We searched Scopus and Web of Science for relevant articles. There were two sets of search terms: one set related to analytics and the other set related to occupational safety. The search terms were developed through consultation with experts and based on results of preliminary attempts to search various databases. The same terms were used across databases, but the search strategy was tailored to the particular syntax of each database. The search string was as follows:

The search was limited to articles published after 2006, and only peer-reviewed journal articles and conference publications were considered. Furthermore, articles had to be written in English, involve the use of analytics techniques, have injuries or near misses as an outcome of interest (i.e., the dependent variable) and use data collected at the enterprise or establishment level. The inclusion and exclusion criteria are listed in Table 1. These studies were excluded to both keep the number of resulting articles to a manageable number and to focus on studies that evaluated injuries or near misses as the outcome variable. Studies may refer to injuries as ‘incidents’ or ‘accidents’ as these terms are often used interchangeably [19]. Examples of studies that were excluded included those that were not concerned with occupational safety (e.g., pedestrian safety, patient safety, etc.); dealt with occupational health issues instead of injuries (e.g., pneumoconiosis, cardiovascular disease, etc.); were concerned with assessing risks or hazards; or were literature reviews, commentaries or other types of non-empirical research papers.

Additional searches, beyond the initial Scopus and Web of Science search, were conducted to ensure that as many qualifying studies as possible were identified. The Safety Science journal had the largest number of qualifying articles identified from the database searches (described earlier). Therefore, we conducted a manual search of the titles and abstracts of every paper published in Safety Science from 2005 to September 2023. Additionally, citation searches were conducted for the studies that met the inclusion criteria. This involved searching the references of a particular study and subsequent citations of that study listed by Google Scholar for inclusion criteria.

At least two co-authors reviewed each citation, and the first author (A.M.F.) reviewed all citations. If two authors disagreed on an article’s inclusion, a third author would break the tie. The searches were conducted in September 2023. A data chart was populated for each article selected for inclusion. The data chart contained the following columns: year, title, authors, affiliation, author keywords, journal, study question(s)/objective(s), description of data source(s), data preprocessing, predictor variables, outcome variables, analytics techniques, analytics-derived insights, challenges described and notes. A thematic synthesis was used to identify and summarize patterns in the included articles and identify knowledge gaps. These findings will be described in the following section.

All studies included in the scoping review attempted to predict injuries or near misses, or, alternatively, to identify common patterns related to injuries or near misses. This is not surprising as one of the inclusion criteria for a study was that it be related to analytics of injuries or near misses. Most of the studies were concerned with exploring data from incident reports, safety inspection reports and other safety data, although there were a few exceptions with more specific research objectives. For example, Marques et al. [20] examined the effects of drug and alcohol testing frequency on the occurrence of injuries. Additionally, Tsang et al. [21] explored the effects of employee characteristics (e.g., age, body mass index [BMI], average heart rate) and environmental/behavioral variables (e.g., ambient temperature, recovery time) on injuries among cold-storage employees.

Many studies did not disclose the sources of their data. This is problematic, as it impedes the ability to replicate or extend that research by others in the field. There were some common themes in the studies that did disclose information on the sources and types of data sources. Twenty-five studies that included information on data sources were conducted in steel plants, and a few studies were conducted in construction organizations or refineries. The datasets usually spanned 3–5 years, with the longest duration of time being 16 years [22]. Various people collected the data, including front-line workers [23,24] and managers [25], and data were typically stored in the organization’s online safety management system (SMS) [23,26–31].

Data reported in the studies featured both structured and unstructured data with a mix of data types (e.g., categorical, numerical, textual). Structured data can be thought of as searchable data that are easily processed by a computer and have a clearly defined organizational system [32]. Unstructured data cannot be easily processed with conventional tools and has no clearly defined organizational system (e.g., the text of written incident reports, photographs, etc.) [32]. Commonly used structured data included the frequency of injuries and incidents. Unstructured data included text-based descriptions of incidents and observations as well as occasional photographic observations.

