Developing an Integrated Construction Safety Management System for Accident Prevention

Risk assessment is a common practice involving the identification of potential hazards and analysis of potential scenarios that could lead to their occurrence. In industries, this process involves recognizing industry-specific risk factors and evaluating the likelihood (frequency) and severity (intensity) of injuries or illnesses caused by these factors. Within the construction industry, the risk assessment similarly focuses on evaluating the occurrence likelihood as frequency, the potential financial damage, or the severity of accidents as intensity (Choudhry et al. 2007; Leu and Chang 2013; Mohamed 2002; Perlman et al. 2014). Leading countries adopt such principles to proactively pinpoint risk factors on construction sites and establish guidelines for addressing potential accidents during construction operations (Halabi et al. 2022; Searle 2023; Suh 2021). These guidelines offer a basis for field inspectors to assess potentially hazardous construction sites in advance. However, existing risk assessment practices often need more reliance on actual accident data. Instead, they heavily depend on designers or clients to subjectively evaluate risks or defer to expert opinions. This lack of objective criteria for field inspectors to evaluate potentially hazardous sites in advance can decrease inspection efficiency and impede effective accident prevention.

Against this background, studies employing actual data to assess risk have also been conducted (Forteza et al. 2017; Fung et al. 2012; Jeong and Jeong 2022; Ryu et al. 2021; Zolfagharian et al. 2014). Fung et al. (2010) developed a risk assessment model (RAM) to assess risk levels at different project stages. RAM utilized data from a single local project and assessed the risk levels of project stages based on the number of workers and accident frequencies at each stage. Jiang et al. (2023) developed a risk assessment model using specific road construction data in China to identify key factors influencing road construction risks. Wuni et al. (2023) classified risk factors affecting modular construction through a literature review and analyzed the impact of these risk factors to rank risks. Chellappa and Salve (2023) assessed risk rankings for scaffold work tasks through fall hazard risk assessment. These studies identified major risk factors through data analysis under specific construction conditions.

Regardless of such research benefits, field inspectors still need to evaluate potentially hazardous sites across a range of construction sites with varying characteristics. Unfortunately, a comprehensive risk assessment methodology for diverse construction sites has not been developed. Therefore, the first objective of this study is to devise a methodology for calculating risk levels of construction sites based on accident data from various construction sites. This methodology will allow for a comprehensive risk-level assessment based on site-specific characteristics. By utilizing the developed methodology, the risk levels of construction sites can be identified, and urgent inspection sites can be determined through risk scores.

Research on preventing construction accidents through the analysis of construction accident data has been pursued in various ways. Initial studies mainly focused on statistical methods to identify factors causing construction accidents (Cheng et al. 2010; Chong and Low 2014; Molenaar et al. 2009; Shapira et al. 2009; Wu et al. 2015). These studies used basic statistical analysis to examine accident frequency based on factors such as worker age, contractor size, and the day of the week to determine the main factors influencing accidents. For instance, López Arquillos et al. (2012) found that most accidents occurred on specific days of the week. These simple statistics often prove overly generalized and insufficient for making a substantial contribution to accident prevention. Thus, other studies introduced more detailed statistical analyses. Molenaar et al. (2009) used structural equation modeling (SEM) to suggest guidelines for construction stakeholders to prevent accidents by analyzing the paths from sources to events. Cheng et al. (2012) utilized a classification and regression tree (CART) to identify factors highly correlated with construction accidents, successfully distinguishing accident situations in which falls and trips occur frequently. Although these studies are valuable for identifying critical factors in construction site accidents, they focus primarily on general accident factors. Such a focus is inadequate for the active prevention of accidents at construction sites, which vary widely in characteristics and, therefore, may not be as practical for on-site application by practitioners.

Some other researchers have conducted accident prevention studies by predicting accidents using machine-learning algorithms (Cheng et al. 2020; Poh et al. 2018; Sadeghi et al. 2020). Goh et al. (2018) conducted a survey targeting 80 workers at a tunnel construction site and analyzed the data using the random forest (RF) algorithms. Through this analysis, they revealed that the most significant factor influencing workers’ unsafe behavior is the belief about social norms. The limitation of their study, though, lies in the reliance on survey data from workers rather than actual accident data. Poh et al. (2018) developed a machine-learning model using RF to calculate accident risks using data collected from construction companies in Singapore. Nonetheless, the reliance on data from a single company diminishes the reliability of the results, and the model needs to explain the causes of accidents, making it challenging for practitioners to utilize the outcomes effectively. Xia et al. (2024) aimed to prevent construction accidents through a cost prediction model based on the eXtreme gradient boosting (XGBoost) algorithm, and Choi et al. (2020) developed machine-learning models for predicting fatal accidents using data obtained from construction workers. Kang and Ryu (2019) proposed a RF model to predict the types of accidents likely to occur based on data collected from construction sites. However, they had difficulties in providing safety managers with information about the specific objects that might cause accidents.

