Clinical decision support systems (CDSS) are designed to assist health care professionals in diagnosis, treatment selection, and patient monitoring. These systems aim to improve patient safety and optimize clinical workflows [1-3]. With increasing digitalization and advances in AI, CDSS have increasingly incorporated AI-based technologies to enhance their potential. These systems can access large volumes of digital health data, analyze historical information, continuously learn from new inputs, and provide real-time, personalized recommendations for clinical practice. By identifying patterns in complex datasets, AI enables more precise and individualized decision-making, thereby improving diagnostic accuracy and treatment quality [4,5].

Despite these promises, the implementation of AI-based CDSS in routine clinical care remains limited. Barriers include not only well-known challenges, such as insufficient interoperability, inadequate integration into existing IT infrastructures, and legal uncertainties (eg, data protection and liability), but also a lack of alignment with actual clinical workflows. Many systems fail to reflect the practices, routines, and decision-making logics of their intended users [6,7]. Specific challenges associated with AI-based systems, such as limited data quality, poor model transparency (the “black-box” problem), and ambiguous accountability in AI-supported decisions, further complicate implementation, especially in nursing contexts where documentation is often fragmented and lacks standardization [8,9]. Studies indicate that insufficient user involvement in system development is a key reason for the failure of technological implementations in health care [10].

Participatory approaches offer a promising strategy to address these challenges by involving users early in the development process and systematically integrating experiential knowledge, everyday practice needs, and organizational conditions [11,12]. Such approaches, particularly participatory and user-centered design, aim to develop technologies in close collaboration with future users. Through mutual learning, iterative adaptation, and shared responsibility, these processes aim to produce systems that are not only technically functional but also contextually relevant, acceptable, and usable [13,14]. Participatory and multiperspective methods have proven effective in increasing acceptance and capturing contextual factors in technology design [15-17]. In the field of AI in health care, recent studies show that participatory approaches can enhance data quality, system design, and clinical applicability by incorporating domain knowledge and user expertise [11,12]. Particularly when interpreting complex clinical data, for example, discipline-specific terminology or documentation practices, the user perspective serves as a crucial point of orientation for technical development [11].

At the same time, the literature emphasizes that participatory approaches are not exclusively viewed as beneficial and have also been subject to critical debate. Participation processes may take on a symbolic character (“tokenism”) when participants are formally involved but have limited influence on actual decision-making processes [18]. Moreover, research in design and human-computer interaction (HCI) highlights that user involvement can be instrumentalized, for example, when participant feedback is used to retrospectively legitimize design decisions that have already been made [19]. In addition, participatory processes may be selective, allowing certain stakeholder groups to dominate while other relevant perspectives remain underrepresented [20].

The application of participatory methods in the development of AI-based CDSS remains an emerging and dynamic field of research, presenting specific opportunities and challenges. How can collaboration between AI developers, clinicians, and other health care professionals be structured to accommodate the complexity of algorithmic systems? Which participatory methods are most appropriate for integrating clinical expertise and ensuring alignment with real-world decision-making processes? And how can ethical, legal, and social implications, such as algorithmic transparency, bias mitigation, and accountability, be addressed during system development? These questions are central not only to research but also to the responsible implementation of AI in clinical care. Establishing an iterative feedback loop in which developers and users regularly exchange insights may help ensure that systems evolve in tandem with clinical needs and remain contextually grounded over time.

This review aims to address a significant gap in the existing literature by offering a comprehensive summary of the approaches, methods, and practices of participatory research and development in the context of AI-based CDSS for clinical staff. To achieve this, the review systematically examines empirical studies, their theoretical frameworks, and applied methods for the participatory design of AI-based CDSS. By reviewing the current state of user involvement in the development, design, and implementation of these systems, the review will provide valuable insights into how participatory approaches have been used to date and how they can contribute to successful development and implementation.

The objective of this review is to provide an overview and systematization of participatory approaches from various disciplines for developing, piloting, implementing, and evaluating AI-based CDSS in clinical settings.

The following research questions were addressed:

The protocol for this scoping review was previously published in JMIR Research Protocols [21]. It outlines the review’s parameters and provides justification and explanations of all methodological steps and decisions. To further ensure rigor, we used the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) checklist [22] and reported it in Checklist 1.

