice []. Nonetheless, some evidence indicates that use intention predicts the actual use of AI-CDSSs [,]. RQ 6 is as follows: what is the relationship between the intention to use AI-CDSSs and their actual use?
The Relative Contribution of the UTAUT Predictors and Additional Predictors in Explaining AI-CDSS Use IntentionExisting empirical research has explored the extent to which the UTAUT predictors account for variance in technology use intention []. For example, performance expectancy has often emerged as the strongest predictor of use intention [-]. Other research has found that trust has a stronger effect on the intention to use AI-CDSSs than performance expectancy []. As the roles of the UTAUT and additional predictors in explaining the intention to use AI-CDSSs remain unclear, we propose the following RQ (RQ 7): what is the relative contribution of the UTAUT predictors and additional predictors in explaining the intention to use AI-CDSSs?
Moderators of the Relationships Between UTAUT Predictors and the Intention to Use AI-CDSSsThe relationships between UTAUT predictors and use intention are proposed to be moderated by age, gender, user experience with the system, and voluntariness of using the system []. First, it has been suggested that younger workers prioritize extrinsic rewards such as improved job performance, thus exhibiting a stronger relationship between performance expectancy and technology use intention []. In contrast, it has been suggested that older workers generally face greater software challenges and are more likely to place increased relevance on social influences. Accordingly, they may rely more on effort expectancy and social influence when deciding to use a technology []. Hypothesis 5 is that the relationship between (1) performance expectancy and the intention to use AI-CDSSs becomes weaker and the relationships between (2) effort expectancy and (3) social influence and the intention to use AI-CDSSs become stronger with increasing age.
Second, the impact of performance expectancy on use intention is expected to be stronger among men, whereas the relationships between effort expectancy and social influence and use intention would be more pronounced among women []. Hypothesis 6 is that the relationship between (1) performance expectancy and the intention to use AI-CDSSs is stronger for men and the relationships between (2) effort expectancy and (3) social influence and the intention to use AI-CDSSs are stronger for women.
Third, according to the UTAUT, limited experience increases the strength of the relationship between effort expectancy and social influence and use intention because individuals with limited experience tend to overestimate the challenges associated with using a new technology and their opinions are more susceptible to social influence []. In contrast, as experience increases, facilitating conditions have been proposed to exhibit a greater impact on actual technology use as more experienced users know better in terms of how to take advantage of facilitating conditions when using the system []. Hypothesis 7 is that the relationships between (1) effort expectancy and (2) social influence and intention to use AI-CDSSs become weaker with increasing experience and the relationship between (3) facilitating conditions and actual use of AI-CDSSs becomes stronger with increasing experience.
Finally, the UTAUT distinguishes between voluntary (ie, individuals decide themselves whether to use a technology) and mandatory (eg, the use of a technology is mandated by the supervisor) adoption settings []. It has been suggested that social influence affects use intention in mandatory situations more because relevant others have the capacity to either incentivize desired actions or penalize noncompliance []. Hypothesis 8 is that the relationship between social influence and the intention to use AI-CDSSs is stronger in mandatory adoption settings.
In addition to the UTAUT moderators, we investigated the influence of additional contextual moderators that are studied in the literature, namely, occupation, type of AI-CDSS, and culture. All additional moderators were selected based on a comprehensive preliminary review of the literature. First, health care practitioners may work in different contexts requiring them to complete different tasks. These differences may influence their perceptions, beliefs, and attitudes toward AI-CDSSs [,]. For instance, one study found differences in the relationship between social influence and perceived risk and use intention between clinicians (eg, surgery and orthopedics) and nonclinicians (eg, radiologists and pathologists). Specifically, for nonclinicians, social influence positively predicted the intention to use AI-CDSSs, whereas perceived risk did not emerge as a significant predictor. In contrast, among clinicians, the reverse pattern was observed []. Second, the type of AI-CDSS likely influences practitioners’ use intention. Specifically, health care practitioners may place greater emphasis on the effectiveness and safety of treatment AI-CDSSs compared to diagnostic AI-CDSSs as an erroneous treatment decision is associated with more severe consequences []. Finally, cultural differences may influence the intention to use AI-CDSSs in health care [,]. For example, one study found perceived ease of use to be a more relevant predictor of the intention to use IT among Taiwanese compared to American physicians []. Accordingly, we propose the following RQ (RQ 8): do (1) the practitioner’s occupation, (2) the type of AI-CDSS, and (3) the cultural background moderate the relationship between UTAUT predictors and the intention to use AI-CDSSs?