To comprehensively prevent accidents on construction sites, it is crucial to analyze not only the objects that act as contributing factors to accidents but also the types of accidents they cause and the severity of those accidents. Construction site managers often receive only partial information on these factors, making accident prevention still challenging. Therefore, this study aims to address these research gaps and enhance accident prevention efforts on construction sites by constructing a model to predict “accident-causing objects,” “accident types,” and “accident severity” as the second specific objective. These two specific objectives are based on the assumption that construction site practitioners will utilize them. Therefore, the third specific objective is to develop a web-based prototype that incorporates these two objectives and validate its practical applicability on construction sites.

Clear definitions are needed for the classes of the accident-causing object and accident-type variables to be utilized as objective variables. In this study, with reference to relevant existing studies, predictive variables are defined with eight classes (temporary facility, materials, tools, machinery, construction member, structure, soil and rock, and others) for accident-causing objects and eight classes (tripped, fell, struck by an object, pinched, slashed, hit, crushed, and others) for accident types (Hwang et al. 2023; Johansen et al. 2023; Kim et al. 2023). However, the preprocessed data output from Section 3.1 exhibit an issue of data imbalance. As can be seen from Table 2, there is a significant disparity in the proportions of data composing each model when comparing the count of predictive variables. This data imbalance could impact the efficacy of the model, making it imperative to achieve data balance.

Oversampling techniques are commonly employed to prevent algorithm degradation (Chang et al. 2023; Shi et al. 2022). This study utilizes the synthetic minority oversampling technique (SMOTE), a prevalent oversampling method designed to solve the problem of overfitting (Douzas and Bacao 2019; Feng et al. 2021; Li et al. 2023). The sampling_strategy parameter, which represents the ratio between the major and minor classes, is set considering the imbalance level of each dataset. For the accident-causing object model and the accident type-model, the sampling_strategy is set to 1.0 to ensure an adequate number of data points for each class. On the other hand, the death-injury model, where the data ratio between the two classes differed by roughly 20 times, has the sampling_strategy set to 0.3 to mitigate overfitting. The data sampling is performed using Python version 3.6.8 and imbalanced-learn library version 0.7.0. Following this, the data is split into training (70%) and testing (30%) datasets. Through this process, a dataset is constructed to be used for the training of each submodel. Table 3 summarizes the results for the training and testing datasets for each submodel.

The results of the sensitivity analysis for the selection of major risk factors are shown in Table 7, and, according to this, the sensitivity ranks for each factor were derived. For example, the sensitivity analysis for ‘facility type’ initially calculated the standard risk scores for all combinations with ‘construction type (minor).’ Subsequently, the calculated standard risk scores were sorted in ascending order, as shown in Fig. 2, from which $Q_3$ and $Q_1$ were obtained, and the sensitivity was calculated. In conclusion, excluding the benchmark variable ‘construction type (minor),’ the major risk factors selected with high sensitivity were ‘facility type’ at 0.264, ‘ordering organization’ at 0.246, ‘construction cost’ at 0.238, and ‘safety management plan’ at 0.220. Therefore, in this study, these five variables were decided as the main risk factors for determining the site’s classification. The site types were categorized into a total of 1,429 with the combinations of the major risk factors.

Risk scores for each site classification were determined according to the combination of five variables, as shown in Table 8. The analysis results of the top 100 site classifications in terms of risk scores for each risk factor are as follows. In the case of ‘construction type (minor),’ ‘reinforced concrete’ accounted for the largest proportion at 19%, followed by ‘temporary facility’ at 13%, and ‘dismantling and demolition’ at 9%. For ‘facility type,’ ‘building’ occupied 60% of the ratio, and ‘civil’ was in second place at 35%. In the ‘safety management plan,’ ‘nontarget’ was 58%, and ‘target’ showed a slight difference at 42%. In the ‘construction cost,’ the range from ‘$1 million to $5 million’ accounted for 26%, ‘$100,000 to $500,000’ was 21%, and ‘$500,000 to $1 million’ was 11%. Finally, for the case of the ‘ordering organization,’ ‘private construction’ was 54%, and ‘public construction’ was 46%.

