Global trends and emerging frontiers of large language models in cancer research

Abstract

Objective

The integration of Large Language Models (LLMs) into cancer research has progressed rapidly, but a comprehensive understanding of global trends, key contributors, and emerging research areas remains lacking. This gap hinders a comprehensive understanding of the development landscape for LLM applications in clinical oncology.

Methods

A bibliometric analysis was conducted using publications retrieved from the Web of Science Core Collection on March 15, 2026. Eligible studies were limited to English-language articles and reviews published till 2025. Records unrelated to LLMs or cancer, duplicates, retracted publications, and those missing complete metadata were excluded. A total of 896 publications were analyzed using VOSviewer, CiteSpace, and R. ClinicalTrials.gov was searched with the same term, obtaining 29 eligible trials.

Results

Publication output increased sharply from 2022 to 2025. The USA and China dominated global output, with Germany demonstrating disproportionate citation efficiency relative to volume, and Heidelberg University and Harvard University leading institutionally. Research hotspots converged on LLM benchmarking, domain-specific fine-tuning, multi-omics integration, and perioperative applications. Among 29 registered trials, application areas spanned patient communication, shared decision-making, and care equity outcomes, reflecting a transition from proof-of-concept toward randomized evaluation.

Conclusions

LLM-driven oncology research has expanded rapidly but remains geographically and institutionally concentrated, with prospective multicenter validation still scarce. Research is transitioning from foundational benchmarking toward fine-tuning, multimodal integration, and clinical deployment. Strengthening cross-institutional collaboration, diversifying trial populations, and developing standardized safety evaluation frameworks are essential for translating bibliometric growth into meaningful advances in cancer diagnosis, treatment, and patient outcomes.

Keywords

large language models (LLMs)artificial intelligence (AI)cancer research bibliometric analysis clinical trials

Introduction

Cancer is one of the leading public health challenges globally and has become a primary cause of death. According to data from the World Health Organization (WHO), there were over 18 million new cancer cases and more than 10 million cancer-related deaths globally in 2023, with projections indicating a further rise by 2050.¹ Breakthroughs are urgently needed in key areas of cancer prevention and treatment, particularly in precise early diagnosis, personalized treatment, and refined patient management. At the same time, cancer research faces the exponential growth of data, the complexity of processing unstructured information, and the difficulty of integrating multimodal data. These challenges underscore the need to develop efficient and intelligent technical tools. Large language models (LLMs) are injecting new vitality into cancer research, with the potential to integrate multimodal data and enable contextual reasoning.

The application of LLMs has progressively permeated several core areas of cancer research. At the clinical decision-making level, LLMs can integrate multimodal data, such as computed tomography (CT) imaging and histopathological results, to support diagnosis and prognostic prediction.^2,3 In terms of comprehensive patient management, LLMs can assist first-time patients in creating a list of questions for their initial consultations,⁴ optimizing pre-clinical examination preparation processes, simplifying imaging and pathology reports,⁵ providing evidence-based postoperative care recommendations, and alleviating patients’ emotional distress through empathetic communication.⁶ In medical education, LLMs can generate visualized teaching content, serving as an efficient tool for training resident physicians.⁷ In basic research, LLMs contribute to drug development, such as predicting drug synergy in rare tissues using few-shot learning.⁸ These applications have significantly improved clinical efficiency and research innovation, driving the full-link optimization of cancer care, from early screening and diagnostic classification to treatment interventions and rehabilitation management.

Globally, research on LLMs in cancer has surged, with ongoing results emerging from clinical trials, technical validation, and application exploration. However, there is still a lack of systematic evaluation of global trends and emerging frontiers. Bibliometric analysis allows for statistical analysis of trends in literature publication, core authors, keyword clustering, and citation networks. It can objectively reveal the implicit patterns of knowledge flow within a discipline, precisely identify high-impact research topics, and help identify patterns and emerging frontiers that may inform future research priorities.⁹ This method has been widely applied in AI-integrated healthcare. However, a gap remains in bibliometric research on LLMs in cancer research.

This study aims to address this research gap by conducting a bibliometric analysis to systematically explore global research trends, core directions, emerging hotspots, and potential challenges in LLMs in cancer research. The findings will provide theoretical support and practical guidance for the deep integration of LLMs into cancer research, accelerate clinical translation, and offer a reference framework for their application in other medical specialties.

Methods

Data source and literature search strategy

Relevant articles were retrieved from the Web of Science Core Collection (WoSCC), restricted to the Science Citation Index Expanded (SCI-EXPANDED) edition, with the search conducted on March 15, 2026. The search strategy was developed collaboratively by all authors in consultation with senior literature search experts (XL and YYX), using LLM-related and cancer-related search terms as detailed in Table 1. The study selection process was conducted in two stages. An initial screening was performed based on article titles, followed by a secondary screening using abstracts. The detailed selection procedure is illustrated in the PRISMA flowchart provided in the Supplemental Materials. To identify cutting-edge advancements in LLM research and enhance the comprehensiveness of literature retrieval for clinical trials, we concurrently expanded our retrieval scope to include the latest clinical trials registered on ClinicalTrials.gov (https://clinicaltrials.gov/). This was followed by an analysis of the retrieved data, conducted on March 15, 2026. ClinicalTrials.gov, maintained by the U.S. National Library of Medicine (NLM), is the world’s largest clinical trial registry, established to enhance transparency and reduce publication bias. It currently contains over 470,000 studies across more than 200 countries and serves as a key resource for accessing global clinical research information.¹⁰ After excluding unrelated studies (n = 21), we identified 29 clinical trials, categorized as completed, ongoing, or not yet recruiting participants. The detailed search query employed is shown in Table 1.

