The governance of health data in the AI era: a scoping review and computational topic model of the global research landscape Authors: L Raymond Guo, Shuimei Liu Published: 2026 ABSTRACT AbstractBackgroundThe integration of artificial intelligence (AI) into health care is critically dependent on vast quantities of patient data, igniting an urgent global debate on data ownership, privacy, and governance. While numerous perspectives exist, the empirical structure and evolution of this scholarly discourse remain uncharacterized. We aimed to systematically map the conceptual landscape of research on AI and health data governance to identify its core themes, temporal trends, and key focus areas.MethodsWe conducted a scoping review according to PRISMA guidelines, searching PubMed, Scopus, and Web of Science for peer-reviewed articles published between January 1, 2018, and May 31, 2025. We performed a descriptive analysis of publication trends. Using Latent Dirichlet Allocation (LDA), we applied computational topic modelling to the abstracts, which serve as concise summaries of each article’s core contributions, to identify latent thematic structures. Topic trends were analyzed using linear regression.FindingsForty-three articles met the inclusion criteria. The volume of publications has increased substantially since 2018. Our LDA analysis identified five distinct research topics: (1) AI Applications & Ownership, (2) AI Models & Data Privacy, (3) Data Sharing Platforms & Technology, (4) Ethical & Legal Concerns, and (5) AI Development & Implementation. Over the study period, research on Ethical & Legal Concerns showed a statistically significant increasing trend in prevalence (slope = 0.023, P = .008), becoming the most dominant topic in recent years.InterpretationThe scholarly discourse on AI and health data has matured, shifting from foundational questions of technical implementation towards a dominant focus on complex ethical and legal challenges. This data-driven evidence signals an urgent need for clinical leaders and policymakers to move beyond theoretical discussions and implement robust, practical governance frameworks. Failure to address this governance gap risks impeding trustworthy AI innovation and eroding public trust, thereby limiting the potential of AI to improve patient outcomes equitably.Keywords: artificial intelligence, data ownership, data governance, health information management, topic model Introduction The integration of artificial intelligence (AI) into health care represents a paradigm shift in the modern era. From enhancing the diagnostic accuracy of radiological scans to personalizing complex oncological treatment protocols and predicting disease outbreaks, AI-driven systems promise to revolutionize patient care, streamline clinical workflows, and accelerate the pace of biomedical research.1 These advanced computational systems, particularly those based on machine learning (ML) and, more recently, large language models (LLMs), are fundamentally data-driven. Their efficacy, safety, and reliability are inextricably linked to the availability of vast, diverse, and high-quality datasets for training, validation, and real-world deployment.2 This foundational dependency has ignited a complex and urgent global debate, moving beyond simplistic notions of data ownership to a more sophisticated discourse on the intertwined frameworks of data stewardship and intellectual property rights that govern these transformative technologies.3 The concept of data ownership in health care is not new, but the advent of AI has amplified its complexity and stakes exponentially. Historically, health data governance was primarily concerned with the privacy and security of records within institutional silos, governed by regulations like the Health Insurance Portability and Accountability Act (HIPAA) in the United States and the General Data Protection Regulation (GDPR) in Europe.4,5 These frameworks, however, were designed for an era of structured data and relatively static uses. The rise of AI, which thrives on massive-scale data aggregation and repurposing for unforeseen future applications, challenges the very foundations of these regulations. The value of data is no longer confined to individual patient care but is now a primary driver of commercial innovation, pharmaceutical research, and public health strategy, creating powerful new incentives and tensions among stakeholders.6 This complex ecosystem has given rise to a burgeoning field of academic inquiry where scholars across medicine, law, ethics, and computer science grapple with how to balance the immense potential of AI with the fundamental rights of individuals. Key questions persist: Do patients own their data, or do they merely have a right to privacy? What new models of consent are needed for data that will be used for countless future, unspecified AI applications? How can benefits derived from AI, which is trained on collective data, be shared equitably with the populations who contributed that data?7 While many editorials, commentaries, and legal analyses have addressed these questions from specific viewpoints, a comprehensive, empirical understanding of the entire scholarly landscape