VOSviewer was utilized to analyze the co-authorship network of key authors in construction accident causality analysis, highlighting collaboration relationships and academic influence. Visualization results are presented in Figure 3. A threshold of at least two publications per author was applied, selecting 26 authors from an initial pool of 353 (Rusydiana et al., 2021). Nodes represent authors, with links indicating collaboration, their thickness reflecting intensity, node size denoting publication count, and color signifying average publication year. For instance, “Jianhong Shen,” “Shupeng Liu,” and “Jing Zhang” share an average publication year of 2024, indicating recent contributions aligned with current industry trends. The overall network density is low, suggesting limited integration across research teams. Most scholars contributed only one or two papers and lack sustained cooperation. This pattern reflects the field’s current developmental stage, where cross-institutional and interdisciplinary collaboration is still emerging.

Citation count is widely regarded as an indicator of research quality (Martins et al., 2024). Key quantitative indicators for the top ten authors by citations are summarized in Table 1. “Ts Abdelhamid” (629 citations), “Chia-fen Chi” (159 citations), and “Michael Behm” (153 citations) demonstrate significant impact. In productivity, “Shengyu Guo” and “Bing Tang” each authored four papers, underscoring their notable contributions. Normalized citation metrics, adjusting for publication timing, highlight “Amir Mahdiyar,” “Saeed Reza Mohandes,” and “Tarek Zayed” for their impactful work within their respective periods.

International collaboration networks were analyzed using VOSviewer to map the global distribution and influence of construction accident causality research, with a focus on transnational knowledge exchange, policy diffusion, and regional research ecosystems, as shown in Figure 4. A threshold of one publication per country was set, including 33 countries. Nodes represent countries, with link thickness indicating collaboration strength. Quantitative indicators for the top nine countries by publication count are presented in Table 2. China, the United States, Australia, and South Korea lead in output and citations, reflecting robust research capabilities. High citation totals are noted for China, the United States, England, and Australia, with recent contributions from Malaysia, Poland, China, and Greece contrasting with earlier peaks (circa 2017) from the United States, England, and Australia.

Collaboration networks among organizations were examined using VOSviewer, selecting 28 from 158 institutions with a minimum of two publications, as depicted in Figure 5. Nodes signify organizations, their size reflecting publication count, and link thickness indicating collaboration strength. Node color denotes activity level, with yellow indicating higher activity. Huazhong University of Science and Technology and China University of Geosciences exhibit strong inter-organizational ties, while Qingdao University of Technology and Wuhan University are notably active, with South China University of Technology also contributing significantly in recent years. However, international collaboration remains limited. Most institutional ties occur domestically within the same country or region. This indicates that while intra-national partnerships are well-developed—especially in China—transnational institutional collaboration is still in its infancy.

Quantitative indicators for the top eight organizations by publication count are summarized in Table 3, incorporating normalized citation metrics for fair impact comparison across time. Huazhong University of Science and Technology (7 publications, 296 citations), City University of Hong Kong (4 publications, 318 citations), Hong Kong Polytechnic University (4 publications, 148 citations), and China University of Geosciences (4 publications, 94 citations) stand out for productivity and influence.

Various research methods for data collection, accident pattern analysis, and causality extraction are employed, each with distinct advantages and limitations. Multiple approaches are often combined to enhance result reliability and comprehensiveness. Traditional methods—surveys, literature reviews, expert interviews, site inspections, and causality extraction from existing data and case studies—are systematically reviewed, with specific steps, limitations, and supporting literature outlined in Table 6 to serve as a reference for future research. As shown in the table, traditional methods for identifying construction accident causation factors exhibit a trade-off between depth of insight and data generalizability. While expert-based and site-specific approaches (e.g., interviews, field investigations) offer detailed and contextualized knowledge, they often suffer from subjectivity and limited coverage. Conversely, methods relying on existing literature or institutional data allow for broader pattern recognition but may be constrained by data quality or research bias.

