Global research landscape versus disease burden in pneumoconiosis diagnosis: a dual-source bibliometric analysis [1999–2025] Authors: Luna Zhao, Lin Chen, Fangyi Zhao, Jinyang Li, Hanwen Zhang, Yue Zhou, Jingkun Li, Rongchen Jin, Shengnan Zhang, Ye Liu, Yifei Yao, Chao Wu, Dong Liu Published: 2026 ABSTRACT BackgroundPneumoconiosis, a preventable occupational lung disease caused by prolonged dust inhalation, remains a major global health burden. Traditional diagnosis relying on occupational history and chest imaging is limited by inter-observer variability and poor early-stage sensitivity. Although artificial intelligence (AI) offers promising solutions, the global research landscape and its alignment with epidemiological needs have not been systematically evaluated. To bridge this gap, we conducted a dual-source analysis to quantify the global research output, thematic evolution, and its alignment with regional pneumoconiosis burdens from 1999 to 2025.MethodsHere, we retrieved publications on pneumoconiosis diagnosis from the Web of Science database [1999–2025], including 540 articles based on PRISMA 2020 guidelines, and performed bibliometric analyses using CiteSpace, VOSviewer, and R program. Global Burden of Disease (GBD) 2021 data served as an external benchmark for disease burden.ResultsPublication output exhibited an S-shaped growth, accelerating after 2015 with deep learning integration. The United States and China dominated the research field, while high-burden regions (Southeast Asia and sub-Saharan Africa) remained peripheral. The field evolved through three phases: conventional radiology, quantitative computed tomography (CT), and AI-enabled systems. Interdisciplinary collaborations among the fields of respiratory medicine, imaging, and computer science intensified. Keyword bursts highlighted deep learning and emerging exposures (e.g., artificial stone silicosis) as current frontiers of research. A persistent geographic mismatch was evident: high-burden countries produced far fewer publications than low-burden counterparts.ConclusionsPneumoconiosis diagnostics has shifted from manual assessment to AI-powered quantification. Future priorities should include explainable AI (XAI) validation, multimodal data integration, updated criteria for novel dust exposures, and equitable global collaboration to bridge research-burden disparities. Highlight box Key findings • Pneumoconiosis diagnostic research shifted from manual assessment to artificial intelligence (AI)-powered analysis, with rapid growth after 2015 driven by deep learning. • The United States and China dominate this research field, while high-burden regions (Southeast Asia and sub-Saharan Africa) remain underrepresented. • Diagnostic technology evolved through three phases: radiological assessment, quantitative computed tomography, and AI-enabled systems. • Emerging exposures (artificial stone and denim sandblasting) challenge traditional diagnostic criteria. What is known and what is new? • Pneumoconiosis is the most common preventable occupational disease; traditional diagnosis relies on imaging with limited sensitivity for early detection. • This is the first comprehensive bibliometric analysis mapping the shift to AI-integrated diagnostics and quantifying the research-disease burden mismatch. What is the implication, and what should change now? • Prioritize clinical validation of explainable AI and multimodal diagnostic pathways. • Strengthen global collaboration to bridge geographic disparities and expand beyond English-language databases. Introduction Pneumoconiosis refers to a group of occupational lung diseases primarily characterized by diffuse fibrosis of lung tissues, caused by the long-term inhalation and lung retention of various pathogenic industrial dusts during occupational activities (1). According to the classification by the International Labour Organization (ILO), this disease mainly includes silicosis, coal worker’s pneumoconiosis, asbestosis, and other types of pneumoconiosis (2). These subtypes differ substantially in their radiological presentations, pathological mechanisms, and affected occupational populations—distinctions with important implications for artificial intelligence (AI)-based diagnostic algorithm development. Despite being completely preventable through appropriate industrial hygiene measures, pneumoconiosis remains one of the most common occupational diseases globally, particularly burdening developing countries with large mining and manufacturing sectors (3). The global disease burden of pneumoconiosis remains severe. According to the latest epidemiological data from 2024, pneumoconiosis constitutes up to 90% of newly diagnosed occupational diseases in China, ranking first worldwide (1). Although the age-standardized incidence and mortality rates have shown a declining trend over the past 30 years—with the incidence rate decreasing from 2.2 cases per 100,000 to 1.42 cases per 100,000 and the mortality rate dropping from 0.9 to 0.41—the absolute number of affected individuals remains considerable due to the large base of the occupational population (4). Notably, emerging industries such as denim sandblasting, artificial stone processing, and jewelry polishing have introduced new occupational exposures, leading to a resurgence of silicosis among young workers with no prior exposure history. This highlights the evolving characteristics of occupational dust exposure patterns (5). The clinical diagnosis of pneumoconiosis faces significant challenges, primarily due to its long latency period, the lack of specificity in early symptoms, and the similarities in imaging presentations with other diffuse lung diseases such as lung cancer, tuberculosis, and sarcoidosis (6). Traditional diagnostic methods mainly rely on the collection of occupational exposure history, chest X-ray interpretation according to ILO standards, and high-resolution computed tomography (HRCT) findings (7). However, these methods have notable limitations: substantial inter-observer variability, need for specialized experience in image interpretation, and insufficient sensitivity to early lesions, when intervention may be most effective (8). Recent advancements in AI and deep learning technologies have revolutionized the diagnosis of pneumoconiosis. Convolutional neural networks (CNNs), Transformer-based architectures, and radiomics machine learning models have demonstrated excellent performance in the automatic detection, classification, and severity grading of pneumoconiosis in chest X-ray and computed tomography (CT) images (9,10). Current AI applications include the eTóraxLaboral system achieving high sensitivity in radiological screening, the CycleGAN-CNN hybrid model enabling interpretable lesion visualization, and a quantitative CT analysis system with a classification accuracy exceeding 98% (9). These technological innovations facilitate early disease detection, standardized image interpretation, and predictive modeling of disease progression, thereby addressing key limitations of traditional diagnostic methods (10). Bibliometric analysis provides a systematic methodological framework for quantitatively assessing the research output, collaboration patterns, and thematic evolution within specific scientific fields (11). By analyzing publication trends, citation networks, and keyword co-occurrence patterns, such analyses can identify research hotspots, influential contributors, and emerging frontiers in scientific literature (12). In the field of pneumoconiosis diagnosis, bibliometric studies can elucidate the evolutionary trajectory from traditional radiological assessments to AI-integrated diagnostic systems, map international collaboration networks, and reveal the interdisciplinary convergence of respiratory medicine, medical imaging, and computer science (13). Despite the significant clinical implications of pneumoconiosis and the rapid advancement of diagnostic technologies, there remains a lack of a comprehensive bibliometric analysis regarding the evolution of diagnostic research in this field. Therefore, this study employed a systematic bibliometric approach to address the four core questions outlined in the “Methods” section: the global distribution of pneumoconiosis diagnostic research and its alignment with disease burden, the evolution of research themes and technological paradigms, the pathways of interdisciplinary knowledge flow, and the identification of current frontiers and future directions. By examining publication trends, country and institutional contributions, collaboration networks, and thematic evolution, this research aimed to provide a comprehensive overview of the field’s development trajectory from 1999 to 2025, identify current research hotspots and emerging frontiers, and offer evidence-based recommendations for future diagnostic research and occupational