The article “Adoption of Artificial Intelligence Technologies in Biomedical Research for Major Diseases: A Bibliometric Analysis” by Duarte-Martínez, Zambrano-Arriaga, and Cobo offers a structured overview of how three major artificial intelligence (AI) families – machine learning (ML), transformers (TF), and generative AI (GenAI) – have been incorporated into biomedical research on five high-impact disease areas: cancer, cardiovascular disease, diabetes, mental health disorders, and COVID-19, over the period 2012–2024. It is essentially a cartography of the AI–biomedicine landscape, mapping not only where research is concentrated, but also how fast it grows, which countries lead it, which applications dominate, and how impactful these publications are in terms of citations.
The authors start from the observation that AI is no longer a niche tool in biomedicine. ML, TF, and GenAI are embedded in diagnostic pipelines, prognostic models, drug discovery workflows, and even operational optimization in health systems. Each of these technologies has a different “entry date” into biomedical research: ML becomes prominent around 2012 with deep learning breakthroughs, transformers enter in 2017 with attention-based architectures, and GenAI rises from 2018 onward with models such as GPT-2. Against this historical backdrop, the article sets out to answer a central question: how have these technologies actually been adopted across major disease areas, and what patterns of scientific production and impact can be observed at a global scale?
To address this question, the authors conduct a large-scale bibliometric analysis using the Scopus database. They focus on publications related to the five selected disease groups that explicitly use one of the three AI families. A key methodological decision is to restrict the dataset to records with a PubMed ID. This filter is meant to ensure that the corpus is genuinely biomedical, excluding purely computational or technical works that are not anchored in health sciences. The trade-off is explicit: this choice likely omits some relevant computer science papers, but it increases the specificity of the biomedical focus.
The time windows are tailored to each technology’s adoption. For ML, the analysis begins in 2012; for transformers, in 2017; and for GenAI, in 2018. Within these windows, the authors construct Scopus queries for each combination of disease and technology. They take care to avoid overlap across technologies: ML queries exclude terms strongly associated with transformers and GenAI; transformer queries exclude GenAI; and so forth. Only three document types are retained: research articles, conference papers, and reviews. Reviews are kept deliberately because they often signal consolidation of knowledge and early recognition of emerging technologies, which is important when studying adoption dynamics.
The data-processing pipeline follows a PRISMA-inspired structure. Initial search results are progressively filtered by year, document type, source type (journals and conference proceedings), and PubMed ID. The final corpus consists of 88,735 publications, down from an initial set of 154,100 records before the PubMed filter. These publications are then enriched with metadata such as title, DOI, year of publication, author and indexed keywords, affiliations, and citation counts.
One important contribution of the study is the classification of each article into one of several biomedical application areas, common to all three AI technologies. Drawing on prior literature reviews, the authors define six domains: diagnosis, prognosis, drug development, clinical outcomes, surgical and organ applications, and fraud and operational applications. They then use BioBERT – a transformer model pre-trained on biomedical text – to assign each article to the most probable application area based on its title, abstract, and keywords. A random 10% sample is checked manually to validate the labels, which provides some reassurance about the robustness of the automatic tagging process.
Another methodological component is the geographic tagging of scientific production. Each publication is associated with a country based on the first author’s affiliation, using the pycountry library to standardize country names. This allows the authors to identify “pioneering” countries in the early years of each technology’s adoption for each disease, and later to characterize the global distribution of AI-enabled biomedical research.
The core analytical strategy has four pillars. First, the authors describe the volume of scientific production by disease and technology. Second, they normalize time by defining the first adoption year of each technology as “year 0,” then compute annual growth rates to compare adoption dynamics across ML, TF, and GenAI. Third, they identify the leading countries in the initial adoption phase. Fourth, they examine the distribution of publications across application areas and calculate the average number of citations per publication for each technology–application combination, thereby linking volume and impact.