The inclusion and exclusion criteria that were applied to the studies are shown in Textbox 2.

We used Zotero (a free and open-source literature management program) [34] as a bibliographic tool. The retrieved references were checked for duplicates and then transmitted to Covidence (software for managing and streamlining reviews, operated by Veritas Health Innovation Ltd) [35] for the screening steps. Moreover, 2 independent reviewers (TR and PG) screened all titles and abstracts separately for inclusion or exclusion. Disagreements were solved by including a third reviewer (CK or PK). The same procedure was then applied to full-text screening, which was carried out by 3 independent reviewers (TR, PG, and CH). Reasons for excluding a study were assessed in each of these steps.

Two reviewers (TR and PG) collaboratively developed a data-charting form to determine which variables to extract. The form comprised seven sections: (1) metadata, (2) description of the contribution, (3) conceptual meaning and attribution, (4) theory and methodology, (5) implementation of participatory or cocreative processes, (6) evaluation and outcomes, and (7) miscellaneous. The data were independently charted by 3 reviewers (TR, PG, and CH), who then discussed the results and continually updated the data charting form in an iterative process. To ensure methodological consistency and quality, all reviewers screened 1 initial publication, discussed discrepancies, and jointly revised the data extraction manual before commencing formal screening. Disagreements were resolved by consensus or, when necessary, through consultation with the remaining authors (PK and CK), until full agreement was reached.

If information on specific participatory procedures, process characteristics, or content addressed within participatory activities was not explicitly described in a publication, this was documented as not reported (N/A) rather than coded as absent. This decision reflects the fact that publications often present only a selective account of the broader development process, and the absence of reporting does not necessarily indicate that a given activity did not take place or that a specific issue was not discussed.

In accordance with JBI guidance, the extraction form was iteratively refined compared with the version published in the review protocol [21,28,29]. Adjustments included expanding metadata categories, providing more detailed documentation of participants, and placing greater emphasis on participatory processes and contextual conditions. Items such as “Generic term for the participatory/co-creative process” and “Degree of participation/Grading” were not retained, as they were either not reported in the studies or would have required substantial interpretation. The final version of the extraction form is available in Multimedia Appendix 2.

Data were extracted on publication characteristics (eg, first author, year, journal, and country), study design (eg, clinical setting, target populations, and project phases), as well as key features of stakeholder engagement (eg, involved roles, recruitment procedures, and participatory or cocreative methods applied). Contextual and technical aspects were documented through detailed descriptions of the nature, function, and intended use of the AI-based CDSS. Barriers and facilitators to engagement were identified based on reported structural conditions and study design characteristics. Where applicable, we also extracted reported outcomes and perceived effects of participatory approaches on system development.

The extracted data were synthesized using a descriptive qualitative approach and organized deductively according to the 4 review questions, enabling a structured comparison of findings across studies.

The systematic search yielded 4421 records from 4 scientific databases, supplemented by 40 additional records identified through citation searching. After removing 143 duplicates, 4318 records were retained for title and abstract screening. Of these, 64 studies were reviewed in full text. Twenty-seven were excluded from the study, primarily due to the lack of participatory elements (n=10) or the absence of an AI component (n=7). A total of 37 studies [36-72] satisfied the inclusion criteria and were thus incorporated into this review (Figure 1).

The 37 included studies represent a diverse range of research projects aimed at developing, testing, or evaluating AI-based CDSS involving clinical staff. Multimedia Appendix 3 provides an overview of key characteristics.

The analysis of the studies reveals considerable conceptual heterogeneity regarding key terms such as participation, AI, and CDSS.

The studies are sometimes not explicitly labeled as “participatory” or “cocreative” [36-47], yet they still report forms of stakeholder involvement in system development. More frequently, they adopt a user-centered design or “iterative or user-engaged” framing [38,39,41,48-59]. Explicit co-design, cocreation, or participatory-design terminology appears in 8 studies [60-67], whereas normative involvement labels such as “participation,” “consultation,” “collaborative partnership,” and “shared leadership” appear in 4 studies [68-71]. Across the included studies, however, explicit reference to participation remained limited, and descriptions of participatory processes were often brief or selective. In many cases, stakeholder involvement was identifiable through reported activities rather than through explicit conceptual framing or detailed methodological reflection. Accordingly, the labels used across studies did not consistently correspond to comparable forms of stakeholder involvement, and in several studies, interviews, focus groups, or usability procedures constituted the primary reported mode of involvement without being embedded in explicit co-design activities. Several included studies would not necessarily qualify as participatory in a strict definitional sense; rather, they reported more limited or consultative forms of stakeholder involvement. This ambiguity was closely related to the often sparse and selective reporting of involvement processes.

