Enhancing Knowledge of Construction Safety: A Semantic Network Analysis Approach

Abstract

The construction industry is recognized as high-risk due to frequent accidents and injuries, prompting extensive research and bibliometric analysis of construction safety. However, little attention has been given to the evolution and interconnections of key research topics in this field. This study applies semantic network analysis (SNA) to examine relationships and trends in construction safety research over the past 30 years. SNA enables quantitative exploration of topic interrelationships that is difficult to achieve with other approaches. Chronological network graphs are evaluated using the number of nodes, edges, density, average clustering coefficient, and average path length. Prominent topics are identified through degree, betweenness, and eigenvector centrality measures. The analysis combines a global overview of the main network, a chronological perspective, and local examination of clusters based on five macro keywords: accident, safety management, worker behavior, machine learning, and safety training. Results show a shift from traditional concerns with mortality and injuries to contemporary issues, such as safety climate, worker behavior, and technological innovations, including building information modeling, machine learning, and real-time monitoring. Topics with lower centrality scores are identified as under-researched. Overall, SNA offers a comprehensive view of the construction safety knowledge system, guiding researchers toward emerging topics and helping practitioners prioritize resources and design integrated safety risk strategies.

1. Introduction

Ensuring the safety of construction workers presents a significant challenge due to the complex and unpredictable nature of construction work. Despite continuous academic research and practical measures to reduce safety risks, the construction sector continues to experience alarmingly high rates of fatalities and injuries. In 2022, the construction industry in the U.S. recorded 1069 occupational fatal injuries [1], while in China, 2020 saw 689 accidents resulting in 794 deaths [2]. The International Labor Organization estimates that, in some countries, 30% of construction workers suffer from back pain or other musculoskeletal diseases [3]. Given the critical importance of safety within the construction industry, extensive research efforts have been undertaken, including bibliographic studies and comprehensive literature reviews. The most prominent of these are an analysis of 977 articles from 2010 to 2016, which examined factors affecting safety performance on construction projects [4], and a systematic review of 753 articles published between 1990 and 2020, which identified structural trends and emerging themes in the domain of construction health management [5].

However, previous literature reviews have primarily relied on qualitative methodologies, such as content analysis and narrative synthesis. This approach offers a limited perspective that prioritizes prominent topics in construction safety yet overlooks the complex interrelationships of various research themes and lacks assessment of their relative importance. Consequently, this qualitative focus falls short in visually and quantitatively identifying emerging topics warranting further investigation. To address these shortcomings, semantic network analysis (SNA) is a robust analytical tool that integrates qualitative and quantitative approaches to develop a comprehensive knowledge map. This enables a more balanced and objective analysis by constructing a domain model through abstracting and mapping a network of entities and their interconnections. This process culminates in creating a network graph that accurately reflects the structural intricacies of the field [6]. Moreover, SNA surpasses other methods in its capability to compute the importance ranking of research topics and measure network structure.

SNA is instrumental in revealing relationships between research literature and associated keywords [7,8] and in elucidating the strength of connections of diverse research objectives [9]. SNA’s metrics, such as centrality, cluster coefficient, and density, play a vital role in identifying research areas that have received significant attention or have been overlooked [10]. Despite its demonstrated effectiveness across various fields, including digital twins [10], entrepreneurship [11], and innovation systems [12], SNA’s application in developing a specialized knowledge base for addressing safety concerns within the construction industry remains limited. This study aims to bridge this gap by utilizing SNA to map the overarching research trajectories and discern annual trends in construction safety. Analyzing 1936 publications spanning from 1991 to 2023, their keywords are examined and a semantic network constructed to provide a comprehensive depiction of the evolution of research on construction safety, thereby illuminating the dynamic landscape of this critical research area. Thus, we propose three research questions in this paper:

• How do the structural characteristics and thematic connections of knowledge networks related to construction safety change over time?
• What knowledge subnetworks form the core focus areas of construction safety research?
• Which research topics are most important within these core focus areas?

2. Background

2.1. Construction Safety Reviews

The construction industry is among the most hazardous sectors, with frequent accidents and numerous risk factors. Studies have explored the safety risks of aging construction workers [13] and developed a vision-based hazard avoidance system to help workers identify hazard sources [14]. Bhandari and Hallowell [15] studied the relationship between safety climate and worker risk-taking behavior in the construction industry. They confirmed that safety climate can reduce organizational risk tolerance and promote more risky decisions at work. Allison et al. [16] proposed an analytical method to reveal the cost of construction accidents in Australia, providing a reference for safety planning. In recent years, the rapid development of artificial intelligence and automation tools, such as virtual reality, robots, and drones, has brought new opportunities to the construction industry [17]. To meet the urgent need to systematically investigate cutting-edge trends in construction safety, previous literature reviews have examined the field from various perspectives, including safety management systems and measurement methodologies [18], safety culture and its implementation on construction sites [19], sensor-based safety management and BIM integration [20], and key factors contributing to accidents and occupational illnesses [21].

While these reviews provide valuable insights, they often overlook the complex interrelations among research topics, their relative significance within the broader knowledge landscape, and their evolution over time. This study addresses these gaps by employing SNA to map topic interconnections, trace their development, and quantitatively assess their prominence. By constructing a semantic network, the analysis clarifies the structure of the construction safety knowledge system and identifies promising directions for future research.

In addition to SNA, several commonly used methods for quantifying construction safety literature exist. Systematic reviews synthesize research evidence [22], traditional bibliometrics quantifies publication and citation patterns and mainly relies on the static perspective of keyword frequency statistics and co-occurrence matrices [23,24], and topic modeling reveals the underlying topic structure [25,26]. These methods quantify saliency and latent topics, while the uniqueness of SNA lies in modeling topic relationships as a complex network and introducing multidimensional indicators, such as topic centrality, network density, and average clustering coefficient. It can simultaneously reveal research hotspots, topic structure, and evolution paths over time and provide more structured explanations through dynamic visualization. In terms of domain applicability, for the rapidly developing field of building safety, SNA is particularly helpful in understanding the flow of ideas and the relationship between concepts, and it complements traditional methods for quantifying construction safety literature [11].

2.2. Semantic Network Analysis

SNA, as defined by Doerfel, applies network analytic techniques to associations rooted in shared meaning, contrasting with behavioral or perceived communication links [27]. In his analysis, Doerfel argued that SNA focuses on word associations in texts, representing inherent meanings within the data [27]. While sharing methods with social network analysis, semantic network analysis emphasizes knowledge representation and semantic connections to clarify concept linkages. For example, Danowski’s message content network analysis research highlighted the concept of co-occurrence, focusing on word proximity [28]. Unlike social network analysis, which uses individuals or organizations as nodes [29,30], SNA uses words or concepts derived from text or knowledge bases. Despite these differences, both analyses find application in information representation, relationship modeling, and web data mining, proving valuable in studies of knowledge corpora [9,31,32]. In this study, keywords function as nodes within the semantic network, with connections (edges) indicated by document co-occurrence, illustrating the strength of relationships. SNA integrates qualitative and quantitative tools to comprehensively analyze thematic trends in research, distinguishing itself from other methodologies that analyze data elements as interconnected components rather than in isolation.