Descriptively, the results show that cancer dominates AI-based biomedical research across all three technologies. Within the final dataset, cancer accounts for 44.51% of ML publications (38,637 papers), 47.11% of transformer publications (701 papers), and 49.44% of GenAI publications (221 papers). Cardiovascular diseases and mental health disorders also receive substantial attention, each contributing between roughly 15% and 19% of publications depending on the technology. COVID-19, despite its obvious contemporary importance, exhibits a more concentrated impact, especially for transformers (15.86% of TF publications), while diabetes tends to be the least represented disease across the three technologies.
The growth-rate analysis highlights different maturity levels. ML emerges as the most stable and consolidated technology. Its annual growth rates, once normalized to year 0, show moderate increases with occasional declines, such as a −41.89% rate for cardiovascular disease and −57.78% for diabetes in year 3, but overall the trajectory is steady. In contrast, transformers exhibit a much more irregular pattern characterized by sharp peaks and troughs. For example, cancer-related TF publications grow by 254.55% in year 3, while cardiovascular disease TF publications record a 500% increase in the same relative year. At the same time, steep early declines appear in certain areas, as with −54.55% for cardiovascular disease and −75.00% for diabetes in year 1. GenAI shows the highest volatility, with dramatic spikes (such as an 800% increase in COVID-19-related work in year 4 and 333.33% for cardiovascular diseases in year 5) followed by sharp drops, reflecting its very recent and opportunistic adoption in specific biomedical niches. In short, ML looks like a mature general-purpose workhorse, whereas TF and GenAI resemble emerging technologies riding successive waves of thematic interest.
The geographic analysis confirms the dominant roles of the United States and China in AI-driven biomedical research. In ML, these two countries lead almost every disease–technology combination, especially in cancer, where China produces 11,548 ML-based cancer publications and the United States 8,565. Other high-income countries such as Germany, India, Japan, Canada, Italy, and the United Kingdom also appear repeatedly among the top contributors, depending on disease area. For transformers, initial adoption is more geographically sparse, with the United States still occupying a central role, but with early contributions also coming from countries like the Republic of Korea, France, and South Africa. For GenAI, the early dataset is too small to form a consistent “top five” across all diseases, but the United States and China appear again as key players, joined by countries such as Norway, Mexico, and the United Kingdom in specific combinations. Taken together, these patterns suggest that AI-driven biomedical innovation is globally distributed but strongly anchored in a small group of research-intensive countries.
The breakdown by application area produces some of the most interesting findings. For ML, the technology is extensively used in surgical and organ applications, particularly for cancer, where 25,310 publications fall into this category. Diagnosis and prognosis are especially prominent for cardiovascular diseases, diabetes, and mental health disorders; for example, diabetes-related ML work shows several thousand papers in diagnostic and prognostic modeling, and mental health ML research has 5,727 diagnostic and 4,791 prognostic publications. In the COVID-19 context, ML is heavily used for diagnosis and operational management, with 3,117 diagnostic and 1,741 fraud or operational applications, reflecting the pressure to rapidly deploy triage tools and optimize health system performance during the pandemic. Fraud and operational applications more broadly are also well represented for cardiovascular diseases, with 4,215 ML publications, underlining the importance of resource optimization and anomaly detection in chronic disease management.
Transformers, by contrast, show a more focused adoption profile. Diagnosis is clearly the dominant area, with 294 transformer-based cancer publications, 116 for COVID-19, 90 for mental health disorders, 81 for cardiovascular diseases, and 34 for diabetes. Other application areas, such as fraud and operational applications, prognosis, and drug development, appear but at much lower volumes. Clinical outcomes are almost absent in TF-based research, indicating that these models are still primarily used for classification and information extraction tasks, often in text or imaging, rather than for long-term outcome modeling or intervention evaluation.
