Healthcare LLM Research Surges From Niche Topic to Multispecialty Volume
The rise of large language models in healthcare has been unusually fast, and the bibliometric record now shows just how quickly the field moved from scattered early papers to a dense, specialty-driven literature. Across the Web of Science Core Collection from 2015 through 2025, 2,226 English-language articles and reviews mapped the growth of this research area, with publication output climbing from four papers in 2022 to 1,327 in 2025. The pace is striking on its face, but the more important point is how unevenly that expansion has spread across clinical disciplines.
Surgery accounted for the largest share of the literature at 31.7%, followed by medicine, general and internal at 24.2%, and radiology, nuclear medicine and medical imaging at 13.0%. Those proportions suggest a field that is not moving along a single track; it is branching into places where language-heavy workflows, documentation burden, image interpretation support, and decision assistance can be tested in different ways. In practical terms, the bibliometric map is already reflecting where clinicians expect these systems to fit first.
U.S. and China Anchor the Publication Base
Geographic concentration remained pronounced. The United States contributed 43.5% of all publications, far ahead of China at 14.3%, Germany at 9.2%, Turkey at 9.0%, and England at 6.8%. Collaboration patterns were centered on the United States, with dense trans-Atlantic ties linking many of the most active groups.
This matters because large language model research is not merely an academic exercise in algorithm design; it is tied to local health systems, data access, regulatory culture, and specialty norms. Countries with stronger publishing footprints are likely shaping which use cases become visible first, which validation standards are emphasized, and which risks receive attention early. The network pattern also suggests that the field remains reliant on a limited number of cross-border hubs rather than a broadly distributed international research base.
Radiology Shows the Fastest Recent Expansion
Among the specialty sub-analyses, radiology, nuclear medicine and medical imaging showed the fastest 2023-2024 growth, with a compound annual growth rate of 265%. That acceleration points to a particularly receptive environment for LLM-adjacent work, even though imaging itself is not text-based in the same way as clinical notes, consult summaries, or discharge documentation. The appeal likely lies in multimodal workflows, report generation, quality control, triage support, and the integration of narrative and visual data.
Medicine and surgery still dominate in volume, but the imaging surge signals an important shift in where technical and clinical interest is concentrating. Radiology has long functioned as a high-throughput specialty with structured reporting needs, so it is unsurprising that generative systems are being tested there early and aggressively. The data imply that adoption is not simply following publication size; it is following workflow density and the possibility of measurable assistance at multiple points in the care pathway.
Co-Citation Structure Points to a Rapidly Organizing Field
The bibliometric analysis used VOSviewer and CiteSpace to examine publication trends, citation patterns, leading countries, institutions, authors, journals, co-citation networks, and keyword structures. That kind of analysis does more than count papers. It reveals whether a field is still exploratory or has begun to crystallize around shared concepts, recurring study designs, and recognizable centers of influence.
Here, the keyword and co-citation structure indicated a research area that is becoming organized around specialty-specific applications rather than a single generalized platform narrative. In other words, the literature is moving from “Can LLMs be used in healthcare?” toward “Where do they fit best, and how should each specialty measure risk, utility, and oversight?” That shift is clinically relevant because different departments face different error tolerances, documentation demands, and regulatory pressures.
DECIDE-AI Frames the Near-Term Clinical Role
The study places its findings in the context of the DECIDE-AI framework, which is designed for developmental and exploratory clinical investigations of decision-support systems driven by artificial intelligence. Used here, the framework helps separate low-risk near-term applications from situations that demand tighter monitoring, stronger human oversight, and more rigorous prospective evaluation. For LLMs, that distinction is especially important because fluency can mask factual error, overconfidence, or hidden bias.
Near-term clinical use is most plausibly in supportive tasks: summarizing text, drafting documentation, helping clinicians navigate guidelines, organizing literature, and reducing administrative friction. Those are not trivial uses. They can influence time burden, communication quality, and access to information at the bedside. Yet the same properties that make LLMs attractive—speed, coherence, and conversational naturalness—also create safety concerns when the model produces a polished but incorrect answer.
A fluent response is not the same as a correct one.
Specialty Awareness Will Matter More Than Generic Adoption
The most useful contribution of this bibliometric study is its specialty-aware baseline. A single adoption strategy will not fit surgery, internal medicine, and radiology equally well. Surgical workflows often involve procedural planning and perioperative communication; internal medicine depends heavily on longitudinal reasoning and synthesis across multiple problems; radiology increasingly faces high-volume narrative reporting and decision support around imaging interpretation. Each of those contexts imposes different standards for evidence, oversight, and failure modes.
Viewed through that lens, the uneven publication growth is not noise. It is a sign that some specialties are already experimenting at scale while others are still defining the questions they need answered. The trans-Atlantic collaboration structure also suggests that standards may be emerging through a relatively compact set of influential research groups, which could accelerate methodological consistency but also narrow the diversity of settings represented in the literature.
Publication Velocity Has Outrun Clinical Validation
The bibliometric curve is steep enough to raise a straightforward concern: research production is moving much faster than clinical validation. With 1,327 papers appearing in 2025 alone, the field is generating knowledge faster than most health systems can assess local safety, bias, workflow impact, and regulatory fit. That gap does not invalidate the research; it defines the next stage of it.
The study does not measure patient outcomes, error rates, or deployment harms. It also does not claim that publication volume equals clinical readiness. What it does provide is a map of where enthusiasm is concentrated and where specialty-specific evidence is beginning to accumulate. For hospitals, departments, and regulators, the question is no longer whether LLMs belong in healthcare research. The sharper question is where each specialty can use them with acceptable risk, and where closer monitoring must come before routine use.