There was a high degree of variability in the types of data collected. Most of the variables included in the studies were reactive or lagging indicators. These are variables such as incident reports, injuries and lost workdays that occur after a safety incident. Figure 4(B) shows the most frequently included incident details. Only 10 (22%) studies included proactive or leading indicators. Leading indicators occur before a safety incident and include preventive variables such as safety audit reports, frequency of toolbox meetings and safety training history. A few studies included human resource management data such as employee age, education and marital status. Figure 4(A) shows the most frequently included employee characteristics. Although some studies in the construction sector included details related to project cost, stage of completion or complexity [33,34], limited studies used data from organizational areas outside safety, such as production or maintenance. For example, Tsang et al. [21] did include production data, but these data were used for analyses separate from those investigating injury.

Some of the studies included upwards of 20 variables in their analyses, and although a majority provided definitions for each variable, 14 (29%) of the studies did not provide any definitions. For example, one study examined several predictor variables, including ‘cause of accident’, but did not describe what causes were captured by that variable [35]. Although some variables may be self-explanatory or not require detailed definitions (e.g., year, day of the week, season), other variables necessitate further explanation (e.g., nature of incident, equipment damage score, etc.). To fully understand and interpret the results of the analyses conducted, the reader needs to know what the variables are measuring so that they may draw informed conclusions from the study results, make comparisons to their own organizational data or compare results across studies.

The preprocessing stage of analytics involves selecting variables to analyze, handling missing data and restructuring text data. This stage typically takes the most amount of time, and there are different methods for approaching the preprocessing steps. Figure 5(A) shows the approaches to variable selection reported in the studies included in this review. Variable selection involves deciding which variables to include in the analysis. For 27% (n = 13) of the studies, there was no description of the process for variable selection [1222, 24,25,34–42]. For studies that described the process, the most common approaches were statistical in nature, including using Boruta feature selection [34,44,45], χ² [27,28,33] and random forest [46], among others [22,24,47–53]. Fifteen studies stated inclusion or exclusion criteria which varied from including all variables [29,31,54–58], the most common variables [59] or variables that aligned with specific International Organization for Standardization recommendations [21], among others [20,60–64]. Last, five studies stated that their variables were selected based on domain knowledge or consultations with experts [30,65–68].

Accounting for and handling missing data is an important step in preprocessing and analysis. The approaches to handling missing data are shown in Figure 5(B). Most of the studies (59%; n = 29) did not include a description of how missing data was handled. Eight studies used a statistical approach to handling missing data, which included imputation [33,50], random forest [27,44–47] and expectation maximization [65]. Seven studies mentioned that missing data were accounted for but did not provide specific details [24,26,29,39,42,57,61]. Four studies stated that data with missing values were omitted from the analysis [35,41,59,60]. One study consulted with experts to inform the techniques to handle missing data, although did not describe the chosen techniques in their manuscript [30].

Free-text data (e.g., incident narratives) can often provide a great deal of information, but it first must be converted into a format that can be analyzed. That is, unstructured, open-ended text (e.g., explanations on an incident report) must be converted to a structured format that can be read by data analysis software. There are numerous methods to convert free text to structured data. The methods to convert free text reported by the studies in this review are shown in Figure 5(C). The most common approach was latent dirichlet allocation [24,27,44,47,53,56] followed by term frequency-inverse document frequency (TF-IDF) [13,24,38,54,64]. Other studies used structural topic modeling [45,57] and expectation maximization [24,28].

