The rapid and widespread integration of artificial intelligence (AI) into health systems (HS) has raised high expectations while also creating significant uncertainties and challenges [1-4]. AI refers to “a machine-based system that can, for a given set of human-defined objectives, make predictions, recommendations, or decisions influencing real or virtual environments” [5]. As AI is increasingly viewed as a lever to improve efficiency amidst workforce shortages [6], health organizations are implementing AI tools to support precision diagnostics, personalized medicine, clinical decision-making, and virtual assistance, among other applications [7,8]. The deployment of these complex technologies introduces new pressures and risks, including ethical, sociopolitical, economic, organizational, professional, and clinical challenges, with particular sensitivities around patient safety and privacy [1,3,4,9]. Additional concerns include bias and inequities across subgroups, regulatory uncertainty, liability, vendor lock-in, technical debt, data quality and interoperability gaps, inadequate management and training, and misalignment between procurement incentives and demonstrated real-world value [10,11]. In response, stakeholders have called for robust and adaptive AI governance frameworks for HS to safeguard the system and enable responsible, sustainable, and effective integration of AI-based technologies [1,12-15].

Although several ethical AI frameworks for HS define fundamental values (eg, trust, transparency, accountability, safety, justice, fairness, nonmaleficence, and privacy) and normative objectives [16,17], they generally do not provide benchmarks or operational guidance for achieving them [17-19]. While the role of ethical frameworks is to articulate principles, the role of governance frameworks is to provide the operational structures, processes, and mechanisms (the “how”) required to translate principles (the “what”) into practice [20-24]. To this end, governance frameworks identify relevant actors, allocate decision-making authority, and clarify responsibilities [20,21,23-25]. They also guide core functions such as data governance, model validation and auditing, and risk assessment and management [20,21,23-25]. To support a practical and consistent application of AI ethical principles [23], governance frameworks should aim to operationalize the following: (1) roles, rules, norms, competencies, and policies applicable throughout the AI lifecycle; (2) value-sharing arrangements, procurement and contracting requirements, remuneration and reimbursement mechanisms, and intellectual property considerations; (3) auditable requirements regarding data governance, human oversight, explainability, model validation, and cybersecurity; (4) postdeployment monitoring and incident reporting; assessments of clinical effectiveness, safety, cost implications, and equity impacts; and liability and risk allocation; and (5) stakeholder engagement, transparent public communication, accountability mechanisms, and criteria for model replacement or retirement [23,26-32].

Although AI governance in HS presents distinct characteristics and requirements, it nonetheless forms an integral component of the broader governance architecture for digital health [1,12,25,33-35]. It is grounded in the institutional, regulatory, and infrastructural arrangements and structures that shape how digital technologies are designed, procured, deployed, used, evaluated, and monitored in real-world care settings and that ultimately determine whether these tools can be integrated safely, effectively, efficiently, sustainably, and equitably at scale [1,12,25,34]. Accordingly, we adopt a governance perspective consistent with digital health governance and define AI governance in HS as “the exercise of political, administrative, and technical authority to manage everything associated with the health information system, in all areas of a national system. The structure of this governance consists of the mechanisms, processes, and institutions through which all stakeholders articulate their interest, exercise their rights, meet their obligations, and resolve their differences” [33]. These governance elements are interconnected, forming a coherent whole that ensures AI-based technologies meet clinical, legal, ethical, and regulatory requirements and are used in alignment with HS strategies, objectives, principles, and values [35].

