This was an ecological, country-level scientometric analysis of the 6 GCC states for the period from 2011 to 2024. Record selection is reported following the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 guidelines [15], and observational reporting follows the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) statement adapted to aggregate bibliographic data [16]. The unit of analysis was the publication; aggregation was to country and country-year. Publications were attributed to each GCC country represented among their affiliations (full counting), so regional totals count a publication once for each GCC country involved; the 3512 country-attributed records, and the record-selection counts in Figure 1, therefore correspond to 3327 distinct publications. Regional intra-GCC collaboration is reported as a share of distinct publications, because full counting would count each such publication at least twice.

Bibliographic records were obtained from OpenAlex, a fully open index of scholarly works with institution, country, concept, and citation metadata [17]. OpenAlex was selected over subscription databases for its open access, programmatic reproducibility, and comparable coverage. We note that OpenAlex concepts (derived from the Microsoft Academic Graph) are no longer actively maintained and have been superseded by OpenAlex topics; because the frozen AI family vocabulary predates the large language model era, works indexed only under newer terms such as “large language model,” “transformer,” or “generative AI” may be undercounted, a limitation we quantify and discuss below. Records were retrieved via the public API on the extraction date (June 2026), filtered to works affiliated with a GCC institution, tagged with the OpenAlex concept artificial intelligence [18], and with a publication year of 2011 to 2024. Records dated as 2025 were present at extraction but were excluded because 2025 was an incomplete, still accruing year subject to indexing lag at the retrieval date (June 2026); restricting the corpus to 2011 to 2024 ensured that every analyzed year was a complete indexing year and that growth estimates were not distorted by a partially indexed final year. To assess robustness to database choice, the same population was independently retrieved from Scopus using a title-abstract-keyword query combining AI and health terms with GCC affiliation and year filters, and the 2 sources were compared on annual output, national distribution, and record-level overlap.

A work was defined as health AI if it included at least one AI family concept (AI, machine learning, deep learning, or artificial neural networks) and at least one health or medicine family concept, each with a concept relevance score of 0.30 or higher. The health or medicine family comprised 6 OpenAlex concepts: medicine, internal medicine, intensive care medicine, family medicine, health care, and public health. The complete concept identifiers and relevance scores for both families are provided in the deposited extraction script [19]. Both concept lists and the threshold were prespecified and are reported as sensitivity parameters. Each work in OpenAlex includes, for every assigned concept, a classifier-generated relevance score in the range from 0 to 1 that reflects the confidence of the automated concept-tagging model [20]; low scores predominantly correspond to weak, incidental assignments. We adopted a 0.30 cutoff to exclude such marginal tags while retaining substantive ones and tested its influence directly: in a sensitivity check, raising the threshold to 0.40 or lowering it to 0.20 changed absolute counts but left the national ranking, growth trajectory, and thematic structure unchanged. To estimate precision, 2 authors manually adjudicated a random sample of 100 included records against title and topic; 94 were unambiguous health AI applications (95% Wilson CI 87.5%‐97.2%), and no clearly off-topic record was identified, with the remainder comprising borderline health informatics works rather than misclassifications. This definition prioritizes specificity over recall and is fully reproducible from the deposited extraction script. Throughout, “health AI” denotes this concept-defined corpus; “clinical AI” is used only for the narrower subset of AI intended for point-of-care clinical use.

For each country, we computed total and annual output, compound annual growth rate (CAGR), output per million population (World Bank midyear estimates) [21], mean and median citations, the proportion of works with at least one non-GCC affiliation (international collaboration) and with 2 or more GCC state affiliations (intraregional collaboration), and the share of national output accounted for by the 3 most prolific institutions. Coauthorship was analyzed at the country level; thematic structure was analyzed by constructing a concept co-occurrence network and partitioning it into clusters through modularity maximization [22]. Thematic evolution was assessed by classifying each work’s primary topic into macrothemes across the period.

To address question 4, national health AI output and per capita intensity were correlated with the equal-weight HIRI [14] and with GDHM maturity phase [6] using Spearman rank-order coefficients. The HIRI is an equal-weight composite index of national health informatics readiness for the 6 GCC states, constructed from ITU, WHO Global Health Observatory, GDHM, and World Bank indicators spanning digital infrastructure, workforce and governance, and financing [14]. The GDHM is a global instrument that classifies countries into ordinal maturity phases across governance, infrastructure, services, and data domains of digital health development [6]. With 6 countries, statistical power was minimal; coefficients are reported descriptively by magnitude and direction rather than against significance thresholds. Analyses used Python (version 3.11; Python Software Foundation; Pandas, NetworkX, and Matplotlib libraries); all data and code are reproducible from the tabulated inputs.