Class imbalances in the data were another concern that was addressed during the data preparation phase in several studies. Class imbalance occurs when one outcome variable occurs disproportionately more often than another, creating bias toward the majority variable [69]. In such datasets, the variable of greatest interest is often the minority variable, such as workplace injuries, compared to more common near-miss events [70]. Data-level techniques are the most popular methods for addressing class imbalance and consist of two approaches, undersampling and oversampling. In oversampling, values from the minority variable are replicated to increase the size of the class, and in undersampling, values from the majority variable are deleted to decrease the size of the class. In the present studies, undersampling was used in one study [50] and oversampling was used in a handful of studies to address class imbalance issues. Oversampling techniques implemented included the synthetic minority oversampling technique (SMOTE) [22,33,34,44], majority weighted minority oversampling technique (MWMOTE) [44], borderline SMOTE (BLSMOTE) [69] and k-means SMOTE (KMSMOTE) [44].

The software used to conduct analysis is not strictly part of data preprocessing. However, there are an increasing number of widely available software programs developed for conducting analytics in which different implementations of the algorithms (e.g., non-convergence rates) within the software lead to slightly different analysis outcomes. We have no reason to suspect that any such issue is related to the outcomes within the studies under review, but it is good practice to report the software used for preprocessing and data analysis. The software used in the reviewed studies are shown in Figure 5(D). Most of the studies (78%) did describe which software was used to analyze data. The most common software was R [25,26,31,33,38,43–47,52,57,66], which is an open-source, free software program. The next two most frequently used software programs were SAS [23,30,58,60,61,67] and SPSS [20,35,37,40,59,68].

Approaches to analytics are sometimes divided into classification and regression tasks [71]. In classification methodologies, the output of a statistical model is assigned to a particular class. In the case of occupational safety, an example of the output may be an injury event or a non-injury event. In regression methodologies, the output of a model is a continuous variable. In occupational safety research, an example of a continuous variable may be the frequency or rate of injuries. For example, Ajayi et al. [33] compared several different machine learning techniques predicting the number of hand injuries in power infrastructure operations workers. For most of the selected studies, the output of the models was assigned to a class (e.g., fatal injury, serious injury, first aid [56]). Most analytic techniques can be used for both classification and regression.

Figure 6 shows the techniques that were used in studies focusing on a single analytics approach or method. The most common analytics techniques among these studies were association rule mining [21,25,30,39,54,59,63,64,72] and Bayesian networks [26,29,36,41,62]. Many of these single-approach studies employed techniques that examine the relations among temporal variables (e.g., antecedent events preceding near misses or injuries) including association rule mining, multiple correspondence analysis, vector autoregression, object role modeling, cause-and-effect diagrams and axiomatic design framework. As an example, association rule mining is a technique that was originally developed to detect patterns of transactions in retail stores to identify the frequency of patterns in data by identifying conditional associations (e.g., if/then or antecedent/consequent relationships). By examining these conditional associations, events or characteristics that are commonly correlated with injury (the ‘then’ or ‘consequent’ portion of the association) can be detected. The resulting rules can vary in the number of items (e.g., a four-item rule could be: young workers [Item 1] with shorter job tenures [Item 2] who work in Department 3 [Item 3] are more likely to have lower limb injuries [Item 4]), although the analysis becomes more complex with additional items. Studies in this scoping review obtained two-item [21,25,39], three-item [20,24,29,38,53,58,62,63,71] all eight studies previously listed, four-item ([21,30,39,40,54,72] and five-item [30,39,54] rules. In general, the researchers were able to identify more specific scenarios in which injuries were likely to happen through association-rule mining. For example, Buddhakulsomsiri et al. [63] found that incidents that resulted in major injuries and high costs were associated with work performed by outside contractors, less experienced workers and workers between 41 and 50 years of age. The studies that used association-rule mining were in oil refining [25], warehousing [21], construction [59], steel manufacturing [39,54,64,67] and car manufacturing [40].

One finding that emerged was that many studies (41%) were comparing the performance of several different analytics approaches. The studies and the techniques compared are presented in Table 2. The objective of those comparisons was to identify techniques that were most accurate in predicting injuries (or near misses, depending on the study). The techniques that were identified as the most accurate in each study are highlighted in Table 2 and included random forest and classification and regression tree techniques. A summary table of the analytics techniques performed in all 49 included studies can be found in the Supplemental data.