is missing. We lack a data-driven map of the key themes preoccupying researchers, how these themes have evolved, and where intellectual energy is being focused. Identifying these patterns is crucial for policymakers, funders, and researchers to recognize critical gaps, anticipate future challenges, and direct resources effectively. This study, therefore, seeks to fill this gap by conducting a scoping review of the literature, augmented by a computational topic modelling analysis. Our primary research questions are: (1) What are the core thematic components of the scholarly discourse on AI and health data governance? (2) How has the prominence of these themes evolved over time? (3) What does this evolution reveal about the maturation and future trajectory of the field? By employing this integrated methodology, we aim to chart the conceptual structure of this global discourse, providing an objective foundation for future research and policy. Methods Methodological framework This study was designed as an integrated, two-phase analysis to provide a comprehensive map of the research landscape. First, we conducted a scoping review, reported in accordance with the Preferred Reporting Items for Scoping Reviews and Meta-Analyses (PRISMA) 2020 statement,8 to rigorously identify the complete corpus of relevant literature. This ensures a reproducible and unbiased selection of articles, forming the empirical bedrock of our study. Second, we used the output of this review, the abstracts of all included articles, as the direct input for a computational topic modelling analysis. As concise summaries of each article’s core contributions, abstracts provide an efficient and powerful dataset for uncovering the latent thematic structure of a field. This second phase moves beyond simple description to uncover the latent thematic structure of the field, revealing the core concepts and ideas that connect disparate publications. Phase 1: Scoping review and data extraction We conducted a comprehensive search of three major electronic databases: PubMed, Scopus, and Web of Science. These databases were chosen to ensure broad coverage across biomedical (PubMed), technical and interdisciplinary (Scopus), and general scientific (Web of Science) literature, thereby capturing the multifaceted nature of the discourse.9 The search was restricted to peer-reviewed articles published in English between January 1, 2018, and May 31, 2025. This timeframe was chosen to capture the contemporary discourse following the widespread recognition of modern AI’s potential in health care. Our search strategy employed a combination of keywords and subject headings built around three core concepts: (1) artificial intelligence, (2) data governance, and (3) health care. The search string was constructed with Boolean logic to maximize sensitivity and specificity. For the AI concept, terms included “artificial intelligence” OR “machine learning” OR “deep learning” OR “large language model.” For the data governance concept, terms included “data ownership” OR “data governance” OR “data sharing” OR “data stewardship” OR “privacy” OR “data ethics.” For the healthcare concept, terms included “health” OR “health care” OR “medical” OR “clinical” OR “patient.” Inclusion criteria were peer-reviewed articles where the primary focus was on the intersection of AI, data governance, and a health-related context. Exclusion criteria were articles without substantive discussion of data governance, purely technical papers, conference abstracts, dissertations, and non-English articles. Titles and abstracts were independently screened by both authors. Potentially relevant articles were selected for full-text review, and any discrepancies in the screening process were resolved by consensus. To ensure consistency during the screening of 60 reports assessed for eligibility, substantive discussion was operationally defined as the presence of dedicated sections or distinct paragraphs addressing at least one of the following: (1) the legal or regulatory frameworks governing data use (eg, HIPAA, GDPR), (2) explicit models of data stewardship or ownership, or (3) formalized ethical oversight mechanisms for AI training datasets. Articles that mentioned governance only as a future requirement in a single sentence without further elaboration were excluded. Data from the included articles were extracted into a standardized spreadsheet, capturing first author, publication year, journal, country of the corresponding author, and the full abstract. Phase 2: Computational topic modelling To identify the latent thematic structures within the literature, we conducted a topic modelling analysis using Latent Dirichlet Allocation (LDA) on the corpus of abstracts from all 43 included articles. The abstract was chosen as the unit of analysis because it represents a concise, author-provided summary of the article’s core contributions, making it an ideal input for identifying the primary themes of a large body of work, a common practice in scientometric studies. LDA is an unsupervised probabilistic model that assumes each document is a mixture of topics, and each topic is a distribution of words, allowing for data-driven discovery of