This work required no ethical approval.

Studies were identified through standard literature searches of PubMed (MEDLINE), Embase, Scopus, Web of Science, the Cochrane Library, and the International Health Technology Assessment (HTA) Database from database inception to March 19, 2026.

Keywords and related terms for “discrete choice experiment” and “artificial intelligence,” including their synonyms and variations, were incorporated into the search strategy (the full search strategy is provided in Table S1 Multimedia Appendix 1). Only studies published in English were included.

After removing duplicate studies using EndNote, 2 reviewers (XZ and YJ) independently screened the titles and abstracts of an initial set of 50 articles to assess eligibility. Discrepancies were discussed until consensus was reached, leading to a refinement of the inclusion and exclusion criteria. The remaining articles were then screened independently by XZ and YJ, with any disagreements resolved by a third reviewer (SL). Studies were included if they (1) used a DCE methodology; (2) evaluated AI-enabled technologies; in this study, AI-enabled health care technologies refer to a broad range of applications that incorporate AI-based algorithms to support or deliver health care, including clinical decision-support systems, mobile health apps, chatbots, wearable devices, and AI-based diagnostic, screening, and management tools; and (3) focused on research in health care settings. Studies were excluded if they (1) did not use DCE or did not investigate AI-related technologies, (2) were review articles, or (3) were not published in English. This approach ensured the systematic identification of primary research capturing stakeholder preferences toward AI-enabled health care technologies.

Two researchers independently extracted detailed data using a standardized data extraction form. The extracted information included study characteristics (publication year, country, sample size, disease area, AI device type, and clinical application), DCE design characteristics (perspective, number of choice sets per block, attributes, and levels, experimental design methods, and econometric analysis methods), as well as the categories of DCE attributes. In addition, study outcome measures were extracted, including attributes identified as the most and least important in the DCE results. Monetary and nonmonetary preference valuation measures reported in the included studies, including willingness to pay (WTP), willingness to save (WTS), and other trade-off estimates, were extracted as originally reported and synthesized descriptively. Any disagreements were resolved through discussion and reexamination of the original articles, and a third reviewer was consulted if necessary.

Reporting quality was assessed using the DIRECT checklist (Checklist 2) [19], which evaluates 7 reporting domains: purpose and rationale, attributes and levels, experimental design, survey design, sample and data collection, econometric analysis, and reporting of results, comprising a total of 26 items. Each item was graded as “yes” or “no”. Two independent reviewers (XZ and YJ) conducted the assessment, with a third reviewer (SL) resolving any disagreement.

A narrative synthesis was conducted to summarize the findings of the included studies. Extracted information was compiled and managed in Microsoft Excel to facilitate comparison across studies. Reported DCE attributes were reviewed and grouped into broader thematic categories to capture the key characteristics. Given the lack of a standardized AI-specific framework, a structured classification system was developed by integrating established health technology assessment (HTA) frameworks, including the HTA Core Model, ISPOR (International Society for Pharmacoeconomics and Outcomes Research) Value Assessment Framework, and NASSS (Non-adoption, Abandonment, Scale-up, Spread, Sustainability) framework. These frameworks were selected due to their widespread use in HTA and shared focus on value-based decision-making and real-world implementation. Attributes were mapped to the integrated framework based on conceptual alignment, and their frequency was calculated according to occurrences in DCE designs. Attributes with potential overlap across domains were independently reevaluated by 2 reviewers based on their conceptual meaning and study context. Discrepancies were resolved through discussion and, if consensus could not be reached, adjudicated by a third reviewer. Attributes that remained conceptually ambiguous were further reviewed through consultation with experts in health preference elicitation, HTA, and AI to ensure consistent and conceptually appropriate domain assignment. Preference patterns were further assessed using relative importance (RI) measures to identify the most and least valued attributes. When RI was not explicitly reported, it was derived from utility coefficients (β) using the utility range method. Specifically, RI was calculated as the range of utility estimates within each attribute (eg, the difference between the highest and lowest β coefficients across attribute levels) divided by the sum of all attribute ranges within the same study. For continuous attributes, RI was calculated by multiplying the regression coefficient by the observed range of attribute levels [20]. Given differences in attribute definitions, coding schemes, and model specifications across studies, RI was used to describe within-study preference structures rather than to enable direct comparison of magnitude across studies.

