Healthcare LLM Research Grew from a Handful of Papers to a Flood in 2025
Large language models moved into healthcare at remarkable speed, and the bibliometric record now shows how quickly that shift took hold. Across the Web of Science Core Collection from 2015 through 2025, 2,226 English-language articles and reviews met inclusion criteria for this analysis of publication trends, citation patterns, leading countries, institutions, authors, journals, co-citation networks, and keyword structure. The annual output rose from four publications in 2022 to 1,327 in 2025, a compound annual growth rate of 592.26%. Few areas of medical informatics have expanded this abruptly.
The growth curve matters because it reflects more than curiosity about a new tool. It captures a field moving from exploratory experimentation toward specialty-specific clinical questions, governance concerns, and implementation problems. The rapid rise after 2022 also mirrors the release and broad uptake of modern LLMs, which lowered technical barriers for drafting, summarization, translation, information retrieval, triage support, and other language-intensive tasks. In healthcare, those capabilities touch nearly every discipline, but the bibliometric footprint is not uniform across specialties.
The United States led the publication output with 43.5% of the total. China followed at 14.3%, then Germany at 9.2%, Turkey at 9.0%, and England at 6.8%. Collaboration networks remained centered on the United States, with dense trans-Atlantic ties connecting many of the most active contributors. In practical terms, the center of gravity for the literature still sits in North American and European academic ecosystems, even as output broadens internationally.
Specialty Patterns Show Uneven Adoption Across Clinical Fields
The specialty breakdown is where the field becomes more clinically legible. Surgery accounted for 31.7% of the publications, medicine, general and internal for 24.2%, and radiology, nuclear medicine and medical imaging for 13.0%. These proportions suggest that LLM development has not spread evenly across the medical curriculum or hospital workflow. Instead, it is clustering in domains where documentation burden is heavy, image-linked reporting is routine, and decision support can be wrapped around structured clinical pathways.
Radiology, nuclear medicine and medical imaging showed the fastest 2023-2024 growth, with a CAGR of 265%. That acceleration likely reflects the specialty’s high-volume interpretive workflow and its longstanding familiarity with computational assistance. Radiology also sits near the interface between text and image, making it a natural testing ground for language models that can summarize findings, generate draft reports, organize differential diagnoses, and integrate clinical context with imaging outputs. The pace of growth in that subgroup suggests the specialty has moved from early curiosity into active experimentation.
Surgery’s larger share of the literature points to a different kind of need. Operative care produces dense notes, handoffs, informed consent discussions, perioperative planning documents, and patient education demands. LLMs may be attractive in surgery not because they replace technical judgment, but because they can reduce the friction surrounding communication-heavy tasks that consume time away from the operating room. In general and internal medicine, the appeal is broader and more diffuse: symptom interpretation, longitudinal documentation, discharge planning, and patient-facing communication all create opportunities for language-based support.
A single sentence captures the pressure point here: healthcare is using LLMs first where language work is already overwhelming clinicians.
DECIDE-AI Frames the Near-Term Clinical Uses and Safety Questions
The analysis links these bibliometric trends to the DECIDE-AI framework, which was designed for developmental and exploratory clinical investigations of decision support systems driven by artificial intelligence. That framework is useful because it prevents the field from jumping too quickly from publication volume to bedside confidence. LLMs can be studied as low-risk tools in early deployment settings, but their clinical legitimacy depends on careful attention to performance drift, workflow fit, human oversight, and equity across patient groups.
This is where bibliometric mapping becomes more than a descriptive exercise. By identifying leading journals, authors, institutions, and keyword structures, the study outlines where evidence is accumulating and where it remains thin. A specialty-aware map can help clarify which applications are close to clinical use and which are still mostly methodological. It also creates a baseline for monitoring whether research effort is being distributed toward the highest-value clinical problems or concentrated in areas with the easiest publishing pathway.
The publication surge also raises a familiar translational problem: volume is not the same as validation. A rapidly expanding literature can create the appearance of readiness even when many studies remain early, small, or laboratory-based. In a field built around generative output, the temptation to equate fluent text with clinical reliability is especially strong. Bibliometric growth therefore has to be interpreted alongside the strength of evaluation methods, the relevance of the endpoint, and the degree of human supervision used in each setting.
One direct bedside consequence follows from that tension. If LLMs are deployed in documentation, triage, or decision support without specialty-specific guardrails, errors may propagate through routine workflow long before they reach formal review.
What the Publication Map Suggests for Clinical Practice and Regulation
The central message is straightforward: LLM research in healthcare has become large enough to display specialty-specific trajectories, and those trajectories are already diverging. Surgery, general medicine, and imaging-related disciplines dominate the current literature, but they do so for different reasons and with different expected uses. Imaging specialties are accelerating quickly, while transatlantic collaboration continues to shape the research agenda. The United States remains the dominant hub, but the international spread of authorship suggests the field is becoming more globally distributed.
For clinicians and health systems, the value of this map lies in prioritization. It points toward domains where near-term use is most plausible and where monitoring should be most intense. DECIDE-AI offers a sensible lens for that process because it emphasizes staged development rather than immediate clinical substitution. The framework suits a technology whose greatest strengths are linguistic fluency, summarization, and pattern recognition in text-rich environments, while its greatest risks remain hallucination, bias, overconfidence, and weak external validation.
The study does not settle the clinical worth of any single LLM application, and it does not measure patient outcomes. Instead, it documents the shape of the field as it expands and identifies where that expansion is strongest. The next question for each specialty is whether research intensity is being matched by rigorous evaluation, transparent reporting, and governance structures that can keep pace with deployment. Without those controls, rapid publication growth can outstrip safe adoption just as quickly as it began.