The top three ‘construction types (minor)’—reinforced concrete construction, temporary construction, and dismantling and demolition construction—are indeed categorized as hazardous tasks at actual construction sites (Hong et al. 2019; Park et al. 2015). Furthermore, in practice, the size of a site is typically classified by construction cost, and it is known that smaller sites, which lack a sufficient number of site managers and are challenging to supervise, have a higher accident occurrence rate (Lim et al. 2018; Sodangi 2023). The analysis results also showed that small-scale sites with a construction cost of fewer than $5 million have high-risk levels, demonstrating that the risk analysis results of this study accurately reflect the actual situation on site.

In summary, the sensitivity analysis conducted in this study identified key risk factors influencing site safety, with ‘construction type,’ ‘facility type,’ ‘ordering organization,’ ‘construction cost,’ and ‘safety management plan’ emerging as major determinants. The subsequent classification of site types based on these risk factors further enhanced our understanding of site-specific risk profiles. Moreover, the analysis of risk scores for each site classification revealed insights into the distribution of risk factors across different construction contexts. Overall, these findings will provide valuable insights for construction management practitioners and safety managers in developing proactive risk mitigation strategies customized to specific site characteristics.

The optimal hyperparameters for each of the four algorithms of each submodel are presented in Table 9. The optimal hyperparameters for submodels were identified as follows. In the accident-object prediction model, the XGBoost algorithm showed optimum performance with a maximum depth of 9, minimum child weight of 3.0, colsample by tree of 0.75, and 200 estimators. CatBoost performed best with a maximum depth of 11, a learning rate of 0.08, and 250 estimators. LGBM achieved optimum results with a maximum depth of 9, 13 leaves, a learning rate of 0.04, and 200 estimators. At the same time, the RF showed the best performance with a minimum sample split of 3, minimum sample leaf of 3, and 300 estimators. For the accident-type prediction model, the XGBoost achieved optimal results with a maximum depth of 11, minimum child weight of 1.0, colsample by tree of 0.75, and 300 estimators. CatBoost had optimal performance with a maximum depth of 11, a learning rate of 0.03, and 250 estimators. The LGBM model showed optimum results with a maximum depth of 11, 13 leaves, learning rate of 0.05, and 200 estimators, while the RF performed best with a minimum sample split of 3, minimum samples leaf of 3, and 300 estimators. Finally, for the death-injury model, the optimal hyperparameters were a maximum depth of 7, minimum child weight of 2.0, colsample by tree of 0.75, and 300 estimators for XGBoost; a maximum depth of 7, learning rate of 0.03, and 200 estimators for CatBoost; a maximum depth of 9, 11 leaves, learning rate of 0.1, and 200 estimators for LGBM; and a minimum sample split of 3, minimum samples leaf of 3, and 200 estimators for RF.

Subsequently, the submodels were trained with the optimal hyperparameters, and the weighted average F1 score was calculated. In the accident-causing object model, XGBoost exhibited the best performance with an F1 score of 0.839, followed by CatBoost with 0.831, RF with 0.829, and LGBM with 0.825. In the accident-type model, XGBoost showed the highest performance with an F1 score of 0.749, followed by LGBM with 0.745, CatBoost with 0.742, and RF with 0.736. Finally, in the binary classification model, death-injury, XGBoost performed best with an F1 score of 0.977, followed by CatBoost with 0.968, RF with 0.963, and LGBM with 0.962. The results are summarized in Table 10 and Fig. 3.

The confusion matrix and performance of the accident-causing object model are shown in Table 10. While it generally showed high performance, the recall values for ‘temporary facility’ and ‘materials’ were notably low, at 0.54 and 0.67, respectively. These two variables often get confused with each other, even by humans, affecting prediction performance. Of the actual accidents caused by ‘temporary facility,’ 144 cases were predicted as ‘materials,’ and of the actual accidents caused by ‘materials,’ 119 cases were predicted as ‘temporary facility.’ Additionally, 97 actual accidents caused by ‘temporary facility’ were incorrectly predicted as ‘others.’ For example, there was an accident where a worker fell from the system scaffolding while trying to avoid a falling Euroform during the scaffolding dismantling process. This accident was incorrectly recorded as being caused by a ‘temporary facility,’ but, in reality, it was an accident caused by the ‘materials,’ which is the Euroform. There was also the case where accidents caused by ‘tools’ were inaccurately predicted as those attributed to ‘temporary facilities.’ To illustrate, during a scaffold clamp-cutting operation, a worker’s finger was severed by a hand grinder. Even though this was clearly an accident caused by a tool, it was predicted as an accident by a temporary facility. This could be because multiple factors can influence a single accident, making it harder for the model to predict the main cause accurately.