SNA enables the relationships between factors to be studied, facilitating comparisons of their relative importance and identifying areas needing attention. In recent years, SNA has found applications across diverse disciplines, including sustainable development, political science, enterprise management, and media and communication. Lee et al. applied dynamic SNA to explore social sustainability, identifying key research topics by analyzing network central keywords [33]. In political science, Barnett et al. used network metrics and big data to analyze social media posts, revealing public perceptions of international relations [34]. In enterprise management, Chelmis et al. integrated social and semantic network analyses to improve internal and external enterprise information management, enhancing organizational decision-making and communication [35]. In media and communication studies, Bayrakdar et al. reviewed advanced technologies for social media data analysis [36], while KangGJ et al. constructed a semantic network to examine public opinions on vaccination through social media, providing insights to improve vaccination campaigns [37].

Moreover, SNA has been effectively utilized within the building and construction sector, addressing various research themes. For example, Zarei et al. utilized SNA to reveal relationships and rankings of factors contributing to construction project delays, thereby advancing delay analysis and management in engineering projects [6]. Yao et al. investigated knowledge sharing in online building safety education and training, constructing semantic networks for ‘building safety’, ‘building health’, and ‘building accidents’ [38]. Noh et al. utilized SNA to extract keywords and form a semantic network to evaluate residents’ satisfaction with housing restoration, identifying critical dissatisfaction factors [39]. Abdul Nabi et al. applied SNA to the literature on modular construction, creating an extensive network diagram to explore factors influencing modular construction and their interconnections [40].

3. Research Methods

SNA was the primary method for constructing a keyword network to achieve the research objectives [41]. Gephi accepts multiple graph data formats, including the .csv format generated by Excel, which was used here as the input keyword co-occurrence matrix. Gephi 0.10.1 was used to construct an undirected network for SNA analysis. Gephi 0.10.1 is an open-source software package that employs a three-dimensional rendering engine to provide expressive and insightful visualizations for large networks [11]. Keywords were selected from curated papers, and pivotal keywords were identified through a network where edges denote the co-occurrence of keywords within papers. A systematic and multi-phased approach was adopted for data analysis, integrating programming techniques to process data and validate network connections [10,42]. Figure 1 illustrates a comprehensive flowchart detailing the research methodology, encompassing five distinct phases: paper collection, paper filtering, network development, calculation of SNA metrics, and network analysis.

3.1. Paper Collection

A literature sample from the Web of Science database was used, which is recognized as a primary database for assessing scholarly performance [11,23]. The publication timeframe was extended until 2023. Filtering criteria included selecting papers under the topic ‘construction safety’, written in English, and published in specified journals, such as Accident Analysis and Prevention, Advanced Engineering Informatics, Advances in Civil Engineering, and others. The initial search yielded a preliminary dataset comprising 5913 English papers with titles, abstracts, and keywords.

3.2. Paper Filtering

Due to the focus on construction worker safety, a rigorous manual filtering process was implemented to exclude papers outside the research scope. Exclusion criteria targeted papers from industries other than construction [43,44], those addressing structural safety rather than worker safety [45,46], discussions unrelated to occupational safety [47,48], and entirely irrelevant content. Additionally, 568 papers lacking author-provided keywords were removed [49,50], as these keywords are critical for conducting SNA. Figure 2 illustrates the complete filtering process. Following this thorough selection process, 1936 papers from 1991 to 2023 were curated to develop the semantic network in the subsequent phase.

3.3. Network Development

The development of the semantic network proceeded through the following steps:

• Processing of author keywords: (1) A total of 10,086 keywords were retrieved, and duplicates were merged, resulting in 5607 unique keywords. During data cleaning, common spelling differences were manually corrected using OpenRefine. (2) Abbreviations were expanded to full forms (e.g., VR to ‘virtual reality’, BIM to ‘building information modeling’, etc.). (3) All uppercase letters were converted to lowercase. (4) Plural words were singularized. (5) Frequencies were identified and calculated for 4214 unique keywords.
• Identification and implementation of macro keywords: This step involved iteratively comparing the 4214 keywords to consolidate similar terms into broader categories known as macro keywords. These macro keywords were derived from groupings commonly used in bibliometric construction safety studies, as cited in references [23,25,51,52,53,54,55,56,57]. To minimize subjectivity, the authors adhered strictly to the original text when consolidating keywords following a systematic approach. Initially, keywords sharing common terms were considered together, and after comparing their similarities and differences, they were grouped under the keyword with the highest frequency. Keywords that remained unclassified underwent further scrutiny. In cases of disagreement, the two authors of the paper resolved discrepancies based on existing categorizations in the literature, aiming to achieve consensus and align perspectives. If consensus was not reached after initial discussions, all four authors reviewed and discussed the remaining discrepancies to reach a collective agreement. This ensured traceability throughout the analytical process. The aim was to consolidate keywords based on existing categorizations wherever possible. For example, keywords such as ‘fall’, ‘coincidental fall’, ‘fall accident’, ‘fall hazard’, and ‘cave in and fall hazard’ were grouped under the macro keyword ‘fall’ to encompass the concept of fall-related risks in the construction industry. Through this method, a total of 1530 macro keywords were identified.
• Shortlisting and encoding of final macro keywords: The occurrence frequency of the 1530 macro keywords within the literature dataset was evaluated. Macro keywords appearing less than ten times were strategically excluded. As such, infrequently occurring macro keywords typically have a lower degree of connectivity and tend to be isolated within the network. This frequency threshold approach has been validated in previous studies for its practicality [11,58]. Additionally, to maintain focus on specific subtopics related to safety in the construction industry, five globally salient words—‘health and safety’, ‘construction safety’, ‘construction’, ‘safety’, and ‘construction industry’—were intentionally omitted from the list. Following this refinement, a concise list of 160 macro keywords was compiled, with the top 40 macro keywords detailed in

  Table 1. Encoding of the top 40 macro keywords in frequency ranking.

  Code	Macro Keywords	Frequency
  K1	accident	254
  K2	machine learning	224
  K3	worker behavior	218
  K4	safety management	204
  K5	safety training	148
  K6	safety climate	131
  K7	injury	125
  K8	real-time technology	119
  K9	hazard identification	106
  K10	ergonomics	102
  K11	fall	101
  K12	risk assessment	98
  K13	accident prevention	96
  K14	safety performance	81
  K15	labor and personnel issues	79
  K16	design	77
  K17	object tracking	76
  K18	building information model	68
  K19	computer vision	65
  K20	risk	64
  K21	safety culture	63
  K22	disease	62
  K23	VR/AR	62
  K24	construction site	60
  K25	musculoskeletal disorder	60
  K26	construction worker	58
  K27	knowledge management	58
  K28	stress	54
  K29	hazard	52
  K30	autonomous safety monitor	51
  K31	structural equation model	51
  K32	risk management	50
  K33	construction management	46
  K34	survey	46
  K35	construction equipment	40
  K36	intervention	39
  K37	risk perception	39
  K38	workplace safety	38
  K39	mental health	37
  K40	communication	37

  . For instance, K1, with a frequency of 254, indicates that various keywords related to ‘accident’ (e.g., accident (69), construction accident (21), occupational accident (20), etc.) were grouped under the macro keyword ‘accident’. The cumulative frequency of these keywords, totaling 254, represents the frequency of the macro keyword ‘accident’ (K1).
• Generation of adjacency matrix Aij: A reference matrix, Rij (160 × 1936), was retrieved to indicate the presence of keywords in literature, where Rij = 1 if keyword i exists in literature j and Rij = 0 otherwise. The adjacency matrix Aij (160 × 160) was computed by multiplying Rij with its transpose and setting the diagonal elements of the resultant matrix to 0. This process yielded a symmetric matrix where both rows and columns represent keywords, and the cells denote the frequency of keyword co-occurrences in sourced papers. Figure 3 illustrates the generation of the adjacency matrix Aij using the reference matrix Rij.
• Visualization of the network: The adjacency matrix Aij was imported into Gephi software for network visualization and layout adjustment.