The pressing need for comprehensive AI governance frameworks in HS has led to a proliferation of proposals over the past decade. Yet, these initiatives often address sociopolitical, ethical, clinical, or data management issues separately, rather than treating them as interdependent dimensions of governance [36]. This literature remains limited by 2 persistent conceptual shortcomings: the absence of a shared definition of AI governance in HS and the frequent conflation of governance with related notions such as ethics, responsibility, or regulation. For example, a scoping review by Stogiannos et al [37] mapped the available literature on AI governance in medical imaging and radiology and proposed a governance framework. However, the authors did not clearly define what they meant by governance; their analysis primarily examined the factors relevant to governance in the United Kingdom’s imaging and radiology sectors, and most of the documents reviewed emphasized ethical and regulatory considerations at the organizational level rather than governance in a broader institutional sense [37]. This fragmentation, combined with a tendency to underestimate the structural and operational demands of governance, has contributed to persistent misalignment between national ambitions for AI integration and the strategies, institutional arrangements, and resources actually deployed to support such integration [36].

To the best of our knowledge, no systematic review has yet identified and synthesized AI governance frameworks in HS, despite the fact that such analysis is essential for identifying recurrent shortcomings and informing strategies for the responsible integration of AI-based technologies into HS [38]. This systematic review aimed to synthesize existing AI governance frameworks for HS and to propose an integrative AI governance model identifying key components to guide AI-related policy, practice, and research in HS.

This systematic review was conducted and reported in accordance with the PRISMA (Preferred Reporting Items for Systematic Review and Meta-Analyses) Literature Search Extension statement (Checklist 1) [39,40]. The full description of methods is provided in the protocol registered on the Open Science Framework prior to data collection [41]. No modifications or additional information were made to the protocol following its registration (PRISMA item 24c). In addition, the following PRISMA checklist items were considered not applicable to this review: 10b, 12, 13e, 13f, 15, 19, 20b, 20d, and 22 (Checklist 1). Given the qualitative nature of this systematic review, these items were not relevant to the methodological approach adopted in this study. Ethical approval was not required for this systematic review.

The evidence map included published peer-reviewed articles or reports proposing a new framework, guideline, standard, or position statement of AI governance in HS. Eligible documents had to be explicitly labeled as AI governance in HS by the authors. Only records published in English, French, Spanish, or Portuguese between November 2014 and March 2026 were included. We selected 2014 as the starting point because the mid-2010s marks the period when AI began transitioning from primarily experimental applications to more concrete and integrated uses within HS [42-44]. This period also coincides with the expansion of large-scale digital health data infrastructures and the growing involvement of major technology companies in health AI ecosystems. From a governance perspective, it corresponds to the emergence of more structured global discussions on AI ethics, responsible AI, and digital health governance, as well as the proliferation of national digital health and AI strategies [43,44]. We therefore chose 2014 to focus the analysis on frameworks most relevant to current policy and implementation contexts. Abstracts, letters to the editor, commentaries, essays, viewpoints, conference proceedings, and unavailable articles were excluded. Two reviewers (RPS and EJP) independently screened all references for inclusion using Covidence (a web-based software platform) [45]. Disagreements were resolved through discussion. If the 2 primary reviewers could not reach a consensus, a third reviewer (HA) adjudicated the final decision.

A comprehensive literature search was conducted in 8 databases (PubMed, MEDLINE [Ovid], Embase [Ovid], ACM Digital Library, Web of Science [Clarivate], Scopus, Social Sciences Abstracts [EBSCO], and PsycINFO [EBSCO]) in October 2024 (updated in July 2025 and March 2026). Database searches were conducted and limited to human studies published since November 2014. The strategy combined free-text keywords and controlled vocabulary terms structured around the concepts of AI (“Artificial Intelligence” or “AI” or “machine learning” or “deep learning” or “natural language processing”) AND governance (“governance” or “regulation” or “framework” or “model” or “policy” or “guideline” or “ethic” or “standard” or “principles” or “regulatory” or “responsible”) AND health system (“health system” or “healthcare system” or “health services” or “health sector” or “hospital” or “healthcare delivery” or “public health systems” or “health sector”). The search strategy was developed by the authors through a rapid literature review and group discussion. We conducted searches across Google Scholar, ProQuest Dissertations and Theses Database (ProQuest), Social Science Research Network (SSRN eLibrary), and Evidence for Informed Health Policymaking (PDQ-Evidence). The electronic search was complemented by manual searches of the reference lists of relevant articles and web portals of institutions including the World Health Organization (WHO), United Nations, the Organisation for Economic Co-operation and Development, World Bank, European Union, National Institute for Health and Care Excellence, and National Institutes of Health. Details on the search strategies are available in Multimedia Appendix 1. All references retrieved from the searches were imported into EndNote20 (Clarivate) to manage citations and remove duplicates. The refined list of citations was imported into Covidence.

