Despite the importance of peer review for grant funding decisions, academics are often reluctant to conduct it. This can lead to long delays between submission and the final decision as well as the risk of substandard reviews from busy or non-specialist scholars. At least one funder now uses Large Language Models (LLMs) to reduce the reviewing burden but the accuracy of LLMs for scoring grant proposals needs to be assessed.

This article compares scores from a range of medium sized open-weight LLMs with peer review scores for a well-researched dataset, 142 Swedish Medical Council post-doctoral fellowship applications from 1994.

Whilst the LLM scores correlate moderately between each other (mean Spearman correlation: 0.34), they correlated weakly but positively and mostly statistically significantly with the average expert scores (mean Spearman correlation: 0.22). The highest rank correlation between expert scores and LLMs was 0.33 for Gemma 3 27 b based on proposal titles and summaries without their main texts, which is about half (56 %) of the correlation between reviewers.

The small sample size, old funding call and heterogeneous evaluation criteria all undermine the robustness of the analysis.

Despite the ability of LLMs to score grant proposals being quantitatively weaker than that of experts, at least in this special case, they may have role in application triage or tie-breaking.

This is the first assessment of the value of LLM scores for funding proposals.

This study aims to address the challenges of medical data exchange across institutions.

A blockchain-based system, UMBD (Unlocking Medical Big Data), is proposed to enable secure and interoperable medical data sharing through decentralized management ensuring data integrity and traceability.

UMBD ensures complete lifecycle management, recording data creation, storage, sharing, utilization, and revocation in the blockchain. Trusted sharing is guaranteed by blockchain, addressing distrust among medical institutions. UMBD provides accountability and traceability for incorrect examination results, tracing errors back to the source and facilitating timely remedial measures. Experimental verification demonstrates UMBD’s high efficiency, making it effective for all involved parties.

Limited experimental data volume may affect real-world generalizability.

UMBD promotes the reliable sharing of medical data among medical institutions, thereby reducing patients’ burden of repeated examinations and supporting fair resource allocation.

The study introduces a blockchain-based system for medical data interoperability, ensuring trusted sharing, full lifecycle management, and accountability within medical big data environments.

Traditional science maps cluster documents into topics but are inherently biased toward clustering certain topics over others. This study investigates the extent to which topic bias can be influenced through the selection of data sources used to construct document networks.

We evaluate the clustering effectiveness of several topic categories using document networks constructed from two traditional science mapping sources (citations and text similarity) and six non-traditional data sources (policy documents, patent families, document authors, Facebook users, Twitter users, and Twitter conversations). Each source is evaluated both independently and in combination with a text similarity network.

Different data sources favor different kinds of topics. Facebook users favor health issues, patent families favor biotechnology topics, policy documents favor government and social issues, Twitter conversations favor food topics, Twitter users favor nursing topics, and document authors favor geographical entities. These findings demonstrate that topic bias can be systematically influenced through data source selection.

The study focuses on biomedical publications and is limited to the topic categories and data sources examined. Additional domains, data sources, and source combination methods may exhibit different patterns of topic bias.

The ability to influence topic bias through data source selection opens up the possibility of creating science maps tailored to different information needs. The reported source-specific biases can support the design of science maps optimized for particular users, tasks, or domains.

This study provides one of the first large-scale investigations of how alternative data sources affect topic emergence in science maps. It introduces an expanded methodology for evaluating topic-level clustering effectiveness and systematically characterizes the topical biases associated with different data sources, providing a foundation for future science map customization.

Digital transformation (DT) is a configurational construct spanning technology, process, organizational, and strategic change, and valid measurement is a prerequisite for understanding it. Existing text-based methods fragment annual reports into keywords or isolated sentences, breaking the cross-sentence semantic structure that expresses the construct. This paper examines how textual representational granularity affects content validity in DT measurement, and proposes a complete-semantic measurement framework.

From 28,823 annual reports of Chinese listed manufacturers (2006–2023), we build a 1,000-report sample through structured sampling and design a three-stage complete-semantic framework (input, training-signal, and model-bearing). The framework is evaluated through comparative experiments, ablation, and sensitivity analysis, then applied to the full sample, with an existing sentence-level measure used for external validation.

Textual representational granularity shapes which forms of the construct enter measurement: a fragmented representation relies on explicit digital keywords and leaves early and implicit transformation unseen. Across the full sample, DT in Chinese manufacturing has widened in breadth and deepened in structure, shifting from localized exploration to cross-departmental integration. An existing sentence-level measure agrees on explicit, deep-stage transformation but diverges in the early stages that fragmentation struggles to capture.