3.4. SNA Metrics

Applying SNA metrics enables a quick understanding of the network’s distribution patterns and competitive dynamics. Key SNA metrics include degree centrality, betweenness centrality, eigenvector centrality, network density, average clustering coefficient, and average path length. Each metric provides distinct insights into the structure and influence within the network, as explained in detail in

Table 2. Overview of relevant SNA metrics.

Target	Metrics	Measures	Equation	Variable Definition	Interpretations
Network node	Degree centrality [10]	The number of edges associated with vertex ‘v’ becomes the degree centrality of vertex ‘v’.	$C_{di}={\displaystyle \sum _j}{A}_{ij}$	$C_{di}:$ the degree centrality of node $i$; $j:$ neighboring nodes of $i$; $A_{ij}:$ adjacency matrix.	Nodes with high degree centrality are the most directly connected nodes in the network.
Network node	Betweenness centrality [42]	A metric that characterizes node importance based on the number of shortest paths passing through a node.	$C_{bk}={\displaystyle \sum _{k,i,j\in G}}{\displaystyle \frac{g_{ij}\left(k\right)}{g_{ij}}},i\ne j\ne k$	$C_{bk}:$ betweenness centrality of node $k$; $g_{ij}\left(k\right):$ number of paths that pass through node $k$; $g_{ij}:$ total number of shortest paths from node $i$ to $j$.	Nodes with high betweenness centrality act as critical bridges or bottlenecks within the network, influencing the flow of information or resources.
Network node	Eigenvector centrality [10]	The importance of a node depends not only on the number of its neighboring nodes (i.e., its degree of centrality) but also on the importance of its neighboring nodes.	$\lambda x=Ax$	$A:$ adjacency matrix; $\lambda :$ largest eigenvalue of $A$; $x:$ eigenvector of $A$ corresponding to $\lambda$, the eigenvector centrality vector [42].	Nodes with high eigenvector centrality are those connected to other nodes that are themselves highly connected or important, thus capturing a more global sense of influence.
Entire network	Density [41]	The ratio of the network’s actual number of edges M to the maximum possible number of edges.	$D\left(G\right)={\displaystyle \frac{W}{W_{max}}}$	$D\left(G\right):$ the density of the network graph $G$; $W:$ the actual number of weighted edges in the network; $W_{max}:$ the number of all possible weighted edges in the network.	A higher density means a more connected network with a larger fraction of all possible edges. A density of 1 indicates a fully connected network (a complete graph), while a density close to 0 indicates a sparse network.
Entire network	Average clustering coefficient [41]	Measures the clustering performance of a node by calculating the ratio of the number of edges between its neighboring nodes to the possible number of edges.	$C\left(G\right)={\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N}{C}_i={\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N}{\displaystyle \frac{2E_i}{k_i\left(k_i-1\right)}}$	$C_i:$ cluster coefficient of node $i$; $E_i:$ the number of edges between neighboring nodes of node $i$; $k_i:$ the degree (neighbors) of node $i$.	A higher average clustering coefficient indicates a greater tendency for nodes to form tightly knit groups (triangles). It reflects the likelihood that the neighbors of a node are also connected.
Entire network	Average path length [41]	The total number of edges on the shortest path connecting two nodes.	$L\left(G\right)={\displaystyle \frac{{\sum }_{i,j\in G}{l}_{ij}}{N}}$	$l_{ij}:$ path length between any node $i$ and node $j$; $N:$ the total number of all possible edges in the network.	A lower average path length indicates a more efficient network, where any two nodes can be reached with fewer steps, facilitating faster communication or transfer of resources.

. These metric values were computed using Gephi software to comprehensively analyze the network’s characteristics.

3.5. Network Analysis

The network analysis involved creating a construction safety knowledge network diagram that evolved. This was achieved by illustrating the network’s small-world characteristics, characterized by specific structural properties balancing a high average clustering coefficient and short average path length. These traits indicate a robustly interconnected network for the respective year, confirming that the network is real rather than randomly generated, thereby verifying the reliability of the results.

Additionally, the network analysis incorporated both global and local descriptive approaches. The global analysis examined all cited literature and macro keywords, providing a comprehensive field overview. In contrast, the local descriptive analysis focused on the global network’s five most significant macro keywords. This analysis identified and isolated these primary macro keywords using an indicator-based ranking system. Subsequently, subnetworks centered around these macro keywords were extracted, enabling a detailed exploration of five key domains.

Significantly, some papers within a subnetwork (e.g., accident) also appeared in other network clusters, highlighting overlapping research communities within the global network. This overlap suggests that the global network comprises interconnected communities, with subnetworks representing specific but interconnected research areas [10]. Therefore, the local descriptive analysis delineated unique cluster attributes and integrated their commonalities, enriching the understanding of the field’s knowledge landscape.

4. Results

4.1. Chronology of Papers

To track the evolution of construction safety research, the analyzed papers were grouped by publication year. Figure 4 shows annual publication volumes in international journals from 1991 to 2023. Output rose gradually in the early years, accelerated from the mid-2000s, and peaked in 2020.

4.2. Time Evolution Analysis

The growth and structural changes of the knowledge network were examined using chronological semantic graphs (Figure 5). Each graph represents one year, with macro keywords shown as labelled nodes. Node size corresponds to degree centrality, and edge thickness reflects co-occurrence strength. For example, in 1996, ‘hazard’ (degree = 6) appears larger than ‘disease’ (degree = 4). The edge between ‘hazard’ and ‘injury’ is thicker than that between ‘hazard’ and ‘disease’ because the first pair co-occurred in two papers, compared with one for the second.

Table 3. Network index for each year.