Table 1.

PRISMA-based search strategy and study selection for web of science core collection (WoSCC) and clinicaltrials.gov.

PRISMA stage	WoSCC (bibliometric analysis)	ClinicalTrials.gov (clinical trials)
Identification	Search strategy: TS = (#1) AND TS = (#2)	Search strategy (Expert Search): (#1) AND (#2)
	#1: LLM-related terms #2: Cancer-related terms	#1: LLM-related terms #2: Cancer-related terms
	Initial records identified: n = 1314	Initial records identified: n = 50
	Limits: Article or Review, English, final publication date before 2026	Initial records identified: n = 50
Manual Screening	Excluded unrelated to LLMs: n = 359	Excluded unrelated to LLMs: n = 12; Excluded unrelated to cancer: n = 9
	Excluded unrelated to cancer: n = 50
	Duplicates removed: n = 3
	Retracted articles: n = 4
	Records without full metadata or citation data: n = 2
Eligibility	Eligible studies: n = 896	Eligible studies: n = 29
Included	Included in final analysis: n = 896	Included in final analysis: n = 29

Note. #1 LLM-related terms: “large language model*” OR “LLMs” OR “LLM” OR “Generative Pre-trained Transformer*” OR “generative pretrained transformer*” OR “ChatGPT*” OR “GPT-4” OR “DeepSeek” OR “Google Gemini” OR “Google Bard” OR “Llama*” OR “Microsoft Copilot” OR “Claude*” OR “Grok” OR “Mistral*” OR “Qwen*”; #2 Cancer-related terms: “cancer” OR “cancers” OR “adenocarcinoma” OR “Neoplasias” OR “Malignancy” OR “Malignancies” OR “oncogene” OR “oncogenic” OR “oncologist” OR “neoplasms” OR “neoplasm” OR “neoplasia” OR “tumor*” OR “tumour*” OR “sarcoma” OR “oncology” OR “carcinoma” OR “hematology” OR “leukemia” OR “multiple myeloma” OR “lymphoma”.

Data analysis

In this study, R (version 4.2.2), VOSviewer (version 1.6.20),¹¹ and CiteSpace (version 6.4. R1)¹² were used to perform the bibliometric analysis. The Bibliometrix R package (version 4.1.4) was used to draw the country contribution graph. Using VOSviewer, we analyzed publication, citation, and keyword counts. Co-occurrence networks from keywords were built and visualized using the clustering algorithm in VOSviewer, based on co-authorship and co-occurrence analysis. It was also used to analyze collaborations among countries, institutions, and authors. VOSviewer employed the full-counting method to quantify occurrences, in which each instance of a term in a publication is counted once, and all co-occurrence links carry equal weight. Similar keywords and identically named organizations merge to concentrate influence and make interactions more precise, as in Table SM1. CiteSpace was utilized to identify keywords that received significant citations over a defined period. The total time span was divided into one-year slices. Node selection criteria were applied to construct the network using the Top N method (N=50) and the g-index (k=25) to better include influential nodes. The Network Pruning Pathfinder method was used to simplify the network and highlight its main structure. Burst detection was conducted using the Kleinberg algorithm, with a minimum duration typically set to 1 year. The online bibliometrics platform enabled us to visualize international collaborations effectively (https://bibliometric.com/). Microsoft Excel was used to analyze the exponential growth in publications in this area. The flowchart of this study is depicted in Figure 1.

Figure 1.

Study workflow illustrating data sources, Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA)-based selection, and bibliometric analysis pipeline.

Results

Publication trends and geographical distribution

Among the 896 articles on LLMs in cancer research, publication output shows a clear upward trend (Figure 2(a)–(b)), with both annual and cumulative numbers rising sharply from 2022 to 2025. Geographical mapping (Figure 2(c)) shows that the United States (USA) and China dominate, with 304 and 241 publications, consistent with their top rankings in Table 2. The USA also leads in citation impact (6,303 vs. China’s 2,340), underscoring its stronger influence. European countries, particularly Germany and the United Kingdom (UK), exhibit robust performance in both output and citations. The cooperation chord diagram (Figure 2(d)) and collaboration network (Figure 2(e)) highlight dense international partnerships, with the USA as the central hub connecting to Canada, South Korea, and Turkey, while China maintains broad links across Asia and the West. European nations (Germany, the UK, Italy, and France) form tight regional clusters. Second-tier countries such as Turkey, Italy, Canada, and South Korea show solid productivity, whereas Saudi Arabia and Switzerland achieve notably high citation impact despite lower publication volumes. The time-overlapping network (Figure 2(f)) indicates that early research was concentrated among a few leaders but has since expanded to emerging contributors, including India, Saudi Arabia, and Egypt.

Figure 2.

Global distribution, temporal trends, and collaborative networks of countries/regions in LLM-driven cancer research. (a) Annual and cumulative publication trends in LLM-driven cancer research; (b) Distribution of publications by major contributing countries over time; (c) Global geographical visualization of national publication output, with darker colors indicating higher volumes; (d) Chord diagram illustrating international collaboration among countries/regions; (e) Clustering network of country/region co-occurrence; (f) Time-overlapping network of country/region co-occurrence.

