This investigation, though rooted in bibliometric principles, adopted a historical lens to observe evolving research practices within the biomedical field. The study’s architecture was underpinned by a custom-built Python script, crafted specifically to handle automated interactions with NCBI Entrez E-utilities via Biopython (version 1.79+). This design wasn’t simply about automation; it allowed for consistent metadata structuring, supporting retrieval efforts spanning multiple decades. The reproducibility this enabled was not incidental, but rather necessary given the scope and scale of the literature involved. Parsing, query segmentation, and batch retrieval were handled programmatically, ensuring the research maintained integrity over time.

Interestingly, the decision to use PubMed as the central corpus stemmed not only from its breadth but from its structural assets. The Medical Subject Headings (MeSH) framework-integrated directly into PubMed indexing-played a crucial role in enabling topical cohesion during the parsing phase. Beyond just standardized classification, MeSH terms provided a stable thematic scaffold. Although other repositories exist, few offer the same granularity or temporal traceability. PubMed, thus, emerged not merely as a data source, but as an analytical foundation capable of reflecting complex temporal and geographic shifts in knowledge generation [5, 9].

Query segmentation wasn’t merely a methodological necessity. It evolved as a strategic decision born from the dual pressures of temporal comprehensiveness and technical restraint. The dataset’s architecture was informed by the uneven distribution of publications across time, particularly during the initial decades (1946–1994), when output relevant to intelligent computational methods remained infrequent. For such intervals, broad temporal windows were acceptable. However, this leniency did not extend into the post-2008 publication surge, which required annual granularity and later month-by-month scrutiny.

Rather than a single continuous sweep, the extraction routine was fragmented. The search syntax-targeting the Title/Abstract field with the query term “artificial intelligence”-was framed using PDAT (Publication Date) parameters to anchor each segment. To enhance transparency and reproducibility, we explicitly defined the search terms applied. The core query used was “artificial intelligence”[Title/Abstract], with supplementary terms such as “machine learning,” “deep learning,” “neural networks,” “natural language processing,” and “radiomics” tested to confirm retrieval sensitivity and thematic coverage. This ensured that the dataset comprehensively reflected AI-related biomedical research indexed in PubMed.

Consistency in this formulation, such as “artificial intelligence”[Title/Abstract] AND (“start_date”[PDAT]: “end_date”[PDAT]), allowed the framework to remain stable even as frequency intensified. Data calls to Entrez.esearch were limited to 1,000 records per batch, and inter-query delays were intentionally programmed not just to conform to Application Programming Interface (API) etiquette but also to preserve response stability across retries. Scalability and error resilience weren’t afterthoughts. They were embedded by design, driven by lessons learned from prior interruptions and latency issues in literature mining.

The metadata phase didn’t follow immediately in a linear fashion-it emerged as a necessary layer once the PubMed Identifiers (PMIDs) had been gathered. Using Entrez.efetch with rettype=”medline” and retmode=”text”, the system retrieved full records in a consistent, readable format. Only then could parsing begin, with Bio.Medline.parse from Biopython performing the heavy lifting-not as an auxiliary step, but as the pivot point for structured bibliometric interpretation.

Parsing was not limited to the obvious. Yes, article titles and abstracts were included-fundamental for thematic indexing-but the process reached further. All records were programmatically retrieved through the NCBI Entrez E-utilities using Biopython (version 1.79+). Each request was limited to 1,000 records per batch, with programmed inter-query delays to ensure compliance with API requirements and prevent retrieval errors. The metadata fields extracted included PMID, title, abstract, author list, journal, publication type, MeSH terms, DOI, affiliations, and country of origin.

The PubMed ID (PMID) served as the anchor, while journal names added disciplinary context. Dates, of course, were extracted-not merely for cataloging but to anchor time-series evaluations. In parallel, publication type classifications such as “Journal Article” or “Review” provided signals about the depth and orientation of the scientific discourse.

Less visible fields proved equally crucial. MeSH descriptors enabled standardized thematic alignment, while author affiliations pointed not just to institutions, but to broader geopolitical patterns in the evolution of computational health studies. DOIs, retrieved from the Locator Identifier (LID) field, were less glamorous but indispensable for persistent referencing. The ‘country’ tag-frequently overlooked-played a quiet yet pivotal role in geographic trend analysis.

Eventually, everything converged into a structured Comma-Separated Values (CSV) repository not just for storage, but for all downstream inferential and descriptive routines.