The systems described in the included studies primarily focused on predictive risk assessment in specific clinical contexts, such as stroke rehabilitation [60], heart failure [41], and wound healing [52]. Technological approaches ranged from machine learning [42,43,48,49,58-62,65,68-70] to deep learning [36,44,46,47,50,51,56,63,64,72] and hybrid models [37-39,55,57,66], while several studies did not specify in depth their underlying approach [41,45,53,54,67,71]. The level of technical detail varied considerably; while several studies provided comprehensive descriptions of model architecture and data sources [36,37,44,47,49,51,56,58,60,62,66,69], others remained vague or focused primarily on application aspects [38-40,42,43,45,46,48,52-55,57,59,61,63-65,67,68,70,71]. Similarly, the term “AI” was used inconsistently, ranging from precise model or explainable AI conceptualizations [36,37,39,44,47,50,51,53,54,56,58,60,62,63,65,66] to broad references to automated or learning-based systems [38,40,42,43,45,46,55,57,59,64,67,68,70,71].

Functionally, the systems were consistently described as tools to support clinical staff in making patient-specific decisions. Reported aims included improving decision quality and accuracy, early or more objective detection, prediction, or diagnostic support [40,42,44-47,49-52,55-60,64-66,68-71], facilitating integration into clinical workflows [38,41-48,51,53-57,59,61-66,68-71], and supporting shared decision-making between physicians and patients [36,43,58,59,67,72], as well as across multiple stakeholder groups, including physicians, patients, carers, and health professionals in integrated care settings [39,42].

Several studies examined homogeneous groups of a single profession, such as psychiatrists [36,72], physicians [44,53,55,57,59], technicians [50], or nurses [52]. In contrast, other studies relied on heterogeneous, multiprofessional samples combining physicians, nurses, resident physicians, or therapists [37,39,41-43,45-49,54,56,58,61,63-67,69-71]. In most cases, the participating professionals mirrored the intended users of the system. Overall, the included studies predominantly involved physicians [36,38-41,43-49,51,53,55,57,61-66,68-72], followed by nurses [39-43,46-49,52,54,56,58,61,70,71]. Less frequently, therapists such as physiotherapists, occupational therapists, psychologists, or speech therapists were included [37,39,42,60,61]. In more detail, for example, Aquino [42] involved physiotherapists, occupational therapists, and speech therapists, Lee et al [60] involved occupational and physical therapists, Lee et al [37] mentioned therapists without specifying their discipline, Nemeth et al [61] involved physical, occupational, and respiratory therapists, and Timotijevic et al [39] included physical, occupational, speech therapists, and psychologists. Similarly, the participant groups also included specialized technicians [50] and, in another single instance, social workers and ventricular assist device coordinators [41], while several other studies further broadened the stakeholder spectrum by involving pharmacists, radiographers, managers, AI or HCI experts, patients, and patient or carer representatives [42,43,46,56,58,63-65,67,71].

Sample composition also varied by study phase. Needs assessments ranged from smaller expert groups or profession-specific samples, such as technicians, therapists, or physicians [50,51,60,69], to broader multiprofessional stakeholder constellations, for example, health and social care professionals, intensive care unit (ICU) and imaging staff, or oncology teams [42,45-47,64,66,70,71]. Physicians and nurses dominated design phases, sometimes also involving therapists [36-39,41,48,49,51,53,60-62,69], and, in some studies, being complemented by technical experts, patient and public contributors, or carers [43,56-59,63,65,67,71]. Evaluation phases ranged from profession-specific assessments to larger multiprofessional samples, involving physicians, nurses, residents, therapists, and other stakeholders [36,37,39-41,43-45,47,50-52,54-56,58-61,69,72]. In addition, some studies also involved nonclinical actors, such as designers, engineers, managers, AI or HCI experts, software developers, and nonclinical stakeholder groups including patient and public contributors, carers, and formal patient or carer representatives [43,51,56,63,65,67-69,71,72].