Table 2.

Top 21 countries ranked by research output and citation impact in LLM-driven cancer research.

Rank	Country	Documents	Citations
1	USA	304	6303
2	China	241	2340
3	Germany	107	2036
4	UK	65	1504
5	Turkey	58	520
6	Italy	54	614
7	Canada	44	523
8	South Korea	38	347
9	Spain	28	441
10	Switzerland	27	510
11	France	27	382
12	India	27	290
13	Australia	25	272
14	Japan	21	347
15	Israel	21	329
16	Saudi Arabia	16	549
17	Singapore	15	145
18	Netherlands	12	221
19	Greece	12	136
20	Belgium	11	167
21	Egypt	10	61

Note. Documents: total number of publications indexed in the Web of Science Core Collection. Citations: total citation count. Countries are ranked primarily by publication output (Documents). Analysis was performed using VOSviewer (version 1.6.20).

Institutional and author analysis

Institutional and author-level analyses (Figure S1) reveal the structural dynamics of LLM-driven cancer research. As shown in Table 3, leading affiliations are dominated by institutions from the USA, Europe, and China. Heidelberg University ranks first (31 publications; 894 citations), followed by Harvard University (25 publications; 685 citations). Chinese institutions such as Sun Yat-sen University and Sichuan University show strong productivity, with Sichuan University achieving a notably high citation impact (784 citations). The organizational collaboration network (Figure S1A) displays clear regional clustering: Western institutions (Harvard University, Mayo Clinic, University of Toronto) form tightly interconnected groups, while Chinese institutions (Sun Yat-sen University, Zhejiang University, Shanghai Jiao Tong University) cluster domestically with fewer cross-regional links. Heidelberg University and the Technical University of Dresden serve as key nodes that bridge European and global networks. In fact, Heidelberg University is characterized by early adoption of clinically integrated AI frameworks and participation in dense collaborative networks. The time-overlapping network (Figure S1B) indicates early activity concentrated in Western institutions, with increasing recent contributions from Asian institutions such as Shanghai Jiao Tong University and Zhejiang University after 2024. At the author level (Figure S1C-D), collaboration remains fragmented. Core authors, including Kather Jakob Nikolas (Table S1), serve as central nodes with high productivity and strong links. His research group is a young and interdisciplinary team recognized for integrating computational modeling with clinical oncology, particularly in cancer genomics, targeted therapy, and immunotherapy. Other contributors, such as Liu Wei and Truhn Daniel, also show notable output and influence, yet most authors form small, loosely connected clusters.

Table 3.

Top 23 global institutions ranked by publication output and citation impact in LLM-driven cancer research.

Rank	Organization	Documents	Citations
1	Heidelberg University	31	894
2	Harvard University	25	685
3	Sun Yat-Sen Univ	21	353
4	Technical University of Munich	21	158
5	Sichuan University	18	784
6	University of Toronto	18	335
7	Mayo Clinic	18	210
8	Stanford University	15	341
9	Shanghai Jiao Tong University	15	81
10	Technical University of Dresden	14	582
11	Memorial Sloan Kettering Cancer Center	14	202
12	Zhejiang University	14	110
13	Central South University	14	87
14	Tel Aviv University	13	268
15	China Medical University	12	115
16	Massachusetts General Hospital	11	528
17	New York University	11	453
18	Icahn School of Medicine at Mount Sinai	11	198
19	Peking University	11	78
20	University of Southern California	10	185
21	Chinese University of Hong Kong	10	112
22	Chinese Academy of Sciences	10	88
23	Sungkyunkwan University	10	75

Note. Documents: total number of publications indexed in the Web of Science Core Collection. Citations: total citation count. Institutions are ranked primarily by publication output (Documents). Analysis was performed using VOSviewer (version 1.6.20).

Journal and citation analysis

As shown in Table S2, the 18 most productive journals collectively account for a substantial proportion of the 896 included publications. Journal of Medical Internet Research ranks first (26 publications), followed by npj Digital Medicine and Scientific Reports (20 each), and Cancers (19). High-impact general journals such as Nature Communications and Radiology each contributed 10 publications, indicating recognition beyond specialized venues. Citation burst analysis (Figure S2) further identifies the most influential publications in this field, with most bursts beginning in 2023 and continuing through 2025, predominantly appearing in leading journals including Nature, Science, Radiology, and Lancet Digital Health. Foundational works by Devlin et al. (2019) and Topol (2019) exhibit strong, sustained citation bursts, while a substantial number of highly cited studies from 2023 focus specifically on ChatGPT and related LLM technologies, covering diverse applications, including clinical decision support, medical imaging interpretation, and patient-physician communication.