This study did not involve human participants, human tissue, or individual-level patient data; it used only aggregate, publicly available bibliographic metadata retrieved from OpenAlex together with publicly reported country-level indicators. Under the applicable national research governance framework and the policies of Saudi Electronic University, secondary analyses of publicly available, nonidentifiable aggregate data of this kind do not require review by an institutional review board or research ethics committee, and no such application was therefore submitted. No informed consent was required as no human participants were involved.

Of 112,480 AI publications affiliated with GCC institutions over 2011 to 2024, a total of 108,968 were excluded for including no health or medicine concepts at the 0.30 threshold, leaving 3512 (3.1%) health AI publications for analysis among all the GCC AI research output throughout the period (Figure 1).

The independent Scopus query retrieved 14,676 records for 2011 to 2024 (15,560 country attributions, counted in the same way as for OpenAlex). The broader keyword strategy returned a larger corpus than the concept-based OpenAlex definition (n=3512), but the 2 databases were highly concordant on aggregate structure (Figure 2). Annual output was almost perfectly correlated (Spearman ρ=0.99; Pearson r=0.99), with both sources showing the same 2019 to 2021 inflection. The national distribution was identical in rank order (Spearman ρ=1.00): Saudi Arabia was first in both (2094/3512, 59.6% of the OpenAlex output and 9463/15,560, 60.8% of the Scopus output), followed by the UAE (768/3512, 21.9% of the OpenAlex output and 3376/15,560, 21.7% of the Scopus output), Qatar, Oman, Kuwait, and Bahrain, with national shares differing by no more than 1.8 percentage points between sources. At the record level, 58.6% (1951/3327) of distinct OpenAlex titles had an exact normalized title match in the larger Scopus set, a lower bound given that title normalization does not capture near matches. The substantive findings reported below are therefore robust to database choice.

Regional health AI output rose from 18 publications in 2011 to 904 in 2024, a CAGR of 35.2%, with 86.4% (3034/3512)of the corpus published from 2020 onward. Output inflected sharply from 2019 to 2021, when annual production more than trebled from 135 to 435 publications (Figure 3).

Output was strongly concentrated: Saudi Arabia produced 59.6% (2094/3512) of the corpus by itself and 81.5% (2862/3512) together with the UAE (Table 1). Normalized to population, the ranking was reordered, with Qatar leading at approximately 136 publications per million population ahead of the UAE (n=81) and Saudi Arabia (n=63)—indicating that absolute volume reflects national scale rather than research intensity.

International collaboration predominated throughout: 72.6% (2551/3512)of the corpus had at least one non-GCC affiliation compared to 4.9% (164/3327) of distinct publications coauthored across 2 or more GCC states. Internationalization rose across the 3 subperiods (70/112, 62.5% in 2011‐2015; 411/604, 68% in 2016‐2020; and 2070/2796, 74% in 2021‐2024), as did intra-GCC collaboration among distinct publications (0/112, 0% in 2011-2015; 6/596, 1% in 2016-2020; and 158/2619, 6% in 2021-2024), indicating regional networks that formed late and remain modest. Intra-GCC collaboration was inversely related to national output, reaching 34% (16/47) in Bahrain and 21.6% (16/47) in Kuwait but only 6.4% (133/2094) in Saudi Arabia, consistent with smaller research systems relying on regional partners while the largest were comparatively self-sufficient. Across 113 partner countries, the principal external collaborators were India (n=598 publications), the United States (n=556), Pakistan (n=410), the United Kingdom (n=393), and Egypt (n=297; Figure 4) [23].

Output was distributed across 3186 distinct institutions. The most prolific were King Saud University (n=328 publications), King Abdulaziz University (n=249), Princess Nourah bint Abdulrahman University (n=195), and Qatar University (n=175). Concentration varied markedly: the 3 leading institutions accounted for 89.4% (328/367)of Qatar’s output but only 34.4% (721/2094)of Saudi Arabia’s and 38.4% (295/768) of the UAE’s (Table 1), distinguishing nationally distributed research systems from those anchored in a single flagship university.

Concept co-occurrence was classified into 4 coherent clusters (Figure 5), each labeled here by its most frequent constituent concepts (in parentheses): classic machine learning clinical prediction (random forest, support vector machines, logistic regression, and internal medicine), medical imaging and oncology (deep learning, transfer learning, segmentation, breast cancer, magnetic resonance imaging, pathology, and radiology), COVID-19 response (SARS-CoV-2, intensive care, and pandemic), and diabetic eye disease (diabetic retinopathy, ophthalmology, and optometry). Temporally, AI-based COVID-19 research surged from 2020 to 2022 before receding as a share, whereas research on oncology and imaging and on general AI in health remained the persistent leaders (Figure 6). The single most cited work in the corpus was a 2017 deep learning study of lymph node metastasis detection with Qatari coaffiliation (3287 citations) [24].