Data were extracted using a standardized form in Microsoft Excel, including (1) document characteristics (author names, year of publication, first author country, objective, type of document, targeted audience, and definition of AI governance); (2) methodology (research question, study design, sample characteristics, and data collection procedures); (3) framework characteristics (name, level of application, design, actors, roles and accountabilities, governance structure functions, processes, mechanisms, barriers, and challenges); and (4) funding sources. Data were extracted by 2 reviewers (RPS and LR) and checked by 2 others (EJP and HA).

Framework quality was assessed using a modified version of the Appraisal of Guidelines for Research and Evaluation for HS (AGREE-HS). The original AGREE-HS and the modifications are presented in Multimedia Appendix 2 [46]. As in the original AGREE-HS, 5 quality items were rated: (1) topic (clarity and relevance of the AI governance challenge); (2) participants (stakeholder inclusion, multidisciplinary representation, and conflicts of interest); (3) methods (transparency and rigor of framework development processes); (4) recommendations (clarity, feasibility, ethical principles, equity, and sociocultural and political factors); and (5) implementability (feasibility, barriers and facilitators, resource requirements, sustainability, and transferability). Regarding modifications, 21 criteria of the original AGREE-HS applicable to frameworks were also included, as no tool currently exists for these.

Additionally, we assigned scores to each criterion in advance to improve assessment reliability. Two authors (RPS and EJP) independently applied the explicit scoring criteria to each item using a Microsoft Excel spreadsheet. Disagreements in each item score were resolved by consensus. Each domain score was calculated by aggregating its items' scores and was rated on a 7-point scale, where 1 indicated the lowest quality and 7 indicated the highest quality. The framework overall score was calculated following the AGREE-HS manual as follows [47]: (obtained score–minimum score)/(maximum score–minimum score)×100%, with a minimum total score of 5 and a maximum total score of 35. We performed a tertile split of the frameworks’ overall scores using the 33rd- and 67th-percentile cut points to classify them as high quality (overall scores >51%), moderate quality (overall scores between >41% and 51%), or low quality (overall scores ≤41%).

Data were analyzed and integrated through a narrative synthesis approach, which allowed for a coherent combination of the varied elements within the AI governance frameworks for HS. Framework levels were categorized as international, national, local, or organizational based on their scope. The analysis was informed by the framework of Van Grembergen et al [48] for IT governance adapted to AI-based technologies. This framework encompasses 3 interconnected dimensions of governance: structures, processes, and relational mechanisms. The structures are the actors and their roles that ensure a clear chain of responsibility for the decisions and operations around AI-based technologies. The processes are the formal methods and workflows for making and monitoring these technologies, including establishing a framework for planning, prioritizing, and monitoring AI initiatives. Relational mechanisms are elements that foster trust, participation, partnership, and mutual understanding in the processes [48]. Whenever it is possible and pertinent, the components of the processes and relational mechanisms identified in the frameworks are presented according to the different stages of the AI lifecycle.

A total of 10,175 records were retrieved from the bibliographic databases, of which 91% (n=9210) remained after removing duplicates. Following title and abstract screening, 163 (2%) documents were full-text screened. Of these, 12 (0.1%) articles proposed an AI governance framework for HS. One (0.009%) article was excluded, as it was a scoping review [37], resulting in the inclusion of 11 (0.1%) peer-reviewed articles. Another 87 additional full-text records were identified through websites and reference lists, adding 8 publications to the systematic review. A total of 19 AI governance frameworks for HS were identified in 15 peer-reviewed articles and 4 reports (Figure 1).