primary themes.10 The analysis was performed using R (version 4.5, R Core Team, 2025). The text preprocessing and topic modelling pipeline relied on several key packages: tm for creating the document-term matrix (DTM), topicmodels for implementing the LDA algorithm, tidytext and dplyr for data manipulation and cleaning, and textmineR for calculating coherence scores. Data visualization was accomplished using ggplot2 for plots and pheatmap for generating heatmaps. The text preprocessing pipeline involved several steps: (1) converting all text to lowercase; (2) removing punctuation and numbers; (3) removing common English stopwords; and (4) removing a custom list of domain-specific but uninformative stopwords (eg, “health”, “data”, “ai”). We then created a DTM from the cleaned text, using both single words and two-word phrases (n-grams). To determine the optimal number of topics (k), we built multiple LDA models with k ranging from 2 to 10 and evaluated them using a coherence score. The coherence score measures the semantic similarity of words within a topic, providing a quantitative method for selecting the most interpretable topic structure.11 The value of k = 5 maximized the coherence score and was selected for the final model. The final LDA model was trained using Gibbs sampling with 1000 iterations. Each of the 5 resulting topics was qualitatively interpreted and assigned a descriptive label by 2 researchers who independently reviewed its top 15 most probable terms. Finally, to analyze the evolution of these topics over time, we calculated the mean topic prevalence (gamma value) for each topic across all articles published in a given year. To quantify temporal shifts and assess the statistical significance, we employed a simple linear regression model, where the independent variable was the publication year (2018-2025), and the dependent variable was the mean annual topic prevalence value. The model was specified y=β0+β1x+ϵ. We calculated the slope (β1) to determine the rate of change in prevalence per year and used F-Statistics to assess whether the trend deviated significantly. A P-value of <.05 was considered statistically significant. Results Study selection and characteristics Our database search yielded 43 articles for inclusion. The PRISMA flow diagram (Figure 1) detailing this process is shown in Figure 1. The included articles were published between 2018 and 2025. As shown in Table 1, the number of publications has grown substantially, from 2 articles in 2018 to a peak of 11 in 2024. The 43 articles were published across 40 different journals, demonstrating a wide and highly fragmented disciplinary interest. Journals represented a mix of medical informatics, bioethics, general medicine, and law (Table 2). The geographical distribution of the research, based on the corresponding author’s affiliation, was led by the United States, followed by the United Kingdom, Canada, and China, as detailed in Table 1. PRISMA 2020 flow diagram detailing the study selection process. The diagram shows the number of records identified, screened, assessed for eligibility, and ultimately included in the scoping review. Table 1.Characteristics of included studies (n = 43).Publication yearResult20181201922020520215202242023820241120257 Top 5 countries of origin a United States15United Kingdom6Canada4China3Germany2aBased on the affiliation of the corresponding author. Not all articles specified a country. Characteristics of included studies (n = 43). Based on the affiliation of the corresponding author. Not all articles specified a country. Table 2.Publishing journals of included articles.IDJournal name 1 Cureus 2 Healthcare Informatics Research 3 Scientific Reports 4 AJOB Empirical Bioethics 5 Academic Radiology 6 The American Journal of Bioethics 7 Arthroscopy: The Journal of Arthroscopic & Related Surgery 8 BMC Medical Informatics and Decision Making 9 Behaviour & Information Technology 10 The Breast 11 Clinical Imaging 12 Current Opinion in Endocrinology, Diabetes and Obesity 13 Digestive Diseases and Sciences 14 Digital Health 15 European Journal of Health Law 16 F1000Research 17 Frontiers in Artificial Intelligence 18 Frontiers in Nutrition 19 Frontiers in Public Health 20 Global Spine Journal 21 Graefe’s Archive for Clinical and Experimental Ophthalmology 22 Human Genetics 23 Indian Journal of Psychological Medicine 24 International Journal of Environmental Research and Public Health 25 Interactive Journal of Medical Research 26 Journal of Community Genetics 27 Journal of Medical Internet Research 28 JMIR Medical Informatics 29 JMIR Public Health and Surveillance 30 JSES Reviews, Reports, and Techniques 31 Medical Teacher 32 NPJ Digital Medicine 33 Nature Genetics 34 Pharmacogenomics 35 Proceedings of the Institution of Mechanical Engineers, Part H: Journal of Engineering in Medicine 36 Radiology: Artificial Intelligence 37 Rural and Remote Health 38 Saudi Medical Journal 39 Surgical Endoscopy 40 iScience Publishing journals of included articles. This reveals that the discourse is heavily concentrated in North America and Europe, with limited contributions from institutions in Africa, South America, and large parts of Asia. This finding points to a significant geographical imbalance in the global research landscape on this topic. Latent topics in AI data governance The LDA model identified 5 distinct topics, summarized in Table 3. Table 3.Ranked summary of identified latent topics from LDA analysis.RankTopic labelOverall prevalenceTop 10 constituent termsInterpretation 1 Ethical & Legal Concerns 30.0% Challenges, ethical, care, patients, concerns, privacy, security, transparency, related, legalEncompasses the broad normative dimensions of AI in health care, including ethics, law, patient rights, consent, bias, and accountability. 