National health AI research intensity (publications per million population) correlated positively with the HIRI (Spearman ρ=+0.60), and absolute output correlated more strongly (ρ=+0.89). Absolute output was also correlated with GDHM maturity phase (ρ=+0.83; Figure 7). In a sensitivity analysis recomputing the correlations across the 19 alternative index specifications defined in the HIRI preprint [14,25], the correlation with absolute output ranged from +0.66 to +0.94 and that with intensity from +0.31 to +0.66 (Multimedia Appendix 1), so readiness was associated more consistently with the volume of national research than with its per capita intensity. Because this analysis was based on only 6 observations, we treat it as exploratory rather than confirmatory; the per-country HIRI and GDHM values are reported in Multimedia Appendix 1. The pattern is consistent with the global association between digital maturity and health system performance [25].

This is the first scientometric baseline of health AI research for the GCC. Earlier bibliometric studies have mapped adjacent slices of Gulf research: Weber et al [3] reviewed eHealth research across the same 6 states; Al-Farsi et al [4] standardized Gulf health research output to population and identified Qatar as the leading country on that basis; and Al-Marzouqi and Arabi [5] normalized regional output to population, gross domestic product, research and development expenditure, and researcher counts, reporting Saudi dominance in absolute volume, Qatar as having the steepest relative growth, and high international collaboration. Our per capita and collaboration findings are consistent with this previous work; the specific contribution here is the restriction of the lens to health AI; the addition of a concept-level thematic structure and an explicit readiness linkage; and, thereby, the characterization of the research base from which locally relevant clinical AI must be developed rather than of Gulf health research as a whole. Four findings stand out. First, output grew rapidly (CAGR=35.2%) but from a low base and with a pronounced post-2019 inflection that coincided with the diffusion of deep learning and with national digital health milestones. Second, the research base is heavily internationalized: nearly three-quarters of the publications involved a non-GCC partner, and the proportion of intraregional collaboration, though rising, remained under 1 in 10 papers. Third, the field was concentrated on medical imaging and, outside Saudi Arabia and the UAE, anchored in a single national institution. Fourth, informatics readiness as captured by the HIRI was associated with both research intensity and absolute output, most consistently with output, as was digital health maturity.

The contrast with the aforementioned European Union panel study is instructive [2]. Where that study found that health AI research capacity did not automatically improve population-level value, the present analysis was deliberately upstream: it measured capacity itself before any outcome linkage. The strong external partner structure—dominated by India, the United States, Pakistan, the United Kingdom, and Egypt—reflects both the GCC’s reliance on an internationally recruited research workforce and the regional and diaspora ties that shape Gulf science [23]. This dependency is invisible to connectivity or maturity indexes and is a distinctive contribution of the scientometric lens. We report indicators in absolute and per capita terms without a formal world, Middle East and North Africa, or matched country baseline; nonetheless, the international collaboration share (2551/3512, 72.6%) was high relative to the roughly one-fifth to one-quarter international coauthorship reported worldwide, and the 35.2% compound annual growth exceeds the global growth of AI publishing over the same period, indicating that the GCC pattern is not merely a reflection of worldwide trends. The absence of a matched external comparator is a limitation and a priority for future work.

The positive association of the HIRI with both research intensity (ρ=+0.60) and output (ρ=+0.89) suggests that informatics readiness and research capacity develop together, although with 6 countries this remains exploratory. The concentration of output on imaging mirrors the global health AI literature and signals an opportunity: domains underrepresented in the GCC corpus—primary care, mental health, health systems, and safety research—are precisely those where locally generated evidence is most needed before deployment.

Four limitations should be considered. The health AI definition is concept based and trades recall for precision; manual audit indicated high precision (94%), but recall was not estimated, and alternative thresholds shift absolute counts but not the comparative structure. In addition, because eligibility relied on the now deprecated OpenAlex concept vocabulary, the corpus underrepresents very recent large language model, transformer, and generative AI work indexed chiefly under newer terms; this most plausibly affects the 2022 to 2024 segment and means that the headline growth curve, if anything, understates the most recent expansion. The primary corpus was derived from OpenAlex; although a Scopus concordance check confirmed that the temporal and geographic structure was robust to database choice (Figure 2), affiliation disambiguation was imperfect, and English-language indexing undercounted Arabic-language output in both sources. The macrotheme classification used for thematic evolution was coarse, leaving a substantial unclassified residual; the concept co-occurrence network is the more rigorous thematic representation. Finally, the 6-country correlation with the HIRI and GDHM is descriptive and cannot support inferential or causal claims.

GCC health AI research has expanded rapidly but unevenly, depends heavily on international collaboration, is concentrated on medical imaging, and—outside Saudi Arabia—remains institutionally narrow. Informatics readiness was associated with both the intensity and the volume of national research output. These results provide a reproducible baseline against which the region’s health AI research capacity can be monitored as GCC health systems move toward the deployment of clinical AI, and they identify regional collaboration and thematic breadth as the clearest priorities for strengthening that capacity.