Finally, we investigated the influence of methodological moderators such as publication year and the scale used to measure AI-CDSS use intention. In a meta-analysis based on the UTAUT, it was found that some effect sizes were stronger in more recent studies []. Moreover, while most studies use the intention to use scale introduced by Venkatesh et al [], some studies use self-developed scales to measure use intention [,]. RQ 9 is as follows: do (1) publication year and (2) the use intention scale used moderate the relationship between UTAUT predictors and the intention to use AI-CDSSs?
Performance and Effort Expectancy as Mediators of the Relationship Between Facilitating Conditions and the Intention to Use AI-CDSSsAccording to the UTAUT, the effect of facilitating conditions on use intention may be explained through performance and effort expectancy []. That is, if the required support infrastructure is provided, a person would perceive the system to be both high performing and easy to use, which, in turn, positively influences their intention to use it. Indeed, effort expectancy has been found to fully mediate the relationship between facilitating conditions and use intention []. Accordingly, we propose the following RQ (RQ 10) to investigate the mediating role of performance and effort expectancy: is the relationship between facilitating conditions and intention to use AI-CDSSs mediated through performance and effort expectancy?
Overview of the Hypotheses and RQsshows all hypotheses and RQs. We omitted the relationship between facilitating conditions and actual use of AI-CDSSs (hypothesis 4.2) as well as the moderators experience (hypothesis 7), voluntariness (hypothesis 8), and occupation (RQ 8.1) from the analyses (see the dashed lines in ) due to the limited number of available independent samples (<3). All deviations from the preregistration are presented in Table S1 in [,,-,,-,,,,,,,,-].
Figure 2. The proposed research model. The dashed lines represent preregistered hypotheses and research questions (RQs) that could not be investigated due to the limited number of available independent samples (<3). RQ 7 is omitted from the figure as it refers to the relative weight analysis. AI-CDSS: artificial intelligence–enabled clinical decision support system; H: hypothesis; UTAUT: Unified Theory of Acceptance and Use of Technology. 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 [] 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 []. We included studies if AI-CDSSs were mentioned alongside other AI-enabled functionalities []. This led to the exclusion of studies that investigated the use intention of other health care technologies, such as telemedicine [] or the Internet of Medical Things []. Notably, one study examined the intention to use explainable and nonexplainable AI-CDSSs in the same sample []. Because only one other study examined explainable AI [], 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 [], 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 []. 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 []. Table S2 in shows a detailed overview of the inclusion criteria per included study.
Search Strategy and Data ExtractionThis 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 []. 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 []. 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 . 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 [] via Google Scholar and backward searches of review articles [-]. 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 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 [,]. We grouped overlapping variables into construct categories (see Table S4 in 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 [], 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.
Meta-Analytic ProceduresAll analyses were conducted using RStudio (Posit Software, PBC) [] using the R packages psychmeta [] and metaSEM [].
Bivariate RelationshipsTo examine the bivariate relationship between the 4 core constructs of the UTAUT (hypotheses 1-4) and the additional predictors (RQs 1-5) with the intention to use AI-CDSSs, a random-effects meta-analysis was conducted []. Effect sizes were based on Pearson product-moment correlations. Composites were calculated if multiple measures of the same construct were reported for the same sample []. Specifically, a variance-weighted composite (across measures of the same construct) was calculated for each independent sample to combine multiple measures of the same construct into a single effect size per independent sample []. Sampling errors were corrected using sample size–weighted correlations. Measurement errors were corrected based on the Cronbach α []. In addition to the sample size–weighted correlation (r) and sample size–weighted and reliability-corrected correlation (rc), the 95% CI and 80% credibility interval (CR) for rc were reported. Finally, we reported the correlation between observed effects and the influence of the study design artifacts.
Relative Weight AnalysisWe conducted relative weight analyses to capture the contribution of the correlated predictors []. Specifically, we calculated multivariate meta-analytic regression models based on the pooled correlation matrices to explore the incremental value of the UTAUT predictors and additional predictor variables in explaining the intention to use AI-CDSSs. We used the harmonic mean of the sample size across the correlations considered as the sample size for the estimated regression models []. In relative weight analysis, raw relative weights are calculated to reflect the proportion of variance explained in the outcome that is attributed to each of the predictors, whereas rescaled relative weights reflect the percentage of the variance that is explained by each predictor variable [,].