2 AI Applications & Ownership 22.0% Potential, issues, ownership, technology, applications, include, decision, making, surgical, developmentFocuses on the use cases of AI in health care and the resulting challenges regarding control and ownership of AI systems and data assets. 3 AI Models & Data Privacy 18.0% Using, privacy, real, models, based, synthetic, monitoring, algorithms, sensitive, detectionConcerns the technical methods for creating AI models, with a strong emphasis on privacy-preserving techniques and the use of synthetic data. 4 AI Development & Implementation 15.0% Development, many, provide, implementation, adoption, heor, llms, improve, clinical, surgicalRelates to the practical lifecycle of AI systems, from initial development and adoption to real-world implementation and impact assessment. 5 Data Sharing Platforms & Technology 15.0% Sharing, management, analysis, digital, ownership, model, covid, platform, access, securityAddresses the infrastructure, platforms, and technologies (eg, federated learning, data trusts) needed to facilitate secure data sharing for AI. Ranked summary of identified latent topics from LDA analysis. Topic 1: AI Applications & Ownership. This topic focuses on the practical applications of AI in healthcare and the resulting questions of technology and data ownership. Its key terms include “potential,” “issues,” “ownership,” “technology,” and “applications.” This theme encapsulates discussions about the promise of AI for clinical decision-making, while simultaneously grappling with the fundamental challenges of asserting control over the technologies and the valuable data assets that are created. Topic 2: AI Models & Data Privacy. This topic centers on the technical and methodological aspects of building AI models while preserving patient privacy. It is characterized by terms like “using,” “privacy,” “real,” “models,” and “synthetic.” This reflects a body of work concerned with privacy-preserving machine learning (PPML) techniques, the generation of synthetic data, and the algorithmic challenges of using AI for remote monitoring. Topic 3: Data Sharing Platforms & Technology. This topic addresses the infrastructure and technologies enabling secure data sharing for AI development. Key terms such as “sharing,” “management,” “digital,” and “platform” highlight a focus on creating ecosystems that facilitate controlled access to data. This includes discussions of federated learning and data trusts or enclaves, which provide a legal and technical framework for governing data access. Topic 4: Ethical & Legal Concerns. This is a broad and central topic covering the normative challenges associated with AI in health care. It is defined by a dense cluster of terms like “challenges,” “ethical,” “care,” “patients,” “concerns,” “privacy,” “security,” “transparency,” “legal,” “bias,” and “accountability.” This theme represents the core ethical and legal debates that have come to define the field, including the adequacy of patient consent, the risk of algorithmic bias, and the need for explainability in AI-driven decisions. Topic 5: AI Development & Implementation. This topic focuses on the practical lifecycle of AI systems, from their creation to their adoption and impact in clinical settings. Key terms include “development,” “implementation,” “adoption,” “HEOR” (health economics and outcomes research), and “LLMs” (large language models). This theme covers the barriers to implementing AI in real-world clinical workflows and the emerging challenges posed by new, powerful technologies like generative AI. Evolution of research topics The prevalence of these 5 topics has shifted significantly over the 7-year study period, as illustrated in the heatmap in Figure 2. Our linear regression analysis of these trends, shown in Table 4, quantified these shifts and revealed a statistically significant increase in the scholarly focus on Ethical & Legal Concerns (slope = 0.0229, P = .008). This topic had a relatively low prevalence in the early years of our study period, but has grown to become the most dominant theme in the literature published since 2022. Heatmap of topic prevalence over time (2018-2025). Each cell represents the mean prevalence of a topic for all articles published in that year. Red indicates higher prevalence, while blue indicates lower prevalence. Table 4.Ranked Temporal Trend Analysis of Topic Prevalence, 2018-2025.RankTopic labelLinear regression slope P-valueTrend direction1Ethical & Legal Concerns0.0229.008Increasing2AI Applications & Ownership0.0126.203Stable3AI Models & Data Privacy–0.0378.168Stable (trending down)4AI