Occupational Health and Safety (OHS) disclosures play a vital role in shaping public perceptions of corporate working conditions. As organizations strive to create safer and healthier workplaces, the voluntary disclosure of OHS procedures has become an essential element of corporate transparency and responsibility (Rikhotsoet al., 2022a). These disclosures not only ensure the well-being and safety of employees but also contribute to the overall sustainability of organizations. Despite its importance, the integration of OHS into sustainability reporting remains a relatively new and evolving practice (Kawashitaet al., 2005). Many companies recognize OHS as both a legal obligation and a key component of their sustainability strategies, using it to demonstrate their commitment to fostering safe work environments and responsible business practices (Alves and Ramos, 2022a). Through comprehensive OHS disclosures, companies provide insights into their efforts to protect and promote the well-being of stakeholders (Olanrewajuet al., 2022). These disclosures may include a variety of measures, such as OHS training programs, safety protocols, risk assessments and open communication regarding safety issues (Abdul Hamidet al., 2020).

Despite the growing importance of OHS disclosures, there is still limited research on how firms systematically incorporate these practices into corporate reporting. While Hahn and Kühnen (2013) investigated sustainability reports and Monteiro (2021), focused on employee-related disclosures, applying advanced methodologies such as Structural Topic Modeling (STM) to OHS disclosure research remains underexplored. STM has been successfully applied in other fields to uncover hidden thematic patterns-for example, baiet al. (2021) used STM to analyze trends in marine transport research, while Aliet al. (2019) applied STM in their analysis of hospitality management literature. Chen and Xie (2020) used STM for bibliometric analysis in sentiment analysis research, demonstrating its versatility across different domains.

The present study builds on this growing body of research by applying Machine Learning (ML) techniques, specifically STM, to explore key themes in OHS disclosure. Using a keyword search in Scopus, 956 articles published between 1974 and 2024 were identified and analyzed through STM and bibliometric analysis. These methods aim to uncover research objectives, affiliations, focal areas and epistemological perspectives related to OHS. Significantly, this study extends traditional bibliometric techniques by applying STM to identify trends in OHS corporate reporting and disclosure practices.

Moreover, this research goes beyond conventional methods by utilizing the findthoughts() function. (Robertset al., 2019) This allows for identifying documents that align closely with thematic constructs derived from the STM model. This innovation enables the researcher to discover specialized publications related to OHS disclosures, thus contributing to a deeper understanding of the current state and future direction of OHS corporate reporting. While prior studies have successfully applied STM in fields like marine transport, hospitality management and sentiment analysis, there remains a significant gap in the literature regarding the systematic use of ML and STM to analyze OHS-specific trends. This study addresses this gap by introducing a novel approach to analyzing OHS disclosure trends, offering theoretical and practical implications for corporate reporting and sustainability practices.

The study builds upon (Baiet al., 2021) by modifying its research process. The first phase involves collecting articles, including selecting the database and keywords to retrieve relevant scholarly articles. The 2nd phase focuses on exploring the dataset and the third phase, STM, encompasses text processing, topic selection, knowledge discovery and identifying research articles related to specific topics derived from the STM model (see Figure 1).

The Scopus database was chosen to collect scholarly articles. A keyword search was performed using the following keywords: TITLE-ABS-KEY (occupation* AND health AND safety OR ohs AND disclos* OR report*.) AND (LIMIT-TO (DOCTYPE, “ar”) OR LIMIT-TO (DOCTYPE, “re”) OR LIMIT-TO (DOCTYPE, “ch”) OR LIMIT-TO (DOCTYPE, “cp”)) AND (LIMIT-TO (LANGUAGE, “English”)) A total of 956 scientific articles were collected from the period spanning from 1974 to 2024, covering 50 years of research papers.

Topic modeling is an unsupervised algorithm that extracts latent themes from textual documents. These themes are learned based on the patterns and relationships within the text data without explicit supervision or predefined labels (Baiet al., 2021). Two commonly used methodologies for topic modeling are Latent Dirichlet allocation (LDA) and STM (Jelodar et al., 2019; Li and Lei, 2021).