Moderation AnalysesModerator analyses were carried out for constructs that were represented in a minimum of 56% (10/18) of the independent samples to ensure adequate coverage of moderator categories []. A total of 5 constructs met this minimum cutoff and were considered for the moderation analyses (ie, performance expectancy, effort expectancy, social influence, trust, and perceived risk). We interpreted categorical moderator effects if each of the levels included ≥3 independent samples. Age was coded as the mean age of study participants, and gender was coded as the percentage of women in the sample. For the type of AI-CDSS, 3 categories were initially identified: diagnostic decision support systems, treatment decision support systems, and systems that combined both diagnostic and treatment decision support. However, the treatment decision support systems category had to be excluded from the moderator analysis because of the low number of independent samples focusing on this type of AI-CDSS (2/18, 11%). Culture was operationalized based on the individualism versus collectivism dimension of the country comparison tool by Hofstede [,]. A higher score denotes stronger individualism. The publication year was coded chronologically. Finally, the scale used to measure the intention to use AI-CDSSs was coded as a categorical moderator. We differentiated between studies using the scale by Venkatesh et al [] and studies using self-developed scales. We conducted moderation analyses that were not preregistered as part of exploratory analyses.
Mediation AnalysisTo test RQ 6, correlation-based meta-analytic structural equation modeling [] based on the 2-stage structural equation modeling approach [,] was performed. In the first step, the sample size–weighted and reliability-corrected bivariate correlation matrices for each independent sample were pooled together. In 2-stage structural equation modeling, the total sample size is used for the estimation of the meta-analytic structural equation model []. In the second step, a path model was fitted to the pooled correlation matrix.
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.
Meta-Analytic ResultsIn the following sections, we report sample size–weighted and reliability‐corrected correlations (rc) for the relationships between relevant antecedent variables and AI-CDSS use intention. In line with Cohen [], we classified our reported effects as weak (rc=0.1), moderate (rc=0.3), and strong (rc=0.5).
Bivariate RelationshipsThe results of bivariate meta-analytic analyses are shown in . The UTAUT predictors performance expectancy (rc=0.66, 95% CI 0.59-0.73), effort expectancy (rc=0.55, 95% CI 0.43-0.67), social influence (rc=0.66, 95% CI 0.59-0.72), and facilitating conditions (rc=0.66, 95% CI 0.42-0.90) exhibited a strong positive relationship with the intention to use AI-CDSSs. The findings support hypotheses 1 to 3 and 4.1. The relationship between facilitating conditions and actual use was not investigated in a sufficient number of independent samples (k<2). Accordingly, we could not address hypothesis 4.2. Regarding the additional predictors beyond the UTAUT, attitude (rc=0.63, 95% CI 0.52-0.73), trust (rc=0.73, 95% CI 0.63-0.82), and innovativeness (rc=0.54, 95% CI 0.43-0.64) exhibited strong positive relationships, confirming RQs 1, 2, and 5. Perceived risk (rc=–0.21, 95% CI –0.35 to –0.08) was weakly negatively related to use intention, supporting RQ 3. Although the estimate for AI anxiety was strong and negative (rc=–0.41), the 95% CI included 0 (–0.98 to 0.15). Accordingly, we cannot conclude that AI anxiety is related to use intention, thus not supporting RQ 4. The 80% CRs for effort expectancy (0.27-0.83), facilitating conditions (0.33-0.99), and AI anxiety (–0.81 to –0.01) were wide, suggesting the presence of moderators [,]. Finally, the intention to use AI-CDSSs was strongly positively related to the actual use of AI-CDSSs, confirming RQ 6 (3/18, 17% of independent samples; N=478; r=0.75; rc=0.85, SD 0.09, 95% CI 0.63-1.00, 80% CR 0.70-1.00; correlation between observed effects and the influence of the study design artifacts=0.44).
Table 1. Bivariate relationships between predictor variables and artificial intelligence–enabled clinical decision support system use intention (N=18).Predictor variableIndependent samples, n (%)Cumulative sample size, Nrarcb (SD; 95% CI)80% CRcCorrelation between r and statistical artifactsPerformance expectancy16 (89)32950.590.66 (0.13; 0.59 to 0.73)0.50 to 0.820.39Effort expectancy15 (83)30580.490.55 (0.22; 0.43 to 0.67)0.27 to 0.830.28Social influence15 (83)30580.570.66 (0.12; 0.59 to 0.72)0.52 to 0.800.46Facilitating conditions6 (33)10480.570.66 (0.23; 0.42 to 0.90)0.33 to 0.990.25Attitude9 (50)20480.510.63 (0.14; 0.52 to 0.73)0.45 to 0.800.43Trust10 (56)18400.660.73 (0.13; 0.63 to 0.82)0.55 to 0.900.35Perceived risk10 (56)2428–0.19–0.21 (0.18; –0.35 to –0.08)–0.45 to 0.020.39Anxiety3 (17)391–0.37–0.41 (0.23; –0.98 to –0.15)–0.81 to –0.010.38Innovativeness5 (28)8430.470.54 (0.09; 0.43 to 0.64)0.46 to 0.610.81aSample size–weighted correlation.
bSample size–weighted and reliability-corrected correlation.
cCR: credibility interval.