A total of 425 studies were identified. After removing duplicates and screening titles and abstracts, 37 full-text articles were assessed. Of these, 10 were excluded (7 not using DCEs, 2 not focusing on AI technologies, and 1 not in health care). Ultimately, 27 studies comprising 28 DCEs (1 study included 2 DCEs [21]) were included (Figure 1). For methodological and preference analyses, independent DCEs were treated as separate units (n=28). Application-context analyses were conducted at the level of individual AI applications (n=30 application contexts), as some DCEs assessed more than 1 application setting. Attribute analyses were based on all extracted attribute occurrences across included DCEs (n=163 attributes). The studies were published between 2021 and 2026 and conducted across multiple countries, with China (n=7, 25.93%) and Australia (n=6, 22.22%) contributing the most. Sample sizes ranged from 39 to >2000. Most focused on general health services (n=10, 37.04%) or cancer care (n=7, 25.93%), with screening (n=8, 29.63%) and decision support (n=7, 25.93%) as the most common applications. Other contexts included diagnosis and disease management.

Most studies evaluated nontangible AI systems integrated into clinical workflows (n=15, 55.56%), including tools, systems, and programs, while others involved apps, wearable devices, and chatbots. Preferences were mainly elicited from the public (n=16, 59.26%), followed by health care professionals (n=7, 25.93%) and patients (n=5, 18.52%). One study implemented 2 DCEs involving both the public and professionals [22], while another study conducted a single DCE across 4 scenarios [23]. Across the 28 DCEs, the number of attributes ranged from 3 to 8, with 6 being the most common (n=13, 46.43%; Table 1).

A 5-domain classification framework was developed through systematic extraction and inductive synthesis, categorizing attributes into AI technology performance, cost, ethical concerns, usability, and clinical applications. The domains of performance, cost, and ethical concerns broadly align with established HTA frameworks (HTA Core Model and ISPOR), with adaptations to reflect AI-specific features such as the “black-box” nature and algorithmic risks. Usability was informed by the NASSS framework, while clinical applications correspond to the “Health Problem and Context” domain, capturing preferences across application settings. Each domain was defined a priori, and detailed operational definitions and examples of included attributes are provided in Table S2 in Multimedia Appendix 1. To minimize overlap, attributes were assigned to a single domain based on their primary conceptual focus as described in the original study context.

A total of 163 attributes were identified and mapped into 5 attribute domains. Table 2 presents the study-level attributes, including their domain classification and preference rankings (most and least important attributes), whereas Figure 2 summarizes the overall frequency of attribute inclusion across domains and subcategories among all included DCEs. Usability-related attributes were most frequently included (n=60, 36.81%), followed by performance-related attributes (n=45, 27.61%). Within these domains, presentation format (n=27, 45%) and effectiveness (n=28, 62.22%) were the most common subattributes.

Preference outcomes were analyzed at the application context level, with 30 disease-application contexts identified across the included studies. Across all included application contexts, performance attributes were most frequently identified as most important (18/30), particularly effectiveness (17/30), whereas usability attributes were most often rated as least important (14/30), mainly driven by presentation format (9/30). Effectiveness consistently emerged as the most important attribute across stakeholder groups. Clinicians additionally emphasized interaction and response time, while patients and the public showed stronger preferences for cost, with the latter also valuing responsibility. Presentation format was the least important attribute across all groups.

When further stratified by clinical application context, among the 30 disease-application combinations, 1 study (Ploug et al [37]) did not report clinical application and was excluded from this analysis, resulting in 29 contexts. These were categorized into screening (n=8, 27.59%), diagnosis (n=7, 24.14%), decision support (n=7, 24.14%), treatment (n=4, 13.79%), and management (n=3, 10.34%). Performance-related attributes remained dominant in diagnostic (6/7), screening (4/8), treatment (2/4), and management settings (2/3), while decision support systems showed greater variability in attribute prioritization, and usability-related attributes (4/7) were most frequently identified as the most important domain. Usability attributes were consistently ranked as least important in diagnostic, screening (4/8), decision support (3/7), and treatment (3/4) settings, while ethical concerns (2/3) were ranked as least important in management-related contexts.

Ten (37.04%) studies quantified the economic value of DCE attributes through monetary WTP (n=8) or WTS (n=2), with time as the trade-off metric. The highest and lowest marginal WTP within each study are extracted in Table 3. Among the 7 studies using monthly or per-visit cost metrics, the highest WTP was associated with performance and usability attributes. Participants were willing to pay an additional US $0.08 to US $0.26 for every 1% increase in diagnostic accuracy, US $2.24-US $3.25 for AI interactions, and up to US $79 for reduced decision uncertainty. Conversely, the lowest WTP was associated with usability (presentation format) and ethical concerns, ranging from US $0.02 to US $2. Reflecting an annual cost scale, one study reported higher baseline values, peaking at US $29,244.8 for outcome integration (performance) and dropping to US $4162.6 for local information (ethical concerns). Two studies reported trade-offs using WTS, and the highest WTS was consistently associated with performance, for 0.7 minutes per 1% increase in accuracy and 111 minutes for maintaining current diagnostic accuracy. In contrast, the lowest WTS was observed for decision uncertainty and ethical concerns, at 32 and 0 minutes, respectively.