Across the studies, the breadth of stakeholder inclusion varied considerably. While several studies involved multiple professional groups relevant to the clinical workflow, others focused on only 1 professional perspective, even in contexts where broader interprofessional involvement appeared plausible. In most cases, participation remained closely aligned with the intended primary users of the CDSS. In contrast, actors indirectly affected by system use or embedded in adjacent care processes were less frequently included.

Multimedia Appendix 4 provides an overview of the participating groups, their roles, and the respective study phases. The column “participating groups” indicates the stakeholder groups involved in the studies, whereas “description of participants” outlines the actual study samples.

The analysis of all included studies shows that user involvement in CDSS development is, in most cases, supported by theoretical, methodological, and empirical foundations. Many studies [36-38,40-42,47-52,54,55,58,60-62,66,68,71] referred broadly to concepts from human-computer interaction or general usability and user-centered design principles to motivate user involvement, typically in a general way and without linking these concepts to concrete participatory procedures. In total, 9 studies [39,56,59,62-65,67,70] explicitly connected theoretical models with the design of participatory processes. Methodological justifications were reported relatively consistently across the studies; however, they primarily referred to user-centered design activities, such as usability testing, think-aloud protocols, or requirements elicitation, rather than to participatory approaches in a stricter sense. In almost all studies [36-43,46-52,54-68,71,72], these methods were justified in functional terms, for example, to gather requirements, improve usability or explainability, or adapt the system to the clinical context. In contrast, explicit rationales for participatory involvement were rarely provided. However, eleven studies [47,51,56,58,59,61,62,64,66,67,71] included broader methodological reflections on the combination or sequencing of methods. Most included studies were empirically grounded, drawing on primary data or existing evidence to substantiate the value of user involvement.

The implementation of participatory and user-centered approaches varied across the studies. Several studies used interviews, focus groups, or workshops to engage clinical experts [36,38,39,41-47,49-51,53-61,63-72]. Additionally, some studies combined qualitative methods with usability and user experience (UX) techniques such as think-aloud protocols, task-based testing, and standardized questionnaires [37,44,45,47,48,50-53,55,58-60,62,65,66,69,72]. Iterative testing and feedback formats with repeated review cycles were also used [38-41,44,50,51,54,56-59,62,63,66,68,71]. Some studies also incorporated observational methods [40,41,50,51,60,61,72]. Finally, systems were evaluated in simulations or near-clinical scenarios [36,44,45,48,52,55,59,62,66,72]. Taken together, the reported activities ranged from formative inquiry (eg, interviews, focus groups, and observations) to more design-oriented collaborative formats and evaluation-oriented involvement (eg, think-aloud procedures, usability testing, and simulation-based feedback).

Reporting on participatory processes themselves was generally limited across the included publications, and details on power relations, possible biases, participation burdens, or supportive measures were often lacking. However, a few studies documented such aspects in more detail. Grant et al [70] reported introductory activities to support participants’ engagement and reflected on the practical burden of aligning participatory activities with clinical work schedules. Jung et al [54] likewise described supportive measures, including adapted session formats and separate walkthroughs for physicians and nurses to reduce hierarchical bias. Samimi et al [58] used structured walkthroughs to help participants engage with interface components step by step, while Fouad et al [63] deliberately balanced expertise across groups. In addition, Jung et al [53] reflected on how different participatory formats shaped participants’ ability to engage with future use scenarios. Overall, these remained exceptions, while most studies provided little information on how participation was enabled, how burdens were managed, or how power imbalances were addressed.

Several supportive conditions were identified. Recurrent facilitators included regular feedback loops or iterative refinement [37,39,41,47-51,54,56,59,62,68,71] and the use of real-world clinical cases [37,44,47,48,51,55,59,60], and interdisciplinary collaboration [38,46,49,51,58,59,63,66,68,69,71]. Some studies have also suggested that representations that are visually intuitive or cognitively accessible are beneficial [40,41,46,49,51,54,57,59,61,66,68]. Additional enabling conditions included the involvement of participants with domain-specific expertise [42,47,58,66], skilled facilitation and methodological triangulation [58,70], training, onboarding, and AI literacy support [53,54,56,70], close collaboration between clinical and technical or informatics teams [46,56,66,71], procedural flexibility to accommodate clinical workloads [46,53,54,71], and previous UX with related systems [47]. In addition, several studies reported compensation or protected time for participation, including paid working time or direct financial compensation [45,55,56,58,59].