Analysis of keywords

The co-occurrence network (Figure 3(a)) provides an overview of the research structure in LLM-driven cancer studies, where node size represents keyword frequency, while distance and link strength indicate association strength. According to Table 4, the most prominent keywords are “artificial intelligence” (419 occurrences), “large language model” (366), and “ChatGPT” (256), with other high-frequency terms including “cancer,” “natural language processing,” “diagnosis,” and “machine learning.” Cluster analysis reveals major thematic groups: the red cluster focuses on core LLM technologies and methodological development (natural language processing, deep learning, prediction, classification); the green cluster on clinical applications, particularly patient education, health literacy, communication, and information dissemination; the blue cluster on clinical decision-making and oncology practice (clinical decision support, survival, treatment-related outcomes); and the yellow cluster on disease-specific applications and guideline-related research (breast cancer, colorectal cancer, medical education). The time-overlapping network (Figure 3(b)) shows early focus on foundational AI techniques (machine learning, model development), shifting toward recent application-oriented themes (“ChatGPT,” “patient education,” “quality,” “healthcare”). The CiteSpace thematic clustering map (Figure 3(c)) identifies key domains, with #0 ChatGPT as the largest and most central cluster; other notable clusters include #1 foundation model, #2 health literacy, #3 natural language processing, #7 lung cancer, and #8 thyroid cancer, with connections demonstrating interdependence among technological development, clinical applications, and disease-specific research.

Figure 3.

Keyword co-occurrence patterns and thematic clustering in LLM-driven cancer research (a) Clustering network of keyword co-occurrence generated by VOSviewer; (b) Time-overlapping network of keyword co-occurrence showing temporal evolution of research topics; (c) Thematic clustering map of keywords generated by CiteSpace, with nodes representing clusters and connecting lines indicating dependency relationships among research themes.

Table 4.

Top 35 keywords ranked by occurrence frequency and total link strength in LLM-driven cancer research.

Rank	Keyword	Occurrences	Total link strength
1	artificial intelligence	419	1094
2	large language model	366	862
3	ChatGPT	256	669
4	cancer	131	383
5	natural language processing	79	231
6	breast cancer	64	166
7	diagnosis	63	229
8	machine learning	59	209
9	deep learning	43	135
10	oncology	43	133
11	information	35	103
12	chatbot	34	114
13	prediction	33	124
14	patient education	33	119
15	classification	32	104
16	Google Gemini	28	83
17	lung cancer	28	80
18	medicine	24	99
19	health	22	75
20	prostate cancer	22	75
21	readability	21	77
22	model	21	70
23	risk	21	58
24	quality	19	65
25	healthcare	18	67
26	pathology	18	60
27	management	18	59
28	care	18	57
29	survival	18	47
30	accuracy	17	60
31	internet	17	60
32	clinical decision support	17	48
33	challenge	16	63
34	therapy	16	51
35	radiation oncology	16	41

Note. Documents: total number of publications indexed in the Web of Science Core Collection. Occurrences: total frequency of each keyword. Total Link Strength: a VOSviewer-derived metric reflecting the cumulative strength of co-occurrence connections within the keyword network. Keywords are ranked primarily by frequency of occurrence. Analysis was performed using VOSviewer (version 1.6.20).

Landscape of clinical trials

A total of 29 clinical trials were identified from ClinicalTrials.gov, comprising 24 interventional and 5 observational studies, of which 3 were completed. Detailed characteristics are presented in Supplemental Table S3. These trials primarily explore LLM applications in clinical diagnosis and treatment decision-making, patient management and care, psychological and emotional support, and patient and healthcare professional education, indicating their potential across the cancer care continuum while remaining largely in exploratory phases.

Discussion

Global research patterns and regional disparities

Our bibliometric analysis identified 896 publications from 2022 to 2025. The USA (304 publications; 6,303 citations) and China (241 publications; 2,340 citations) dominated global output, followed by Germany (107 publications; 2,036 citations) and the UK (65 publications; 1,504 citations). Heidelberg University (31 publications; 894 citations) and Harvard University (25 publications; 685 citations) led institutional productivity and citation impact. Collaboration networks were hierarchical, with the USA as the central hub, China as a bridge between Eastern and Western communities, and European nations forming tight regional clusters. These patterns reflect meaningful structural differences in how research capacity translates into academic influence. Although the USA and China contribute comparable volumes of output, the substantially higher citation impact per publication observed in the USA indicates qualitative rather than merely quantitative leadership. At the author and institutional levels, collaboration remains partially fragmented, with many groups operating within localized clusters and limited cross-regional integration. This structural constraint may attenuate the field’s overall translational efficiency. Several interconnected factors underlie these regional disparities. The USA’s citation dominance reflects earlier institutional engagement with foundational LLM technologies, preferential access to high-impact journals, and the structural advantage of large academic medical centers with integrated clinical and computational research programs. China’s high publication volume, combined with comparatively lower citation impact, is consistent with a rapidly expanding research base still maturing in international network integration and journal access, a pattern previously observed in other emerging biomedical fields. The growing participation of contributors from Southeast Asia and the Middle East further signals an ongoing globalization of the field, driven in part by the increasing accessibility of open-source LLM frameworks. Strengthening cross-institutional collaboration will be essential for accelerating knowledge exchange and ensuring equitable distribution of research benefits across diverse healthcare systems.

Notably, Germany ranks third globally in publication output but generates a citation impact that approaches that of China, yielding a citations-per-publication ratio approximately 2.3-fold higher than China and comparable to that of the USA. At the institutional level, this pattern is exemplified by Heidelberg University and the Technical University of Dresden, both of which demonstrate disproportionate citation influence relative to output. Notably, Heidelberg University ranks first globally in institutional productivity, and researcher Kather Jakob Nikolas from Clinical Artificial Intelligence at TU Dresden (https://digitalhealth.tu-dresden.de/) appears as a central node in our author collaboration network, suggesting that a small number of highly connected, methodologically rigorous teams may be associated with Germany’s outsized influence. Germany’s impact-oriented profile may be associated with several structural characteristics, including tight integration between university hospitals and AI research institutes, participation in dense intra-European collaborative networks, and a funding landscape that has historically prioritized translational rigor. Collectively, these features suggest that Germany’s citation efficiency could be partly explained by a deliberate, structurally supported research strategy, one that may offer instructive lessons for communities seeking to maximize translational impact in LLM-driven oncology.