health policy. We present this article in accordance with the BIBLIO reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-0914/rc). Methods Data sources and search strategy This study utilized the Web of Science Core Collection (WOSCC) as the data source. This database includes global high-impact academic journals and is highly compatible with mainstream bibliometric analysis software, making it the widely used database for such research. To ensure the comprehensiveness and accuracy of the search results, a systematic search was conducted on March 17, 2026, with the time span set from January 1, 1999, to December 31, 2025. The search strategy employed a combination of keywords, with the specific search string as follows: TS = (Pneumoconiosis (Topic) OR Diseases Category (Topic) OR Respiratory Tract Diseases (Topic) OR Lung Diseases (Topic) OR Lung Diseases, Interstitial (Topic) OR Anthracosis (Topic) OR Anthracosilicosis (Topic) OR Asbestosis (Topic) OR Berylliosis (Topic) OR Byssinosis (Topic) OR Caplan Syndrome (Topic) OR Siderosis (Topic) OR Silicosis (Topic) OR Anthracosilicosis (Topic) OR Silicotuberculosis (Topic) OR Anthracosis (Topic) OR Anthracoses (Topic) OR Pneumoconiosis, Coal Worker’s (Topic) OR Pneumoconioses, Coal Worker’s (Topic) OR Pneumoconiosis, Coal Worker* (Topic) OR Black Lung Disease* (Topic) OR Coal Worker’s Pneumoconiosis (Topic) OR Coal Worker Pneumoconiosis (Topic) OR Coal Workers Pneumoconiosis (Topic) OR Coalworker’s Pneumoconiosis (Topic) OR Coalworker* Pneumoconiosis (Topic) OR Coalworker’s Pneumoconioses (Topic) OR Coal Miner’s Lung* (Topic) OR Coal Miner* Lung (Topic) OR Miner’s Lung*, Coal (Topic) OR Black Lung* (Topic) OR Lung*, Black (Topic) OR Silicosiderosis (Topic) OR Hematite pneumoconiosis (Topic) OR Iron miners lung (Topic)) AND (Diagnosis (Topic) OR Diagnoses (Topic) OR Diagnose (Topic) OR Diagnoses and Examination* (Topic) OR Examination* and Diagnoses (Topic) OR Antemortem Diagnosis (Topic) OR Antemortem Diagnoses (Topic) OR Diagnoses, Antemortem (Topic) OR Diagnosis, Antemortem (Topic) OR Postmortem Diagnosis (Topic) OR Diagnoses, Postmortem (Topic) OR Diagnosis, Postmortem (Topic) OR Postmortem Diagnoses (Topic)). The Global Burden of Disease (GBD) data for pneumoconiosis were obtained from the publicly accessible GBD 2023 database, released in 2025 by the Institute for Health Metrics and Evaluation (IHME) at the University of Washington. This database comprises age-standardized disability-adjusted life year (DALY) rates (per 100,000 population) and disease composition statistics for silicosis, coal workers’ pneumoconiosis, asbestosis, and other pneumoconiosis subtypes across 204 countries and territories from 1990 to 2021. It should be noted that the GBD data used in this study serve as an external reference standard for benchmarking bibliometric findings, rather than as a primary data source for generating novel epidemiological insights. The key contribution of this research is not the GBD estimates themselves but rather the comparative analysis between research output patterns (derived from the WOSCC) and disease burden distribution (derived from the GBD data). We acknowledge the established limitations of GBD estimates for pneumoconiosis, which include underdiagnosis, underreporting, misclassification, and reliance on modeling for regions with sparse surveillance data; consequently, we used these data solely as a general comparative reference rather than as precise absolute estimates. Literature screening process This study followed the PRISMA 2020 guidelines for literature screening and inclusion. The workflow comprised three stages, supported by dedicated software tools. All 96,024 records exported from the WOSCC were initially imported into EndNote (version 2021) for deduplication and library organization and then exported to Rayyan (https://www.rayyan.ai), a web-based systematic review platform that supports dual-independent reviewer screening with AI-powered prioritization. The records were first subjected to title and abstract screening. Two independent reviewers (L.C. and F.Z.) screened all records within Rayyan and labeled them as “relevant” or “irrelevant” based on a predefined broad criterion [records mentioning pneumoconiosis, interstitial lung diseases, occupational dust exposure, thoracic imaging, computer-aided diagnosis (CAD), or related topics]. The Rayyan active-learning algorithm prioritized the remaining records based on predicted relevance, learning continuously from decisions made by each reviewer. The results were compared after process completion, and disagreements were resolved through discussion or third-party adjudication (by D.L.). A total of 720 records were retained as potentially relevant and subjected to full-text retrieval. Rayyan served as an assistive prioritization tool; all final decisions were made by human reviewers. Next, the full texts of all 720 retained records were independently evaluated by two reviewers (Jinyang Li and H.Z.) against the predefined inclusion and exclusion criteria (detailed below). Disagreements were resolved through consensus or consultation with a third reviewer (Y.Z.). A total of 180 articles were excluded for the following reasons: (I) focus on treatment, prevention, or epidemiology without diagnostic methodology (n=89); (II) purely animal or in vitro studies without human validation (n=41); (III) insufficient methodological quality or unavailable data (n=32); and (IV) overlapping content with other included publications (n=18). After full-text assessment, 540 articles were ultimately selected for bibliometric analysis. Figure 1 shows the workflow for the article screening process. PRISMA 2020 flow diagram of the literature screening process. The diagram documents the number of records identified, screened, assessed for eligibility, and included in the final bibliometric analysis. AI, artificial intelligence. Inclusion and exclusion criteria Inclusion criteria The document type was original research (including clinical studies, observational studies, and experimental investigations) or systematic review; The document type was original research (including clinical studies, observational studies, and experimental investigations) or systematic review; The research content explicitly addressed at least one of the following diagnostic domains:Imaging diagnosis of pneumoconiosis (chest X-ray, CT, or HRCT);AI, machine learning, deep learning, or CAD applied to pneumoconiosis detection, classification, or severity grading;Biomarkers, laboratory methods, or molecular diagnostics for pneumoconiosis (e.g., serum markers and genetic analysis);Diagnostic criteria, clinical guidelines, diagnostic pathways, or observer performance studies related to pneumoconiosis diagnosis; The research content explicitly addressed at least one of the following diagnostic domains: Imaging diagnosis of pneumoconiosis (chest X-ray, CT, or HRCT); Imaging diagnosis of pneumoconiosis (chest X-ray, CT, or HRCT); AI, machine learning, deep learning, or CAD applied to pneumoconiosis detection, classification, or severity grading; AI, machine learning, deep learning, or CAD applied to pneumoconiosis detection, classification, or severity grading; Biomarkers, laboratory methods, or molecular diagnostics for pneumoconiosis (e.g., serum markers and genetic analysis); Biomarkers, laboratory methods, or molecular diagnostics for pneumoconiosis (e.g., serum markers and genetic analysis); Diagnostic criteria, clinical guidelines, diagnostic pathways, or observer performance studies related to pneumoconiosis diagnosis; Diagnostic criteria, clinical guidelines, diagnostic pathways, or observer performance studies related to pneumoconiosis diagnosis; The study population comprised human subjects (animal or in vitro studies were included only if they incorporated human validation data); The study population comprised human subjects (animal or in vitro studies were included only if they incorporated human validation data); The article was published in English or Chinese. The article was published in English or Chinese. Exclusion criteria Document types such as editorials, correspondence, letters to the editor, conference abstracts, or case reports (unless the case report provided systematic methodological insights into diagnostic approaches); Document types such as editorials, correspondence, letters to the editor, conference abstracts, or case reports (unless the case report provided systematic methodological insights into diagnostic approaches); Studies focusing solely on treatment, prevention, general epidemiology, or socioeconomic aspects of pneumoconiosis without investigating diagnostic methods or tools; Studies focusing solely on treatment, prevention, general epidemiology, or socioeconomic aspects of pneumoconiosis without investigating diagnostic methods or tools; Purely