STM offers several advantages over LDA (Baiet al., 2021). Firstly, STM allows for incorporating document-level covariates, which provide additional context and can improve the accuracy of topic modeling. For example, by including metadata about the author, publication date, or source of a document, STM can account for topic prevalence or word use variations based on these covariates. STM facilitates a better understanding of the relationship between topics and the underlying factors influencing them. Secondly, STM models give correlations between topics, providing deeper insights into how topics relate.

Additionally, STM allows for modeling variations in word distributions within a topic based on covariates and this means that the words associated with a particular topic may vary depending on the context provided by the covariates. This flexibility allows for more precise and context-specific topic modeling, leading to a better understanding of the data. Furthermore, STM provides the advantage of directly assessing the impact of covariates on topic prevalence or word use within a topic (Baiet al., 2021) and it provides more insights into how different factors impact the distribution of topics and the usage of words within those topics (Slade, 2016). (Figure 1) shows the 4 computational steps for STM modeling: text processing, selection of topics, knowledge discovery (Baiet al., 2021) and results.

Inbuild text processor () function in the STM model, which will automatically create a corpus, convert to “Lower Case,” “remove punctuation,” “remove stop words,” “remove numbers,” “stemming,” and Create output for the following procedure (Küsters and Garrido, 2020).

It is critical to find the appropriate number of topics. However, no single, generally acknowledged method exists for determining the ideal quantity of topics for STM estimations. (Ravenda et al., 2022). The search K () function determines the ideal number of topics for the model. It combines statistical methods with human judgment to find the optimal balance. By integrating coherence scores and expert input, this approach ensures that the model accurately reflects the complexity of the data, enhancing its interpretability and usefulness (Robertset al., 2019). We begin by applying rigorous statistical diagnostic tools to examine dynamic fluctuations in semantic coherence and other relevant metrics to assess the goodness of fit of STM estimates. These evaluations are conducted throughout a spectrum of topic quantities ranging from 3 to 25 (Figure 2).

The held-out likelihood score increases as the number of topics increases. The scores rise steadily from k=3 to k=11, but after k=11, they fluctuate. The score improves again at k=16, with k=19 providing the best model fit. However, semantic coherence tends to decrease as the number of topics increases, with fewer topics generally having higher coherence (Robertset al., 2019)This does not fully capture the model’s complexity. A better goodness of fit correlates with lower residuals and higher lower bounds. We chose k=11 for further analysis, as it best fits the first phase. Beyond this, the model risks overfitting and generating redundant or meaningless topics. Thus, k=11 balances fit and semantic coherence, outperforming k=19 in topic quality. While semantic coherence is important, it only measures internal consistency (Mimnoet al., 2011) We followed Roberts (2014, 2019) and Wang (2022) to finalize the optimal number of topics, use labelTopics() and plot() to ensure meaningful topic labeling.

After selecting k=11, we fit the model accordingly. The plot Models () function averages all topics from each model run and plots them numerically. The model with favorable qualities in both semantic coherence and exclusivity will be chosen, with models scoring in the upper right corner of the plot considered for further analysis. (Robertset al., 2019)The plot Models () function also displays each topic’s quality in the model (Figure 3).

Various methodologies can be used comprehensively to investigate the output of topic models, including exploring word associations with topics and examining the relationships between metadata and topics. (Robertset al., 2019). Researchers frequently employ several key functions from the Structural Topic Model (STM) toolkit to obtain and visualize results, such as label Topics (), cloud (), topic Corr () and estimate Effect ().