Development & Implementation0.0084.223Stable5Data Sharing Platforms & Technology–0.0062.735Stable Ranked Temporal Trend Analysis of Topic Prevalence, 2018-2025. Conversely, while not reaching statistical significance, Topic 2: AI Models & Data Privacy was most prominent in 2018 and 2019 but has seen its relative share of the discourse decrease in subsequent years. The other topics have remained relatively stable in their prevalence. The distribution of these topics across the top 10 publishing journals and countries is visualized in Figures 3 and 4. Heatmap of topic prevalence across the top 10 publishing journals. This figure illustrates the thematic focus of the most frequent publication venues for research in this area. Heatmap of topic prevalence by country of corresponding author. This figure visualizes the geographical distribution of research themes, highlighting the concentration of different topics in various countries. Discussion This scoping review and computational topic analysis provide an empirical map of the conceptual structure of the scholarly discourse on AI and health data governance. Our analysis reveals a field that is not only rapidly growing but also undergoing a significant thematic evolution. The central finding is a data-driven confirmation of a “turn to ethics”: the conversation has matured from an initial focus on technical feasibility to being increasingly dominated by complex ethical, legal, and societal challenges. This shift, likely driven by a combination of high-profile cases of algorithmic bias in the public sphere, the implementation of influential regulations like the GDPR, and a growing awareness within the research community itself, has profound and actionable implications for clinical practice, health policy, and the future of digital health innovation. The turn to ethics and its consequences: a widening governance gap The statistically significant rise of the Ethical & Legal Concerns topic is the most important finding of this study. It serves as a clear mandate for action. For hospital executives and clinical IT leaders, it signals that AI procurement and implementation strategies must evolve beyond technical specifications to prioritize ethical due diligence, risk management, and robust governance. It is no longer sufficient to ask if an AI tool is accurate; leaders must now ask if it is fair, transparent, accountable, and developed using data that was governed ethically and with meaningful community engagement.12 Our findings suggest that failing to address these issues poses a direct threat to successful clinical integration and exposes institutions to significant legal, financial, and reputational risk. This turn to ethics, when viewed alongside the steady pace of work on AI Development & Implementation” (Topic 5), reveals a widening governance gap. As Figure 2 illustrates, the technologies are advancing relentlessly while the unresolved ethical and legal questions multiply. This phenomenon, often described as the pacing problem in law and technology, is particularly acute in healthcare due to the high stakes involved.13 This gap is exacerbated by the highly fragmented nature of the discourse, which we found to be spread across 40 different journals. Such fragmentation can slow the development of consensus on standards and create disciplinary silos where, for example, legal scholars and informaticians do not benefit from each other’s crucial insights. This fragmentation is not only disciplinary but also geographical. Our finding (Figure 4) that the research is concentrated in North America and Europe reveals a crucial opportunity for the global informatics community. To develop truly robust and universally applicable governance frameworks, it is essential to broaden the current discourse. The next phase of research must actively engage scholars, policymakers, and communities from Asia, Africa, and South America to co-design solutions that respect diverse contexts and ensure the benefits of AI in health are shared equitably. Delving deeper into this governance gap reveals fundamental challenges in applying traditional legal doctrines to AI-generated insights and health data. The concept of data ownership itself is increasingly seen as inadequate, with scholars proposing more nuanced frameworks that consider a bundle of rights and stakeholder interests rather than a single, proprietary owner.7,14 This complexity is compounded when considering the outputs of AI systems. The question of who owns, or can even hold a copyright for, an AI-generated diagnostic algorithm or treatment recommendation is a frontier legal issue. Navigating the gap: the dual challenge of stewardship and intellectual property The stable presence of the Data Sharing Platforms & Technology topic points towards a constructive path forward. The focus on concepts like federated learning and data trusts suggests the research community is actively exploring technical and organizational solutions that move beyond simplistic, individualistic notions of data ownership.16 These models, which treat data as a collective resource to be managed for societal benefit, may hold the key to resolving some of the tensions between individual privacy and