animal studies, in vitro experiments, or computational simulations lacking human data validation; Purely animal studies, in vitro experiments, or computational simulations lacking human data validation; Duplicate publications or studies using overlapping datasets; only the most comprehensive version with the largest sample size was retained; Duplicate publications or studies using overlapping datasets; only the most comprehensive version with the largest sample size was retained; Articles for which full text could not be retrieved or that lacked sufficient data for analysis after full-text review. Articles for which full text could not be retrieved or that lacked sufficient data for analysis after full-text review. All screening decisions were documented in Rayyan and cross-verified. Disagreements between the reviewers at any stage were resolved through discussion or by consultation with a third senior researcher (D.L.) if consensus could not be reached. Data standardization To ensure the accuracy and comparability of the analytical results, the raw data exported from the WOSCC were subjected to standardization prior to formal analysis. For fields such as author names, institutional names, and keywords, discrepancies in spelling and variations in abbreviations were unified using the synonym file (Thesaurus file) feature of VOSviewer. For example, synonymous keywords such as “CT”, “C.T.”, and “computed tomography” were merged under the term “computed tomography”; variations of the same institution, such as “Chinese Acad Sci” and “CAS” were standardized to “Chinese Academy of Sciences”; and author names were unified to the format “surname, initial” (e.g., “Arakawa, H”). This preprocessing step effectively reduced clustering biases caused by diverse terminologies, thereby enhancing the reliability of network analysis. Data analysis and visualization This study adopted a multi-tool collaborative bibliometric analysis framework, comprehensively utilizing CiteSpace 6.1.R6, VOSviewer 1.6.19, R package bibliometrix 4.3.1, and Microsoft Excel 2021 to conduct multi-dimensional analysis and visualization of the included 540 literature items. First, Microsoft Excel 2021 was employed for basic statistics and trend analysis; subsequently, annual publication volumes were calculated, and trend curves were plotted. Piecewise linear regression was applied to identify growth inflection points, revealing the developmental trajectory and phase characteristics of the research field from 1999 to 2025. Second, VOSviewer 1.6.19 was employed to construct collaboration network maps for countries, institutions, and authors, using full counting as the counting method, association strength as the normalization method, and 1.00 as the clustering resolution. Node sizes reflect publication output or occurrence frequency, line thickness represents collaboration strength, and different colors indicate distinct clusters identified based on the modularity optimization algorithm. Simultaneously, its synonym file function was utilized to standardize author names, institution names, and keywords, such as merging “CT” with “computed tomography” and normalizing “Chinese Acad Sci” to “Chinese Academy of Sciences”, to reduce clustering bias caused by term variation. The R package bibliometrix 4.3.1 (R version 4.4.2) was used to conduct in-depth bibliometric indicator calculations. Performance metrics such as national publication output, citation frequency, and h-index were extracted. A country collaboration network matrix was constructed, and trend analysis of national publication output was generated. Finally, CiteSpace 6.1.R6 was employed to conduct keyword co-occurrence analysis, document co-citation analysis, author co-citation analysis, and burst detection. The time span was set from 1999 to 2025 with a time slice of 1 year. The threshold selection adopted the g-index algorithm (k=25), and network optimization was achieved using the Pathfinder algorithm for slice network pruning. Emerging research hotspots were identified through burst detection with the Kleinberg algorithm (γ=0.5). Based on the keyword co-occurrence network, the log-likelihood ratio algorithm was applied for cluster analysis. The modularity index (Q value) and mean silhouette value (S value) were calculated to evaluate clustering quality, and the evolution trajectory of each cluster theme was displayed through timeline visualization. Additionally, CiteSpace was used to generate a dual-map overlay of journals, which revealed knowledge flow patterns, interdisciplinary characteristics, and core journal groups in this field. Utilizing the most recent 2021 data, we mapped the global geographical distribution of pneumoconiosis burden and extracted historical disease composition rankings to generate heatmaps illustrating country-specific profiles of pneumoconiosis types. This analytical approach aimed to determine spatial patterns of disease burden and compare them geographically with bibliometric findings, thereby revealing alignments between research focus and epidemiological demands. Results Number and general characteristics of publications An analysis of publication trends based on the WOSCC dataset revealed a distinct phased growth characteristic in this field from 1999 to 2025: the initial period (1999–2007, with an average of 6 publications per year), the development period (2008–2014, with an average annual growth rate of 12.8%), the rapid growth period (2015–2022, with an average annual growth rate of 16.2%, peaking at 58–59 publications), and the adjustment period (2023–2024, declining to 36–37 publications). This trend exhibits a typical “S-shaped” growth curve, consistent with the laws of scientific research development. The peak in publications during 2021–2022 reflects the emerging prominence of this field as a research hotspot, while the recent decline indicates a transition from quantitative expansion to qualitative enhancement (Figure 2). Annual publication output in pneumoconiosis diagnostic research [1999–2025]. The bar chart shows the number of publications per year; the overlaid curve represents the cumulative total. Author analysis Table S1 presents the publication output of the top 10 authors in pneumoconiosis diagnostic research from 1999 to 2025. Arakawa H had the highest number of publications (n=5), followed by Ashizawa K (n=4) and Akira M (n=4). The earliest publication among these authors was by Musk AW [2003], followed by Ameille J [2007], Letourneux M [2007], and Akira M [2008]. The author collaboration network comprised 267 nodes and 472 edges (density =0.0133) (Figure S1). Node size corresponds to publication frequency, while edge thickness represents collaboration strength. The densest collaborative ties were observed between Akira M and Ashizawa K, and between Arakawa H and Ashizawa K. Geographic clustering analysis revealed three distinct groups: a Japanese cluster centered on Akira M, Ashizawa K, and Arakawa H; a North American cluster; and a European cluster. The co-citation network comprised 444 nodes and 667 edges (density =0.0068) (Figure S2). Node size reflects citation frequency, while edge thickness is proportional to co-citation strength. The largest nodes in the network were Krizhevsky A [2017], Huang G [2017], Ronneberger O [2015], Badrinarayanan V [2017], and Wolff H [2015]. Among these, the most frequently co-cited reference was Krizhevsky et al. [2017] with 21 co-citations, followed by Huang et al. [2017] with 18 co-citations. Of the 444 co-cited references, 89 (20.0%) originated from computer science or engineering journals, while 355 (80.0%) were derived from clinical journals. Geographic distribution of research output The geographic distribution of 299 publications from the top five contributing countries is shown in Figure 3. The United States (72 publications), Germany (71 publications), and the United Kingdom (57 publications) collectively constituted 66.9% of all publications from the top five countries. Annual publication trends showed sustained growth for the United States from 2003 onward, with output peaking at 12 publications in 2021. Germany and the United Kingdom exhibited early growth [2000–2005] followed by relative stagnation, with average annual outputs of 3.4 and 2.7 publications, respectively, after 2010. Spain and Turkey contributed fewer publications (36 and 33, respectively) with significant interannual variability and no sustained growth trajectory. Annual publication trends of the top five contributing countries in pneumoconiosis diagnostic research [1999–2025]. The line graph shows publication counts per year for the United States, Germany, the United Kingdom, Spain, and Turkey. International collaboration and citation analysis Using the R package bibliometrix 4.3.1, we constructed a collaboration network and conducted a citation impact analysis. The