The label Topics () function provides a detailed overview of the characteristics of each topic, including the highest probability words, FREX (Frequency×Exclusivity) words, lift words and score words (see Table 1). While the highest occurrence probability for words conditional on the topic offers insights into word association, certain common terms such as “study,” “research,” “management,” and “data” may lack specificity and provide limited value for topic interpretation (Baiet al., 2021)The FREX metric is crucial to address this as it identifies more distinctive and meaningful associations between terms and topics. (Baiet al., 2021; Bischof and Airoldi, 2012; Robertset al., 2019) By applying the harmonic mean, FREX balances the influence of frequency and exclusivity, ensuring a more refined and relevant assessment of a word’s contribution to a specific topic domain.(Bischof and Airoldi, 2012). Equation 1 illustrates the FREX formula, as referenced by Baiet al. (2021) and W. Wanget al. (2023).

In addition to examining word associations, the cloud () function generates word clouds for each topic, providing a visual representation of the most frequent terms (Figure 6). The word cloud highlights the theme classified within each topic, with words sized according to their probability of occurrence. This visualization can help quickly identify the dominant themes across topics.

Moreover, the topic Corr () function allows for the exploration of topic correlations, shedding light on how topics are related to each other within the model. Such correlations can reveal underlying thematic linkages that may not be apparent through isolated topic analysis.

Figure 1:
Research outline.

The estimate Effect () function enables researchers to examine the impact of metadata (such as publication year, region, or author affiliation) on topic prevalence. By doing so, the function provides insights into how the prominence of certain topics evolves over time or varies across different contexts. This study also leverages bibliometric analysis to assess publication trends, such as the most frequently cited journals and the most significant contributing journals within the domain of OHS disclosure. This analysis complements the STM results, providing a comprehensive understanding of the scholarly landscape related to OHS reporting.

Cumulative scholarly article production is a key metric for identifying the most influential journals in the field. Appendix 1 highlights the top 20 journals contributing significantly to OHS reporting and disclosure research. The Journal of Cleaner Production leads with 23 articles, representing 3% of the total output. Safety Science and Sustainability (Switzerland) follow, each with 19 articles, making up 2.12% of the total. Notably, out of over 704 journals identified, most have contributed only one article, while 34 journals have published more than three articles each

Figure 2:
Determination of number of topics.

Figure 3:
Semantic coherence exclusivity and topic quality.

Figure 4 shows the trend in annual scientific production on occupational health and safety from 1974 to 2024, measured by the number of documents published each year. From 1975 to 1990, production was low and stable, with fewer than 20 documents annually. The early 1990s marked the start of an upward trend, peaking around 1995, followed by a slight decline and a subsequent rise in the late 1990s. The 2000s saw a gradual increase in document production, with fluctuations but a general upward trajectory. This trend accelerated sharply around 2010, leading to a significant rise in annual publications. The most notable surge occurred in the early 2020s, peaking at over 100 documents in 2022 before a decline. Overall, the data reflect substantial growth in occupational health and safety research, particularly from the 2000s onward, with a peak in the early 2020s.

Appendix 2 highlights the most cited articles, including “Corporate Social and Environmental Reporting: A Review of Literature and a Longitudinal Study of the UK” by Grayet al. (1995), published in the Accounting, Auditing and Accountability Journal. The article explores integrating OHS concerns into corporate social and environmental reporting, focusing on the UK. It emphasizes the importance of OHS reporting for meeting stakeholder expectations and regulatory requirements. It examines trends in OHS reporting over time, its impact on corporate reputation and performance and challenges like data collection and standardization. The next most cited work is Patton’s (2016) paper, “Our Future: A Lancet Commission on Adolescent Health and Well-Being.”

The analysis of scholarly articles highlights the prevalence of specific key terms in their abstracts. The most frequently occurring words include “health,” “social,” “safety,” “research,” “study,” “CSR” (Corporate Social Responsibility), “management,” “work,” “environmental,” “corporate,” “information,” “companies,” “occupational,” “sustainability,” and “rights.”

To visually represent the frequency distribution of these terms within the corpus, the Tree map () function was used. This visualization provides a detailed depiction of the relative occurrence of each term, offering valuable insights into the predominant themes and focal points within the scholarly discourse on occupational health and safety disclosure (Figure 5).