Reporting quality was assessed using the DIRECT checklist (Figure 3). Across the 27 studies, the mean completion rate for the 26 items was 84.47% (SD 11.49%), ranging from 50% to 100%. Four studies achieved complete reporting (100%), 2 of which explicitly referenced the DIRECT checklist.

Among the 7 reporting domains, “Purpose and rationale” and “Attributes and levels” were fully reported in all studies (n=27, 100%), while the “Sample and data collection” reporting domain also showed high completeness (n= 25, 92.59%). In contrast, several items were less consistently reported. The lowest reporting rate was observed for item 8 (identification of design effects) in the experimental design domain (n=10, 37.04%). The reporting of randomization in the survey design domain was also limited (n=14, 51.85%). Other less frequently reported items included model performance (n=15, 55.56%) and the use of pilot study information (n=17, 62.96%). Overall, although reporting quality was generally high, items related to experimental design and survey procedures were less consistently reported. Detailed results are provided in Checklist 2.

This review provides the first comprehensive synthesis of DCEs on stakeholder preferences for AI-enabled health care technologies, with a marked increase in studies since 2021. Across different stakeholders, effectiveness consistently emerged as the primary driver of preferences, underscoring the importance of performance in the acceptance of AI in health care. While reporting completeness was generally high, the findings suggest that further improvements are needed to better align study designs with stakeholder priorities.

The consistent prioritization of effectiveness across stakeholder groups suggests that clear clinical benefit is essential for the acceptance of AI-enabled health care technologies [30,34,35]. This may be due to ongoing uncertainty about their real-world performance and reliability. Across the included DCE studies, differences in the most important attributes appear to reflect role-specific concerns. Clinicians tend to value interaction and response time, likely because AI systems need to fit into clinical workflows without increasing the time burden [48]. In contrast, patients and the public place more emphasis on cost, reflecting concerns about financial burden and value [49]. Although effectiveness was most frequently identified as the dominant attribute, this finding should not be interpreted as indicating a universal preference across AI-enabled health care technologies. Our review suggests that preference structures vary according to application context. Performance-related attributes were consistently prioritized in diagnosis, treatment, and disease management, whereas usability assumed greater importance in decision-support applications. These findings indicate that the relative importance of AI attributes is shaped by the intended use of the technology rather than stakeholder type alone. Other contextual factors, including AI functionality, target population, and health care setting, are also likely to influence preferences, although current evidence remains insufficient to evaluate their effects systematically. Future DCEs should examine these sources of contextual heterogeneity to improve the generalizability and applicability of preference evidence.

This study suggests that attributes commonly included in AI-related DCE designs may not fully align with those most strongly valued by stakeholders. Usability-related attributes are frequently incorporated but are often of limited importance, whereas effectiveness consistently emerges as the primary driver. This pattern is consistent with prior AI-related DCE evidence, such as studies in screening contexts where usability-related attributes, such as interaction, are commonly included [18]. However, compared with traditional health care DCEs, where attribute selection and preference evaluation have generally focused more consistently on performance-related dimensions such as clinical outcomes, AI-related DCEs appear to place greater emphasis on usability and process-oriented attributes [50-52]. This pattern was also reflected in WTP findings, where higher monetary or time trade-offs were more frequently associated with performance-related improvements, particularly diagnostic accuracy and reduced uncertainty, whereas presentation-related attributes generally attracted lower values. AI-related DCEs appear to place greater emphasis on usability and process-oriented features at the design stage. This suggests a limitation in current DCE practice in AI-enabled health care, where attribute selection may not adequately reflect stakeholder priorities [53].

Additionally, the lower importance of usability attributes may partly reflect design-related effects in some studies. In DCEs, attribute importance depends on both attribute selection and the utility variation implied by attribute levels. For example, in a study by Liu et al [34], usability attributes, such as response time, were defined using relatively narrow level ranges (eg, diagnosis time categories of 0, 15, and 30 minutes), compared with wider ranges for performance attributes (eg, diagnostic accuracy from 60%-100%), which may reduce observed utility differences and attenuate their estimated importance. These design factors should be considered when interpreting DCE-based preference evidence. Importantly, observed differences across stakeholder groups may also reflect variation in study context, attribute selection, AI application type, and sample composition, rather than solely true stakeholder-level preference heterogeneity.