Year	Number of Nodes	Number of Edges	Density	Average Clustering Coefficient	Average Path Length
1991	2	1	1	——	1
1992	2	1	1	——	1
1993	7	13	0.619	0.924	1.381
1994	6	3	0.200	——	1
1995	4	6	1	1	1
1996	13	20	0.256	0.870	1.469
1997	15	27	0.257	0.829	1.571
1998	14	17	0.187	0.885	1.675
1999	9	9	0.250	1	1
2000	4	4	0.667	0.778	1.333
2001	10	8	0.178	1	1
2002	27	47	0.134	0.803	2.516
2003	31	56	0.12	0.884	1.969
2004	31	64	0.138	0.901	1.919
2005	16	17	0.142	0.730	2.173
2006	21	35	0.167	0.885	1.760
2007	9	11	0.306	0.833	1.312
2008	39	54	0.073	0.587	3.415
2009	52	101	0.076	0.778	2.891
2010	74	166	0.061	0.732	2.828
2011	45	60	0.061	0.828	2.062
2012	80	168	0.053	0.689	3.244
2013	98	309	0.065	0.616	2.833
2014	97	211	0.045	0.697	3.554
2015	103	324	0.062	0.657	2.849
2016	101	372	0.074	0.630	2.626
2017	116	401	0.060	0.633	2.595
2018	113	342	0.054	0.596	2.866
2019	129	468	0.057	0.538	2.646
2020	131	554	0.065	0.564	2.564
2021	121	510	0.070	0.514	2.550
2022	134	574	0.064	0.558	2.621
2023	122	453	0.061	0.557	2.634

summarizes the main characteristics of these yearly networks. Between 1991 and 1999, networks were sparse, with few nodes and weak connections. From 2000 to 2008, they became denser, and node sizes varied more, indicating topic differentiation. After 2009, the network expanded rapidly: more nodes and edges emerged, but density fell, reflecting a broader range of themes. This structural resemblance to real-world networks supports the small-world network theory [59], emphasizing the concentrated ties of main keywords and clusters within the construction safety knowledge domain.

Network density dropped from 1.0 in 1991, 1992, and 1995 to around 0.3 in 2007 and below 0.1 after 2008. This is typical of large-scale networks where many keywords are indirectly linked. Despite this, high clustering coefficients (>0.5) and short path lengths indicate a small-world structure, with keywords forming tightly connected clusters. Even with a 60-fold increase in node count from 1991 to 2023, average path length increased by only 1.634, suggesting that influential bridging keywords link otherwise distant themes.

4.3. Main Semantic Network

Figure 6 contains 160 macro keywords.

Table 4. Macro keywords with top 20 centrality scores in the main semantic network.

Macro Keyword	Degree Centrality	Macro Keyword	Betweenness Centrality	Macro Keyword	Eigenvector Centrality
accident	126	accident	818.386	accident	1.000
safety management	120	safety management	656.125	safety management	0.979
worker behavior	102	worker behavior	457.737	machine learning	0.880
machine learning	102	machine learning	391.738	worker behavior	0.858
safety training	95	safety training	362.064	safety training	0.819
labor and personnel issues	89	injury	266.737	labor and personnel issues	0.812
accident prevention	85	labor and personnel issues	261.320	accident prevention	0.781
risk assessment	82	risk assessment	239.381	risk assessment	0.751
injury	80	accident prevention	229.454	real-time technology	0.720
real-time technology	78	risk	225.635	fall	0.707
fall	77	construction worker	214.085	hazard identification	0.693
risk	76	hazard	198.572	construction site	0.690
hazard identification	74	fall	195.993	injury	0.685
construction worker	73	safety climate	183.144	risk	0.674
safety climate	73	hazard identification	180.016	hazard	0.655
hazard	71	real-time technology	175.751	safety climate	0.651
construction site	71	ergonomics	170.379	construction worker	0.642
ergonomics	70	construction site	149.530	ergonomics	0.635
building information model	67	safety performance	147.586	construction management	0.634
construction management	67	building information model	140.563	building information model	0.630

ranks the top 20 by degree, betweenness, and eigenvector centrality. Five topics—‘accident’, ‘safety management’, ‘worker behavior’, ‘machine learning’, and ‘safety training’—appear at the top in all metrics. ‘Machine learning’ ranks above ‘safety training’ in degree and eigenvector centrality, showing broader connections. In contrast, ‘worker behavior’ ranks higher in betweenness, highlighting its role as a bridge between research domains.

4.4. Subnetworks of Core Macro Keywords

Subnetworks centered around the five core macro keywords—‘accident’, ‘safety management’, ‘worker behavior’, ‘machine learning’, and ‘safety training’—provide a detailed framework for analyzing the construction safety knowledge system. Clusters were identified using a heuristic approach [10]. First, papers containing each core keyword were selected. Subnetworks were then constructed based on the co-occurrence relationships of keywords within the selected papers. This approach allows for focused exploration of both global and local structures in the construction safety research landscape.

4.4.1. Accident Cluster

The ‘accident’ cluster contains 121 macro keywords across 231 papers (Figure 7,

Table 5. Macro keywords with top 15 centrality scores in the macro keyword ‘accident’ subnetwork.

Macro Keyword	Degree Centrality	Macro Keyword	Betweenness Centrality	Macro Keyword	Eigenvector Centrality
safety management	42	safety management	2091.969	safety management	1.000
injury	25	injury	789.980	machine learning	0.712
machine learning	24	machine learning	738.797	injury	0.548
risk	20	labor and personnel issues	588.881	risk	0.523
safety training	18	hazard	524.711	labor and personnel issues	0.507
labor and personnel issues	18	risk management	505.678	safety training	0.500
risk management	17	risk	478.088	risk assessment	0.492
risk assessment	16	cause 1	456.271	risk management	0.481
fall	15	safety strategy 1	421.714	data analysis 1	0.474
accident prevention	14	construction management	414.135	prediction 1	0.459
construction worker	14	safety training	381.957	fall	0.431
cause 1	14	fall	354.196	hazard identification	0.430
object tracking	14	risk assessment	343.431	object tracking	0.415
safety performance	13	safety factor 1	333.474	safety performance	0.409
hazard	13	construction site	293.886	accident prevention	0.403

¹ Note: macro keywords ‘data analysis’ (K42), ‘cause’ (K46), ‘prediction’ (K64), ‘safety factor’ (K69), and ‘safety strategy’ (K120) were manually added as they are not shown in Table 1.

). Across all centrality metrics, ‘safety management’, ‘injury’, and ‘machine learning’ consistently emerge as the most central topics. Notably, ‘injury’ rises significantly in local centrality rankings compared to global values (degree centrality ranked three locally versus nine globally), reflecting a research emphasis on injury prevention and assessment of accident severity.

Degree centrality highlights topics with many direct connections to others, showing that ‘safety management’, ‘injury’, and ‘machine learning’ are heavily linked within the cluster. Betweenness centrality identifies topics acting as bridges between other keywords. In this cluster, the three central topics serve as connectors linking methodological advances to applied safety issues. Eigenvector centrality measures the influence of topics through connections to other important nodes. High eigenvector scores indicate that these topics co-occur with other central themes, helping shape the broader research agenda. Together, these metrics demonstrate a concentrated, interconnected emphasis on injury prevention using data-driven approaches.

4.4.2. Safety Management Cluster

The ‘safety management’ cluster encompasses 120 macro keywords across 198 papers (Figure 8,

Table 6. Macro keywords with top 15 centrality scores in the subnetwork of the macro keyword ‘safety management’.

Macro Keyword	Degree Centrality	Macro Keyword	Betweenness Centrality	Macro Keyword	Eigenvector Centrality
accident	42	accident	1666.769	accident	1.000
construction site	29	construction site	977.336	labor and personnel issues	0.850
worker behavior	28	worker behavior	670.280	accident prevention	0.838
labor and personnel issues	24	safety climate	472.857	risk management	0.805
accident prevention	23	hazard identification	427.036	hazard identification	0.797
hazard identification	23	labor and personnel issues	411.278	worker behavior	0.760
safety culture	22	risk management	401.861	safety climate	0.754
safety climate	22	safety culture	389.859	safety culture	0.722
risk management	22	machine learning	387.949	machine learning	0.713
machine learning	22	accident prevention	378.816	construction site	0.687
safety performance	18	building information model	350.569	safety training	0.681
safety training	17	risk assessment	273.848	risk perception	0.659
hazard	17	safety performance	273.847	hazard	0.650
risk assessment	17	Hong Kong 1	271.289	risk assessment	0.567
real-time technology	17	construction project 1	235.716	safety performance	0.555

¹ Note: macro keywords ‘construction project’ (K49) and ‘Hong Kong’ (K70) were manually added as they are not shown in Table 1.