At the same time, the studies highlighted several recurring challenges. Time and organizational constraints [36,38-41,48,51,53,54,60-62,65,66,69,70] frequently limited the depth of user involvement. Strategies to mitigate this included condensed feedback formats, asynchronous data collection or adjusted study designs to work around clinical schedules [36,38,41,45-47,53,54,56,61,69]. Skepticism and uncertainty toward AI, especially in early development phases, were also reported in the studies [39-42,44,45,47,51,59,62,62,67]. These concerns were addressed through explainability-oriented model or interface design, stakeholder-oriented design processes, and professional training or information adaptations [39-42,44,46,47,51,54,58,59,62,63,67]. Other methodological limitations or concerns included small or single-site samples, artificial test environments, recruitment bias, and prototype-stage evaluation [36,38,40,41,43-48,50,52-55,58-60,64,66,71,72]. These limitations were acknowledged in most cases, with some authors suggesting follow-up studies in real-world or multicenter settings [41,44-47,50,53,55,59,66,72]. In some studies, the inclusion of different disciplinary perspectives also led to coordination challenges and tensions arising from the diversity of stakeholder groups or the incomplete involvement of relevant groups [38,39,45-48,59,61,64,66,67,70,71]. Balancing model performance, explainability, and usability also emerged as a recurring trade-off [40,41,44-47,51,53,59,62,66,67].

Despite these challenges, many studies reported positive experiences with participatory development. User involvement supported the identification of practice-relevant requirements [36,38,41-49,51,53,57-59,61-67,70-72], facilitated contextual adaptation [39,41,42,44,46,47,50-52,55,58-61,63-66,69-71], and helped improve usability and interpretability through iterative testing and feedback [36,37,40,43,44,47,48,51,53-55,58,59,62,63,65,66,72]. In some studies, active cocreation appeared to help foster trust [37,39,40,44,51,56,58,63,65]. Additionally, several studies reported joint prioritization of features [38,39,47,49,51,53,57-59,63,66,67]. Several studies further described participatory processes as useful for surfacing mismatches between technical assumptions and clinical reasoning and for translating these into concrete design revisions, including shifts in system logic or interface structure [45-47,49,50,58,59,66,68].

Overall, the analysis highlights both the potential and the limitations of participatory development in the context of clinical AI. Although some studies offer practical strategies such as adaptive design processes or structured communication, persistent challenges—limited clinical validation, restricted generalizability, and structural resource constraints—show that important underlying tensions remain unresolved in the development of participatory AI systems in health care.

The analysis shows that participatory methods can contribute in multiple ways to the development and refinement of AI components in CDSS. In 6 studies [37,46,50,56,64,69], the structured involvement of clinical professionals supported the identification of practice-relevant variables for model training. This was achieved through interviews, observations, journey mapping, contextual inquiry, workflow analysis, co-design workshops, focus groups, think-aloud procedures, iterative design processes, prototyping, and user studies. Additionally, participatory formats were used to co-develop visual and explanatory components that made AI outputs more transparent, interpretable, and clinically meaningful [37-39,41,44,48-55,58-60,62,63,65,66,72]. Methods used for this purpose included interviews [38,44,50,59,69], focus groups and iterative design sessions [41,49,57,63], think-aloud, observations and user studies [44,48,50,55,66], as well as simulation- or workshop-based formats (eg, using paper prototyping or probing cards [36,38,58,63,65], usability or cognitive walkthroughs [54,66], and heuristic evaluation [66]). Together, these approaches were used to elicit clinical reasoning and information needs, refine explanatory displays and interface elements, and test their comprehensibility and practical fit in more practice-oriented settings. Participatory methods also supported the alignment of AI-based components with clinical workflows and implementation contexts. Illustrative examples include simulation-based testing with standardized patients [36], which informed tool refinement and clinician training materials; focus groups and iterative design sessions [49], which shaped GUI adaptations; interviews and a workshop [38], which informed task models and user requirements; and iterative design activities across phases [41], which supported prototype refinement toward routine care.

