What if workday loss in coal mines is not just a list of hazards, but a connected causal network? This study uses causal artificial intelligence to trace which factors most strongly shape injury-related absence. The study’s purpose is to establish the causal relationship network between various factors leading to workday loss in underground coal mines using a novel causal artificial intelligence method.

In other words, the target is not merely predicting absence, but mapping how risk factors connect to workday loss. Coal mine safety is a universally recognized global concern, and underground mining is considered a high-risk occupation in the United States.

Despite the overall improvement in safety, the mining industry continues to be associated with a high degree of uncertainty. Variations in operational risk stem from mine size, workforce size, coal production, and investments in monitoring and safety systems. Many risk assessment models have been developed to improve employee safety and identify potential hazards in the mining industry.

Earlier work used machine learning to categorize coal mines based on mining personnel injuries, and Bayesian networks to evaluate the risk of an underground gas explosion. A crucial capability is adapting and improving dynamically over time, so risk assessment can respond to emerging risks that may not have been previously observed or reported.

The National Institute for Occupational Safety and Health maintains records covering injury counts, mining site details, employment statistics, production figures, and work hours for every mine operating in the United States. These data are stored in two distinct databases: the Address/Job database and the Accident/Injury/Illness database.

The study uses a novel causal artificial intelligence method to establish the causal relationship network of risk factors affecting the duration of work absence among miners due to job-related injuries. Workday loss has been recognized as an indicator of mine risk, representing the duration during which miners are absent from work due to work-related injuries.

That makes the duration of absence a measurable outcome for studying mine risk. Earlier analysis of Mining Safety and Health Administration data spanning from 2003 to 2007 revealed a statistically significant association between age, total mining experience, and workday loss.

A review of nine mining industry articles identified antecedents that can impact workday loss, including the mining work environment, gender, age, mining equipment, organizational size, disease status, job training, recovery time, pre-injury health, and access to medical care.

The study uses an innovative causal artificial intelligence algorithm with NIOSH Address/Employment and Accident/Injury/Illness databases to establish a causal relationship network through a directed graph approach. The mine is treated as a fully connected system, with personal characteristics, working characteristics of the mine, and mine-specific employee attributes as variables of equal importance.

Causal discovery results are evaluated using adjacency precision, adjacency recall, arrowhead precision, and arrowhead recall, with the goal of identifying causes of employee sickness or injury absences and guiding coal mine risk assessment and risk management. The study used NIOSH address and employment records together with accident, injury, and illness records to analyze mining operations data.

The dataset included mine ID, county, workday loss, employee counts, employee hours, clean coal production, accident information, mining experience, job title, source of injury, degree of injury, and shift time hours. The analysis focused on coal mining operations from 1990 to 2020, excluding earlier years because of concerns about inconsistent reporting and technological advances that improved working conditions.

A total of 101,010 data entries were analyzed, each corresponding to annual employment information reported to NIOSH; mines with missing data or contractor-operated mines were excluded. During preprocessing, coal mining experts helped encode non-numeric variables into categorical variables.

The figure’s lower branch shows non-numeric data passing through encoding before joining the numeric-data stream as category data. At the all-numeric-data stage, the combined set enters expert review, and the final selected-data block visibly contains both the blue numeric portion and the green encoded portion.

Causal discovery, or network modeling, uses data to uncover the underlying causal relationship between variables within a dataset. Traditional causal discovery methods demand a large sample size and complete information about the variables, which can lead to inaccurate conclusions or illogical causal relationships in real-world datasets.

To overcome these limitations, the study introduces Grouped Greedy Equivalence Search, or GGES, which leverages prior knowledge to enhance identification of a local optimal solution. GGES is a score-based method that stems from the Greedy Equivalence Search method.

The method requires the dataset to satisfy the Markov condition and the faithfulness hypothesis. GGES calculates the score of the graphical model using the Bayesian Information Criterion score and a generalized score. Its implementation has three steps.

First, all variables are divided into two groups: causal variables and predict variables. The predicted variables are assumed not to influence other variables within their group and only to accept influence from other variables. The causal variables can produce and receive effects on each other, and the grouping requires prior knowledge about which variables represent the outcomes.

GGES starts with an initial empty graph and sets the score of the whole graph to zero. Edges are added from nodes sequentially, and the graph score is calculated and compared to select the model with the highest score. Nodes assigned to the outcome variable group are skipped according to the constraint placed at the beginning of the method.

Figure two outlines GGES as a sequential graph-search procedure. Variables are divided into causal nodes and a predictor node, then the method starts with an empty graph at score zero, adds edges one by one, and compares the resulting scores to select a highest-scoring graph.