Most of the studies provided detailed findings from the analytics conducted. Of the studies that reported detailed findings, many of the relations identified were rather complex and were unlikely to be detected by using more traditional, simpler analytical approaches. For example, Ajayi et al. [33] identified complex interactions among project complexity, time of year, worker age and experience, and task type that were more predictive of injury. Sarkar et al. [44] developed 19 specific safety rules that described scenarios under which incidents were more likely to happen in different divisions of a steel manufacturing facility. Lingard et al. [43] identified complicated interactions among safety indicators over time that suggested a cyclical relationship between the behavior of management and the occurrence of injury. Through a χ² automatic interaction detector (CHAID) analysis, Marques et al. [20] were able to identify an optimal schedule for drug and alcohol testing of employees that reduced injuries but would not place undue burden on either the organization or its employees. Although the findings of some studies were somewhat unsurprising (e.g., attributes like incident types and primary causes were related to injury risk) [22,28,67], many analyses in these studies made use of free-text, narrative data which often go unanalyzed. Finally, six studies were methodologically focused and only reported information related to model fit (e.g., accuracy, robustness) [28,36,38,41,45,66].

Although a desirable end goal of any analytics effort is to reveal actionable insights, the findings may be less actionable than desired. In fact, they may reveal more fundamental or underlying problems with a company’s existing data reporting systems. For example, through an expectation-maximization-based clustering analysis of free-text injury narratives, Verma et al. [67] identified misunderstandings on the part of workers about operational definitions for incident variables. The analysis revealed that workers were categorizing some events related to falling materials as slip/trip/fall events instead of ‘struck by’ events. This study demonstrates that the analytics process is not unidirectional and can also inform and improve processes related to data collection. For example, it can lead to unexpected outcomes, like the improvement of data collection systems, by identifying unclear or incorrect instructions for incident reporting.

As analytics in occupational safety is a relatively new area of exploration, most of the studies were designed as methodological demonstrations or considered pilot studies from which future studies could be based. Only one study in the review described actions taken based on the analytics that were conducted. Tsang et al. [21] reported that because of the implementation of the analytics software that was developed, actions such as allocating additional recovery time from low-temperature environments or wearing at least three layers of clothing insulation decreased the frequency of employee injuries from 12 injuries per week to five.

There were many limitations identified in the included studies. Two categories of limitations were cited most frequently: model choice and data integrity. Whereas all models had some degree of success in predicting safety incidents, some data did not meet the assumptions of the chosen model which undermined validity. In one example, which concerned text analytics, researchers found that the use of short-hand narratives to describe accidents led to too sparse an amount of contextual data to allow for valid conclusions [36]. To mitigate this shortcoming, Sarkar et al. [28] used more advanced methods such as expectation-maximization algorithms in follow-up studies. The authors also stated that more in-depth analyses using similar statistical techniques may yield richer insights. For another example, Guo et al. [59] suggested using additional multidimensional association rule mining of unsafe behavior to learn more about behavior patterns.

Other common limitations identified by the studies in the review included data collection methods, data integrity and data variety/breadth. Recurring problems with collection methods included biased observers, a lack of subject matter experts and the underreporting of injuries; a frequently proposed solution within the identified studies to these problems is the use of standardized surveys made by subject matter experts. Common data integrity issues included confidentiality concerns and inconsistency of record-keeping across different departments. The final limitation within datasets concerned the breadth of data. In one case, while building a rule mining database, the researchers found that their case study was not sufficiently large or broad enough to allow for generalizable results [59]. Another study found that the sample size limited their ability to extract decision rules from the decision tree they had generated [65].