). ‘Accident’ consistently ranks as the most interconnected keyword in all centrality measures. Its local betweenness centrality (1666.769) is more than twice the global value (818.386), highlighting its central role as a major connector within this thematic cluster.

‘Construction site’ ranks second in both degree and betweenness centrality. This indicates it is widely co-mentioned across studies and serves as a routing hub linking diverse subthemes. In contrast, ‘labor and personnel issues’ ranks second in eigenvector centrality but lower in degree and betweenness, suggesting that it preferentially connects to already influential keywords rather than forming multiple direct links or acting as a bridge. These patterns highlight the complementary roles of site-specific hazards and workforce-focused issues, demonstrating how different mechanisms shape the cluster’s research agenda.

4.4.3. Worker Behavior Cluster

The ‘worker behavior’ cluster includes 209 macro keywords across 209 papers (Figure 9,

Table 7. Macro keywords with top 15 centrality scores in the subnetwork of the macro keyword ‘worker behavior’.

Macro Keyword	Degree Centrality	Macro Keyword	Betweenness Centrality	Macro Keyword	Eigenvector Centrality
safety climate	30	safety management	929.846	safety climate	1.000
safety management	28	safety climate	825.261	safety management	0.985
structural equation model	21	structural equation model	509.591	construction site	0.777
machine learning	19	labor and personnel issues	412.665	structural equation model	0.674
construction site	18	safety culture	405.711	machine learning	0.664
real-time technology	17	safety training	377.034	risk perception	0.583
safety training	15	machine learning	320.423	real-time technology	0.536
risk perception	15	construction site	318.111	knowledge management	0.533
labor and personnel issues	14	real-time technology	307.959	risk assessment	0.524
risk assessment	14	risk perception	236.732	labor and personnel issues	0.498
safety culture	13	risk assessment	215.039	computer vision	0.493
knowledge management	12	knowledge management	206.241	human factor 1	0.435
computer vision	12	theory of planned behavior 1	157.891	accident prevention	0.431
intervention	11	qualitative research 1	150.761	intervention	0.419
theory of planned behavior 1	11	association rule 1	150.004	theory of planned behavior 1	0.407

¹ Note: macro keywords ‘qualitative research’ (K51), ‘human factor’ (K75), ‘association rule’ (K125), and ‘theory of planned behavior’ (K149) were manually added as they are not shown in Table 1.

). ‘Safety climate’ and ‘safety management’ emerge as the most central topics. Safety climate exhibits higher local centrality than global, indicating that the ‘worker behavior’ cluster places greater emphasis on this topic than other clusters. Its top degree and eigenvector centrality values show that it co-occurs widely with many subtopics and connects to other influential topics, anchoring worker-focused research. Safety management ranks highest in betweenness but second in degree and eigenvector, indicating that it primarily functions as a cross-cutting connector linking methodological and applied topics within the cluster.

4.4.4. Machine Learning Cluster

The ‘machine learning’ cluster contains 100 macro keywords across 179 papers (Figure 10,

Table 8. Macro keywords with top 15 centrality scores in the macro keyword ‘machine learning’ subnetwork.

Macro Keyword	Degree Centrality	Macro Keyword	Betweenness Centrality	Macro Keyword	Eigenvector Centrality
computer vision	25	real-time technology	631.737	real-time technology	1.000
real-time technology	25	accident	546.671	computer vision	0.972
accident	24	computer vision	489.718	autonomous safety monitor	0.919
autonomous safety monitor	22	worker behavior	459.635	accident	0.904
safety management	22	safety management	444.153	risk assessment	0.890
risk assessment	20	risk assessment	428.928	safety management	0.857
worker behavior	19	autonomous safety monitor	342.632	fall	0.795
ergonomics	19	data analysis 1	295.788	ergonomics	0.757
data analysis 1	18	ergonomics	251.690	data analysis 1	0.732
fall	16	text mining 1	247.238	worker behavior	0.654
object tracking	14	fall	209.578	construction site	0.608
prediction 1	13	risk	203.479	activity recognition 1	0.594
text mining 1	13	safety compliance checking 1	197.548	unmanned aerial system 1	0.560
activity recognition 1	13	hazard identification	156.686	object tracking	0.550
risk	12	object tracking	153.705	prediction	0.544

¹ Note: macro keywords ‘data analysis’ (K42), ‘unmanned aerial system’ (K63), ‘prediction’ (K64), ‘activity recognition’ (K65), ‘safety compliance checking’ (K77), and ‘text mining’ (K106) were manually added as they are not shown in Table 1.

). ‘Computer vision’ and ‘real-time technology’ are the most central topics in terms of degree centrality. However, ‘real-time technology’ has higher betweenness centrality than ‘computer vision’ and ‘accident’, serving as a key conduit connecting otherwise separate topics. Notably, ‘computer vision’ does not directly link to ‘accident’; the connection is facilitated by ‘real-time technology’. Degree centrality indicates that both ‘computer vision’ and ‘real-time technology’ (degree = 25) are locally well connected, reflecting frequent co-occurrence in studies. Overall, ‘real-time technology’ functions as both an anchor and a mediator for applying computer vision techniques to accident-related research.

4.4.5. Safety Training Cluster

The ‘safety training’ cluster comprises 90 macro keywords across 123 papers (Figure 11,

Table 9. Macro keywords with top 15 centrality scores in the subnetwork of the macro keyword ‘safety training’.

Macro Keyword	Degree Centrality	Macro Keyword	Betweenness Centrality	Macro Keyword	Eigenvector Centrality
hazard identification	24	VR/AR	699.044	safety management	1.000
VR/AR	21	accident	693.441	safety culture	1.000
accident	18	hazard identification	688.628	hazard identification	0.993
safety culture	17	worker behavior	555.074	labor and personnel issues	0.919
accident prevention	17	visualization 1	508.357	hazard	0.916
safety management	17	accident prevention	429.174	accident prevention	0.858
building information model	16	hazard	428.892	accident	0.855
hazard	15	building information model	330.839	construction management	0.847
worker behavior	15	safety culture	266.296	risk perception	0.833
object tracking	15	case study 1	252.000	VR/AR	0.794
construction management	14	human factor 1	220.462	risk	0.696
labor and personnel issues	13	object tracking	196.530	personnel management 1	0.696
risk perception	12	safety management	196.344	risk management	0.696
knowledge management	11	injury	173.536	building information model	0.604
risk	9	fall	170.000	object tracking	0.602

¹ Note: macro keywords ‘case study’ (K43), ‘human factor’ (K75), ‘visualization’ (K76), and ‘personnel management’ (K134) were manually added as they are not shown in Table 1.