the common good of AI-driven research. However, our findings also send a clear message: the increasing dominance of ethical concerns underscores that purely technical solutions for data sharing will fail if they lack social legitimacy. For these innovative stewardship models to succeed, they must be co-designed with the public to build trust, ensure that governance structures reflect societal values, and establish clear mechanisms for equitable benefit-sharing with the communities that provide the data.17 The future of data stewardship is not just about better cryptography or more secure platforms; it is about building trustworthy relationships between institutions and the public they serve. Furthermore, the governance gap creates profound uncertainty for the intellectual property rights that underpin innovation. Uncertainty remains regarding who holds the copyright or patent for an AI-generated treatment algorithm, a creation derived from patient data, institutional resources, and developer ingenuity. Recent guidance suggesting that works generated autonomously by AI are not protectable by copyright creates instability in the very incentive structures that drive investment in AI for health.15 Resolving the tension between responsible data stewardship and viable intellectual property models is a central, unresolved challenge for the field. An agenda for future research Our analysis of the literature’s structure and evolution points to several critical gaps that must form the basis of a future research agenda. First, there is an urgent need for empirical research evaluating the real-world implementation of novel data stewardship models like data trusts and data cooperatives. While theoretically promising, their effectiveness, scalability, and public acceptability remain largely unproven. Second, research must move beyond identifying the problem of algorithmic bias to developing and validating practical, scalable methods for bias detection and mitigation throughout the AI lifecycle. Third, a critical research frontier is the development of meaningful frameworks for equitable benefit-sharing. This requires moving beyond abstract principles to create concrete economic and non-economic models that return value to the patients and communities whose data fuels AI innovation. Finally, all these research avenues must actively decenter the current North American and European focus and prioritize the inclusion of perspectives, values, and governance challenges from the Global South to ensure a truly equitable global framework for AI in health. Strengths and limitations of this study This study’s primary strength is its integrated, data-driven methodology. However, several limitations must be acknowledged. First, our search was limited to three databases and English-language publications. Second, our analysis was based exclusively on abstracts. While this provides a high-level map of the field’s primary themes, it may not capture nuanced arguments or secondary topics discussed only in the full text of the articles. Third, the number of articles (n = 43) is relatively small, reflecting the emergent nature of this research niche. Conclusion The scholarly discourse on AI and health data governance is at a critical inflection point. The clear and measurable pivot towards ethical and legal concerns signals a collective recognition that the path to trustworthy AI in health care must be paved with robust governance, transparent processes, and a steadfast commitment to patient rights and equity. This study provides a data-driven roadmap of the current landscape, highlighting for clinical leaders and policymakers the urgent need to close the governance gap and providing a clear agenda for the future research required to build a more just and trustworthy era of digital medicine. The work ahead requires not only translating academic debate into practical standards for data stewardship and intellectual property but also ensuring these standards are inclusive, culturally competent, and designed to build a more just and trustworthy era of digital medicine for all. Contributor Information L Raymond Guo, School of Pharmacy, West Virginia University, Morgantown, WV 26506, United States. Shuimei Liu, School of Juris Master, China University of Political Science and Law, Beijing 100088, China. Author contributions L. Raymond Guo (Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Software, Visualization, Writing—original draft, Writing—review & editing) and Shuimei Liu (Conceptualization, Data curation, Formal analysis, Investigation, Writing—original draft, Writing—review & editing) Conflicts of interest The authors declare that they have no competing interests. Funding This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. Data availability The data underlying this article will be shared on reasonable request to the corresponding author. Protocol and registration The protocol for this scoping review was not pre-registered. The full study protocol is available from the corresponding author upon reasonable request. References Associated Data Data Availability Statement The data underlying this article will be shared on reasonable request to the corresponding author.