A further contribution of participatory processes concerned data quality. Although data quality concerns were addressed in many studies, they were usually not framed as a distinct methodological focus of participation. In several studies [37,39,41-43,46-48,50,51,53,55,56,58,59,64,68,70,71], participants contributed to assessing the relevance, adequacy, and limitations of data used for AI development, for example, by identifying concerns related to clinically meaningful features, incomplete or unreliable routine data, erroneous inputs, limited dataset diversity, and problems of coding, documentation, or labeling. Methodologically, these data-related aspects were elicited through interviews and semistructured interviews [37,39,42,43,46,51,59,63,64], observations, journey mapping, contextual inquiry, and workflow analysis [46,50,64], individual evaluation sessions, user studies, and think-aloud-based testing [37,48,55], as well as iterative design, workshops, and fieldwork-based evaluation formats [41,50,56,58,59,64,70]. Additional contributions included therapist annotation [37], computational modeling and vignette-based assessments [39], and questionnaires and survey-based assessments [53,71].

Interpretability and explainability constituted a third area of participatory impact. In various studies [38,41,44,48-52,54,55,57-59,62,63,65,66,69,72], participants contributed to addressing these aspects within participatory processes. In several studies, explanatory components—such as visualizations, highlight features, or other explanation-oriented displays—were developed or revised with users through think-aloud–based evaluation and individual testing [44,48,55,66], focus groups and iterative design sessions [49,57,63], structured interviews and user observations combined with iterative design [50,59], and co-design sessions [58,62,65]. In the remaining studies [38,41-43,45,51-53,64,69,72], interpretability and explainability were addressed more indirectly by eliciting user requirements regarding transparency, data presentation, algorithmic reasoning, and the need for explanatory support. This was achieved through interviews [42,43,51,69], interviews and workshops [38,64], simulation-based testing and interviews [72], user feedback during prototype testing [52], iterative evaluation formats [41], and surveys or questionnaires combined with workshops or interviews [45,53].

In sum, participatory processes contributed not only to user-friendly interface design but also to the clinical relevance, data integrity, and transparency of AI components, provided that they are methodologically tailored to these aspects. Formats that engage clinical users not only retrospectively but proactively and iteratively in design and development decisions appear particularly effective.

The studies included in this review address a variety of ethical, legal, and social implications (ELSI) emerging from the participatory development, testing, and evaluation of AI-based CDSS. While the term ELSI was rarely used explicitly, most studies included theoretical references and empirical insights into relevant aspects. In many cases, however, ELSI-related issues were only indirectly identifiable, as they were embedded in broader discussions rather than explicitly framed as ethical, legal, or social implications.

A recurring ethical principle is that AI systems should support, not replace, clinical decision-making—an aim explicitly stated in several studies [40,41,45,47,50,59,60,65,67,69]. The role of explainable AI is repeatedly emphasized, particularly in relation to transparency, trust, and the interpretability of automated recommendations [37,39,42-44,46,47,49,51,53,54,56,58,59,62,63,65-67,70]. Other common themes include adaptability to clinical workflows [40-42,46,59,60,68,69,71] and the potential of AI systems to enhance the visibility of underrepresented professional roles [40,42,46,56,60]. In addition, several studies highlighted the need for AI-based support to remain clinically meaningful and context-sensitive, for example, by adapting outputs to different users, tasks, and care settings rather than relying on overly generic recommendations [42,46,47,54,59,65]. Closely related to this, some studies also emphasized the importance of preserving clinician autonomy and avoiding overreliance on AI systems, so that AI was experienced as supportive rather than directive [36,45-48,50,61,66,67,72].