Current hotspots

The bibliometric analysis using CiteSpace revealed 10 distinct clusters, reflecting the structural complexity of the research field. However, to provide a more integrated and insightful overview, the following discussion employs a higher-level narrative synthesis rather than an exhaustive report of each individual cluster. We have synthesized the most prominent trends and core concepts into four primary thematic domains, guided by keyword co-occurrence patterns, citation burst detection, and a systematic review of the retrieved literature (Figure 4). This approach allows for a more focused exploration of the underlying logic and future directions of the field, spanning from foundational model benchmarking to real-world clinical deployment.

Figure 4.

Current research hotspots of LLMs in cancer studies. This figure presents a synthesized overview of four primary thematic domains that were derived from the authors’ integration of the ten CiteSpace clusters identified through bibliometric analysis, together with keyword co-occurrence networks, citation burst detection, and an in-depth review of the retrieved literature. It is intended to illustrate the overall translational pipeline from foundational model evaluation to clinical application, rather than to directly reproduce the automated clustering outputs of CiteSpace.

Comparative studies between benchmarking methods

Benchmarking and comparative evaluation of mainstream LLMs, including GPT-3.5, GPT-4, Claude, Gemini, Microsoft Copilot, and DeepSeek, represents the most prominent and rapidly expanding research cluster identified in our analysis.^13,14 Evidence consistently shows that newer models outperform earlier chatbots, yet performance remains domain-dependent. GPT-4 demonstrated superior performance over GPT-3.5 and Claude in breast oncology across treatment decision-making, prognosis assessment, and psychosocial support, though occasional inaccuracies persisted.¹⁵ A network meta-analysis of 168 studies confirmed specialty-specific leaderboard variation: GPT-4/4o ranked first in dentistry, oncology, and radiology by SUCRA, while other models ranked higher in orthopedics or urology.¹⁶ In gynecological cancer TNM staging, Gemini 1.5 achieved accuracies of 0.994 (T stage) and 0.993 (N stage), and Qwen2.5 72B reached 0.971 and 0.923, respectively, both outperforming manual entry.¹⁷ However, hallucination and citation fabrication remain persistent limitations across all models.¹⁸

Fine-tuning domain-specific LLMs for personalized tasks

A growing body of work has focused on adapting general-purpose LLMs to specialized oncological tasks through fine-tuning and retrieval-augmented generation (RAG).¹⁹ Representative examples illustrate the breadth of this approach. BioBERT, fine-tuned for biomedical text mining, enabled more efficient extraction and analysis of oncology literature,²⁰ while Med-PaLM demonstrated strong medical question-answering capability by excelling in the USMLE.²¹ In cancer imaging, RAG-augmented NotebookLM achieved 86% accuracy in lung cancer CT staging, outperforming GPT-4o,²² and LLMSeg, a multimodal model integrating image- and text-based clinical information, demonstrated superior generalization over traditional unimodal models in three-dimensional target volume delineation for radiation oncology.²³ In clinical decision support, retrieval-enhanced LLMs delivered oncology treatment recommendations 4 to 20 times more effectively than pre-trained baselines in clinical trial matching for head and neck tumors, with performance gains amplified by prompt specificity and dataset quality.²⁴ Collectively, these studies demonstrate that domain adaptation, spanning zero-shot prompting, RAG, and fine-tuning, substantially enhances LLM performance across staging, imaging, and treatment recommendation tasks, though hallucination and variable generalization remain unresolved challenges.²⁵

Multi-omics integration at the proteomic and genomic levels

Multi-omics integration represents an emerging frontier in LLM-driven oncology. DrBioRight 2.0 exemplifies this approach, integrating nearly 8,000 TCGA patient datasets and 900 Cancer Cell Line Encyclopedia samples into an LLM-powered platform that enables natural-language querying of proteogenomic data, survival analysis, and biomarker correlation, effectively lowering the technical barrier for large-scale functional proteomics research.²⁶ At the genomic level, NLP-based LLMs have demonstrated the ability to extract tumor mutation burden from unstructured clinical documents and link genomic profiles with clinical outcomes across 13 cancer types, while a proposed framework for responsible LLM integration into precision oncology, validated in uterine carcinosarcoma, offers a structured pathway for translating these capabilities into clinical practice.²⁷ In drug discovery, LLMs are being applied to predict drug synergy and optimize therapeutic strategies. CancerGPT used few-shot learning to predict drug synergy in rare tissue types, outperforming traditional models with minimal training data,⁸ while the BAITSAO model leveraged GPT-3.5 embeddings to generalize across drug synergy and single-drug inhibition prediction tasks.²⁸ Together, these examples demonstrate LLMs’ capacity to accelerate high-throughput drug sensitivity screening and support personalized treatment decisions, though validation at scale and interpretability remain key challenges.