The reviewed studies primarily analyzed safety-related data. Most of the data consisted of lagging indicators such as the details of incidents and near misses. More recent studies have incorporated leading indicators, including safety audit reports, safety training records and the frequency of toolbox talks. By incorporating these leading indicators into safety analytics, organizations can assess how the effort put toward proactive safety programs is borne out quantitatively in improvements to lagging indicator metrics. Additionally, by conducting the somewhat onerous data preprocessing stages for leading indicators, organizations may be able to identify areas in need of improvement or gaps in their safety programs. Extending data sources beyond safety-specific data to other divisions of an organization could have the potential to detect new relations among variables. For example, incorporating production data (e.g., production pressure), human resource management data (e.g., overtime) or weather data (e.g., temperature, humidity, precipitation) could provide a more detailed picture of the potential causes of injury.

Providing more information about the variables analyzed is another area that could be improved. More than one-quarter of the studies did not provide definitions for the variables in the analysis, and on a few occasions the names of the variables were shorthand labels used in the data analysis programs, making interpretation by the reader more difficult. Explicitly stating the variable type, whether it was Boolean, numeric, categorical or free text, would also improve the understanding of the data, including if and how the data were converted from one format to another.

As with the sources of data, more information about the data preprocessing stages could improve future manuscripts. A large proportion of the studies did not include information related to how variables were selected for analysis or how missing data were dealt with. In some cases, variable selection may be straightforward if the variables are limited to those collected as part of incident reporting. In any case, explicitly stating what variables were included and excluded would help the reader better understand the authors’ analytics process. Regarding missing data, there are a variety of approaches that can be implemented, including omitting files with missing data from the analysis or addressing them through imputation [73]. More than half of the included studies did not mention missing data, and of those that did only 11 described the specific procedures used. Given that missing data is an inevitability, particularly with data collected by managers and workers, how this is handled is an important part of the data preprocessing details that should be included.

There were a variety of analytics techniques implemented across the included studies, but often the reasons for selecting techniques were not described. Stating a rationale for the chosen methodology would assist others in deciding what approach to take with their own organizational data. Similarly, providing more detailed information about the software and packages utilized would provide further assistance to the research community. For example, providing a list of the R packages selected for data analysis would help others replicate an approach with their own data.

The analytics conducted in the included studies resulted in findings that often would not be obtained with approaches to injury data analysis that solely involve descriptive statistics. In many cases, more complex relations indicating the specific circumstances under which types of incidents occurred were detected and described. One potential area of improvement is to more clearly describe detected relationships among variables. Jargon that is specific to an organization may be difficult for the reader to interpret, and thus better-defined terms would improve understanding. This includes ensuring that all acronyms are spelled out and adequately defined within tables, figures and text. Further, studies often failed to describe their findings in great practical depth. While technique comparisons were common, these models were not compared to the base rate of safety predictions without modeling [45]. Studies also failed to report the practical application of their models; trained models were excellent at predicting another already-gathered dataset, but there was no mention of accident reduction when implemented at a facility [41,65].

Other studies generated conclusions using an inductive approach, with generation of their hypotheses post hoc [33]. While this method allows scientists to be more liberal in their investigations, a lack of a-priori hypothesis generation is a principal component that leads to reproducibility issues and to ‘just so’ explanations that overgeneralize results [74]. Once again, without follow-up studies confirming the reduction in outcome variables pre and post model integration, it is difficult to assess whether these inductive conclusions are valid.

Studies often attributed blame to safety observers, citing a lack of informative data entry [30,67] or data integrity issues related to a diversity of reporting methods and different storage mediums [63]. While there very well may be data integrity issues associated with method of entry, a productive safety culture dissuades blaming individuals, instead focusing on aligning goals [75].

Future studies can overcome the limitations described in the previous section with additional focus on prescriptive rather than simply predictive outcomes. Although models have shown success in predictive power, follow-up studies demonstrating reduction in real-world injuries would better inform practitioners in the field and demonstrate the utility of analytics. Model comparisons to more traditional analytics tools, such as multiple regression, could offer practitioners better perspective on the comparative advantages of more sophisticated models. Additionally, 37 out of 49 studies focused on steel manufacturing or construction. Expansion of these models to new sectors could both improve safety in those industries and possibly offer new insights into current predictive models.