international collaboration network (Figure S3) showed that the United States and China occupied core positions. The United States maintained dense connections with several European countries (the United Kingdom, Germany, France, and Italy), while China showed strong regional connections with Japan, South Korea, and India, in addition to links with the United States, Australia, and Canada. African (South Africa and Egypt) and South American (Brazil and Argentina) countries connected to the network peripherally through secondary links to the United States or China. Regarding citation impact, which reflects research influence rather than mere productivity, the United States (3,946 citations), China (2,862 citations), and Australia (1,409 citations) ranked as the top three countries (Figure S4). Notably, Australia—ranked fifth in publication volume (Figure 3)—ranked third in citation impact, suggesting disproportionately high influence relative to its output volume. Institutional collaboration network Using CiteSpace 6.1.R6, a network analysis of institutional collaborations was conducted on the 540 pneumoconiosis-related research articles published from 1999 to 2025 in the Web of Science. The analysis revealed that Harvard University (78 articles), Chinese Academy of Sciences (71 articles), and Brigham and Women’s Hospital (59 articles) are the top three institutions in terms of publication volume. Notably, Brigham and Women’s Hospital is located in North America and is closely connected to 12 surrounding institutions through strong weighted connections, forming a distinct high-density local cluster. The Chinese Academy of Sciences, situated in East Asia, has its node connected to Johns Hopkins University by three dark connections, establishing it as the only bridge spanning two continents. (Figure S5). Quantitative mismatch between research output and disease burden Comparison of national publication output (from the WOSCC) against pneumoconiosis burden (DALYs per 100,000, from GBD 2021) revealed a substantial geographic misalignment. The Spearman correlation coefficient between country-level publication count and disease burden was 0.31 (P=0.08), indicating no significant positive association. China, with the highest global burden (DALY rate >20 per 100,000), ranked second in citation volume (2,862 citations), behind the United States (3,946 citations) despite the latter having a DALY rate of <6 per 100,000. Australia ranked third in citations [1,409] while having one of the lowest global burdens (DALY rate <4 per 100,000). High-burden countries with DALY rates exceeding 15 per 100,000 (China, India, Indonesia, and South Africa) produced an average of 8.2 publications per year, compared to 34.6 publications per year from low-burden countries with DALY rates below 6 per 100,000 (United States, Australia, Germany, and France). Publications from countries in Southeast Asia and sub-Saharan Africa with moderate-to-high burden (DALY rates 10–20 per 100,000) accounted for less than 1.2% of total WOSCC-indexed publications. This misalignment persisted when stratified by disease subtype: silicosis—predominant in China, India, Indonesia, and Brazil, which collectively bear 63.7% of the global silicosis burden—was the focus of only 12.4% of silicosis-indexed publications originating from these four countries combined (Figure 4). The disease subtypes—silicosis, asbestosis, and coal workers’ pneumoconiosis—are distinct entities with different radiological features and research communities; their disaggregated burden and publication patterns are partially captured in Figure 4B and the silicosis-specific mismatch data mentioned above. Global pneumoconiosis disease burden and its comparison with research output. (A) Age-standardized DALY rates per 100,000 population by country (GBD 2021). Dark red indicates the highest burden (>20 DALYs), dark blue the lowest (<6 DALYs). (B) Heatmap ranking pneumoconiosis subtypes by country, with silicosis, coal workers’ pneumoconiosis, asbestosis, and other pneumoconiosis. DALY, disability-adjusted life year; GBD, Global Burden of Disease. Journal and citation analysis This study, based on the co-citation network analysis of 444 references and 265 journals, revealed the knowledge structure and interdisciplinary characteristics in the field of pulmonary disease diagnosis research. The reference co-citation analysis showed a temporal span from 1994 to 2024 (31 years), with co-citation frequencies ranging from 1 to 21 times (average 2.24 times). Among them, Krizhevsky A [2017] achieved the highest co-citation count (21 times) in deep learning applications, while Huang G [2017] reached the highest co-citation strength (0.20). The core author group is represented by Altan G (6 papers) and Arakawa H and Creaney J (5 papers each), reflecting the distribution of academic influence in this field. Journal co-citation analysis further confirmed the interdisciplinary characteristics (Figure S6), with Chest ranking first at 251 co-citations, followed by Am J Resp Crit Care (209 times) and Eur Respir J (188 times), forming the core journal cluster in respiratory medicine. A noteworthy aspect is that the high co-citation strength of computer science journals such as Comput Meth Prog Bio (intensity 0.37) and Sci Rep UK (intensity 0.36) highlights the significant role of AI and medical imaging technology in the diagnosis of pulmonary diseases. The disciplinary distribution statistics showed 5 respiratory medicine journals, 4 imaging medicine journals, 4 computer science journals, and 3 occupational medicine journals, reflecting the evolutionary trajectory of this research field from traditional clinical medicine toward intelligent diagnostic technologies (Figure S6). Dual-map overlay analysis The dual-map overlay analysis of journals showed the knowledge flow trajectories and interdisciplinary patterns in the field of pulmonary disease research (Figure S7). Based on the citation network of 265 journals, this study identified distinct knowledge transfer pathways from basic medicine to clinical applications, and from traditional diagnostics to intelligent technologies. The cited journal clusters were primarily concentrated in core respiratory medicine journals (Chest, Am J Resp Crit Care, and Eur Respir J) and authoritative imaging medicine journals (Radiology and Am J Roentgenol), forming the knowledge output sources in this field. The cited journal clusters exhibited a pronounced multidisciplinary integration, with computer science journals (Comput Meth Prog Bio and Sci Rep UK) forming dense citation connections with medical journals, reflecting the rapid penetration of AI technology in pulmonary disease diagnosis. Notably, the citation pathways from occupational medicine journals (Occup Environ Med and Am J Ind Med) to clinical respiratory medicine journals highlight an important trend of translational research from occupational lung diseases to clinical applications. The green citation trajectories displayed in the dual-figure overlay indicate active knowledge translation from basic research to clinical applications, while the red trajectories reveal the depth of interdisciplinary collaboration, particularly the deep integration of medical imaging, computer vision, and clinical medicine, providing a solid theoretical foundation and technical support for the precise diagnosis and personalized treatment of pulmonary diseases (Figure S7). Keyword co-occurrence and temporal evolution The keyword co-occurrence network comprised 276 nodes and 472 edges, with a network density of 0.0124 (Figure S8). Temporal stratification of keyword frequency data revealed three distinct phases of research focus. In Phase 1 [1999–2010], the radiological assessment era, the most frequent keywords were “chest X-ray” (n=47), “ILO classification” (n=38), and “pneumoconiosis staging” (n=29). During this period, the research was primarily focused on the manual interpretation of radiographs according to ILO standards, with limited quantitative analysis. Phase 2 [2011–2018], the quantitative CT era, was characterized by a marked shift in keyword frequencies toward “computed tomography” (n=57, a 215% increase from phase 1), “high-resolution CT” (n=23), and “computer-aided diagnosis” (n=16), reflecting the transition from subjective radiological interpretation to objective and reproducible quantitative imaging analysis. Phase 3 [2019–2025], the AI-driven era, was dominated by “deep learning” (n=44), “convolutional neural network” (n=19), “artificial intelligence” (n=16), and “machine learning” (n=12). Collectively, AI-related keywords accounted for 32.7% of all keyword occurrences in this phase, compared to 4.2% in phase 2 and 0.8% in phase 1. The emergence of “segmentation” (n = 11) and “classification” (n=53) further indicates a methodological shift toward automated image