). ‘VR/AR’ ranks first in betweenness centrality within the cluster, highlighting its emerging role in worker safety training. However, it is not prominent in the global network (not listed in Table 4). Conversely, ‘hazard identification’ has higher centrality locally than globally (betweenness 688.628 versus 180.016), underscoring its importance in the context of training. VR/AR’s interactive and immersive features enhance workers’ hazard recognition. Its low eigenvector centrality (0.794) indicates that these connections are primarily direct, rather than linking to the core discourse (e.g., safety culture). Together, these patterns suggest that VR/AR is an emerging, bridging innovation in construction safety training, with potential long-term impact.

4.4.6. Structural Characteristics of Subnetworks

Table 10. Main network and subnetwork metrics.

	Number of Nodes	Number of Edges	Density	Average Clustering Coefficient	Average Path Length
Main network	160	2982	0.234	0.481	1.771
Accident	121	398	0.055	0.533	2.799
Safety management	120	443	0.062	0.540	2.694
Worker behavior	100	302	0.061	0.603	2.689
Machine learning	100	367	0.074	0.534	2.601
Safety training	90	262	0.065	0.693	2.955

shows that all subnetworks exhibit low density, indicating sparse connections among most keywords. Nevertheless, high clustering coefficients and low average path lengths indicate small-world characteristics. The ‘safety training’ cluster, in particular, shows the highest clustering coefficient, reflecting tightly knit thematic groups. Despite segmentation among clusters, interconnections persist, revealing a structured, non-random organization of the knowledge system.

Overall, approximately 75%, 75%, 63%, 63%, and 56% of nodes in the main network are contained within the subnetworks of ‘accident’, ‘safety management’, ‘worker behavior’, ‘machine learning’, and ‘safety training’, respectively. These results emphasize the pivotal role of the core macro keywords in shaping the construction safety knowledge domain. Overlapping keywords such as ‘machine learning’ appear in multiple subnetworks, providing diverse analytical perspectives and enriching understanding of the field through multifaceted analyses.

5. Discussion

5.1. Global Descriptive Analyses

Over the past three decades, the knowledge network in construction safety has evolved into a complex and expansive structure, reflecting growing interest and advancements in the field [60]. This evolution is marked by the adoption of new research methodologies and emerging influential macro keywords, which have integrated diverse nodes into a cohesive network characterized by high clustering coefficients and thematic coherence. Analysis of network indicators from Figure 5 and Table 3 reveals a significant increase in the number of nodes and a decrease in network density over the last 15 years. These trends indicate a robust and potentially expansive growth trajectory in construction safety knowledge. In terms of research focus, studies over the past 30 years have shifted from traditional concerns, such as mortality and injuries, to contemporary topics, such as safety climate, worker behavior, and technological advancements. These shifts reflect broader societal and technological changes influencing research priorities within the construction safety domain. Key milestones in this progression include the following:

• 1991–2003: The primary focus was on mortality [61], injuries [62], common incident types (e.g., musculoskeletal disorders [63]), and specific trades (e.g., carpenters [64] and electricians [65]).
• 2004: The spotlight turned to Hispanic and migrant workers [66], indicating a growing awareness of how demographic factors (e.g., language and culture) influence construction safety performance.
• 2007: Research began emphasizing safety climate and worker behavior [67].
• 2008: Attention was drawn to the residential sector [68], suggesting that the nature of different construction sectors (e.g., residential vs. heavy civil) affects safety.
• 2013: Object tracking and real-time technologies [69] emerged as significant focus areas.
• 2014: Labor and personnel issues gained substantial attention [70].
• 2015: Building information modeling became a key area of interest [71].
• 2018: Machine learning started to play a significant role in research [72].
• 2019: Computer vision enriched the application of artificial intelligence in construction safety [73].
• 2023: Compared to previous years, the centrality of machine learning significantly improved, becoming the most frequent macro keyword. Mental health became a key area of interest [74].

This timeline reflects an evolving understanding of construction safety, incorporating broader demographic considerations, the impact of organizational and group-level safety climate and culture, and technological advancements to enhance safety outcomes.

In the main network, five pivotal macro keywords—‘accident’ (K1), ‘safety management’ (K4), ‘worker behavior’ (K3), ‘machine learning’ (K2), and ‘safety training’ (K5)—emerge as central influencers, consistently holding critical positions across various centrality metrics. Their significance stems from their integral roles in advancing construction safety objectives. ‘Safety management’ (K4) and ‘safety training’ (K5) are pivotal in accident prevention, with ‘worker behavior’ (K3) playing a key role in these strategies. ‘Machine learning’ (K2), widely applied in diverse construction domains, is a key enabler of intelligent and digital advancements. These macro keywords are intricately interconnected with other significant terms, highlighting their fundamental roles. For instance, ‘risk assessment’ (K12) maintains robust connections with all five core macro keywords and ranks prominently in centrality metrics, underscoring its important function in engineering safety. It serves as a pivotal node, linking disparate themes, such as the integration of ‘machine learning’ (K2), with ‘real-time data’ (K8) for enhanced risk evaluation methodologies. Similarly, the strong association between ‘safety climate’ (K6) and ‘worker behavior’ (K3) signifies their intertwined nature, supported by extensive research demonstrating safety climate’s profound impact on construction workers’ behaviors [15,67]. These connections enhance understanding within the construction safety domain and pave the way for interdisciplinary research avenues and innovative collaborations.

5.2. Local Descriptive Analyses

The initial subnetwork centered around the macro keyword ‘accident’ (K1) includes 121 macro keywords, with Table 5 detailing centrality measures within this cluster. This subnetwork highlights the intersection of ‘safety management’ (K4) and engineering ‘accidents’ (K1), emphasizing ‘injuries’ (K7) sustained by workers, innovative applications of ‘machine learning’ (K2) in accident analysis, and ‘risk management’ (K20, K32) associated with ‘labor and personnel issues’ (K15). Reason’s ‘Swiss Cheese Model’ emphasizes that accidents occur only when the ‘holes’ in multiple layers of defense (management measures, technical measures, and individual behavior) align [74]. In the subnetwork, ‘safety management’ (K4) and ‘engineering accidents’ (K1) represent the first line of defense at the management level, ‘labor and personnel issues’ (K15) correspond to the third line of defense at the on-site operation level, while technical approaches, such as ‘machine learning’ (K2), can be regarded as the second line of defense, leveraging big data and intelligent analysis to identify potential hazards promptly. Research has explored various contributing factors in the domain of injuries, such as noise and contact accidents, impacting worker health. Heinrich’s domino theory posits that accidents result from a series of ‘precursor events’—unsafe acts and unsafe conditions—that evolve step by step. In our subnetwork, these unsafe factors are the ‘dominoes’ that can be targeted for reinforced intervention before an accident occurs [75]. Hinze’s [76] analysis highlights prevalent non-fatal injuries, including tears, lumbar spine issues, and eye injuries. Accident analysis within construction safety management has investigated multiple causal factors, such as falls, fatigue, traffic incidents, exposure to harmful substances, and contact or collision incidents. Falls are particularly prominent as a primary cause, being extensively documented in the field [77,78,79]. Accident analysis is integral to risk management, significantly influencing project cost and schedule risks [16,80]. The industry’s emphasis on proactive safety management has driven the adoption of diverse accident analysis methodologies, from interviews [81] and surveys [82] to advanced techniques, such as neural network predictions [83], regression models [84], and construction accident causation models [85]. Recent advancements in natural language processing [86] and big data analysis [87] have expanded the scope of this research area. However, certain topics within the accident subnetwork, such as ‘electrical safety’ (K71), ‘worker compensation’ (K107), ‘diseases’ (K22), and ‘simulation techniques’ (K53), show limited connectivity, suggesting opportunities for further exploration. For example, focusing on ‘worker compensation’ (K107) in relation to fall accidents indicates potential for broader research into other accident types, which can enhance the understanding and management of construction-related accidents and safety.