Topic labeling can often be simplified by selecting words with the highest probabilistic association with a given topic. However, this approach may limit the quality of the labels. To address this issue, Bischof and Airoldi (2012), introduced the FREX metric, which combines word frequency within a topic and topic exclusivity. FREX is calculated as the harmonic mean of a word’s rank by probability within the topic and its rank by the distribution across topics (Baiet al., 2021). Previous literature reviews using Structural Topic Modeling (STM) have applied both metrics to enhance the accuracy of topic labeling (Baiet al., 2021; Daset al., 2017; Kuhn, 2018).

A word cloud for each topic is generated using the cloud () function (Figure 6), which visualizes words based on their probability of occurrence within the topic. Words with a higher probability appear in larger font sizes, indicating their prominence in the topic. Researchers often use label Topics () and cloud () functions to determine topic labels through expert judgment.

Figure 4:
Annual production of scholarly scientific articles.

Figure 5:
word frequency tree map.

Figure 6:
Word cloud- the visual representation of topics.

Table show that most articles (14%) focus on occupational safety and health regulation. Promoting occupational health and safety education is closely followed, accounting for 13% of the research. In contrast, the least researched area is Pharmaceutical Hazard Management, with only 2% of articles dedicated to this topic.

Topic, one addresses the OHS Auditing in Construction Supply Chains; the highest frequency words are “construction” and “industry,” and FREX words are “chain,” “supplies,” “auditing,” and “techniques.” Topic two and Topic eleven discuss the corporate role in OHS; topic two addresses issues related to sustainable business practices with the high-frequency words “social”, “sustainability”, “CSR”, “corporate”, “response”, “companies” and the FREX words are “CSR”, “corporate”, “sustain”, “social”, “response”, “mine”, “stakeholder” furthermore, topic 11 related to Corporate Reporting and Disclosure Practices (8%). The highest frequency words are “report”, “disclosure”, “inform”, “company”, “studio”, “paper”, “indic” FREX: “annual”, “sdgs”, “report”, “gri”, “sdg”, “disclosure”, “legitimacy”. Topic three consists of the highest frequency words such as “health”, “care”, “profession”, “people”, “mental”, “support” and “communities” FREX words are “sex”, “privacy,”, “disable”, “access”, all together we can label the topic three as healthcare accessibility and support for mental health and disabilities.

Topic four is related to Pharmaceutical Hazard Management with the highest frequency of “food”, “research”, “product”, “consumer” and “medical,” and the FREX words are “grant”, “Pfizer”, “nanoparticle,”, “HIA”, this is least researched area in the model with 2% of research has done on this field. Topics five and eight address the common area of regulation frame related to OHS; the topic addresses research related to the regulatory framework and global rights. The highest frequency words are “right,” “global”, “human”, “regular”, “state”, “reserve,” and FREX words are “actor”, “law”, “cyber secure”, “privacy,”, “regulation”, “reform”, “state” and topic 8 is about Occupational Safety and Health Regulations with high-frequency words are “safety”, “health”, “occup”, “worker”, “risk”, “hazard”, “chemic” FREX: “chemic”, “hazard, “OSHA”, “avail”, “accid”, “safety”.

Topic six is about health studies and safety measures; the highest frequency words are “health”, “exposur”, “safeti, “patient,” and FREX: “ipv”, “self-report”, “preval”, “smoke”, “vaccin”, “injuri”, “stress”. Topic seven is related to environmental risk assessment with higher frequency word such as “environment”, “chang”, “risk”, “system”, “assess”, “climate”, “sector”, FREX: “lean”, “pollut”, “climate”, “waste”, “energi”, “ocean”, “plastic”. Topic Nine addresses Employee Well-being and Workplace Dynamics hight highest frequency words are “employe”, “work”, “studi”, “relationship”, “effect”, “workplace”, “organiz” FREX: “mediat”, “organiz”, “employe”, “compens”, “negat”, “creativ”, “employees”.