The clinical value of LLMs in perioperative management

Cancer surgery presents uniquely complex perioperative challenges, including the management of patients undergoing multimodal treatment sequences like neoadjuvant therapy, radical resection, and adjuvant therapy, that demand real-time integration of oncological guidelines, imaging findings, and laboratory data. Emerging evidence suggests LLMs can address these challenges across three domains. In risk stratification, GPT-4 Turbo predicted ICU admission and in-hospital mortality from preoperative free-text notes with F1 scores of 0.81 and 0.86, respectively,²⁹ and LLM-based models demonstrated substantial AUROC and AUPRC gains over traditional word-embedding approaches for postoperative complication prediction.³⁰ In clinical decision support, the PEACH system integrated 35 institutional protocols into a Claude 3.5-based framework, achieving 96.7% overall accuracy with minimal hallucination and expediting decision-making in 95% of cases.³¹ In patient education, RAG-augmented NeuroBot delivered personalized perioperative counseling across a broad range of patient queries.³² Collectively, these examples suggest that LLMs can complement classical risk-scoring systems and standardize complex oncological perioperative workflows, though prospective validation in dedicated cancer surgical populations remains necessary.

Practical implications and guidance for future research

Our clinical trial landscape analysis of 29 registered clinical trials (Table S3) provides insight into the current translational priorities. LLMs have achieved demonstrable clinical utility across domains. In patient communication and education, registered trials are actively evaluating LLM-based counseling for anxiety and depression management (NCT06854315, completed), shared decision-making in prostate cancer (NCT06856694), personalized question-prompt generation for hematologic malignancies (NCT07226934), and AI-assisted chemotherapy side-effect management (NCT07198581), suggesting a rapid transition from proof-of-concept to randomized evaluation. In perioperative and radiotherapy workflows, LLMs are being deployed to collect symptom data during radiotherapy (NCT06525181) and to support thyroid cancer management (NCT07234539). Several structural gaps warrant targeted investment. Despite 29 registered trials identified in our analysis, prospective multicenter RCTs remain scarce relative to the number of bibliometric outputs. Some trials are in early recruitment, or not yet recruiting (NCT06856694, NCT07226934), and a few are designed with oncology-specific primary endpoints that would satisfy regulatory standards for clinical integration. Hallucination and variable cross-domain performance, identified as persistent limitations across benchmarking studies, must be addressed through domain-specific fine-tuning and standardized safety evaluation frameworks before broader deployment in trials. Notably, trials such as SAFE. AI (NCT07410689) and TALK-U (NCT07325617) signal an emerging priority beyond clinical task performance, addressing financial toxicity and health literacy.

For researchers entering this field, we recommend closely following the output of high-impact institutions identified in our analysis, Heidelberg University, Harvard University, and major Chinese academic medical centers, including Sun Yat-sen University and Sichuan University, as these hubs define current methodological frontiers. Priority research directions include domain adaptation and RAG-augmented systems, LLM evaluation across diverse language and healthcare settings, multimodal integration of imaging and genomic data, and prospective trial design with oncology-specific endpoints. For clinician-researchers, the strongest evidence currently supports LLMs as decision-support adjuncts in documentation, staging, and patient communication, with the greatest opportunity in studies prospectively validating structured LLM outputs against clinical outcomes.

Limitations

Although this bibliometric study offers valuable insights into the use of LLMs in cancer research, several methodological limitations should be considered. Relying solely on English-language literature indexed in WoSCC may introduce selection bias, potentially missing regional innovations. Despite all advances, LLM applications remain predominantly academic and retrospective, with limited integration into real-world clinical workflows.^33,34 Most evaluations rely heavily on quantitative datasets and automated methods, without involving human raters. The current limitations of LLMs in cancer-related decision-making, including their limited domain knowledge and reliance on human supervision, underscore the need for open datasets and standardized evaluations to enhance their reliability.³⁵ In the meantime, both the bibliometric and trial registration data show only academic interest and planned studies, while real clinical utility requires completed prospective trials with clearly defined patient-centered endpoints, a standard most registered trials haven’t yet met.³⁶

Conclusion

This bibliometric analysis of 896 publications and the landscape of 29 registered clinical trials map the rapid evolution of LLM-driven oncology research. While the USA and China dominate output, citation efficiency disparities suggest that collaborative network density may contribute to academic influence alongside publication volume. Research has progressed from benchmarking toward fine-tuning, RAG-augmented systems, and multi-omics integration, with clinical trials now evaluating LLMs across patient communication, decision support, and care equity. Hallucination, limited prospective validation, and geographic concentration remain key barriers. Implementing recommended collaborative and technical improvements is a necessary step for turning bibliometric growth into real progress in cancer treatment.

Supplemental material

Supplemental material - Global trends and emerging frontiers of large language models in cancer research

Supplemental material for Global trends and emerging frontiers of large language models in cancer research by Dianzhe Tian, Zhixuan Xie, Zixuan Hu, Zuyi Yang, Hu Tian, Youxin Chen, Haitao Zhao, Shunda Du, Fengdan Wang, Lei Zhang, Yiyao Xu and Xin Lu in Digital Health.

Supplemental material

Supplemental material - Global trends and emerging frontiers of large language models in cancer research

Supplemental material

Supplemental material - Global trends and emerging frontiers of large language models in cancer research

Footnotes

Acknowledgments

The authors thank CiteSpace, VOSviewer, and all relevant R packages for providing free access to researchers.