Clinicians voiced diverse expectations, concerns, and requirements. Trust in the technology emerged as a central theme; users demanded validated models, transparent recommendations, and adaptive visualizations, especially in situations of clinical uncertainty [39,40,42-44,47,49,51,53,54,56,58-60,62-67]. Preservation of professional autonomy was emphasized, with AI perceived as a source of reassurance or means of broadening perspectives rather than as an authoritative decision-maker [40,43-47,50,58,60,64,66,67,69]. Some studies reported that AI systems may help reduce subjective variability in clinical assessments [50,60]. Relatedly, participants in several studies described AI more concretely as a form of second opinion, an additional safety layer, or a supportive backup in situations of uncertainty, fatigue, or limited staffing [43,46,49,51,60,64]. Participants also stressed that useful AI support depends on contextual fit; outputs and interfaces needed to be clinically meaningful, interpretable, and adaptable to different users, tasks, and care settings, rather than generic or overly abstract [38,39,42,46,48-50,53,54,58-60,63-65,68,71]. Several studies further highlighted resistance to one-size-fits-all solutions and pointed to the need for role-specific interfaces or explanations tailored to different professional groups [47-50,54,58,61,65].

In most studies, fairness, bias, discrimination, and nondiscrimination were not explicitly discussed with users during participatory activities. Where these aspects were considered, they were mainly implicit, through design choices or technical adaptations. More direct participant-related considerations were evident in 4 studies [37,45,46,58] in which participants raised concerns about consistent, unbiased outputs; anchoring on AI recommendations and over-reliance; overly generic recommendations that may not fit diverse users and contexts; and uneven model performance across patient groups or clinical situations. Several other studies [36,38,40,51,52,60,61,72] showed more indirect consideration, for example, adapting functions to different experience levels [40,51], role-specific customization [38,60], integrating explainability features [52,60], or ensuring broad accessibility [36,61], or through accessibility-oriented design, safeguards against biased interpretation, dataset-related reflections, or broader ethical framing [42,47,56,59,63,67].

Questions of accountability and responsibility in CDSS-supported decision-making were seldom explicitly discussed with users. Where present, they were often addressed indirectly, for example, by emphasizing clinician autonomy or implementing design features to prevent over-reliance on AI [37,38,41,48,72]. More explicit reflections on role delineation, clinician responsibility, and the boundaries of AI support were evident in 9 studies [38,41-43,45,46,54,66,67], for example, through explicit reflections on role delineation, clinician responsibility, and the boundaries of AI support. Several studies [36,40,41,44,47,50,58-61,63] addressed the topic more implicitly, for example, by enabling users to review or reject AI suggestions [36,40], framing the CDSS as a supportive tool [36,47,58-60], or embedding mechanisms for critical review [41,44,61,63]. In Grant et al [70], related concerns focused mainly on clarifying responsibilities, routing documents, and tracking who submitted which data, rather than on accountability for AI-supported decisions in a narrower sense.

In total, 14 studies explicitly reported obtaining approval from a research ethics committee [36,38,46-49,55,56,59,61,64,69,71,72]. Funding sources were predominantly from public research grants [37-39,41,43-45,47-49,51-58,60,61,64-67,69-71]. Some studies additionally drew on philanthropic, foundation-based, or university-internal support [36,40,56,64,66,67,71,72], while Fouad et al [63] reported workshop-specific support from the Alan Turing Institute, although broader project funding was not specified. Moreover, 4 studies combined public funding with contributions from industry partners [36,50,62,72], whereas Staes et al [59] reported direct industry support without a clearly stated combination with public funding.

Of the total, 4 studies reported substantial conflicts of interest, as several authors were directly involved with the company developing the respective CDSS, either as shareholders, employees, directors, or scientific advisors, and some additionally received honoraria, sponsorship, or research funding from pharmaceutical companies [36,59,71,72]. In total, 13 studies explicitly declared that no conflicts of interest existed [38,39,43,45,48-50,54,56,57,64,69,70].

To our best knowledge, this is the first review to systematically examine approaches to involving clinical staff in the development, piloting, and evaluation of AI-based CDSS. By analyzing not only when involvement occurred across project phases but also how it was organized and who was engaged, the review provides a differentiated understanding of the forms of involvement. This highlights both the diversity of approaches and existing asymmetries in representation. The review thus offers a comprehensive overview and insights into how involvement can shape data, models, and normative principles, supporting the development of more inclusive and clinically meaningful AI systems.

This review has several limitations that should be considered when interpreting the findings.