The second subnetwork focuses on ‘safety management’ (K4), a critical theme in construction safety discourse. Figure 8 illustrates the keyword network, while Table 6 details centrality scores. As can be seen, ‘accident’ (K1) emerges with the highest centrality score, highlighting its pivotal role in safety management discussions, particularly concerning ‘construction site’ (K24) safety and ‘worker behavior’ (K3). Research within this subnetwork has used real-time on-site monitoring data to advance ‘machine learning’ (K2), ‘accident prevention’ (K13), and ‘hazard identification’ (K9). For instance, Guo et al. [88] applied Bayesian networks to analyze unsafe worker behavior dynamics, while Tran et al. [89] utilized 4D BIM for enhanced hazard identification through spatiotemporal analysis. The integration of cutting-edge technologies, including the Internet of Things and machine learning, has significantly driven the evolution of safety management systems. Resilience engineering theory emphasizes an organization’s adaptability and recovery capability in dynamic and complex environments. With real-time monitoring and IoT data from construction sites, organizations can employ adaptive scheduling to respond to fluctuations in worker behavior and emerging safety hazards [90]. ‘Safety climate’ (K6) holds substantial centrality, reflecting worker perceptions influencing organizational development [91]. Efforts to enhance safety climate and productivity have been a priority, with methodologies such as the 6S system [92], international comparisons of safety climate improvement systems [93], and the application of safety climate data in ANN for predicting worker behavior [94]. Over the past 15 years, safety management has become integral to reducing construction costs, expanding into areas such as management accounting systems [95], and addressing challenges such as managing aging employees [13]. The socio-technical systems theory holds that safety performance depends on the coordinated optimization of technical and social systems (people, organizations, and culture). Studying how technologies such as 4D BIM (technical subsystems) can be integrated with construction team incentive mechanisms and safety culture (social subsystems) can enable a genuine upgrade in safety management [96]. Areas with lower centrality scores, such as ‘lean construction’ (K123) linked solely to ‘risk assessment’ (K12) and ‘regression analysis’ (K128) associated with ‘migrant workers’ (K85), suggest potential research avenues. Integrating lean construction into broader safety management discussions and applying quantitative methods, such as regression analysis, to assess the risks of migrant workers [97] are promising directions for future exploration in safety management.

The third subnetwork, centered on ‘worker behavior’ (K3), investigates integrating safety management strategies with employee behavioral patterns within the construction safety domain. Analysis of Figure 9 and Table 7 highlights the utilization of ‘structural equation modeling’ (K31), ‘real-time technology’ (K8), and ‘machine learning’ (K2) as robust analytical tools for examining various facets of employee behavior. This subnetwork also explores the impact of ‘safety culture’ (K21) and ‘safety training’ (K5) on worker behavior. The network analysis reveals strong interconnections of core clusters, such as safety management systems, machine learning, safety training, and worker behavior. The theory of planned behavior posits that behavioral intention is shaped by attitude, subjective norms, and perceived behavioral control [98]. Therefore, safety training should incorporate interventions to shape attitudes and norms and measure changes in workers’ attitudes and compliance intentions after the training. Recent developments have expanded research into areas including ‘behavior prediction’ (K64), ‘communication’ (K40), ‘safety compliance’ (K77), ‘risk assessment’ (K12), ‘real-time technology’ (K8), and machine learning-based behavior classification [99,100]. Furthermore, the analysis highlights less central macro keywords, such as ‘noise’ (K57), ‘workload’ (K136), ‘decision-making’ (K52), and ‘contractor behavior’ (K103), signaling potential avenues for further exploration. For example, there is a noticeable research gap regarding the impact of stakeholder behavior on engineering safety. Similarly, the decision-making processes influencing worker behavior remain underexplored, with current studies only linking decision-making to ‘agent-based modeling’ (K150) within this cluster. Expanding research on the interaction between worker behavior and decision-making processes could provide new insights and strategies to enhance safety management in construction environments.

The fourth subnetwork focuses on ‘machine learning’ (K2), a critical component of artificial intelligence that significantly enhances productivity and efficiency within the construction sector. In the field of construction safety, machine learning encompasses various methodologies, such as deep learning (DL), graph convolutional networks (GCN), support vector machines (SVM), random forests (RF), naive Bayes (NB), K-nearest neighbors (KNN), and artificial neural networks (ANN) [101]. Analysis from Figure 10 and Table 8 indicates that ‘computer vision’ (K19) and ‘real-time technology’ (K8) are central nodes across all centrality metrics in this subnetwork. Key macro keywords, including ‘autonomous safety monitor’ (K30), ‘safety management’ (K4), ‘accident’ (K1), and ‘worker behavior’ (K3), also maintain high centrality rankings. Machine learning applications in construction predominantly focus on real-time monitoring of workers, structural conditions, and equipment to enhance proactive safety management. Noteworthy examples include DL models, such as those by Ghafoori et al. [102], which utilize wearable devices to monitor workers’ postures in real time, predict potential health risks, and prevent accidents. Beyond ‘ergonomics’ (K10), machine learning extends to tracking ‘worker behavior’ (K3) and ‘equipment status’ (K35) through advanced techniques, such as autonomous safety monitoring (K30) and point cloud analytics. The evolution of machine learning in construction has seen significant growth since 2009, initially focusing on optimization algorithms for path planning [103] and Bayesian networks for predicting worker behavior [104]. Recent research broadens this scope to include ‘computer vision’ (K19), ‘text mining’ (K106), and comprehensive data analysis (K42), highlighting machine learning’s widespread application across various construction phases, from planning and design to supervision, control, and post-project analysis. Macro keywords with lower centrality scores, such as ‘safety protection’ (K68), ‘collision’ (K126), ‘safety inspection’ (K94), and ‘near-miss events’ (K97), suggest potential areas for further exploration and application of machine learning within construction safety. These areas represent fertile ground for future research and development efforts aimed at enhancing safety practices and reducing risks in construction environments.