The study focuses on Chinese manufacturing; its conclusions need testing in other industries, settings, and disclosure contexts, and on new model architectures and cross-construct tasks.

The framework requires little annotation, supports local deployment, and aligns with national standards, offering a reproducible tool for research, policy, and firm-level analysis.

The contribution lies not in using large language models (LLMs) for text classification, but in integrating the cross-sentence semantic structure on which a configurational construct depends. The paper brings textual representational granularity into content-validity assessment, extending construct validity from “item–construct domain” coverage to “textual representation–construct structure” alignment, and offering a testable lens for measuring complex organizational constructs from disclosure text.

The translation of science into policy is frequently oversimplified by linear assumptions, leaving the specific micro-level drivers of policy adoption unclear. This study aims to develop a systematic and explainable framework to predict and interpret how structural, academic, and dissemination factors influence the policy citation of scientific papers.

This study constructs a large-scale dataset linking scientific research to policy documents, and evaluates multiple machine learning classifiers to predict policy citations, employing SMOTEENN to address class imbalance. To uncover the decision-making mechanisms, the SHAP framework is applied to the XGBoost model. This enables an in-depth analysis of global feature importance, contribution directions, and non-linear dependencies.

The results show that policy citation is moderately predictable and is better captured by ensemble-based machine learning models than by simpler baselines. SHAP analysis reveals a differentiated and non-linear selection process. Citation Count is the strongest predictor, while Patent Count and Tweet Count also provide substantial but distinct predictive information. The dependence plots further identify saturation effects, power-law-like decay, threshold effects, and bounded novelty. These findings support a Validation–Knowledge Filtering Mechanism, in which scientific papers are more likely to be cited in policy publications when they are both externally validated and suitable for policy-oriented knowledge use.

The focus on articles from science and Overton-indexed policy documents may limit the generalizability of the findings. Additionally, the binary classification of citations overlooks the substantive degree and contextual depth of the extent and context of how scientific evidence is used in policy documents.

Enhancing policy impact requires more than academic impact, institutional prestige, or media exposure alone. Researchers should strengthen both evidence credibility and policy-oriented usability, while policy organizations should avoid over-reliance on simple prestige, scale, or visibility signals.

This study advances scientometrics by shifting from descriptive correlations to mechanistic prediction. It identifies the drivers of policy citation and proposes a Validation–Knowledge Filtering Mechanism for explaining scientific paper selection in policy.

Most minimally invasive surgery (MIS) literature analyses focus on minimally invasive diagnosis and treatment of specific diseases or single procedures, lacking a systematic depiction of the field’s overall structure and evolution, while traditional methods suffer from strong subjectivity and limited data scale. This study aims to employ a topic modeling approach to comprehensively map the knowledge graph and evolutionary pathway of the MIS field.

A combination of latent dirichlet allocation (LDA) and dynamic topic model (DTM) was used to analyze all MIS articles published between 1985 and 2024 in 37 high-quality, authoritative PubMed journals (including both specialized surgical and comprehensive journals), examining development maturity, chronological distribution, country distribution, author distribution, hot topic identification, and evolution.

The number of MIS publications is projected to grow continuously, with the United States leading and China ranking second; a cohort of highly productive scholars has formed. The field can be grouped into five major topic clusters, among which “minimally invasive diagnosis and treatment of digestive system (gastrointestinal) diseases” currently has the highest influence and attention. Core subtopics running through the hot topics include minimally invasive diagnosis and treatment of colorectal diseases, hernia, and bariatric/metabolic conditions. The research focus of colorectal minimally invasive diagnosis and treatment has shifted from general laparoscopic techniques to refined operations for specific diseases, deeply integrating with precision oncology. Hernia minimally invasive diagnosis and treatment has extended from safety evaluation of basic surgical procedures to instrument innovation, simulation training, and complex hernia management. Bariatric/metabolic minimally invasive diagnosis and treatment has expanded from simple weight-loss surgery to multidisciplinary intersecting areas such as management of metabolic comorbidities, prevention and control of postoperative complications, and health economics. Overall, MIS research is continuously evolving toward greater precision, more systematic integration, and deeper interdisciplinarity.

The limited journal scope may affect the generalizability of the findings.

This study reveals the hot topics and evolutionary patterns of MIS research, helping researchers quickly grasp frontier dynamics and providing empirical evidence for hospitals to optimize resource allocation and formulate disciplinary development strategies.