ORCID iDs

Dianzhe Tian

Hu Tian

Xin Lu

Author contributions

D.T.: Conceptualization, Methodology, Software, Writing- Original draft preparation; Z.X.: Writing- Original draft preparation; Z.H.: Writing- Original draft preparation; Z.Y.: Visualization, Writing- Reviewing and Editing; S.D.: Supervision; H.Z.: Supervision; F.W.: Supervision, Project administration; L.Z.: Writing- Reviewing and Editing; Y.X.: Supervision, Project administration; X.L.: Supervision, Project administration

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This article was partially sponsored by Beijing Natural Science Foundation (Grant No. 7262093), the National High-Level Hospital Clinical Research Funding (2022-PUMCH-C-049, 2022-PUMCH-B-067), the Youth Fund of the National Natural Science Foundation of China (Grant No. 82001900), the CAMS Innovation Fund for Medical Sciences (2021-I2M-1-051), and the Beijing Students’ Funding for Innovation and Entrepreneurship Training Program (2025dcxm165).

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data Availability Statement

The datasets analyzed during the current study are available in public databases.*

Declarations of AI use

We confirm that no AI tools were used in preparing this manuscript. All aspects of the work were completed independently by the authors.

Guarantor

Xin Lu.

Supplemental material

Supplemental material for this article is available online.

References

GBD 2023 Cancer Collaborators . The global, regional, and national burden of cancer, 1990-2023, with forecasts to 2050: a systematic analysis for the Global Burden of Disease Study 2023. Lancet Lond Engl 2025; 406(10512): 1565–1586, PMID:41015051.

Kehl

Gusev

, et al. Artificial intelligence-aided clinical annotation of a large multi-cancer genomic dataset. Nat Commun 2021; 12(1): 7304. https://doi.org/10.1038/s41467-021-27358-6

Park

Patterson

Acitores Cortina

, et al. Enhancing EHR-based pancreatic cancer prediction with LLM-derived embeddings. NPJ Digit Med 2025; 8(1): 465, PMID:40691317. https://doi.org/10.1038/s41746-025-01869-8

Wang

Chen

Zhang

, et al. Large language models could be applied in personalized out-of-hospital management for breast cancer: a prospective randomized single blind study. Sci Rep England 2025; 15(1): 33589, PMID:41023208. https://doi.org/10.1038/s41598-025-18759-4

Prucker

Bressem

Peeken

, et al. A Prospective Controlled Trial of Large Language Model-based Simplification of Oncologic CT Reports for Patients with Cancer. Radiology United States 2025; 317(2): e251844, PMID:41251553. https://doi.org/10.1148/radiol.251844

Akdogan

Uyar

Yesilbas

, et al. Effect of a ChatGPT-based digital counseling intervention on anxiety and depression in patients with cancer: A prospective, randomized trial. Eur J Cancer Oxf Engl 1990 England 2025; 221: 115408, PMID:40215593. https://doi.org/10.1016/j.ejca.2025.115408

Rangarajan

Manivannan

Singh

, et al. Simulation training in mammography with AI-generated images: a multireader study. Eur Radiol Germany 2025; 35(2): 562–571, PMID:39134745. https://doi.org/10.1007/s00330-024-11005-x

Shetty

Kamath

, et al. CancerGPT for few shot drug pair synergy prediction using large pretrained language models. Npj Digit Med 2024; 7(1): 40. https://doi.org/10.1038/s41746-024-01024-9

Yang

Tian

Zhao

, et al. Evolutionary patterns and research frontiers of artificial intelligence in age-related macular degeneration: a bibliometric analysis. Quant Imaging Med Surg AME Publishing Company 2025; 15(1): 813–830. https://doi.org/10.21037/qims-24-1406

10.

Zarin

Tse

Williams

, et al. The ClinicalTrials.gov Results Database — Update and Key Issues. N Engl J Med Massachusetts Medical Society 2011; 364(9): 852–860. https://doi.org/10.1056/NEJMsa1012065

11.

van Eck

Waltman

. Software survey: VOSviewer, a computer program for bibliometric mapping. Scientometrics 2010; 84(2): 523–538, PMID:20585380. https://doi.org/10.1007/s11192-009-0146-3

12.

Chen

. CiteSpace II: Detecting and visualizing emerging trends and transient patterns in scientific literature. J Am Soc Inf Sci Technol 2006; 57(3): 359–377. https://doi.org/10.1002/asi.20317

13.

Tordjman

Liu

Yuce

, et al. Comparative benchmarking of the DeepSeek large language model on medical tasks and clinical reasoning. Nat Med United States 2025; 31: 2550, PMID:40267969.

14.

Koga

Martin

Dickson

. Evaluating the performance of large language models: ChatGPT and Google Bard in generating differential diagnoses in clinicopathological conferences of neurodegenerative disorders. Brain Pathol 2024; 34(3): e13207. https://doi.org/10.1111/bpa.13207

15.

Deng

Wang

Yangzhang

, et al. Evaluation of large language models in breast cancer clinical scenarios: a comparative analysis based on ChatGPT-3.5, ChatGPT-4.0, and Claude2. Int J Surg 2024; 110(4): 1941–1950. https://doi.org/10.1097/JS9.0000000000001066

16.

