A comprehensive bibliometric and content analysis examining research on artificial intelligence–based computer-aided diagnosis and clinical decision support systems shows that while publication output surged approximately 21-fold over a decade, the vast majority of influential empirical publications focused on initial algorithmic development rather than real-world clinical implementation. Researchers led by Shuo-Chen Chien, Chien-Chu Huang, and Chun-Chung Huang conducted the investigation to assess publication growth, geographic distribution, conceptual trajectories, public search trends, and translational depth across scholarly literature published between 2016 and August 2026. Evaluating nearly twenty thousand indexed records alongside detailed reviews of citation-ranked full texts, the authors observed that scholarly work has heavily concentrated on algorithmic image analysis and diagnostic tasks. However, rigorously documented external validation, workflow integration studies, and direct patient-outcome assessments were extraordinarily rare among the most visible publications.
To capture the evolution of artificial intelligence in diagnostic and clinical decision contexts, the investigators queried the Web of Science Core Collection for journal articles and proceedings papers published from 2016 through August 2026. The initial database query retrieved 24,226 records, which was narrowed through systematic exclusions including the removal of 1,092 records published before 2016, five dated after 2026, six duplicate entries, 43 non-English items, 2,593 reviews, 418 meeting abstracts, 148 editorials, and 112 records categorized as other material. This screening yielded a primary analytical corpus of 19,809 English-language publications. The research team examined this body of work using Bibliometrix and Biblioshiny in R, combined with Python scripting to execute time-series analyses. They tracked corpus-wide annual and monthly growth rates, evaluated journal and institutional contributions, analyzed author-keyword frequencies, mapped citation networks, and conducted a parallel post hoc sensitivity check using the Scopus database to test whether geographic publication patterns remained consistent across indexing platforms.
The quantitative growth trajectory demonstrated a dramatic expansion across the study period, rising from 193 published records in 2016 to 2,876 in 2024 and reaching 4,082 papers in 2025. This annual increase corresponded to an increase in mean monthly output from 16.1 records in 2016 to 239.7 in 2024 and 340.2 in 2025. Incomplete indexing through August 2026 accounted for another 3,378 records, which the authors analyzed separately to avoid distorting annual projections. Despite this massive expansion in scholarly output, publication volume remained geographically concentrated. Among the 19,797 records in the primary dataset with identifiable corresponding-author locations across 107 countries or territories, China accounted for 8,015 publications, or 40.49 percent of the corpus, followed by the United States with 2,485 records, or 12.55 percent, and India with 1,641 records, representing 8.29 percent. Together, these three nations contributed 61.33 percent of all identified output, yielding a Gini coefficient of 0.869 that confirmed strong geographic concentration.
Institutional productivity reflected this geographic concentration, with prominent contributions from research centers in East Asia and North America. Standardized institutional output was led by Harvard University with 452 documents, followed by the Chinese Academy of Sciences with 420 documents, Shanghai Jiao Tong University with 374, Harvard University Medical Affiliates with 359, and the University of California System with 313 documents. At the journal level, multidisciplinary and open-access venues published the greatest share of research. Scientific Reports published the largest volume with 916 included papers, followed by IEEE Access with 664 papers, and Biomedical Signal Processing and Control with 425 papers. Citation performance did not mirror publication volume, as specialized venues such as Medical Image Analysis achieved substantially higher mean global citations per paper despite producing fewer total articles. A sensitivity analysis assessing complete publication years from 2016 through 2025 in the Scopus database demonstrated high rank-order agreement with Web of Science (Spearman rho of 0.911), confirming that country rankings were broadly robust across major bibliographic databases.
Mapping author keywords and co-occurrence patterns revealed that the conceptual structure of the field has been heavily dominated by specific computational techniques. Across the full 19,809-record corpus, the most common normalized author keywords were machine learning, appearing in 5,666 publications, deep learning in 4,022, computer-aided diagnosis in 1,512, artificial intelligence in 1,449, and convolutional neural network in 1,120 publications. Other prominent terms among the top twenty included risk prediction, breast cancer, feature extraction, transfer learning, classification, prognosis, and radiomics. Co-occurrence network analysis clustered these computational methods alongside medical imaging modalities, such as magnetic resonance imaging, and oncology targets. Bibliographic coupling partitioned 19,731 connected publications into 15 distinct research communities, dominated by clinical prediction, diagnostic imaging, deep-learning image processing, molecular oncology, and medical image segmentation, while reference co-citation identified seven underlying clusters centered on foundations in statistical learning and deep neural network architectures.
To investigate the relationship between academic publishing and public visibility, the authors conducted a longitudinal comparison tracking monthly bibliometric author-keyword frequencies against Google Trends relative search volumes from January 2016 through July 2026. Examining ten leading author-keyword concepts, the researchers generated composite Google Trends series by weighting five coordinated search terms per concept and standardizing trajectories using min-max scaling. In unadjusted level-series comparisons, nine of the ten concepts exhibited statistically significant correlations with Google Trends interest, which the authors characterized as descriptive long-term co-trending reflecting general secular expansion. However, when the authors evaluated first-differenced monthly series to remove underlying secular trends, Kendall’s tau-b coefficients ranged only from negative 0.104 to positive 0.079. None of the ten first-difference associations remained statistically significant following Bonferroni correction, and all six-month block bootstrap confidence intervals crossed zero, indicating that short-term monthly shifts in public search interest do not reliably mirror academic publication fluctuations.