An overview of included documents is presented in Multimedia Appendix 3. Six documents resulted from primary studies [50-55], and only 6 provided a definition of AI governance [1,56-60]. The geographical distribution of first authors and the scale of intended audiences are presented in Figure 2. Most frameworks (13/19, 68%) were published between 2022 and 2024 [50-57,59,61-64]. Ten (53%) frameworks were developed by authors based in North America—7 (37%) from the USA [50,55,57,59-61,63] and 3 (16%) from Canada [58,62,65]—3 (16%) in Australia [30,53,64], 2 (10%) in Switzerland [1,56], 1 (5%) in the United Kingdom [51], 1 (5%) in Germany [52], 1 (5%) in New Zealand [54], and 1 (5%) in Qatar [66]. Details of the included frameworks are presented in Multimedia Appendix 4. They targeted audiences at 4 levels: 3 international [1,51,56], 5 national [30,59,64-66], 3 local [53,54,62], and 8 organizational [50,52,55,57,58,60,61,63]. Ten frameworks were developed through collaboration between academics and practitioners [1,50,55,56,59,60,63-66].

Table 1 presents the results of the adapted AGREE-HS. Seven frameworks were rated as high quality [1,30,50,51,56,60,62], 6 were rated as moderate quality [53,54,58,61,63,66], and 6 were rated as low quality [52,55,57,59,64,65].

While 9 frameworks received a score of 6 or higher for the information about the framework development team (Table 1—Participants) [50,51,53,54,56,61-63,65], only 7 reached this threshold for describing the AI governance challenge, its causes, and the priority attributed to it (Table 1—Topic) [1,51,56,58,60,62,66]. Most of the authors do not provide a complete description of systematic and transparent methods to review the evidence and to agree upon the framework components. Therefore, 14 frameworks scored 3 or lower on the dimension assessing the methods used to develop them (Table 1—Methods) [1,30,50,52,53,55,57,59,61-66]. Finally, most authors did not include information on the framework implementability (17 scored 3 or less; Table 1—Implementability) [1,30,51-56,58-66].

AI governance structures encompass functions and the roles and responsibilities of involved actors. Multimedia Appendix 5 presents authors’ statements on the functions of AI governance structures. Seventy-four percent (14/19) of frameworks explicitly define these functions [1,30,50,51,53,54,56,59-63,65,66]. Two international-level frameworks emphasize that these structures are intended to support the achievement of national health objectives [1,51].

According to Morley et al [51] and the WHO [1], AI governance structures also play a crucial role in establishing shared standards of safety, privacy, and efficacy to ensure compliance with minimum requirements. Furthermore, the WHO indicates that these structures can help prevent regulations that could create unfair competitive advantages or disadvantages for both companies and governments. Such structures also ensure that governments develop regulatory frameworks that uphold ethical principles, human rights, and international law [56].

At a national level, 4 frameworks indicate that AI governance structures have the function of establishing guidelines to evaluate AI-based technologies intended for use in the HS [30,59,65,66]. According to Reddy et al [30], Parker et al [59], and Solaiman [66], these guidelines help to ensure that AI-based technologies are implemented safely, ethically, and in compliance with regulatory standards. Likewise, for 3 local-level frameworks, the function of AI governance structures is to ensure adherence to ethical principles in the deployment of AI, such as equity, privacy, and transparency [53,54,62]. Arnaout et al [62] and Whittaker et al [54] highlighted that these structures also play a crucial role in identifying gaps for AI solutions and ensuring the appropriateness, safety, and effectiveness of AI development and deployment. Additionally, 6 organizational-level frameworks underscore the pivotal role of AI governance structures in providing guidance and support throughout the AI lifecycle [50,57,58,60,61,63]. For some, these structures are responsible for supporting the development of AI-based technologies [57,58,60] and for establishing evaluation, monitoring, and adoption processes [50,57,58,60].