Once the initial set of PMIDs had been assembled, attention shifted to the more nuanced task of metadata extraction. Rather than beginning with thematic analysis, the process first required careful structuring of bibliographic records. The Entrez.efetch utility-configured with the parameters rettype=”medline” and retmode=”text”-served as the conduit for retrieving entries in a format that supported consistent parsing and interpretation.

Parsing was not undertaken arbitrarily. The Biopython module, particularly the Bio.Medline.parse function was leveraged at this stage, allowing the retrieval process to unfold through a controlled pipeline capable of dissecting structured MEDLINE entries with minimal data loss. Cleaning procedures included deduplication of records, standardization of author names using a semicolon delimiter, cross-checking missing abstracts and DOIs, and validation of country fields for geographic analysis. These steps ensured a consistent and reproducible dataset for downstream bibliometric analysis.

Interestingly, it was not just the volume of records, but the variability across fields-titles, abstracts, and DOIs-that introduced complexity into the extraction procedure.

Crucially, each entry yielded multiple elements-some expected, others surprisingly inconsistent across records. Article identifiers (PMIDs), titles, and abstracts were prioritized for their central role in guiding later thematic segmentation. In parallel, author lists were standardized using a semicolon delimiter, which proved helpful when mapping collaborative networks. Journal names, less frequently used in preliminary analytics, were nonetheless retained for evaluating publication venues and disciplinary reach.

The publication date for each article was captured not as a timestamp, but as a temporal anchor for mapping shifts in output over defined eras. DOIs-often buried in the LID field-were parsed to ensure stable linkage for referencing purposes. Additionally, classification into publication types, whether as original research, reviews, or other forms, was maintained, recognizing that the structure of scholarly contributions could influence citation dynamics and dissemination pathways.

The overall quality of the dataset was high, but some disparities in metadata coverage emerged. Abstracts were absent in approximately 9% of articles (n = 6,912), often in older or brief report formats. MeSH term coverage was significantly lower, missing in 41.6% of entries (n = 31,902), particularly in recent publications that had not yet undergone full indexing. DOI identifiers were lacking in about 2.6% of records (n = 2,021), again predominantly among older studies. In contrast, fields such as PMID, publication type, date, and country were consistently populated, indicating a strong foundation for bibliometric analysis Table 1.

The volume of AI-related biomedical research has grown dramatically over the past decades. Before 2000, AI in this domain remained marginal, with fewer than 800 total articles published across four decades. Between 2000 and 2009, a modest uptick occurred, yielding 595 publications. From 2010 to 2014, growth continued steadily, with 424 papers published, reflecting the early integration of AI into biomedicine (Fig. 1).

However, the period following 2015 marked a paradigm shift. Between 2015 and 2019, output surged to 4,136 articles, and in 2020 alone, 4,550 AI-related studies were indexed. The trajectory steepened further: 7,577 articles were published in 2021, 10,839 in 2022, 14,076 in 2023, and 20,135in 2024-the highest annual total on record. By the partial year of 2025, 13,617 studies had already been logged. This explosive growth likely corresponds with advancements in deep learning, increased computational resources, and heightened interest following the COVID-19 pandemic.

According to the place of publication, the global profile of AI-driven biomedical research shows dominance by high-income nations, though emerging economies are steadily increasing their contributions. The United States leads unequivocally, contributing 24,958 articles-more than one-third of all indexed studies. This reflects both a rich research ecosystem and deep investment in digital health. England ranks second with 17,280 publications, supported by the UK’s emphasis on AI in public healthcare transformation (Fig. 2).

Switzerland’s output (n = 13,936) is particularly notable given its smaller population, likely reflecting concentrated innovation hubs and global pharmaceutical leadership. Other major contributors include the Netherlands (n = 4,657), Germany (n = 4,397), and Canada (n = 2,264), each backed by established biomedical and AI research infrastructures. Ireland (n = 1,045), France (n=629), and Australia (n =568) also demonstrate strong engagement. China (n = 1,355) was the 6^th globally but number one in Asia. In the rest of Asia, Korea (n = 709), India (n = 625), and Japan (n = 581), reflect growing capacity, aligning with national strategies in AI and healthcare digitization. This distribution underscores both a global concentration of research in wealthier nations and a growing inclusivity as more countries invest in AI for health (Fig. 3).