The final subnetwork centers on ‘safety training’ (K5), with analysis from Figure 11 and Table 9 highlighting ‘hazard identification’ (K9), ‘safety management’ (K4), ‘safety culture’ (K21), and VR/AR technology (K23) as central themes, underscoring their pivotal roles in shaping the construction safety domain. This subnetwork explores the impact of safety training on ‘risk’ (K20), ‘accident prevention’ (K13), and ‘worker behavior’ (K3), integrating technologies such as VR/AR (K23), building information modeling (K18), 3D modeling (K130), artificial intelligence (K60), and 4D modeling (K108). Discussions within this cluster encompass various facets of safety training, including business aspects, communication strategies, educational interventions, personalized training methods, training transfer effectiveness, and planning. Safety training initiatives promote compliant and health-conscious worker behaviors, enhance hazard identification skills, and encourage active participation in safety protocols. From the perspective of Heinrich’s domino theory, safety training acts as an upstream intervention that removes unsafe acts and unsafe conditions before they trigger the accident sequence [75]. In addition, the theory of planned behavior provides a useful framework for understanding training effectiveness. Attitude toward safety can be shaped through immersive VR experiences that make hazards tangible. Subjective norms can be reinforced by embedding team-based training modules where peer expectations influence safety compliance. Perceived behavioral control can be enhanced via AI-powered adaptive training platforms that allow workers to practice hazard responses until mastery [98]. Historical literature dating back to 1992 has addressed safety training, with Baker et al. [105] focusing on training programs and quality plans for hazard identification participation. Prior to 2009, research predominantly covered general topics, such as ‘communication’ (K40), ‘musculoskeletal disorders’ (K25), ‘accident prevention’ (K13), ‘safety protection’ (K68), ‘contractors’ (K103), ‘qualitative research’ (K51), ‘supervision’ (K62), and ‘electrical safety’ (K71) [65,106,107,108]. In the past 15 years, advanced topics have been incorporated, including safety ‘data analysis’ (K42), ‘target tracking’ (K17), ‘accident’ (K1), and ‘visualization’ (K76) [99,109,110], paralleling technological advancements where VR/AR/MR technologies, 360-degree panoramas, and BIM technologies [17,111] are increasingly applied in safety training methodologies. Less emphasized macro keywords, such as ‘design’ (K16) for safety in safety training, ‘carpenter’ (K132), ‘worker age’ (K66), ‘gender’ (K105), and ‘supervisor’ (K62), suggest potential avenues for future research in safety training. Exploring these topics could bridge significant gaps in the construction industry’s current understanding and application of safety training practices.

5.3. Prospective Research Topics

The following four prospective research topics are proposed and discussed based on the findings of this paper:

• Total safety culture vs. resilience safety climate: Based on the comprehensive analysis of the main semantic network and its subnetworks from 1991 to 2023, it is evident that the research topics over the past three decades align well with the total safety culture (TSC) framework [112]. The TSC framework encompasses three key dimensions: person (e.g., knowledge, skills, and personalities), environment (e.g., equipment and tools), and behavior (e.g., compliance, coaching, and actively caring). TSC emphasizes the voluntary involvement of all individuals in fostering a culture where safety is a core value rather than a situational priority. It underscores the importance of effective feedback and encourages everyone to serve as safety coaches willingly. Another pertinent concept, the resilience safety climate (RSC) [113], operates at the organizational level and similarly emphasizes the engagement of all personnel, from top management to frontline workers. RSC advocates for proactive safety behaviors, such as anticipation and awareness, and focuses on success rather than failure, akin to the principles of TSC. A crucial theoretical question arises: What is the relationship between TSC and RSC? Specifically, is RSC a subset of TSC? This can be one future research topic. In addition, the analysis of subnetworks over the years reveals that safety culture and climate frequently appear. However, the distinction between safety culture and safety climate within the construction industry remains unclear.
• Personal factors: This paper has identified several personal factors impacting the safety performance of construction workers, such as gender, age, Hispanic ethnicity, and migrant status. The TSC framework emphasizes three critical personal factors: empowerment, self-esteem, and belongingness. While RSC and related research on psychological capital have addressed empowerment [114], the effects of self-esteem and belongingness on the safety behaviors of construction workers remain underexplored. Yang et al. investigated the relationships between abusive supervision, social standing uncertainty, belongingness need satisfaction, and workplace safety using survey data from manufacturing technicians and airline pilots [115]. Conducting similar studies on construction workers, particularly given their harsh work environments, would be highly valuable. Understanding how supervisor leadership style, contractor type (e.g., employee-owned companies versus contractors with different subcontracting practices), social standing uncertainty (e.g., a new hire unsure of their role within the team), and belongingness influence safety performance in the construction industry would provide crucial insights.
• Technology vs. communication: Changing construction workers’ safety perceptions is vital for influencing their safety behaviors. One promising approach is through innovative training methods. Kazar and Çomu demonstrated that safety training using serious games outperforms traditional safety training methods [116]. Their experiments with 42 civil engineering students showed significant improvements in safety knowledge and engagement with serious game-based training. This suggests that integrating advanced technologies, such as VR/AR and serious games, into safety training programs could change workers’ safety perceptions and behaviors more effectively than conventional training methods. These technologies offer immersive and interactive experiences that can better capture workers’ attention, making the learning process more engaging and impactful. Future research should explore the development and implementation of these technologies in construction safety training, assessing their effectiveness in real-world settings and across diverse worker populations. Additionally, understanding the factors influencing workers’ receptiveness to safety training and their perceptions of safety will be crucial in designing more effective communication strategies and training programs.
• Emerging topics: Several emerging topics warrant further attention in construction safety, such as temperature stress (both cold and heat stress) and fire safety on construction sites. In the United States, there has been an increase in the number of fires in structures under construction (https://www.nfpa.org/education-and-research/research/nfpa-research/fire-statistical-reports/fires-in-structures-under-construction (accessed on 7 January 2025)). Investigating these issues across different countries or regions, such as comparing developed and developing countries, would provide valuable insights.

6. Conclusions

This study utilized SNA to examine 1936 publications on construction safety from 1991 to 2023, mapping the field’s thematic structure and its evolution over time. The analysis revealed an exponential growth in the knowledge network’s nodes, accompanied by a consistently high clustering coefficient and only marginal increases in path length. This tightly connected structure facilitates interdisciplinary integration and fosters innovation within the field. By employing centrality analysis, five core topics were identified: accidents, safety management, worker behavior, machine learning, and safety training. These topics represent the foundational pillars of current research, with centrality rankings within clusters further clarifying the most and least prioritized areas, thus highlighting both dominant themes and underexplored research gaps.

The theoretical contributions of this study are significant. By introducing SNA metrics, the research provides a quantitative perspective that mitigates the subjectivity inherent in qualitative reviews. Diachronic network analysis illuminated long-term trends, revealing a shift from traditional concerns, such as mortality and injuries, toward contemporary issues like safety climate, worker behavior, and technological advancements. Visual network representations enhance the interpretability of thematic relationships, making complex connections more accessible. Additionally, the study identifies critical research gaps, such as total safety culture and construction site fire safety, which warrant further investigation to advance the theoretical understanding of the field.

Practically, the findings offer actionable insights for stakeholders in construction safety. High-centrality topics, such as safety climate and artificial intelligence, should be prioritized for investment to drive impactful advancements. The identification of frequently co-occurring risks, like heat stress and prolonged outdoor work, underscores the need for integrated safety planning to address compound risks effectively. Furthermore, network-based metrics, such as network density or target-topic centrality, provide a framework for evaluating intervention effectiveness and optimizing resource allocation, enabling more strategic decision-making in safety management practices.

Despite these contributions, the study has limitations that should be addressed in future research. The reliance on the Web of Science Core Collection excludes non-English or non-indexed publications, potentially limiting the scope of the findings. Variations in keyword processing thresholds may also influence the results. To enhance comprehensiveness, future studies should incorporate broader literature sources, explore additional subnetworks, and investigate how the strength of keyword connections shapes the network structure. These efforts will further refine the understanding of construction safety’s evolving knowledge landscape and support more robust safety interventions.