In the accident-type prediction model, three accident types—’tripped,’ ‘fell,’ and ‘struck by object’—showed low recall values of 0.45, 0.52, and 0.59, respectively, as presented in Table 11. Unlike other variables, these three often got confused with each other due to the similarities in the accidents they represent. For instance, an accident where a worker trips over an object on the floor and falls off a building could involve both ‘tripped’ and ‘fell’ accident types. An accident where a worker trips and injures an ankle after being struck by a collapsed structure was reported as ‘struck by object,’ despite also involving the ‘tripped’ accident type. Another accident where a worker, who was not wearing a safety belt, fell to the ground while installing a system scaffold involved both ‘tripped’ and ‘fell’ accident types but was only entered as ‘fell.’ These problems arise because the accident data entry system only allows one accident type to be selected. This highlights a need for improvements in the data entry system. Possible solutions could be to unify the confusing variables into one or establish guidelines that clearly specify the accident type for each accident.

Finally, the death-injury prediction model performed well, as shown in Table 12. However, given the approximately 20-fold difference in the number of instances for two variables in the original data, it is difficult to eliminate the problem of overfitting completely, even after performing sampling to address this imbalance and achieving very high F1 score. Nonetheless, considering that construction accident history data started to be collected from 2019, the number of data points is relatively limited at this point. With the expectation of collecting more data in the future, this issue is anticipated to be naturally resolved.

In summary, three submodels exhibited excellent performance, yet several issues need to be resolved. Both accident-causing object and accident-type models had ambiguities in the prediction of certain variables. Notably, in the accident-type prediction model, considerable confusion was observed with some variables, which highlighted problems in the data entry process. Thus, there is a need to improve the data entry process for the construction accident data being collected. It is also essential to provide better data entry guidelines to accident reporters to improve the quality of the data. In the death-injury model, one downside that was identified was the short data collection period, which led to an insufficient amount of data. This issue can be resolved as the data volume increases with more collection over time.

To validate and ensure the practical utility of the initial design and functionalities of the construction safety management system, feedback was collected from construction practitioners. Most practitioners were very positive about the potential on-site application of the system proposed in this study. However, they emphasized that there were areas needing improvement for actual site implementation. The key opinions and requirements derived from the interviews to enhance the system’s usability on-site are as follows: (1) a feature that allows for an overall view of the data utilized by the system is needed; (2) a ‘filtering’ feature that enables users to select and input variables is necessary since manually typing in data is inconvenient for practitioners; and (3) visualization of prediction results is needed to check the outcomes of the accident prediction model easily. Therefore, reflecting the practitioners’ feedback, a user interface (UI) and user experience (UX) for a web-based prototype of the safety management system were developed. The prototype was created via Tomcat, an open-sourced web application framework based on Java. To make it easy for construction site workers to use, the prototype offers several features developed based on research, with key features including accident data exploration, construction site risk assessment, and onsite accident prediction. The functionalities of the developed system prototype were introduced, and to verify the suitability of each function in real-world settings, comparisons and validations were conducted between the results obtained through variable inputs and actual data.

The accident data exploration interface displays the CSI accident history data collected up to the present in an accessible manner, as shown in Fig. 4. The “(1) filtering” function is designed to allow users to select variables directly, enabling them to search for the data they want. The “(2) accident history data browsing” function presents the accident history corresponding to the category selected by the user. As the data comprises a plethora of variables, the design is such that only representative variables are shown to enhance user convenience. The “(3) site description” function enables users to view detailed information about the site where the accident occurred when they select data marked in yellow in the data browsing function. Details include the ordering organization, facility type, construction type, location, and number of workers. The “(4) accident description” function allows users to view detailed information about the accident. Details include the date of the accident, time of the accident, object causing the accident, type of accident, number of fatalities, and number of injuries.