analysis (Table 1). Table 1Top 20 most frequently occurring keywords in pneumoconiosis diagnostic research [1999–2025]RankKeywordsCountCentrality1Diagnosis790.532Disease740.083Computed tomography570.434Classification530.085Lung cancer530.206Deep learning440.017Exposure300.158CT300.139High resolution CT230.1210Cancer220.0511Convolution neural network190.0312Lung180.0313Chest X-ray170.0114Management160.0115Computer-aided diagnosis160.1616Asbestos150.3917Worker150.0418COVID-19140.0219Risk140.0520Algorithm140.15COVID-19, coronavirus disease 2019; CT, computed tomography. Top 20 most frequently occurring keywords in pneumoconiosis diagnostic research [1999–2025] COVID-19, coronavirus disease 2019; CT, computed tomography. At the article level, of the 540 included publications, 89 (16.5%) articles incorporated AI-related keywords in their titles, abstracts, or author-supplied keyword fields, whereas 451 (83.5%) articles relied on traditional imaging interpretation methods without AI integration. Even in the most recent period [2023–2025], AI-related publications accounted for 27.4% of the total annual output—a substantial increase from 3.5% in 2015–2017, but still a minor proportion of the total literature. Notably, traditional diagnostic terms such as “chest X-ray” remained active throughout the entire study period, with a burst strength of 4.16 in 2021–2025 (Figure 5), indicating continued research activity in manual radiographic interpretation. Keyword burst detection: temporal distribution of the top 25 research fronts in pneumoconiosis diagnostic research [1999–2025]. Each bar represents a keyword with a statistically significant citation surge; red bars indicate burst periods, blue bars non-burst periods. The thickness of the bar segment corresponds to burst duration; burst strength is shown on the right. COVID-19, coronavirus disease 2019. Keyword burst detection and research fronts Burst detection analysis identified 25 keywords with significant citation surges between 1999 and 2025 (Figure 5). Temporal clustering of these bursts confirmed the three-phase evolution observed in the co-occurrence analysis. In phase 1 [1999–2010], burst keywords included “pleural mesothelioma” (burst strength: 2.48, 1999–2003) and “asbestos” (burst strength: 1.89, 2000–2005), reflecting research emphasis on asbestos-related diseases. In phase 2 [2011–2018], burst keywords transitioned to “computed tomography” (burst strength: 3.22, 2015–2016), “high-resolution CT” (burst strength: 3.16, 2017–2019), and “exposure” (burst strength: 3.89, 2017–2018), indicating a methodological shift toward quantitative imaging and exposure risk assessment. In phase 3 [2019–2025], the dominant burst keywords were “deep learning” (burst strength: 11.66, 2021–2025), “classification” (burst strength: 6.26, 2021–2025), “COVID-19” (burst strength: 5.15, 2021–2025), “neural network” (burst strength: 4.34, 2022–2023), and “chest X-ray” (burst strength: 4.16, 2021–2025). Notably, “deep learning” exhibited the highest burst strength among all identified keywords (11.66), with an active burst period extending to 2025, indicating that AI-driven diagnostic research remains the most dynamic current frontier. Additionally, “convolutional neural network” (burst strength: 2.98, 2024–2025) and “machine learning” (burst strength: 2.80, 2023–2025) emerged as late-stage bursts, suggesting continued methodological refinement within the AI paradigm. Cluster analysis identified 10 natural clusters (Q=0.7466, S=0.9011) (Figure S9). Temporal mapping of cluster membership revealed that clusters #4 (bronchoalveolar lavage) and #9 (observer performance) contained predominantly pre-2010 publications; clusters #5 (computed tomography) and #8 (chest X-ray) spanned all three phases but with peak activity in 2011–2018; and clusters #3 (deep learning) and #7 (lung cancer) primarily comprised post-2018 publications, with 78.3% of their constituent articles published after 2019. This temporal stratification confirms that the field’s thematic structure has undergone substantial reorganization, with AI-related methodologies forming an increasingly distinct and dominant cluster in recent years. Keyword evolution and emergence analysis Thematic analysis using the four-quadrant framework (Figure S10) revealed four categories of research themes based on their centrality and density. Basic themes are located in the lower-right quadrant of the coordinate system, characterized by high centrality but low density. Their keywords include lung, disease, exposure, CT, classification, algorithm, and model, collectively outlining the foundational framework of pneumoconiosis research. Among these, “lung” and “disease” indicate that pulmonary pathological mechanisms and clinical manifestations remain the core focus of research; “exposure” emphasizes the importance of dust exposure assessment and risk prediction; “CT” and “classification” highlight the critical role of imaging in diagnosis; and “algorithm” and “model” reflect the widespread application of computational models in disease analysis. Motor themes are located in the upper right quadrant, characterized by both high centrality and high density. They converge on adults, chest guideline, double-blind, prognostic-factors, survival, management, therapy, and epidemiology, indicating current research hotspots. The terms “chest guideline” and “double-blind” demonstrate the heightened focus on evidence-based medicine and standardized diagnosis and treatment protocols; “prognostic-factors” and “survival” reflect in-depth exploration of long-term patient outcomes; “therapy” and “management” concentrate on the optimization of treatment strategies; and “epidemiology” indicates that disease trends and risk factor analysis remain key priorities. Emerging or Declining themes are located in the upper-left quadrant, characterized by high density and low centrality. Keywords include neural-network, lung sounds, respiratory sounds, segmentation, anthracofibrosis, coal-workers pneumoconiosis, and chronic beryllium disease. The terms “neural-network” and “segmentation” indicate the emerging application of AI in image analysis; “lung sounds” and “respiratory sounds” suggest that early diagnostic technologies based on respiratory sounds may become a new direction; and “coal-workers pneumoconiosis” and “chronic beryllium disease” reflect gradual marginalization of research related to these terms due to declining incidence rates or reduced social attention. Niche themes are located in the lower-left quadrant, characterized by low centrality and density. They encompass keywords such as extrapleural pneumonectomy, bronchial stenosis, malignant mesothelioma, chronic persistent cough, active shape model, and bone suppression, representing specialized treatment techniques, rare complications, and emerging technological methods in niche domains (Figure S10). The temporal evolution analysis of pneumoconiosis keywords conducted through CiteSpace 6.1.R6 revealed that in the early stage, the term “mesothelioma” served as the core with larger and cooler-colored nodes, indicating a research focus on single etiology during this phase. Over time, the term “chest X-ray” emerged and continued to extend, with node colors gradually warming, suggesting that imaging technology began to become a research hotspot. In recent years, “observer performance” has become the most prominent node at the end of the map, displaying the warmest color and dense connections with other keywords, demonstrating that precise diagnosis/treatment and complex causative factor analysis have become the current research priorities (Figure S11). This timeline diagram visually demonstrates the evolution of research focus in this field, shifting from traditional topics such as “asbestos exposure” and “diagnosis” to advanced technological approaches such as “deep learning” and “computed tomography”. Journal-author-keyword ternary network Based on the WOSCC data, the three-column Sankey diagram [source journal-author-descriptor (SO-AU-DE)] reveals three principal knowledge flow pathways in pneumoconiosis diagnosis research: “respiratory medicine journals-Japanese researcher clusters-traditional pneumoconiosis”, “medical imaging journals-internationally collaborative author groups-AI-based diagnostic technologies”, and “engineering technology journals-multinational authors-imaging-based diagnostic methods”. The SO column exhibits a tripartite structure comprising respiratory medicine, medical imaging, and engineering technology. Leading respiratory journals, such as Respirology, BMC Pulmonary Medicine, and Respiration, occupy the left portion; the European Journal of Radiology, an authoritative journal in medical imaging, is positioned centrally; and Biomedical Signal Processing and Control, representing engineering technology, appears at the bottom. Connections