A central limitation concerns the inconsistent and often selective reporting of stakeholder involvement, which hampered comparability and frequently obscured the extent, continuity, and influence of participatory processes. Understandings of involvement also varied across disciplines and contexts, underscoring the need for clearer methodological frameworks and standardized reporting practices. Another challenge concerned classification: because only a few studies explicitly described their approach using participatory terminology, inclusion decisions had to rely partly on reviewer interpretation, which carries a risk of subjectivity. Furthermore, most of the included publications were descriptive or conceptual, limiting the ability to draw quantitative inferences about the effectiveness of participatory approaches.

In addition, no formal critical appraisal of study quality was conducted. This approach is consistent with the exploratory purpose of scoping reviews, which aim to map the breadth and characteristics of a research field rather than assess the effectiveness of interventions. However, it implies that studies of varying methodological rigor were considered equally in the synthesis, which may influence the strength and interpretation of the conclusions.

A further limitation relates to the scope of the review. The search was restricted to English- and German-language publications from 2012 to 2026, which may have introduced language bias. Participatory design research has strong traditions in regions such as Scandinavia, and relevant studies published in other languages may therefore not have been captured. Moreover, the relatively small number of included studies compared to the initial search yield reflects the specificity of our review focus on AI-based CDSS with reported stakeholder involvement. While numerous studies address related areas such as digital health, health information systems, or co-design in adjacent technologies, participatory processes are often not explicitly described or are insufficiently reported. This underreporting itself represents an important finding of this review. At the same time, the focused scope of this review may have excluded relevant insights from adjacent domains, including participatory design in non-AI health technologies or broader co-design practices in health care. Future research may benefit from integrating these perspectives to develop a more comprehensive understanding of participatory approaches in the design of AI-based clinical systems.

Although this review covered the full spectrum from development to evaluation, none of the included studies provided systematic evidence on real-world implementation. The findings, therefore, mainly reflect early phases of development, piloting, and testing. The lack of implementation studies is likely related to the fact that AI-based CDSS have so far rarely been integrated into routine care [107], among other reasons due to restricted data access and regulatory hurdles [108], and broader socio-technical barriers that complicate implementation in health care settings [107]. Nonetheless, the review provides a comprehensive overview of existing practices and gaps across multiple disciplines.

The development of AI-based CDSSs should integrate participatory processes continuously throughout all stages of system design and evaluation, rather than restricting them to isolated phases. Participation should address usability, data quality, bias, and ethical, legal, and social considerations. Co-design workshops, iterative prototyping, and shared evaluation dashboards can strengthen communication between developers and clinical users. Moreover, the reporting of participation and co-design remains insufficiently standardized, even though transparent documentation would be essential for contextualizing and critically reflecting on the research process. Although no reporting guideline currently exists specifically for participatory AI-CDSS development, existing guidance may offer useful orientation. This includes the Best Practices in the Reporting of Participatory Action Research [109] as well as the reporting guideline for the participatory development and evaluation of digital health interventions (ParDE-DHI) [87]. Such guidance underlines the need for studies to articulate more clearly how participation is conceptually understood, which theoretical assumptions underpin it, and how these assumptions are translated into concrete procedures. In addition, the use of technical reporting standards, such as TRIPOD+AI [110], can enhance transparency in model development and validation. Integrating participatory and technical reporting standards may help to make both user involvement and model transparency systematically visible, comparable, and reproducible. In addition, embedding deliberative and ethical reflection formats (eg, scenario discussions, data ethics cafés) can help translate high-level AI ethics principles into concrete design choices. It is equally important to involve diverse professional groups, particularly those with nursing and therapeutic expertise, to ensure that systems reflect the full spectrum of clinical routines and decision-making logics. As discussed above, participation has largely been framed through an HCI-oriented view of clinicians as “users,” yet their expertise would also be valuable for addressing implementation questions.

Future research should develop validated frameworks for participatory AI design that integrate ethical, organizational, and technical perspectives, thereby providing methodological coherence across studies and advancing the field. Longitudinal and comparative research could examine how different levels and forms of participation influence adoption, trust, and patient outcomes. In parallel, implementation studies in real clinical environments are needed to understand how to sustain participatory processes beyond project phases and embed them in institutional governance structures. Creating such frameworks and gathering empirical evidence will be essential to moving participatory design from a methodological aspiration to a practical standard for developing safe, responsible, and context-sensitive AI-CDSS.