Advancements in technology and data processing have led to the adoption of sophisticated methods in construction accident causal analysis. Digital tools enable risk factor extraction from unstructured reports (Liu et al., 2024). Techniques such as natural language processing (NLP), deep learning, and statistical modeling, alongside complex system analysis, are utilized to deepen causal insights. NLP and machine learning models (e.g., KeyBERT, BERT) paired with clustering techniques extract key factors from extensive accident reports (Liu et al., 2024), while text mining reveals accident patterns in large datasets (Shen et al., 2024; Luo et al., 2021). Novel NLP, dimensionality reduction, clustering, and large language model (LLM) prompts are integrated to identify accident types and causes (Smetana et al., 2024). Probabilistic models like HFACS, refined with qualitative comparative analysis, and the modified loss causal model establish causal chains and relationships (Li and Wen, 2022; Wang et al., 2022b; Chan et al., 2022). Deep learning approaches, such as the HANN model (Zhang, 2022), handle nonlinear modeling of complex datasets, while system models like ConAC and the Constraint-Response Model uncover intrinsic accident mechanisms (Winge et al., 2019; Behm and Schneller, 2013; Suraji et al., 2001). Fault tree analysis (FTA) combined with Boolean algebra and case studies further refines causal factor identification (Chi et al., 2014). These advanced methods surpass traditional approaches in depth and applicability, with ongoing technological progress promising enhanced accuracy and a stronger foundation for safety management.

Traditional methods, such as surveys, literature reviews, and expert interviews, are valuable for identifying accident causality factors but often face limitations in terms of accuracy and generalizability due to small sample sizes and subjectivity. These methods are typically more interpretable, providing direct insights, but they may overlook underlying patterns that advanced techniques can uncover. In contrast, advanced methods such as natural language processing (NLP) and deep learning offer enhanced accuracy by analyzing large datasets and identifying complex causal relationships. However, these methods often lack transparency, making it difficult for practitioners to fully understand the rationale behind the results. While traditional methods remain crucial for site-specific analysis and expert judgment, advanced methods are increasingly relevant as they enable more comprehensive, data-driven insights. The integration of both approaches can provide a more robust and effective framework for construction accident prevention.

Classification methods for construction accidents are predominantly based on multiple dimensions, such as direct causes, physical manifestations, and regional factors. Accidents are attributed to human unsafe behaviors, unsafe object conditions, environmental factors, and management deficiencies (Shen et al., 2024), or categorized by manifestations like falls from heights, object strikes, and collapses (Yang et al., 2024). With construction companies increasingly operating internationally, heightened safety risks and uncertainties are encountered (Jin et al., 2021). Risks in global projects stem from countries, partners, companies, and project-specific factors (Zhu et al., 2022), with Heinrich’s theory applied to classify accidents into traditional and non-traditional safety risks, environment- and health-related incidents, and socio-cultural conflicts, aiding cross-cultural safety management (Jin et al., 2021). Regional studies highlight distinct risk profiles, offering valuable insights for multinational projects. Innovative classifications, such as work area intersections (high-altitude, ground, or combined cross-working) (Chen et al., 2023) and severity levels (general, major, serious, particularly serious) (Chen N. et al., 2022), refine risk characterization, noting that falls, collapses, strikes, and lifting injuries dominate (∼80% of incidents). Additional categories—entrapment, electrocution, chemical exposure, temperature extremes, fires, explosions, and traffic accidents—are also recognized (Zhang J. et al., 2020). Advanced tools like graph neural networks are employed to classify unstructured reports efficiently (Pan et al., 2022), while the Task-Competence-Interaction model links accident risk to mismatches between task demands and worker capabilities (Kartam and Bouz, 1998). These multidimensional approaches, evolving toward standardization and dynamism, provide robust theoretical support for safety management, with future efforts urged to align classification with practical safety enhancements.