Topic ten is about Promoting Occupational Health and Safety Education; the highest frequency words are “book”, “research”, “student”, “provide”, “social”, “chapter” and “ethic” FREX: “edit”, “student”, “handbook”, “tourism”, “sociology”, “reader”, “chapter”. A key learning objective for sustainability that contributes to creating a safe environment (Moyoet al., 2022).

Figure 7 shows the correlation between topics. Topic 1, “OHS Auditing in Construction Supply Chains,” is closely linked with Topic 2, “Sustainable Business Practices,” and Topic 11, “Corporate Reporting and Disclosure Practices.” This suggests a connection between auditing in construction supply chains and broader sustainability efforts, highlighting the importance of transparent reporting in promoting sustainable practices. Additionally, Topic 5, “Regulatory Frameworks and Global Rights,” is strongly related to Topic 10, “Promoting Occupational Health and Safety Education,” emphasizing the link between regulatory frameworks and educational initiatives to enhance awareness and compliance, especially in a global context. Similarly, Topic 4, “Pharmaceutical Hazard Management,” and Topic 7, “Environmental Risk Assessment,” are correlated with Topic 5, underscoring the broad scope of regulatory frameworks that cover areas like pharmaceutical hazard management and environmental risk assessment, all tied to global rights and standards.

Figure 7:
Topic correlation.

The impact of publication year on topic prominence is a key aspect of using STM in academic research. STM enables the analysis of interactions between variables like publication year and topics of interest. By including publication year as a covariate, researchers can track and compare the prevalence of topics over time, offering insights into their temporal dynamics.

Figure 8 illustrates these dynamics, showing trends in various topics over the years. Topics like “OHS Auditing in Construction Supply Chains,” “Sustainable Business Practices,” “Environmental Risk Assessment,” “Regulatory Frameworks and Global Rights,” “Employee Well-being and Workplace Dynamics,” “Promoting Occupational Health and Safety Education,” and “OHS Corporate Reporting and Disclosure Practices” have shown an upward trend.

Figure 8:
topic trends.

In contrast, research focus on “Occupational Safety and Health Regulations” and “Health Studies and Safety Measures” has declined over the analyzed period. This shift suggests that earlier OHS research focused on regulations and safety measures, but as stricter OHS laws were implemented, interest has shifted toward newer topics within the field.

Roberts (2019) used the find thoughts () function to examine documents highly associated with topics and the researcher used the same function to find papers highly associated with the topics, resulting in a niche result. Topic 11 in the topic solution is associated with OHS-corporate reporting and disclosure practices and the results in (Appendix 3) show the scientific articles related to this topic. Research using this function to explore the relevant paper related to OHS-corporate reporting and disclosure is shown in (Table 2) though2<- find Thoughts (select Model, texts=data$Title, n=20, thresh=.25, topics=11) though2.