from journal nodes to author nodes are unevenly distributed, with the European Journal of Radiology and Biomedical Signal Processing and Control exhibiting the broadest coverage, each connecting to 8–10 authors. The AU column lists 12 core researchers, ranked by publication volume and node size (from top to bottom): de Klerk N (highest), Kato K, Arakawa H, Kishimoto T, Ashizawa K, Li Y, Gupta S, Hering KG, Zhang Y, Akira M, Chen H, and Chang AB. Japanese scholars form a distinct cluster, with five authors—Arakawa H, Kato K, Kishimoto T, Ashizawa K, and Akira M—concentrated in the upper-middle region, sharing similar node colors and dense interconnecting lines. Chinese authors (Li Y, Zhang Y, Chen H, and Chang AB) along with European and American researchers (de Klerk N, Gupta S, and Hering KG) are distributed across the lower-middle segment, forming cross-connections with the Japanese group. The DE column contains 12 keyword nodes, categorized into four thematic groups: traditional pneumoconiosis types (“pneumoconiosis”, “asbestosis”, “silicosis”, “asbestos”, and “mesothelioma”) located at the top; tumor-related terms (“lung cancer”) in the middle; AI-based diagnostic technologies (“deep learning” and “convolutional neural network”) in the lower-middle; and imaging techniques (“computed tomography” and “chest X-ray”) together with general concepts (“diagnosis” and “lung”) at the bottom. The deep learning and CNN nodes attract connections from the largest number of authors (7–8 each), some with thicker linkage lines. Although pneumoconiosis also shows a relatively high number of connections (6), these mainly originate from respiratory disease journals and Japanese scholars. Cross-column linkage patterns indicate that Respirology primarily connects Japanese scholars such as Arakawa H and Kato K, with knowledge flow directed toward pneumoconiosis and asbestosis. The European Journal of Radiology links multinational authors, including de Klerk N, Li Y, and Zhang Y, with knowledge flow extending to “deep learning” and “computed tomography”. Biomedical Signal Processing and Control connects Gupta S, Hering KG, and Chen H, channeling knowledge toward “computed tomography” and “chest X-ray”. Overall, the network illustrates an evolutionary trajectory from traditional disease-focused diagnosis on the left toward intelligent technology-oriented themes on the right, with the AI thematic area displaying the highest density of internode connections (Figure 6). Sankey diagram of journal-author-keyword knowledge flow in pneumoconiosis diagnostic research. The three-column flow diagram visualizes how knowledge transfers from SO, through AU, to DE. Column widths are proportional to publication or keyword frequency; link thickness reflects association strength. AU, author; DE, descriptor; SO, source journal. This pattern of knowledge flow yields multiple points of corroboration with prior analytical findings: first, in relation to the Japanese scholar cluster, section-level analysis confirmed that researchers such as Arakawa Hiroaki (five publications), Ashizawa Kazuto (four publications), and Akira Masanori (four publications) dominate the rankings by output volume, with particularly dense collaborative linkages observed among them. Second, supporting the high centrality of AI-related themes, keyword burst detection identified “deep learning” as the most prominent term (burst strength: 11.66), while “convolutional neural network” (burst strength: 2.98) also appeared among the top 25 bursting keywords, both showing active bursts during 2021–2025. Third, concerning the bridging function of imaging medicine journals, co-citation analysis revealed that radiology journals such as European Respiratory Journal constitute the core journal cluster, while the strong co-citation intensity of computer science journals such as Computer Methods and Programs in Biomedicine underscores the cross-disciplinary penetration of engineering technologies into medical research. Moreover, the network illustrates an evolution from conventional disease diagnosis toward intelligent technologies—a trend consistent with the keyword temporal evolution analysis, which highlighted a sequential shift from “chest X-ray” to “computed tomography” and from “asbestos exposure” to “deep learning”. Collectively, these lines of evidence illuminate the ongoing distinctive technology-driven transformation in this research domain. Discussion Research overview and key findings This bibliometric analysis of 540 publications [1999–2025] delineates the evolutionary trajectory of diagnostic research on pneumoconiosis. Key findings reveal S-shaped growth with post-2015 acceleration driven by deep learning; a United States-China dual-core landscape, with high-burden regions remaining peripheral; three paradigm shifts from conventional radiology to AI-enabled systems, with current emphasis shifting toward clinical validation and explainable AI (XAI); and emerging occupational exposures—artificial stone and denim sandblasting—that challenge conventional diagnostic criteria. Global research inequities The “dual-core China-United States paradigm” identified in this study is shaped by distinct but complementary epidemiological and policy-driven contexts. China leverages its massive clinical caseload—the highest global pneumoconiosis burden—to achieve rapid translational impact, with institutions such as the Chinese Academy of Sciences playing a critical bridging role between East Asian and North American research networks (14,15). In contrast, the United States capitalizes on its strengths in foundational computer science and industrial translation to lead innovation in algorithm development and diagnostic criteria formulation, with high-density collaborative hubs at institutions such as Harvard University and Brigham and Women’s Hospital (16). This division of labor is not merely a matter of resource allocation; it reflects different epistemic cultures: China’s output is clinically driven and epidemiologically grounded, whereas the contribution of the United States is technologically innovative and methodologically diverse. A notable divergence between publication volume and citation impact further underscores that research influence is not a simple function of productivity. Australia ranked fifth in publication volume but ranked third in citation impact, suggesting that its research output—likely focused on high-impact topics such as artificial stone silicosis and biomarker validation—carries disproportionately high international visibility and translational relevance. Conversely, Germany’s high volume but lower citation impact may reflect a more regionally focused research agenda or delayed recognition in the global discourse. This divergence highlights the importance of network integration, international collaboration, and topic selection in determining research influence (17). The near-zero correlation between country-level DALY rates and publication counts confirms that current research is decoupled from epidemiological need—a finding most starkly illustrated by the silicosis subtype analysis. This pattern reflects the well-documented “10/90 gap” in global health, wherein research resources disproportionately address the health concerns of high-income populations (18). However, this geographic mismatch is not merely a passive outcome of resource constraints; it is actively perpetuated by structural factors including English-language publication requirements, Western-centric journal indexing [e.g., WOSCC’s limited coverage of journals from lower-middle-income countries (LMICs)], and funding priorities that favor technological innovation over implementation science (19). The recent outbreaks of artificial stone-associated silicosis in many countries, including Israel, the United States, Australia, and several European nations, further emphasize that occupational health threats are increasingly borderless, requiring more inclusive international collaboration frameworks (20,21). Australia’s 2024 ban on artificial stone products exemplifies the type of policy action that must be accompanied by equitable technology transfer to high-burden, resource-limited settings (22). Without deliberate intervention, the geographic misalignment documented here is likely to persist, further marginalizing populations with the greatest need for diagnostic innovation as AI technologies continue to advance. Furthermore, the AI-driven literature itself is unevenly distributed: although AI-related publications grew rapidly after 2018, they represent a minority of the total literature. This finding suggests that the technological transition observed in the literature reflects the activities of a relatively small, well-resourced segment of the research community, rather than a broad transformation of the field. For the majority of researchers