Accident type frequency and proportion in construction vary by research perspective and data source. Falls from heights, object strikes, and collapses consistently predominate, with falls leading across studies (Zakaria et al., 2023; Betsis et al., 2019; Jabbari and Ghorbani, 2016), while electrocution, fires, and explosions, though less frequent, remain significant risks (Rafindadi et al., 2023; Hwang et al., 2023). These patterns inform targeted safety measures. Common accident types are refined based on physical manifestations and prior research, as detailed in Table 7. Scientific classification underpins safety management by linking accident types to specific work environments, operational methods, and management factors, enabling precise prevention strategies. Major categories—falls, strikes, collapses, electrical shocks, and fires/explosions—alongside specific incidents like falling object strikes, traffic accidents, chemical exposure, and temperature extremes, are delineated to enhance risk assessment and training. Future refinements, integrating dynamic analysis and modern technology with occurrence data, are expected to bolster safety management guidance.

Traditional statistical methods are employed as foundational tools for identifying accident causes and exploring simple relationships in construction accident analysis. Their limitations in addressing complex causal interactions are acknowledged, as detailed in Table 8. Combining these with advanced techniques is recommended to achieve more comprehensive insights. This table highlights the complementary nature of traditional statistical methods in causation analysis. Descriptive statistics provide an accessible overview of factor distributions, while correlation analysis enables deeper exploration of pairwise relationships. Meta-analysis serves to integrate findings across studies but may obscure contextual variability. Together, these methods offer a layered understanding of causation patterns, though each carries limitations in handling complex or multifactorial interactions.

Advanced statistical methods are increasingly applied to uncover risk factors and their interrelationships in construction accident causation, enabling the development of robust analysis models for improved prevention and management accuracy. Challenges related to applicability, data needs, and computational complexity are outlined in Table 9. The table reflects a growing tendency toward the use of more sophisticated analytical techniques in causation studies. These approaches offer expanded capabilities for exploring hidden patterns, complex relationships, and uncertainty within the data. While they hold promise for producing more nuanced insights, their effective application often hinges on appropriate methodological design and data conditions.

Network- and structure-based methods are utilized to reveal complex factor interactions, pinpoint key elements, and trace influence pathways in construction accident analysis. Their strengths in managing complex systems and dynamic interactions are highlighted, though high demands on network construction, data quality, and computation are noted in Table 10.

With digital technology advancements, advanced data analysis methods are leveraged to process large-scale data and uncover intricate relationships in construction accident causation through sophisticated algorithms. These methods support accident prevention by identifying underlying patterns, yet face challenges such as limited interpretability, high costs, and reliance on specialized expertise, as presented in Table 11. The table illustrates a gradual shift toward structural and network-based thinking in causation analysis. These methods help to represent complex interdependencies and visualize cause-effect linkages in more systematic ways. Although promising in theory and structure, their effective use still relies on suitable data inputs and careful model interpretation.

Traditional statistical methods offer strong interpretability and ease of use but struggle to capture the multifactorial and nonlinear relationships often present in construction accidents. In contrast, advanced statistical and network-based approaches demonstrate superior accuracy and structural insight, yet they typically require large datasets and involve complex modeling processes, limiting their practical applicability. Advanced data-driven techniques—such as machine learning and deep learning—excel at handling large-scale data and detecting hidden patterns, but often suffer from limited transparency and high computational and expertise demands. Therefore, a trade-off exists between interpretability and analytical capability across different methods. To address this gap, an integrated analytical framework combining traditional, advanced statistical, and data-driven methods can be adopted. This approach enhances the robustness and comprehensiveness of causal factor analysis by leveraging the strengths of each method, ensuring more accurate and holistic insights into construction accident causation. Collaborative approaches, such as combining statistical and machine learning techniques (Rafindadi et al., 2023; Li and Wen, 2022; Nguyen et al., 2024) or Bayesian networks with SNA, are adopted to enhance accuracy and model propagation paths, despite challenges like computational demands and integration barriers. Methods or combinations are selected based on research goals, data traits, and resources to yield precise, comprehensive causation insights.