The study also assessed documentation within the funding-acknowledgement fields of 16,431 records published during the complete calendar years 2016 through 2025. Across this period, 11,633 records, or 70.80 percent, contained nonblank funding information, with documented shares increasing from 66.32 percent in 2016 to 72.51 percent in 2025, peaking at 73.05 percent in 2024. In an ecological country-level analysis of 51 nations with at least twenty publications, funding-acknowledgement proportions correlated positively with national median gross domestic product per capita at purchasing power parity (Spearman rho of 0.658) and the 2024 Global Health Index (Spearman rho of 0.474). Documented publications per million inhabitants also showed strong positive associations with gross domestic product per capita (rho of 0.835) and the Global Health Index (rho of 0.705). The authors cautioned that these ecological patterns reflect country-level reporting contexts rather than causal effects of national funding policies, noting that total publication output per million residents ranged widely from 0.181 in Russia to 15.199 in South Korea.
To probe study quality and translational maturity, the investigators performed a structured content analysis of 215 citation-ranked publications published between 2016 and 2023, representing approximately the top 1.09 percent of the corpus by local citation count. Among these 215 influential papers, 191 were empirical research investigations, comprising 185 quantitative descriptive studies, three randomized controlled trials, and three nonrandomized studies, with the remaining 24 categorized as nonempirical frameworks or reviews. An examination of the highest observed evaluation stage across the 191 empirical papers revealed that 138 studies (72.3 percent) ceased evaluation at internal validation. In contrast, external validation served as the highest stage in 32 publications (16.8 percent), user or workflow evaluation in 13 (6.8 percent), and comparative impact trials in only three (1.6 percent). Initial model development without validation accounted for two papers (1.0 percent), prospective silent testing for one (0.5 percent), and routine clinical deployment was documented in just two publications (1.0 percent).
Detailed methodological extraction from these 191 empirical studies highlighted substantial translational deficits. Retrospective cohort designs predominated, accounting for 170 studies (89.0 percent), while convolutional neural networks represented the primary computational architecture in 104 studies (54.5 percent), and diagnostic classification formed the central task in 96 studies (50.3 percent). Although nearly all empirical papers evaluated discriminative performance, with 190 of 191 reporting metrics such as area under the curve or accuracy, only 29 papers (15.2 percent) reported model calibration. Prospective evaluations appeared in only 11 studies (5.8 percent), and measured clinical workflow or process outcomes were tracked in just 12 (6.3 percent). Rigorous human evaluations of algorithmic explanations occurred in only two of 172 applicable studies (1.2 percent). Strikingly, zero publications reported measured patient health outcomes, explicit fairness metrics, or postdeployment algorithmic monitoring, illustrating an pervasive gap between model performance and real-world clinical action across the field’s most cited literature.
An additional year-stratified content analysis of 130 publications from 2024, 2025, and partial 2026 examined recent advances, including the emergence of generative models. Among 117 studies evaluating an artificial intelligence system, confirmed generative applications were documented in three publications, including masked-image reconstruction during pretraining for lung cancer screening and retrieval-augmented large language model recommendations for clinical testing. However, translational safeguards remained scarce in recent work: independent external validation appeared in 26 studies, calibration in 39, and prospective testing in just one. Summarizing their findings, the authors emphasized that the rapid expansion of academic literature and public interest should not be mistaken for clinical readiness. They noted that future progress in computer-aided diagnosis requires shifting resources from retrospective algorithm tuning toward independent multisite validation, explicit fairness assessments, prospective workflow testing, and continuous postdeployment monitoring.
Subject of Research: Bibliometric and content analysis of AI-based computer-aided diagnosis and clinical decision support systems research
Article Title: Mapping Research on AI-Based Computer-Aided Diagnosis and Clinical Decision Support Systems: a Bibliometric Study Integrating Google Trends and Content Analysis
Article References: Chien, S.-C., Huang, C.-C., & Huang, C.-C. (2026). Mapping Research on AI-Based Computer-Aided Diagnosis and Clinical Decision Support Systems: a Bibliometric Study Integrating Google Trends and Content Analysis. Journal of Medical Systems, 50(1), Article 161. https://doi.org/10.1007/s10916-026-02488-w
Image Credits: AI Generated
DOI: 10.1007/s10916-026-02488-w
Keywords: artificial intelligence, clinical decision support systems, computer aided diagnosis, bibliometric analysis, machine learning, deep learning, convolutional neural networks, external validation, model calibration, Google Trends, medical imaging, retrospective studies, algorithmic fairness, workflow integration, translational medicine, health informatics
News Source: Blake Davidson. (November 16, 2026). Study Finds Clinical AI Research Surged While Real-World Evaluation Remained Sparse. Scienmag.