and clinicians worldwide—particularly those in LMICs—access to AI-interpreted images remains a distant prospect (23), and traditional manual interpretation continues to be the prevailing practice. This observation reframes the “AI revolution” as a highly stratified phenomenon, with implications for global health equity. Diagnostic technology evolution The three-phase evolution identified in our analysis—from chest X-ray and ILO classification, through quantitative CT and CAD, to deep learning-enabled systems—illustrates a progressive trajectory toward objective, reproducible, and automated diagnostics. However, this evolution is not characterized by replacement but by integration and superposition. HRCT provides high-quality input for deep learning, while AI mitigates the high inter-observer variability inherent in HRCT assessments (24,25). This complementary relationship underscores that AI should be viewed as a decision support tool rather than a standalone diagnostic arbiter. Nevertheless, several critical challenges remain. First, the absence of an internationally standardized, AI-compatible HRCT classification system for pneumoconiosis severely limits cross-study comparability and clinical translation. While the ILO classification has served chest radiography well, it was not designed for CT imaging or for AI-based interpretation. Different regions using varying classification systems—such as Japan’s four-type system—further compound this heterogeneity. This gap must be urgently addressed, drawing lessons from parallel efforts in the broader interstitial lung disease community, where deep learning algorithms have been validated against international guideline criteria (25). Second, the “black box” nature of deep learning models impedes clinical trust. Although CNNs can achieve high accuracy in pneumoconiosis classification, the lack of interpretability limits clinician acceptance and regulatory approval (26,27). Moreover, the performance of generative multimodal AI models (e.g., GPT-4o, GPT-5, and MedGemma) for ILO profusion scoring remains moderate, and these models have not yet achieved expert-level performance equivalent to certified B Readers. This variability across AI architectures—with distinct strengths and weaknesses—highlights the need for rigorous, multimodel validation before clinical deployment. Crucially, AI does not eliminate but transforms variability. Human variability is stochastic and context-dependent, whereas AI variability is systematic and predictable. While systematic AI biases can be identified and corrected through rigorous validation, stochastic human variability is more difficult to standardize. This distinction has important implications for clinical deployment: AI tools require continuous performance monitoring and carefully calibrated thresholds that vary by use case (e.g., screening versus compensation adjudication) (28,29). The emerging emphasis on XAI techniques—such as gradient-weighted class activation mapping (Grad-CAM), SHapley Additive exPlanations (SHAP), and local interpretable model-agnostic explanations (LIME)—represents a necessary response to these challenges, enhancing clinician trust and diagnostic accuracy by visualizing decision bases and quantifying feature contributions (30,31). Third, the depth and breadth of interdisciplinary integration remain limited. Current AI-based research primarily focuses on image classification and segmentation tasks, with relatively scarce longitudinal studies on silicosis progression monitoring, treatment response assessment, and prognostic prediction (32). Moreover, bioinformatics approaches—particularly imaging genomics—are emerging as a promising frontier, correlating fibrosis-related gene expression with imaging phenotypes to enable early risk stratification rather than post-disease identification (33). This multimodal integration, combining HRCT features with pulmonary function parameters, serum biomarkers (e.g., KL-6 and SP-D), and occupational exposure timelines, could establish a comprehensive management pathway spanning prevention, early warning, and prognostic evaluation across the full disease cycle (34-36). The translation gap A critical distinction must be made when interpreting the surge in AI-related publications: the difference between efficacy—performance against a ground truth reference under controlled conditions—and field effectiveness—performance and value under real-world occupational health conditions, including cost-effectiveness. Most of the AI-related studies report efficacy, often using curated datasets from well-resourced clinical settings. Consequently, these studies largely miss a crucial aspect of field effectiveness: how these tools perform in routine occupational health surveillance, particularly in LMICs, and whether they are cost-effective in those settings. This gap has practical consequences. In most LMIC settings, CT imaging is not a feasible screening option, leading to the adoption of chest X-ray as the central diagnostic tool (37). However, challenges persist even for chest X-ray-based AI. Current algorithms for silicosis have high sensitivity but suboptimal specificity, particularly in tuberculosis-endemic settings where radiological features overlap. This results in high false-positive rates, which may have severe socio-economic consequences for individuals—potentially affecting job security or leading to unnecessary additional investigations. The public health goal of screening for pneumoconiosis is needed most in LMIC settings, including artisanal and small-scale mining, where access to expert readers is limited and where the consequences of both false positives and false negatives are amplified. Nonetheless, there are notable exceptions to the general scarcity of field effectiveness research. For instance, Zambart and Delft Imaging have introduced CAD4Silicosis in Kitwe, Zambia—one of the few World Health Organization (WHO)-endorsed AI tools for lung disease detection—and are conducting real-world validation studies to generate local evidence on accuracy, feasibility, and policy relevance (38). Similarly, the Southern Africa ex-miners CAD research program has documented that locally developed CAD systems face challenges such as dataset representativeness and biases, which hinder utility in diverse populations (39). These initiatives provide critical lessons for future implementation science, emphasizing that locally validated AI tools developed with representative datasets are essential for translating research innovation into population health impact. Clinical challenges and new threats The timeline analysis reveals a notable shift from a predominant focus on algorithm development toward renewed emphasis on “observer performance” and “clinical decision support”. This transition reflects the academic community’s growing recognition that the true value of AI lies not in standalone metrics, but in its ability to enhance observer performance—particularly by reducing diagnostic errors among less-experienced physicians and improving consistency across different clinical settings (28,29,40,41). This reorientation aligns with the broader trend in medical AI toward human-AI collaboration rather than automation. Concurrently, shifts in global industrial patterns are reshaping the occupational disease landscape. Keyword burst analysis indicates that pneumoconiosis linked to emerging industries such as artificial stone and denim sandblasting has become an accelerating research frontier. Two decades ago, Turkey’s denim sandblasting industry triggered a silicosis crisis affecting thousands of young workers; today, artificial stone—containing over 90% crystalline silica—is driving a new and more aggressive epidemic, with accelerated silicosis developing after exposure periods as short as 4 years (21,42). Radiologically, this form of silicosis can mimic sarcoidosis or other interstitial lung diseases; however, it progresses more rapidly and is frequently complicated by emphysema, spontaneous pneumothorax, and end-stage respiratory failure requiring lung transplantation (21,43). Conventional protective measures—such as dust masks, wet cutting, and ventilation systems (44)—have proven insufficient against such high-intensity exposure, while existing ILO classification criteria struggle to accurately assess disease severity under these novel exposure conditions (45). This evolving epidemiological profile underscores the urgent need for diagnostic technologies with enhanced early-warning capabilities. Developing specific diagnostic criteria for novel dust exposures and constructing dynamic risk prediction models using machine learning have emerged as critical interdisciplinary priorities. The clinical reorientation we observe—from “disease description” to “intelligent diagnosis”—is therefore not merely a technological shift but a response