The literature on actors and their roles and responsibilities is notably more heterogeneous. Across the 19 frameworks, actor composition within governance structures at all levels varied substantially, reflecting differences in contextual scope and sectoral representation. These actors are categorized by subgroup and scale in Figure 3; however, the categories overlap. All frameworks underlined the crucial role of multidisciplinary bodies or teams in governing AI within HS across these different levels. Notably, only the international-level frameworks explicitly highlight the importance of synergy between governance structures operating at different scales [1,51,56].

Owing to the large number of actors involved, the frameworks present significant variations in the roles and responsibilities assigned to each actor, with many responsibilities shared among multiple actors. For instance, while Morley et al [51] suggested that “collaboration and multidisciplinary working by decision-makers, technologists, healthcare professionals, and academics are needed to ensure appropriate expertise throughout the AI life cycle,” the WHO [56] stated that “governments could require developers to ensure that the design and development of a general-purpose foundation model achieve certain outcomes throughout its life cycle.” For this reason, it was not feasible to synthesize the roles and responsibilities comprehensively. Other examples of actors’ roles and responsibilities are provided in Multimedia Appendix 6.

Fifty-three percent (10/19) of the frameworks identified barriers and challenges to AI governance for HS [1,50,51,54,55,57,59,60,63,66]. International-level frameworks highlighted challenges arising from the disconnection between different yet highly interrelated policy domains across governance levels, the lack of international coordination in AI governance for HS [51], and concerns about the potential emergence of a “market” in which individuals can “buy and sell health data” [1]. Across other levels, 6 frameworks emphasized challenges related to the resources (eg, costs, infrastructure, and time) required for the effective implementation of AI governance [54,57,59,60,63,66]. Additionally, 3 organizational-level frameworks raised concerns about the ongoing resource demands necessary for the continuous monitoring of AI-based technologies [55,57,60].

None of the reviewed frameworks provided a comprehensive approach that fully captures the multidimensional and dynamic nature of AI governance in HS. We therefore synthesized the dimensions and components reported across the 19 included frameworks to propose an integrative AI governance model that identifies key elements to guide AI-related policy, practice, and research in HS. The resulting model is presented in Figure 4. The model advances the field by explicitly articulating AI governance as a multilevel, interrelated system, in contrast to existing health information system or responsible AI governance frameworks that typically address only selected components or specific settings. Importantly, this integrative model reflects the current state of the literature and is not intended as a ready-to-implement governance framework for decision-makers. Rather, it provides a foundational scaffold that requires further refinement through expert validation, stakeholder and end user consultation, pilot testing, and evaluation.

The integrative AI governance model for HS shows how the key governance components identified across the included frameworks are deeply interconnected, as well as the permeability of different levels of AI governance. The framework comprised 2 governance dimensions (structures and processes) and 4 relational mechanisms (ethical principles and/or values, education and training, communication, and standards and regulations) in 4 interconnected levels (organizational, local, national, and international). The darker blue at the organizational level reflects a higher concentration of well-established and more specific AI governance frameworks for HS. For instance, the AI governance functions vary across levels but follow a complementary pattern.

The double-headed arrows highlight the dynamic nature of interactions between the dimensions and their components. For instance, the bidirectional arrow linking the structural elements (functions, actors, and roles and responsibilities) illustrates a reciprocal relationship. AI governance functions are shaped by actors who hold distinct yet complementary (and sometimes conflicting) roles and responsibilities within multidisciplinary bodies. These actors play a decisive role in defining, operationalizing, and adapting governance functions. Conversely, the governance structure and its functions can also shape, redefine, or shift the roles, relative importance, and influence of these actors over time. Similarly, the 2-way arrows between data governance, risk assessment and management, and validation and/or evaluation illustrate the interdependence among these components across the entire AI lifecycle. These 3 components are intended to ensure that AI-based technologies fulfill their goals while safeguarding public health, and individual privacy and safety.