It pioneers the introduction of topic modeling to systematically mine the vast MIS literature, expanding its application in healthcare services and offering an objective and dynamic cognitive perspective.

The discovery, interpretation, and transformation of knowledge into practical insight are being revolutionized by generative artificial intelligence. In light of these developments, this conceptual paper critically reexamines human-guided KDD and knowledge discovery in databases. As a revised framework for understanding knowledge discovery in human AI environments, it suggests Generative Knowledge Discovery in Databases, or Gen-KDD.

Using a conceptual approach based on envisioning, the paper seeks to advance theory. In order to create a layered framework that incorporates process, agency, governance, and epistemological reflection, it synthesizes literature on KDD, knowledge creation, human AI collaboration, responsible AI, and generative AI.

Knowledge discovery is rethought as a generative and reflexive human AI ecology by the suggested Gen-KDD framework. Generative contextualization, autonomous data curation, generative feature engineering, generative discovery, and co-evolutionary sensemaking are its five stages. These stages are arranged according to human-dominant, AI-dominant, and hybrid layers, making it clear where AI can take the lead, where human judgment is still crucial, and where shared sensemaking is necessary.

Rather than providing actual evidence, the paper provides a conceptual framework. Future studies should operationalize Gen-KDD across domains and investigate how it affects organizational learning, accountability, transparency, fairness, and decision quality.

Gen-KDD’s design guidelines can be used by organizations that want to ethically incorporate generative AI into analytics and decision-making. It demonstrates how human oversight, ethics, and governance can be integrated into discovery processes rather than being added as external controls.

This paper advances a theory-oriented framework that integrates KDD, knowledge-creation theory, human-AI collaboration, responsible generative AI, and epistemological reflection into a single-layered ecology. Gen-KDD offers a richer foundation for knowledge discovery than linear process models by explicitly addressing the changing roles and limits of both human and machine cognition.

Although citation-based indicators are widely used, they are not useful for recently published research, directly reflect only one of the three common dimensions of research quality, and have little value in some social sciences, arts and humanities. Large Language Models (LLMs) may address some of these weaknesses.

This article reports a science-wide assessment of the research quality scoring capability of ChatGPT-4o mini, ChatGPT-4o, and ChatGPT-5 mini. It correlates ChatGPT scores, averaged over 5 repetitions, with departmental average quality scores for 107,212 UK-based journal articles.

ChatGPT-4o is marginally better than ChatGPT-4o mini in most of the 34 field-based Units of Assessment (UoAs) tested. ChatGPT-4o scores have a positive correlation with research quality in 33 of the 34 UoAs, with the results being statistically significant in 31. ChatGPT-4o scores had a higher correlation with research quality than long term citation rates in 21 out of 34 UoAs and a higher correlation than short term citation rates in 26 out of 34 UoAs. The most substantial exception is Physics, for which citations are more useful. ChatGPT-5 mini has even stronger correlations overall and for departmental averages, it correlates more strongly with quality scores than do citations in 31 out of 34 UoAs, with correlations reaching 0.905.

All articles assessed are from the UK. The practical value of LLM scores for decision making is not assessed. Only the Normalised Log-transformed Citation Score (NLCS) was tested against ChatGPT rather than other citation rate indicators.

ChatGPT can be considered as a research quality indicator to support expert judgement in almost all academic fields.

The results give science-wide evidence that ChatGPT-4o mini, ChatGPT-4o, and ChatGPT-5 mini are competitive with citations as new research quality indicator sources, and technically better in most fields.

This study aims to improve fine-grained understanding of citation contexts in scientific literature by jointly modeling two complementary aspects of citation behavior: citation intent and citation evaluation.

We propose a joint multi-task learning (Joint-MTL) framework that simultaneously models citation intent classification and citation evaluation classification through a shared encoder with task-specific prediction heads. To support the evaluation task, we construct a new manually annotated dataset, CiteEva. The model is trained using an alternating optimization strategy across the two tasks, enabling the shared encoder to learn representations that capture both functional and evaluative characteristics of citation contexts.

Experiments conducted on the public SciCite dataset and the newly constructed CiteEva dataset show that the joint learning framework consistently outperforms strong single-task baselines and several recent multi-task or feature-based approaches. Additional analysis reveals a moderate representation overlap between the two tasks, indicating that they share meaningful semantic signals while maintaining task-specific distinctions. These findings empirically support the effectiveness of jointly modeling citation intent and citation evaluation.