Mean reporting completeness was relatively high (84.47%, SD 11.49%), consistent with prior AI-related health economic reviews reporting rates of 66% to 77.4% [54-56]. However, key methodological elements, particularly those related to experimental design, were insufficiently reported. For example, the identification of design effects (n=10, 37.04%) and the reporting of randomization procedures (n=14, 51.85%) were limited, aligning with previous findings of substantial gaps in analytical planning, with reporting rates as low as 0% and 10.52% [54,55]. Similar reporting deficiencies have also been identified in non–AI-related DCE reviews, where the ISPOR checklist was applied to evaluate reporting quality and revealed insufficient reporting of key experimental design elements [57,58]. It is important to emphasize that the DIRECT checklist assesses reporting completeness rather than methodological quality. Accordingly, missing reporting does not imply that key design features such as randomization, pilot testing, design-effect assessment, or model diagnostics were not undertaken in the original studies. Rather, incomplete reporting limits the ability to fully appraise the link between study design and estimated preferences, which may reduce confidence in the interpretation of trade-offs and WTP estimates [59]. This issue may be further exacerbated in AI-related DCEs, where increasing model complexity may outpace reporting transparency, thereby constraining the evaluation of methodological justification [60,61].

Model performance was reported in only 55.56% of the studies, restricting the evaluation of model fit and comparison across specifications. However, this finding reflects limitations in reporting rather than direct evidence of inadequate model performance or inappropriate model selection. Most studies relied on conditional and mixed logit models, with limited use of multiple models to assess robustness or explore heterogeneity. While some studies accounted for preference heterogeneity, scale heterogeneity was rarely addressed [34,59,62]. This is particularly relevant in AI contexts, where system-embedded and less observable technologies may increase the difficulty of attribute interpretation, leading to greater response variability. As a result, observed variation may partly reflect response noise rather than true preference heterogeneity, potentially biasing results [60,61,63]. The use of more flexible and complementary modeling approaches may improve robustness [64-66].

In the absence of an AI-specific evaluation framework, we integrated the HTA Core Model, ISPOR Value Assessment Framework, and NASSS framework to categorize attributes. HTA and ISPOR capture clinical, economic, and value dimensions, while NASSS addresses the sociotechnical complexities of technology adoption [67-69]. This integrated approach is well suited to AI, as it incorporates features such as limited explainability, decision uncertainty, and algorithmic fairness that extend beyond traditional domains. Although AI-related DCEs share similarities with conventional health care studies, they additionally reflect interaction dynamics and “black-box” concerns. By combining value assessment with implementation complexity, this framework provides a structured and relevant basis for synthesizing DCE evidence and informing the design of AI-enabled health care services.

These findings also have implications for health care decision-making. For developers and health care providers, understanding which AI attributes are most valued may support user-centered technology design and implementation strategies. For regulators, payers, and HTA agencies, preference evidence may help identify technology characteristics that are most relevant to stakeholder acceptance and real-world uptake. Incorporating stakeholder preferences into the evaluation of AI-enabled health care technologies may therefore facilitate more patient-centered and context-sensitive decision-making.

Several limitations should also be noted. First, the proposed attribute classification framework, although informed by the HTA Core Model, ISPOR, and NASSS frameworks, may not fully capture all relevant dimensions of AI-enabled health care. In addition, as the framework is derived from attributes reported in existing DCE studies, it reflects what has been operationalized in preference research rather than the full spectrum of characteristics of AI-based health care services. Second, reporting quality was assessed using the DIRECT checklist, which focuses on transparency rather than methodological quality. As such, while reporting gaps can be identified, this approach may not fully capture all sources of bias in study design and analysis. Third, only English-language publications were included, which may have resulted in the omission of relevant studies published in other languages. Finally, this review synthesizes stakeholder preferences based on DCE evidence. Although DCEs are well suited for eliciting trade-offs, they represent only 1 preference elicitation approach. Moreover, all included studies elicited stated preferences using hypothetical choice scenarios, and none examined the relationship between stated preferences and may not fully reflect real-world decision-making, particularly for complex and less observable AI technologies. Nevertheless, this review provides a structured synthesis of stakeholder preferences and reporting practices in AI-related DCEs. Future research should expand attribute identification beyond existing DCE designs and adopt complementary methods to better capture the complexity and real-world characteristics of AI-enabled health technologies.

This review identifies a mismatch between DCE attribute design and stakeholder preferences in AI-enabled health care. Effectiveness generally emerges as a key determinant of preferences, particularly in clinically oriented settings, while usability and performance attributes vary in importance across AI applications. Preference structures are context-dependent rather than uniform across technologies. Although overall reporting completeness was high, key methodological elements such as design effects and randomization were inconsistently reported, indicating room for improvement in study design and reporting transparency. Strengthening methodological rigor and aligning attribute selection with decision contexts may enhance the interpretability of DCE evidence and support the development of AI technologies that better reflect real-world health care decision needs.