GenAI exhibits a narrow but distinctive pattern. Diagnosis and fraud and operational applications are present but relatively small in absolute numbers. Notably, cardiovascular diseases and mental health disorders each have 39 diagnostic GenAI publications, while COVID-19 and diabetes lag behind with 13 and 24 papers respectively. Fraud and operational applications are most visible in cancer (85 publications) and mental health disorders (22 publications), but adoption is modest in other disease areas. The most striking feature is the concentration of GenAI in surgical and organ applications for cancer, with 135 publications in this single combination, suggesting intense experimentation with generative models in surgical planning, simulation, and imaging for oncology. Prognosis and drug development using GenAI are still rare, reflecting an early, exploratory stage.
The citation analysis adds another layer of interpretation by comparing scientific impact across technologies and application areas. For ML, clinical outcomes achieve the highest average citations per publication at 51.08, despite representing a relatively smaller volume of papers. This indicates that when ML-based research focuses directly on patient-centered outcomes, it tends to attract strong attention from the scientific community. Fraud and operational applications (27.91) and surgical and organ applications (27.83) also have high average citation counts, similar to diagnosis (26.46), while prognosis (22.92) and drug development (24.71) show slightly lower but still substantial impact. Thus, ML not only dominates in volume across multiple application areas but also produces highly cited work, particularly when linked to tangible clinical benefits.
For transformers, overall impact is more modest but with a clear peak in prognosis, where the average number of citations per publication reaches 18.28. Diagnosis and drug development have intermediate citation levels (11.72 and 12.95 respectively), while fraud and operational applications and clinical outcomes are cited less often. GenAI, surprisingly, obtains some competitive averages in specific niches. Fraud and operational applications record 20.79 citations per publication on average, which is higher than TF’s performance in many domains and not far from ML’s mid-range values. Drug development (16.25) and prognosis (12.65) also show meaningful impact, whereas diagnosis (9.37) and surgical and organ applications (11.46) remain lower. These figures suggest that, even with a low publication volume, GenAI is already generating influential work where it is applied strategically, particularly in administrative optimization and early drug discovery tasks.
In the conclusion, the authors argue that ML has reached a level of maturity and breadth that makes it an entrenched component of biomedical research, spanning from surgery and organ applications to fraud detection and operational optimization. Transformers and GenAI, in contrast, are still in emergent phases, characterized by rapid growth, thematic concentration (especially on diagnosis for TF and on specific cancer applications and operations for GenAI), and smaller but potentially high-impact clusters. A cross-cutting gap identified by the authors is the relatively limited use of transformers and GenAI for clinical outcomes. This represents an opportunity for future work to move beyond diagnostics and operational cases toward more outcome-oriented evaluations and long-term patient trajectories.
The paper also emphasizes the policy and strategic implications of these findings. Mapping global adoption patterns can inform national digital health strategies by highlighting where capacities already exist, where there are underexplored opportunities, and which countries or institutions could be targeted for capacity building. For example, the observed concentration of innovation in a handful of countries suggests a risk of widening global inequities in AI-enabled healthcare, unless deliberate efforts are made to support AI research ecosystems in low- and middle-income contexts. Similarly, the strong but under-impactful presence of AI in certain domains implies that research funders and health system leaders could prioritize areas where added value for patients remains underdeveloped, such as clinical outcomes for TF and GenAI.
Overall, the article offers a panoramic, quantitatively grounded view of how different AI technologies are shaping biomedical research on major diseases. It combines a large-scale bibliometric approach with modern text-mining tools and clear application-area definitions to reveal adoption patterns, geographic leadership, and impact asymmetries. For researchers, policymakers, and health technology strategists, it functions as both a map of where the field stands today and a compass pointing toward gaps and opportunities for future work.
Reference: Duarte-Martínez, V., Zambrano-Arriaga, J., & Cobo, M. J. (2025). Adoption of artificial intelligence technologies in biomedical research for major diseases: A bibliometric analysis. In Proceedings of the 11th International Conference on eDemocracy & eGovernment (ICEDEG 2025) (pp. 278–285). IEEE. https://doi.org/10.1109/ICEDEG65568.2025.11081535