The gradient from blue to white at the center of the processes represents the adaptation and the interaction between the relational mechanisms arising from the interplay among all components of the AI governance framework for HS. Thus, establishing ethical principles, standards, and regulations, together with clear, context-sensitive, and transparent communication strategies, as well as tailored education and training programs and initiatives adapted to different audiences (eg, health professionals, policy-makers, managers, patients, and the general population), is essential at international, national, local, and organizational levels to guide the responsible development and deployment of AI-based technologies within HS [1,30,56,64], particularly given their potential to reflect and reproduce structural inequities [67,68]. Nevertheless, their adoption and adaptation over time can differ depending on the context characteristics and resources, potentially giving rise to new relational mechanisms within AI governance processes. The 4 relational mechanisms underpin all processes, especially need and/or problem identification, maintenance and monitoring, and integration, which are critical to ensure AI and HS sustainability.

This study synthesized existing AI governance frameworks for HS published over the last decade and reviewed 19 frameworks distributed across international, national, local, and organizational levels. The frameworks articulated operational arrangements meant to translate ethical principles and legal and regulatory requirements into actionable structures, processes, and mechanisms. Consistent with prior literature reviews [10,69,70], our findings highlighted that despite the growing number of AI governance frameworks for HS, quality and implementability are largely lacking.

While the proposed integrative model constitutes an important synthesis of knowledge about the governance of AI in HS, there remains significant scope for improvements.

To the best of our knowledge, this systematic review is the first to synthesize existing AI governance frameworks across 4 governance levels while simultaneously assessing their quality. Yet, this study has certain limitations. First, although our search strategy was comprehensive and extended to gray literature, it remains possible that frameworks in progress or those disseminated in other languages were not identified. Regarding language restrictions, we acknowledged that relevant frameworks may exist in other languages, particularly in countries that are advanced in AI development (eg, China and Germany). However, we did not conduct literature searches in these languages as our team does not possess the linguistic expertise required to analyze documents in these languages. Relying on automated translation tools was not considered an appropriate option, as these tools pose a substantial risk of misinterpreting concepts and terms whose meaning depends heavily on context. Consequently, our model carries a Western bias, as the underlying evidence is primarily sourced from the Global North. This disparity underscores a critical need for future research to investigate AI governance within contexts of the Global South.

Second, a limitation arises from the use of an adapted version of AGREE-HS to evaluate the methodological quality of each included framework. Although this tool is the best option to evaluate whether the framework’s authors explicitly define the topic (what), participants (who), and methods (how), several AGREE-HS items are interpretative in nature requiring contextual judgment. However, we ensured the robustness and consistency of data extraction and appraisal through independent coding, calibration, and consensus-based adjudication. This approach is commonly adopted in reviews incorporating qualitative appraisal, where calibration, discussion, and consensus are recognized as appropriate strategies to ensure methodological rigor [112]. Finally, the heterogeneity of this body of literature may have constrained the extent to which the distinctive nuances of each framework were fully captured in the results. To mitigate this limitation, our review and our model were developed systematically, following rigorous methodological guidance [39], to better examine diverse frameworks spanning multiple AI governance components and contexts.

This systematic review of AI governance frameworks for HS highlights key elements of AI governance structures, processes, and mechanisms distributed across international, national, local, and organizational levels. Rather than focusing on specific governance dimensions or settings, the review innovates by integrating all governance components and their interactions across contexts. This study thus differs from other reviews by contributing a multilevel analysis that elucidates the strengths and gaps within current frameworks, adding to this expanding body of scholarship. The study brings to the field an integrative AI governance model for HS. The proposed model may inform future scholarly work as well as real-world policy and practice-oriented discussions on ensuring the responsible and sustainable integration of AI-based technologies into HS.