The current study focuses on two citation-related tasks and evaluates the framework on a limited set of datasets and citation categories. Future work could extend the approach to additional citation analysis tasks and larger-scale corpora.

The proposed approach can support applications such as scientific impact assessment, literature review assistance, and scholarly knowledge mining by enabling more nuanced interpretation of citation roles and evaluative stances.

This study provides empirical evidence for the benefit of jointly modeling citation intent and citation evaluation and introduces CiteEva, a new high-quality dataset for citation evaluation research, contributing to more comprehensive citation context analysis.

As technological competition increasingly becomes the focus of global attention, identifying key core technologies (KCTs) is of great significance for seizing the technological high ground, promoting national economic development, and safeguarding national security.

Based on a clear definition of the concept and characteristics of KCTs, this study proposes a field–topic–technology (FTT) identification method, which follows a progressive, multi-method fusion approach that combines primary and auxiliary recognition paths. The FTT method is empirically applied to patents in the field of small-molecule targeted therapies for lung cancer. The primary path of the method proceeds as “core subfield → core topic → KCTs”, while the auxiliary path follows “core subfield → stage-based topic evolution trends → auxiliary identification of KCTs.” Through the combination of primary and auxiliary paths and a progressive hierarchical approach, this method achieves systematic identification of KCTs. Meanwhile, the LightGBM model is employed for indicator weighting, enhancing the scientific validity and objectivity of the weighting process.

The study successfully identifies six KCTs in the field of small-molecule targeted therapies for lung cancer. Comparative analyses with other identification methods as well as with official documents and high-impact publications verify the scientificity, feasibility, effectiveness, and robustness of the FTT method. Empirical evidence demonstrates that this method is suitable for fields with abundant patent data and relatively clear technological evolution pathways.

Although this study constructs a KCTs identification framework, several limitations remain in the identification process. First, the characteristics of KCTs are multidimensional and complex; this study selects indicators only from four dimensions, which makes it difficult to fully capture their intrinsic attributes. Second, the constructed indicator system mainly relies on quantitative indicators, with insufficient representation of qualitative factors such as tacit knowledge and engineering experience. Third, this study is primarily based on patent data and does not fully integrate multi-source information such as academic publications, industry standards, and policy documents, which may lead to certain biases in the identification results. In addition, the empirical validation is conducted in a single field, and the cross-field applicability of the method still requires further examination. Future research may enhance the comprehensiveness and robustness of KCTs identification by expanding feature dimensions, integrating multi-source heterogeneous data, and conducting cross-field validation.

The proposed FTT approach supports technology intelligence and strategic decision-making through KCT identification and technological evolution tracking. It can be extended to other patent‑rich technology fields to support R&D planning and policy formulation.

Addressing existing research gaps – such as limited conceptual dimensions, fragmented identification methods, and lack of focus – this study defines the concept and attribute features of KCTs from multiple dimensions and constructs a hierarchical, multi-dimensional, and multi-method KCT identification framework that progresses from the macro field level to the micro technology level, integrating both primary and auxiliary recognition paths. Compared with previous studies that relied on single methods or linear processes, the proposed approach demonstrates stronger logical hierarchy, systematic structure, and methodological rigor. Furthermore, by introducing a machine learning-based approach for dynamic learning and nonlinear optimization of indicator weights, this study overcomes the limitations of traditional weighting methods and enhances the scientificity and objectivity of the identification results. Overall, this research achieves conceptual and methodological innovation in KCT identification, offering new perspectives for hierarchical system construction and indicator weighting optimization, and providing valuable references for subsequent studies and practical applications.

This study examines how data-driven research (DDR) has diffused across U.S. higher education institutions and investigates its relationship with scientific impact as measured by citation counts. It aims to clarify how institutional context, disciplinary affiliation, and publication characteristics shape the visibility of data-intensive research.

Using publications indexed in the Web of Science from 2013 to 2023, DDR is identified through a transparent text-mining approach based on abstract-level keywords related to artificial intelligence, machine learning, big data, and data science. Disciplinary, institutional, and demographic patterns in DDR output are analyzed. Citation counts are modeled using Zero-Inflated Negative Binomial regression, Random Forest, eXtreme Gradient Boosting, and Support Vector Regression. Feature variables capture DDR status, institutional classification, population served, and publication-level characteristics.