Emerging technologies, including big data, artificial intelligence (AI), and the Internet of Things (IoT), are harnessed to support real-time monitoring, accident prevention, and causal analysis in construction, significantly enhancing safety management (Deng et al., 2024). AI and large models are increasingly applied to identify risk factors, analyze causal relationships, and predict risks, offering intelligent, data-driven solutions. Text mining combined with Bayesian networks extracts risk factors from reports using algorithms like TF-IDF and TextRank, with improved Bayesian models identifying key factors despite limitations in semantic and temporal dynamics (Shen et al., 2024). Dynamic Bayesian networks and N-K models are utilized to assess risk coupling in deep excavation near tunnels, aiding on-site safety decisions (Jiang et al., 2022). Metro accident causality networks, built via data mining and network theory, reveal topological links between accidents and factors, optimizing safety resource allocation (Deng et al., 2024). The BERT and TAN model achieves high performance (AUC 0.938) in risk factor extraction (Liu et al., 2024), while GPT-3.5, paired with NLP and clustering, analyzes OSHA data for real-time safety advancements (Smetana et al., 2024). ARM uncovers risk factor combinations—e.g., management, site conditions, and behavior in Malaysia (Rafindadi et al., 2023)—and patterns in cross-operations (Chen et al., 2023) and fall accidents (Luo et al., 2021). Text classification and causal modeling are advanced through Word2Vec and hybrid neural networks (Zhang, 2022), convolutional bidirectional LSTM (C-BiLSTM) for OSHA narratives (Antoniou and Merkouri, 2021), Copula Bayesian networks for collapse accidents (Chen et al., 2024), and NLP with Accimap for systematic risk analysis (Ma and Chen, 2024). Fault tree analysis with Bayesian networks dynamically assesses subway accident risks (Qie and Yan, 2022), and contextual semantic networks (CCNet) enhance classification using deep learning and attention mechanisms (Gupta et al., 2022). Tunnel accident databases highlight geological unpredictability (Sousa and Einstein, 2021), while knowledge graphs integrated with BIM identify high-risk factors like unsafe behavior and poor management (Li et al., 2024). Python-based processing of complex data aligns with deep learning for future intelligent systems (Nowobilski and Hoła, 2023). These technologies collectively bolster analysis efficiency, accuracy, and decision-making in construction accident causation.

While the AI models mentioned show potential in risk factor identification and accident causation analysis, their effectiveness, scalability, and feasibility in real-world applications require further evaluation. Practical implementation depends on data quality, model adaptability, and integration with existing safety management systems. Validation in real-world projects is essential to ensure accuracy and reliability. Scalability remains a concern, especially in resource-limited smaller projects. A comprehensive evaluation of model performance is crucial for optimizing their broader application.

Limitations in data, models, and methods are encountered in current research, as detailed in Table 12. Advances in AI, large models, and data mining offer new opportunities for construction accident causation analysis. Future research directions, outlined in Table 13, emphasize improving risk identification accuracy and intelligence. The integration of these technologies is poised to deliver precise, efficient solutions, though ongoing efforts in data quality, model optimization, and interpretability are required to ensure practical, scalable outcomes for construction safety management. Particularly, the challenges related to data quality, such as incomplete or biased datasets, can significantly impact model performance and generalization across diverse construction contexts.

In addition to technological advancements, it is crucial to consider the ethical and policy implications of AI-based accident prediction models. Issues such as data privacy, algorithmic fairness, and transparency in decision-making should be carefully addressed to ensure that these models are used responsibly and effectively in construction safety management. Furthermore, policy frameworks may need to be updated to support the integration of AI in safety protocols, ensuring that the models are aligned with industry standards and regulatory requirements.