Wang

Zhuang

, et al. Accuracy of Large Language Models When Answering Clinical Research Questions: Systematic Review and Network Meta-Analysis. J Med Internet Res 2025; 27: e64486. https://doi.org/10.2196/64486

17.

Ishida

Murakami

Yamanoi

, et al. Real-world application of large language models for automated TNM staging using unstructured gynecologic oncology reports. Npj Precis Oncol Nature Publishing Group 2025; 9(1): 366. https://doi.org/10.1038/s41698-025-01157-4

18.

Rajjoub

Arroyave

Zaidat

, et al. ChatGPT and its Role in the Decision-Making for the Diagnosis and Treatment of Lumbar Spinal Stenosis: A Comparative Analysis and Narrative Review. Glob Spine J 2024; 14(3): 998–1017. https://doi.org/10.1177/21925682231195783

19.

Zhu

Lin

Jiang

, et al. Large language model trained on clinical oncology data predicts cancer progression. NPJ Digit Med Berlin: Nature Portfolio 2025; 8(1): 397. https://doi.org/10.1038/s41746-025-01780-2

20.

Lee

Yoon

Kim

, et al. BioBERT: a pre-trained biomedical language representation model for biomedical text mining. Bioinformatics 2020; 36(4): 1234–1240. In: Wren J (ed). https://doi.org/10.1093/bioinformatics/btz682

21.

Harris

. Large Language Models Answer Medical Questions Accurately, but Can’t Match Clinicians’ Knowledge. JAMA 2023; 330(9): 792–794. https://doi.org/10.1001/jama.2023.14311

22.

Tozuka

Johno

Amakawa

, et al. Application of NotebookLM, a large language model with retrieval-augmented generation, for lung cancer staging. Jpn J Radiol 2024; 43: 706–712. https://doi.org/10.1007/s11604-024-01705-1

23.

Park

Byun

, et al. LLM-driven multimodal target volume contouring in radiation oncology. Nat Commun Nature Publishing Group 2024; 15(1): 9186. https://doi.org/10.1038/s41467-024-53387-y

24.

Hung

TKW

Kuperman

Sherman

, et al. Performance of Retrieval-Augmented Large Language Models to Recommend Head and Neck Cancer Clinical Trials. J Med Internet Res JMIR Publications 2024; 26: e60695. https://doi.org/10.2196/60695

25.

Tian

Jiang

Zhang

, et al. The role of large language models in medical image processing: a narrative review. Quant Imaging Med Surg AME Publishing Company 2024; 14(1): 1108–1121. https://doi.org/10.21037/qims-23-892

26.

Liu

Tang

, et al. DrBioRight 2.0: an LLM-powered bioinformatics chatbot for large-scale cancer functional proteomics analysis. Nat Commun Nature Publishing Group 2025; 16(1): 2256. https://doi.org/10.1038/s41467-025-57430-4

27.

Mathes

Ferber

Dreyer

, et al. Collaborative framework on responsible AI in LLM-driven CDSS for precision oncology leveraging real-world patient data. Npj Precis Oncol Nature Publishing Group 2025; 10: 15. https://doi.org/10.1038/s41698-025-01180-5

28.

Liu

Chu

Luo

, et al. Building a unified model for drug synergy analysis powered by large language models. Nat Commun 2025; 16(1): 4537. https://doi.org/10.1038/s41467-025-59822-y

29.

Chung

Fong

Walters

, et al. Large Language Model Capabilities in Perioperative Risk Prediction and Prognostication. JAMA Surg 2024; 159(8): 928–937. https://doi.org/10.1001/jamasurg.2024.1621

30.

Alba

Xue

Abraham

, et al. The foundational capabilities of large language models in predicting postoperative risks using clinical notes. Npj Digit Med 2025; 8(1): 95. https://doi.org/10.1038/s41746-025-01489-2

31.

Jin

Elangovan

, et al. Real-world deployment and evaluation of PEri -operative AI CHatbot (PEACH): a large language model chatbot for peri-operative medicine. Anaesthesia 2025; 81: 16755–16771. https://doi.org/10.1111/anae.16755

32.

Guan

Mok

PK-L

, et al. Development and Validation of a Large Language Model–Powered Chatbot for Neurosurgery: Mixed Methods Study on Enhancing Perioperative Patient Education. J Med Internet Res 2025; 27: e74299. https://doi.org/10.2196/74299

33.

Kolla

Parikh

. Uses and limitations of artificial intelligence for oncology. Cancer Hoboken: Wiley 2024; 130(12): 2101–2107. https://doi.org/10.1002/cncr.35307

34.

Busch

Hoffmann

Rueger

, et al. Current applications and challenges in large language models for patient care: a systematic review. Commun Med Nature Publishing Group 2025; 5(1): 26. https://doi.org/10.1038/s43856-024-00717-2

35.

Hao

Qiu

Holmes

, et al. Large language model integrations in cancer decision-making: a systematic review and meta-analysis. Npj Digit Med Nature Publishing Group 2025; 8(1): 450. https://doi.org/10.1038/s41746-025-01824-7

36.

Shen

Bäck

, et al. Knowledge-Empowered, Collaborative, and Co-Evolving AI Models: The Post-LLM Roadmap. Engineering 2025; 44: 87–100. https://doi.org/10.1016/j.eng.2024.12.008

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.22 MB

0.46 MB

0.00 MB

0.36 MB