The construction site risk assessment interface displays the results of a risk analysis based on user input, as seen in Fig. 5. In the “(1) input variable selection,” users can choose the detailed categories of the site they are interested in, with the variables composed of the five major risk factors identified in this study. The “(2) risk assessment result” displays the risk analysis results for the site categories selected by the user. The risk analysis results provide a score based on the overall site and indicate its rank out of 1,429 site categories. Additionally, suppose the site in question falls into the category of small-scale sites with low construction costs. In that case, the system also provides its ranking among site categories with construction costs of less than $5 million. This design allows construction practitioners to more easily determine the risk of the site. The “(3) top 20 risk categories” provide information on the top 20 risk categories in terms of risk score. This feature can be used as a supplementary tool to support safety inspections at actual construction sites. For instance, if there are 10 sites requiring inspection, the risk rankings of these sites can be calculated by inputting their site classifications, helping to prioritize urgent inspections. This can enhance the efficiency of site inspection practitioners and contribute to effective accident prevention.

For comparison and verification with actual data, accidents corresponding to the same site category as shown in Fig. 5 (ordering organization = private; facility type = building; safety management plan = nontarget; construction $\mathrm{cost}=\$1–\$5\mathrm{million}$; construction type (minor) = reinforced concrete) were sorted. Out of a total of 9,671 accident cases, there were 91 cases classified under this category, involving one fatality and 90 injuries. Considering a total of 1,429 site categories, 91 cases represent a very high frequency. However, considering that there was only one fatality out of 91 accidents, it indicates a relatively low severity of accidents. This aligns with the results of the analysis in the system prototype (standard $\mathrm{frequency}=0.827$; standard $\mathrm{damage}=0.054$). These findings support the reliability of the developed construction site risk assessment functionality, along with its suitability for real construction sites. Such practicality helps aid construction practitioners in making informed decisions about site safety. From academic perspectives, the interface incorporates a risk assessment methodology to generate risk scores and rankings for various site categories while allowing quantitative assessments of safety on specific site conditions. This supports facilitating comparative analyses and benchmarking studies across difference construction contexts.

The construction site accident prediction interface displays construction accident prediction results based on user input, as depicted in Fig. 6. In the “(1) input variable selection,” users can enter detailed information about the construction site, which is organized around variables used in model training. Given the abundance of variables, it could be more inefficient for users to type in the variables manually. Therefore, the interface was designed to categorize these detailed variables, facilitating an easier selection process for the users. “(2) Accident prediction result” provides the probability of an accident occurring in accordance with the three submodels developed in this study: accident-causing object, accident type, and death-injury. In construction sites, identifying multiple risk factors in advance can significantly influence accident prevention. Thus, the system determines the accident-causing object and accident type, providing predictive probabilities for each accident cause. This approach provides a total of four accident scenarios and the corresponding probabilities of death and injury for each scenario. Additionally, the right side of the screen explains detailed results through statistical analysis of the predicted risk objects. For instance, if the prediction result were ‘construction materials,’ then the interface would illustrate statistical rankings for the detailed objects belonging to the subcategory of construction materials, such as rebar, pins, and deck plates. In the “(3) result visualization,” the previous prediction results are visualized in the form of a graph. Through this feature, practitioners can check the risky objects, accident types, and the probabilities of death and injury and can predetermine the risk of an accident. Additionally, because the input of variables is flexible, it is possible to identify in detail the hazardous objects and types of accidents for various tasks within a single site. This allows site managers to conduct customized training for workers involved in each task.

The practical usefulness of the construction site accident prediction feature was validated by testing it with random variables, as direct comparison with actual data is not feasible due to its predictive nature. A total of 100 tests were conducted, resulting in an average response time of 2.32 s. Generally, website users do not experience discomfort with response times below 2.5 s (Equation Research 2011). Therefore, the construction site accident prediction feature meets the acceptable standard. Considering that a predictive model is embedded, this also indicates that the developed prototype demonstrates good usability. Additionally, there was no instance of malfunction or incorrect output, and the visualization results for each scenario were accurately displayed. Through this process, the utility of the feature in real-world scenarios was validated.

The development and implementation of the accident prediction interface not only offer practical benefits for construction management but also make significant academic contributions to the field of construction safety research. The interface’s user-friendly design, with categorized input variables and intuitive visualization features, improves a better understanding of the underlying mechanisms of accidents in construction sites. Researchers can leverage the interface to conduct in-depth analyses of accident patterns and trends, leading to the development of more effective safety management strategies and policies. Furthermore, the interface’s practical usefulness through rigorous testing and performance evaluation contributes to the validation of predictive modeling techniques in real-world construction scenarios. Overall, the findings of this study serve as a valuable tool for both practitioners and researchers, offering insights, solutions, and opportunities for advancing knowledge and practices in construction safety management.