OHS disclosure has evolved, with early studies like Chan (1979b) highlighting the need for uniformity in reporting within Fortune 500 companies. Initial research focused on basic metrics such as workplace injuries but lacked depth. Geographical and industry-specific studies, such as those by Kawashitaet al. (2005) and Nagataet al. (2017), demonstrated that OHS disclosure varies by industry and company size, with larger firms and high-risk industries like manufacturing disclosing more. In contrast, Japanese firms were found to report less OHS data compared to other CSR elements. Research by Rahmanet al. (2019) on Indonesian trade unions and Evangelinos et al. (2018b) on multinational corporations revealed significant gaps between the union and corporate OHS communication, with reporting often lacking depth despite detailed management approaches. OHS is increasingly integrated into broader CSR frameworks, as studies by Monteroet al. (2009), Searcyet al. (2016) and Kumaret al. (2018) indicate. However, it is often secondary to other dimensions like environmental sustainability. Legal requirements, rather than voluntary sustainability initiatives, frequently drive OHS reporting, as Mariappanadar et al. (2021) pointed out. The table showcases the evolution of OHS disclosure across industries, countries and frameworks, reflecting a trend towards integrating OHS into CSR and sustainability practices. Larger firms, certified companies and those in high-risk industries tend to disclose more OHS-related information, driven by both stakeholder pressure and legal requirements. However, gaps in standardization, voluntary reporting and sector-specific challenges continue to hinder uniform reporting despite its positive impact on financial performance and corporate reputation. Various factors influence OHS disclosure, including industry profile, company size, MNC operations and financial performance, as noted by Cahayaet al. (2017) and Zhang et al. (2023b) while board involvement and corporate debt may negatively impact transparency. High-risk sectors like construction show improved transparency due to stakeholder pressure, according to Fotiadis et al. (2023b) and Oakmanet al. (2024). Stakeholder and institutional influences also play a role in disclosure practices, with studies by Tsalis et al. (2018b) and Roland et al. (2020b) suggesting that certifications like OHSAS 18001 enhance transparency. Alves and Ramos (2022a) further explored how certifications and external assurance improve disclosure quality in European companies. The distinction between voluntary and mandatory OHS reporting is critical, as studies like Gardiner (2022) and Oakmanet al. (2024) highlight, with mandatory standards generally driving more comprehensive disclosures. Health-related OHS issues are also industry-specific, with studies like Park and Kim (2023) and Rikhotso et al. (2022b) focusing on noise-induced hearing loss and injury rates in sectors such as mining and manufacturing. OHS disclosure positively impacts financial performance and corporate reputation, as Zhang et al. (2023b) found, though factors like institutional investor shareholding can negatively influence it. Despite these developments, challenges remain, including inconsistent reporting practices and focus on injury rates rather than broader OHS indicators, as discussed by Shimizu et al. (2022b) and Aliet al. (2021).

An intensive study on OHS reporting and disclosure has been conducted. The sheer number and diversity of scholarly articles in the OHS disclosure make it time-consuming and challenging to identify the significant research themes and trends accurately. This review paper focuses on OHS reporting and disclosure of scholarly published articles from 1974 to 2024 in the journal indexed on the Scopus database. This study found 956 articles focused on OHS disclosure and reporting, most published between 2020 and 2022.

This review paper delves into OHS reporting and disclosure, employing a methodologically rigorous approach to analyze a vast array of literature spanning nearly five decades. Leveraging advanced computational techniques, particularly STM, the study uncovers latent themes and trends within the OHS landscape. “OHS regulation” is the predominant theme and noteworthy relationships have been identified. Firstly, a significant connection exists between “OHS auditing in the construction supply chain,” sustainable business practices and OHS corporate reporting and disclosure practices. This relationship highlights the integration of safety audits with broader corporate sustainability and transparency efforts. Secondly, promoting OHS education is closely linked to the regulatory framework and global rights, underscoring the necessity of educational initiatives to support compliance and protect workers’ rights. Thirdly, pharmaceutical hazard management is tied to regulatory frameworks, global rights and environmental risk assessment, indicating a comprehensive approach to managing health risks within the pharmaceutical industry.

The emerging topics in OHS include “OHS corporate reporting and disclosures,” “employee well-being and workplace,” “sustainable business practices,” and “OHS auditing in the construction supply chain.” Conversely, there is a noticeable decline in interest in topics such as “health studies and safety measures,” “OHS regulation,” “pharmaceutical hazard management,” and “healthcare accessibility and support for mental health and disabilities.” This shift reflects evolving priorities and focal points within the OHS field. The findthought function of the STM package assists researchers in identifying articles highly associated with specific topics, such as OHS corporate reporting and disclosure. Through this process, the researcher has pinpointed 27 highly relevant articles to OHS in corporate reporting.

Thank you to Sudhakar from Pondicherry University for his help coding the Structural Topic Modeling (STM) tool, which made this study successful. I also thank Pondicherry University for giving access to Scopus, which was essential for data collecting and analysis. Finally, I thank my mentors, colleagues, and Pondicherry University’s academic environment for inspiring and enabling this effort.