Results indicate that DDR output expanded across all research areas, with the strongest growth in computer science and engineering and disproportionately large contributions from R1 institutions. Across modeling approaches, publication age, number of authors, and DDR status consistently emerge as the most influential predictors of citation impact.

DDR identification relies on keyword-based text classification, which may omit relevant studies or include unrelated work. The Web of Science database emphasizes English-language and high-impact journals, potentially limiting coverage. In addition, the observational design supports association rather than causal inference.

The findings inform research evaluators, policymakers, and institutional leaders about patterns of methodological diffusion and citation visibility. The proposed framework can support evidence-based assessment of data-driven scholarship and guide capacity-building efforts, particularly in less-resourced institutions.

This study integrates bibliometric analysis, text mining, and machine-learning methods to examine data-driven research at scale. It provides a reproducible approach for identifying DDR and offers new empirical evidence on how institutional and disciplinary contexts shape research visibility, contributing to science-of-science and research evaluation literature.

Research on inverse problems of geophysics is extensive. Thus, there is a need to conduct in-depth and comprehensive bibliometric research to capture the overall development of the field.

We conduct a bibliometric analysis using CiteSpace to reveal dynamic trends in collaboration, co-citation, and co-occurrence in this field.

We find that publications in the field have shown a fluctuating upward trend since 2004, accompanied by a gradual increase in its proportion within the broader geophysics’ literature. In terms of institutional collaborations, Chinese and French research institutions occupy an important position in the global collaboration network. From a co-citation perspective, Geophysics is the most cited journal, and Araya-Polo Mauricio’s article (Araya-Polo, M., J. Jennings, A. Adler, and T. Dahlke. 2018. “Deep-Learning Tomography.” The Leading Edge 37 (1): 58–66) is the most cited reference. From the perspective of keyword co-occurrence, the keyword clusters show the diversity of geophysical inversion research. The research topics cover fundamental theory, method development, and practical applications. In addition, future research may focus more on research based on data-driven workflows and artificial intelligence.

Only the SCI and SSCI editions in the Web of Science database were used as the data source, ignoring the literature in other databases. We focused on the literature from 2004 to 2023, which excludes the latest literature in 2024 and earlier literature before 2004.

We provided a comprehensive and specific analysis of the literature in the field over the last 20 years to help readers gain a comprehensive understanding of the current state of research, hotspots, evolution, and trends in the field. We discussed current research challenges and possible future research directions to help researchers identify appropriate research directions and conduct subsequent research more efficiently.

Compared with previous related studies, this study is innovative as we provide a comprehensive and specific analysis of the literature in the field over the last 20 years to help readers gain a comprehensive understanding of the current state of research, hotspots, evolution, and trends in the field and discuss current research challenges and possible future research directions to help researchers identify appropriate research directions and conduct subsequent research more efficiently.

This study explores the patterns of critical thinking in human-AI co-creation tasks and how they relate to creative performance, aiming to clarify the relationships between these patterns and creative performance and deepen understanding of the cognitive processes.

The dataset comes from human-AI dialogue data for two level creativity tasks, text summarization and creative writing, on the Hugging Face platform. A critical thinking coding scheme is developed to analyze critical thinking skills, then lag sequential analysis (LSA) and frequent sequence mining (FSM) are used to identify the critical thinking patterns. Statistical methods are employed to explore associations between these patterns and creative performance.

Six task-dependent critical thinking patterns are identified in human-AI co-creation: Explanatory consolidation (EC), analytical structuring (AS), evaluative specification (ES), evaluative decomposition (ED), evaluative regulation (ER) and inferential self-loop (IS). Whereas low and medium-order thinking skills (patterns EC, AS, ES) dominated the low-creativity task, the high-creativity task incorporated low, medium and high-order skills (patterns ES, ED, ER, and IS). Furthermore, these patterns show associations with creative performance in the creative writing task: the ER pattern supports comprehensive creativity, the ED and ES patterns facilitate partial creativity, the IS pattern acts as a limiting factor.

Reliance on merged large-scale datasets analyses may constrain interpretation of human-AI co-creation, and future studies should validate these findings through experimental approaches and single-dataset replication.

The findings can help educators design tasks at different levels of creativity to elicit students’ critical thinking, and help writers design personalized prompts that enhance specific dimensions of creativity (originality or elaboration) in line with their goals.

By adapting frameworks of critical thinking to human-AI co-creation, the study contributes to a comprehensive understanding of critical thinking patterns, compares these patterns across different creative tasks, and uncovers their associations with creative performance.