It then deletes edges sequentially, again comparing scores, before producing the output causal graph; predictor-group nodes are skipped under the initial constraint. In the final GGES step, each edge is deleted in turn from the graph model with the highest score, and a new graph model is calculated.

The graph model with the highest graph score is selected as the final causal relationship network result. GGES constructs a causal relationship network showing connections between multiple independent variables and the actual workday loss experienced by mining employees.

The approach analyzes causal relationships between all variables and workday loss time. It then assesses variables directly associated with workday loss to discern the causality of variables exerting the most significant influence at each stage.

The study uses the intervention do-calculus adjustment score, or IDA score, to evaluate contributions between variables. By integrating independent component analyses with an overarching analysis, the final model provides a representation of the most suitable causal model.

This approach gauges the relative impact and interplay of variables influencing workday loss among mining employees. The model underwent validation using ten-fold cross-validation: the training dataset generated a causal relationship network, and the test dataset assessed the causal outcomes.

Evaluation used adjacency precision, adjacency recall, arrowhead precision, and arrowhead recall. The Total Causal Effects value described the magnitude of relationships between variables, with higher absolute Total Causal Effects values indicating stronger associations.

All statistics were performed in R three point four point one and Python, and a p-value less than zero point zero five was considered statistically significant. The dataset consisted of 101,010 data entries from 3,982 mines from 1990 to 2020, including 64,597 entries from 1990 to 1999, 22,009 from 2000 to 2009, and 14,404 from 2010 to 2020.

The final data cohort included 119 variables from 2,917 miners. After removing non-significant variables, 21 variables were identified, and 15 original data type variables were analyzed for data distribution. Using all available variables, the study employed GGES to construct a comprehensive causal relationship network.

After removing variables unrelated to workday loss, the result was a clear causal relationship network containing 17 variables. The identified direct causes included mean underground employees, mean employees per mine, mean office employees, source of injury, total mining experience, accident type, total mining experience in years, mean surface employees, job title, and county.

In the graph, workday loss is the yellow node at the bottom, and the incoming arrowheads around it show links arriving from several variables, including total mining experience, source of injury, and accident type. The dense web of directed arrows across the figure indicates that these relationships are embedded in a broader network rather than appearing as isolated pairings.

Total mining experience emerged as the most influential cause, while mean employees per mine exhibited the least influential impact on workday loss. In the TCE Value column, the visible entries range from negative zero point one nine three for mean employees per mine to negative one point eight seven two for mean surface employees.

The p-value column also separates the rows: most shown entries are below zero point zero five, while mean surface employees is listed at zero point six three zero. The evaluation produced an adjacency precision of zero point six four six, adjacency recall of zero point six two eight, arrowhead precision of zero point three six eight, and arrowhead recall of zero point three five four.

The causal relationship network from 1990 to 2005 was highly consistent with the causal relationship network from 1990 to 2020. For 1990 to 2005, the direct causes of workday loss were mean office employees, mean underground employees, source of injury, accident type, and total mining experience in years.

Look at the arrows converging on the workday loss node in the lower-left corner: the network shows incoming links from several variables, while most other nodes are connected through additional intermediate paths. The authors also report agreement measures of zero point six four six and zero point six two eight, alongside corresponding values of zero point three six eight and zero point three five four.

For the later period, the direct causes of workday loss included total mining experience, mean office employees, mean underground employees, county, and total mining experience in years. Total mining experience was again the most influential cause, while mean employees per mine was the least influential cause.

In the network, follow the arrowheads into the workday-loss node at the bottom: several paths arrive there from the lower-left and lower-right employee and experience variables, as well as from county. The dense web of curved and straight arrows also shows that some variables influence workday loss indirectly through intermediate nodes, rather than through a single isolated pathway.

The study developed a causal discovery algorithm called Grouped Greedy Equivalence Search and applied it to coal mine data across the United States using the NIOSH database. To account for the impact of coal mine safety implemented in 2006, the data were divided into pre-2006 and post-2006 periods.

This approach established a causal relationship network for employee workday loss and offered new perspectives for risk assessment and management strategies in the industry. The study has several limitations, including the fact that other potential factors affecting workday loss were not taken into account.

The impact of mining equipment and methods, such as the longwall method or hand method, was not considered because the information was difficult to obtain and differentiate. The data spanned 31 years, during which changes in regulation and safety practices may have affected accuracy and completeness, and incomplete employee injury reports were discarded when mines failed to report employment or accident information.

For categorical variables not represented by data, expert opinions were used to reduce the number of categories, which may still affect the results. Despite these limitations, the study reports that it captured causes of workday loss involving individual employee characteristics, mine job characteristics, and employee mine characteristics.

The network identifies total mining experience as the most influential cause of workday loss, while also showing that mine structure, injury source, accident type, job title, and location matter. The result is useful, but limited by missing factors and incomplete reporting.