When considering the author affiliations, the landscape of global contributions to AI-driven biomedical research shifts slightly, highlighting deeper international collaboration. The United States remains the most prominent contributor, with approximately 35,000 mentions, reinforcing its central role in AI and health innovation ecosystems. China follows closely, reflecting its increasing integration into global scientific networks. The United Kingdom ranks third, driven by strategic investments in AI for healthcare delivery and academia-industry partnerships. Other significant contributors include Germany, Italy, and India, while countries such as Saudi Arabia, Switzerland, and the Netherlands also appear in the top 15, emphasizing the expanding diversity of participating nations. This distribution suggests a growing globalization of AI in medicine, with research efforts no longer concentrated solely in a few traditional scientific powerhouses.

AI-focused biomedical research is distributed across a wide spectrum of journals, signaling the field’s interdisciplinary nature. Sensors (Basel, Switzerland) tops the list with 1,919 publications, indicating a strong representation of studies involving health-related sensing technologies and diagnostic tools. Scientific Reports (n = 1,832), Diagnostics (Basel, Switzerland) (n=1096), followed by Cureus (n=990) (Fig. 4).

Journals like Diagnostics (n = 1096) and Cureus (n = 990) emphasize clinical translational work. Computational-focused publications such as Frontiers in Artificial Intelligence (n = 487488), Computers in Biology and Medicine (n = 947), and Computational Intelligence and Neuroscience (n = 450) provide dedicated spaces for technical innovation. Specialty journals-Frontiers in Oncology (n = 474), and Journal of Clinical Medicine (n = 635)-demonstrate the application of AI within domain-specific contexts. Journal of Medical Internet Research (n = 760) reveals the role of AI in informatics, behavior, and mental health. These findings confirm that AI research is not siloed in computational science but has penetrated core clinical and translational journals.

When focusing specifically on AI-related Medical Subject Headings (MeSH) and keywords-such as machine learning, deep learning, and neural networks-the analysis reveals the core technological themes shaping contemporary biomedical research. Machine learning stands out as the most frequently cited concept, appearing in 16,281 articles, underscoring its foundational role in predictive modeling, diagnostics, and clinical decision support systems. Deep learning follows with 12,297 mentions, reflecting widespread adoption of complex neural architectures, particularly in imaging, genomics, and signal interpretation. Neural networks, a broader term encompassing earlier architectures, remain prevalent (n = 7,117), particularly in foundational studies. Meanwhile, natural language processing (n = 1,681) continues to gain traction, driven by applications in electronic health record analysis and clinical documentation. Other rising domains include radiomics (n = 1,440), central to precision oncology, and emerging trends like transformers, GPT, Large Language Model (LLMs), and AI ethics, albeit with lower frequencies-signaling a new wave of interest in generative models and ethical considerations (Fig. 5).

Terms like “transformers” (n = 1,358) and “GPT” (n = 1,641) signaled a strong pivot toward large language models (LLMs) in recent years. Computer vision (n = 1,137) captured AI applications in diagnostics and surgery. Chatbots (n = 1,008) and general LLMs (n = 832) illustrated the expanding presence of AI in patient engagement and digital triage. Notably, ethical concerns were comparatively scarce, with only 110 articles indexed under AI ethics. This limited presence suggests that while technical advancements are accelerating, scholarly engagement with regulatory, equity, and safety considerations remains disproportionately low.

The insights from this analysis resonate across several domains of health and science policy. For researchers, the findings clarify where innovation is happening, which journals are most receptive, and what thematic areas are surging. This can help shape publication and collaboration strategies. For institutions and funders, the data spotlight not only strengths but also blind spots-regions and topics that warrant greater investment or capacity-building. Clinicians and educators can draw on these insights to integrate AI literacy into medical training, support clinical trials using AI tools, and promote evidence-based AI adoption. Policymakers, perhaps most crucially, must heed the signal that ethical AI is underdeveloped in the literature, and global participation remains uneven. Investments in regulation, governance frameworks, and equitable infrastructure are now as essential as algorithmic refinement.

Looking ahead, the field would benefit from deeper bibliometric efforts, including longitudinal topic modeling to map conceptual evolution, network analysis of co-authorships to chart scientific collaboration, and tracking how AI regulation impacts publication trends and clinical implementation. AI in medicine is no longer emerging. It is already here. Ongoing bibliometric monitoring will be essential to ensure that its trajectory is not only innovative but also inclusive, ethical, and aligned with real-world healthcare needs.