to a changing occupational health landscape that demands more sensitive, specific, and context-adaptive diagnostic tools. Interdisciplinary knowledge transfer The dual-map overlay analysis and Sankey diagram delineate three distinct knowledge flow pathways, illustrating the field’s evolution from traditional clinical medicine toward intelligent diagnostic technologies. The pathway through respiratory medicine journals (e.g., Respirology) primarily connects Japanese scholars to traditional pneumoconiosis topics, representing the classical “respiratory medicine-clinicians-disease ontology” pathway. In contrast, the pathway through imaging journals (e.g., European Journal of Radiology) links international hybrid author groups to AI diagnostic technologies, reflecting the emerging “imaging medicine-transnational researchers–intelligent methods” pathway. The engineering technology pathway (e.g., Biomedical Signal Processing and Control) bridges multinational authors and imaging diagnostic techniques, exemplifying the translational “engineering technology-clinician-engineer collaboration-technical tools” pathway. However, the depth and breadth of interdisciplinary integration remain limited. Current AI research primarily focuses on image classification and segmentation tasks, with relatively scarce longitudinal studies on silicosis progression monitoring, treatment response assessment, and prognostic prediction (32). Moreover, the application of advanced methodologies such as federated learning and transfer learning in multicenter data collaboration and model generalization remains exploratory (46). The strong co-citation intensity of computer science journals underscores the deepening integration of AI with medical imaging, although the translational pathway from computational innovation to clinical practice remains fragmented. Future efforts should focus on creating structured mechanisms for clinician-engineer collaboration, developing shared data standards, and fostering interdisciplinary training programs to bridge these epistemic divides. Future research directions Based on the results of the present study, future research and policy should prioritize the following directions. First, efforts should concentrate on advancing the clinical translation of XAI by developing standardized interpretability metrics and conducting real-world trials that assess its impact on diagnostic accuracy and workflow efficiency. Second, a multimodal diagnostic pathway integrating HRCT, serum biomarkers, pulmonary function tests, and occupational exposure history should be established to facilitate early warning, precise diagnosis, and prognostic evaluation across the full disease cycle. Third, diagnostic criteria and risk prediction models must be urgently updated to address novel occupational exposures—particularly accelerated silicosis associated with artificial stone processing and denim sandblasting—which present distinct radiological features and rapid progression patterns that challenge current ILO frameworks. Fourth, the development of an internationally accepted, AI-compatible HRCT classification system for pneumoconiosis should be prioritized. The ILO classification was designed for chest radiography, not CT imaging or AI-based interpretation. A unified standard would enable cross-study comparability, facilitate regulatory approval, and accelerate clinical adoption. Fifth, standardized benchmarking and validation frameworks for AI systems are urgently needed, including model-specific variability metrics, robust learning methods to mitigate label noise from inter-reader variability, and continuous performance monitoring protocols for real-world deployment. The observed variability across AI architectures highlights the critical importance of rigorous, multimodel validation before clinical deployment. Sixth, international collaboration and capacity-building must be reinforced to narrow the research-disease burden gap, particularly in high-prevalence, resource-limited regions. This includes strengthening South-South networks (e.g., the ASEAN Occupational Health Network and the African Union’s occupational health framework) to facilitate knowledge exchange among countries facing similar exposure profiles and diagnostic challenges. Additionally, future bibliometric studies should adopt a multi-database approach integrating WOSCC, Scopus, PubMed, and China National Knowledge Infrastructure (CNKI) and incorporate grey literature to achieve more representative coverage of global research output and reduce language and indexing biases. Limitations This study has several limitations that should be considered when interpreting its findings. First, the exclusive reliance on the WOSCC introduces substantial language, geographic, and publication-type biases. Notably, a significant volume of Chinese-language clinical research indexed in the CNKI—as well as grey literature from the WHO, ILO, and national surveillance systems—was excluded. Other high-burden regions (Southeast Asia, South Asia, sub-Saharan Africa, and South America) are similarly underrepresented due to English-language publication requirements and limited indexing of regional journals. While we included both English- and Chinese-language publications, the proportion of Chinese-language publications in our dataset (2.2%) likely underrepresents the true volume of relevant research. Second, the use of GBD data as a comparative benchmark entails well-documented limitations, including underdiagnosis, underreporting, and misclassification—particularly in LMICs where occupational health surveillance is weak. However, because we used GBD data as a relative comparative benchmark rather than as precise absolute estimates, these limitations do not materially affect our core conclusion. The gross misalignment between research output and disease burden is very large and cannot be reversed by applying plausible corrections to underlying estimates; moreover, true burden in LMICs is likely higher, strengthening our conclusion. Third, bibliometric methods inherently suffer from time lags. Cutting-edge advances emerging in late 2024–2025 may not yet have accrued sufficient citation impact. Keyword co-occurrence analysis relies on author-provided keywords, which may present terminology standardization issues despite our use of synonym-merging functions. CiteSpace’s burst detection algorithm is sensitive to time window selection, and burst strength of recent literature may be underestimated. Fourth, our analysis focuses on quantitative characteristics—publication volume, citation frequency, and collaboration networks—rather than qualitative evaluations of research quality, clinical translation rates, or health economic benefits, which require subsequent systematic reviews and real-world studies. Additionally, the proportion of AI-related publications reflects research activity within a select group of well-resourced institutions, not the adoption of AI technologies in routine clinical care, particularly in high-burden regions with limited AI infrastructure. Fifth, our temporal trend analysis was not stratified by pneumoconiosis subtype. Given the distinct radiological features, risk populations, and research communities associated with silicosis, asbestosis, and coal workers’ pneumoconiosis, this aggregation may obscure subtype-specific patterns in AI adoption and research focus. Sixth, and critically, our bibliometric analysis captures the surge in AI efficacy research but does not reflect the relative scarcity of field effectiveness studies. The few publications that incorporate AI-related keywords predominantly report performance under controlled conditions using curated datasets. Studies evaluating AI performance in real-world occupational health settings—particularly in LMICs, where the need is the greatest—remain scarce. This distinction between efficacy and effectiveness should be considered when interpreting the “AI dominance” narrative, as the bibliometric surge does not necessarily translate into clinical or public health impact. Conclusions This bibliometric analysis confirms that pneumoconiosis diagnostic research has undergone a paradigm shift from manual qualitative assessment to AI-powered quantitative analysis, driven by deep learning and interdisciplinary convergence. However, the persistent geographic decoupling between research capacity and disease burden—with high-burden regions remaining peripheral—reveals significant inequities in global occupational health. Future priorities should focus on clinical validation of XAI, multimodal diagnostic integration, updated criteria for emerging occupational exposures, and equitable resource allocation to bridge the gap between innovation and real-world impact. Supplementary The article’s supplementary files as Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-0914/prf