Relative Weight AnalysisIt was not possible to explore all 9 predictors in a single relative weight analysis because they were not investigated together in a sufficient number of independent samples (Table S5 in ). Accordingly, to answer RQ 7, we analyzed 1 model with only the UTAUT predictors () and 4 separate extension models consisting of 5 to 6 predictors (). In the initial model with only the UTAUT predictors, the combined effects of performance expectancy, effort expectancy, social influence, and facilitating conditions explained 50% of the total variance in the intention to use AI-CDSSs. Performance expectancy was the dominant predictor, accounting for 31% of the total variance explained, followed by social influence (28%), facilitating conditions (26%), and effort expectancy (15%). In the extension models, trust emerged as the most influential overall predictor of use intention (between 29% and 35% of the total variance explained). In all 3 models including trust, performance expectancy was the second most influential predictor (between 19% and 24% of the total variance explained). Facilitating conditions (between 20% and 25%) and social influence (between 14% and 21%) consistently explained additional variance in all extension models. In the extension models including trust and perceived risk as well as trust and anxiety, the regression estimate of effort expectancy became negative. Finally, AI anxiety and perceived risk negatively predicted use intention and accounted for 10% (AI anxiety) and 2% (perceived risk) of the total variance explained.
Table 2. Multiple regression models and relative weights for the Unified Theory of Acceptance and Use of Technology predictorsa.PredictorBb (SE)t test (df)P valueRaw RWcRSd RW (%)Performance expectancy0.31 (0.02)13.97 (1732)<.0010.1631.19Effort expectancy0.08 (0.02)3.56 (1732)<.0010.0815.2Social influence0.27 (0.02)12.29 (1732)<.0010.1427.91Facilitating conditions0.21 (0.02)9.33 (1732)<.0010.1325.7aF4,1732=429.28 (P<.001); R2=0.498.
bRegression estimate.
cRW: relative weight.
dRS: rescaled.
Table 3. Multiple regression models and relative weights for the Unified Theory of Acceptance and Use of Technology (UTAUT) and additional predictors.PredictorBa (SE)t test (df)P valueRaw RWbRSc RW (%)UTAUT extension (attitude and perceived risk; F6,1284=222.31; P<.001; R2=0.509)aRegression estimate.
bRW: relative weight.
cRS: rescaled.
Moderation Analysesshows the results of the meta-regression for continuous moderators. Regarding age, older participants showed a weaker relationship between social influence and use intention (B=–0.01, 95% CI –0.01 to –0.00), thus contradicting hypothesis 5.3, according to which this effect would become stronger with increasing age. The moderation effect is shown in Figure S2 in . Age did not moderate any other relationship, thus not confirming hypotheses 5.1 and 5.2. Gender did not moderate any of the relationships, thus not confirming hypotheses 6.1, 6.2, and 6.3. Experience and voluntariness of use were not investigated in a sufficient number of independent samples. Accordingly, we were unable to address hypotheses 7 and 8. Cultural individualism (RQ 8.3) as a contextual moderator that was measured continuously did not influence any of the relationships. Finally, publication year (RQ 9.1) as a methodological moderator that was measured continuously did not influence any of the relationships.
Table 4. Results of the meta-regression (N=18).Predictor variable and moderatorIndependent samples per moderator, n (%)B (SE; 95% CI)P valuePerformance expectancyThe Wald-type pairwise comparisons for each level of categorical moderators are presented in . We could not investigate RQ 8.1 because information about occupations was not provided in a sufficient number of independent samples. Regarding RQ 8.2, the type of AI-CDSS (diagnostic AI-CDSSs versus diagnostic and treatment AI-CDSSs) did not moderate the relationship between performance expectancy and use intention nor did it moderate the relationship between social influence and use intention. However, the positive relationship between effort expectancy and use intention was stronger for diagnostic AI-CDSSs than for AI-CDSSs that combined diagnostic and treatment recommendations (mean difference=–0.31, 95% CI –0.58 to –0.04). Finally, regarding RQ 9.2, we observed no differences between studies using the scale by Venkatesh et al [] and those using other measures.
Table 5. Wald-type pairwise comparisons of categorical moderators (N=18)a.
留言 (0)