Journal of Safety Research最新文献

{"title":"A large language model framework to uncover underreporting in traffic crashes","authors":"Cristian Arteaga,&nbsp;JeeWoong Park","doi":"10.1016/j.jsr.2024.11.009","DOIUrl":"10.1016/j.jsr.2024.11.009","url":null,"abstract":"<div><div><em>Introduction:</em> Crash reports support the development of traffic safety countermeasures, but these reports often suffer from underreporting of crucial crash factors due to miscoded entries during data collection. To rectify these issues, the current practice relies on manual information rectification, which is time consuming and error prone, especially with large data volumes. To address these hurdles, we develop a framework to analyze traffic crash narratives and uncover underreported crash factors by capitalizing on the capabilities of Large Language Models (LLM). <em>Method:</em> The framework integrates procedures for prompt definition, selection of LLM generation parameters, output parsing, and underreporting determination. For evaluation, we present a case study on identification of underreported alcohol involvement in traffic crashes. We investigate the framework’s identification accuracy in relation to different underlying LLMs (i.e., ChatGPT, Flan-UL2, and Llama-2), prompt framings (i.e., explicit vs. implicit matching), and generation parameters (i.e., sampling temperature and nucleus probability). Our validation dataset consists of 500 crash reports from the State of Massachusetts. <em>Results:</em> Analysis results demonstrate that the developed framework achieves a recall and precision of up to 1.0 and 0.93, respectively, indicating a successful retrieval of underreported instances. These findings indicate that the developed framework addresses a critical gap in the existing traffic safety analysis workflow by enabling safety analysts to uncover underreporting in crash data efficiently and accurately, without the need for extensive expertise in natural language processing. <em>Practical Applications:</em> Thus, the developed approach offers unprecedented opportunities to maximize the quality and comprehensiveness of traffic crash records, paving the way for more effective countermeasure development.</div></div>","PeriodicalId":48224,"journal":{"name":"Journal of Safety Research","volume":"92 ","pages":"Pages 1-13"},"PeriodicalIF":3.9,"publicationDate":"2024-11-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142662539","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Drivers’ long-term crash risks associated with being ticketed for speeding","authors":"Darren Walton ,&nbsp;Ross Hendy","doi":"10.1016/j.jsr.2024.10.009","DOIUrl":"10.1016/j.jsr.2024.10.009","url":null,"abstract":"<div><div><em>Introduction</em>: This research analyzes the relationship between police-issued tickets for speeding and the crash risk of those drivers, in New Zealand, between 2015–2019. <em>Method</em>: The main data are constructed through data-matching license details of crash outcomes with all officer-issued tickets for speeding between 2015–2016 (N = 534,935). The sub-group of drivers that accumulate tickets is compared to a coarsened exact matched set of drivers of the same age. <em>Results:</em> There is a strong relationship between the number of tickets a person has in a two-year period (2015–16) and the likelihood of a crash outcome (2017–2019). However, the accumulation of tickets is not the best predictor of crash likelihood. A combination of the excess in speed <em>and</em> the accumulation of tickets increases the relative odds of a subsequent crash. These results are discussed considering the threshold at which New Zealand criminalizes alcohol-relating offending (notionally 4.2 times the base rate crash risk). The same rate of elevated crash risk exists when a driver has one ticket for being 10 km/h over the speed limit and has another speeding ticket within two years.</div></div>","PeriodicalId":48224,"journal":{"name":"Journal of Safety Research","volume":"91 ","pages":"Pages 431-436"},"PeriodicalIF":3.9,"publicationDate":"2024-11-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142653096","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Factors influencing behavioral intentions to use conditionally automated vehicles","authors":"Sjaan Koppel ,&nbsp;David B. Logan ,&nbsp;Xin Zou ,&nbsp;Fareed Kaviani ,&nbsp;Hayley McDonald ,&nbsp;Joseph F. Hair Jr ,&nbsp;Renée M. St. Louis ,&nbsp;Lisa J. Molnar ,&nbsp;Judith L. Charlton","doi":"10.1016/j.jsr.2024.10.006","DOIUrl":"10.1016/j.jsr.2024.10.006","url":null,"abstract":"<div><div><em>Background:</em> This study explored factors influencing the acceptance of conditionally automated vehicles among Australian drivers by extending the Technology Acceptance Model with the Technology Readiness Index. <em>Method:</em> Data from an online survey of 844 participants were analyzed using partial least squares structural equation modeling (PLS-SEM). <em>Results:</em> Perceived usefulness had the strongest direct effect on behavioral intention (0.469, p &lt; 0.001), followed by attitude (0.318, p &lt; 0.001). Innovativeness positively influenced behavioral intention (0.183, p &lt; 0.001), while insecurity had a negative impact (−0.071, p &lt; 0.01). Optimism and discomfort were not significant. Perceived usefulness also had significant indirect effects through attitude (0.156, p &lt; 0.001) and trust (0.072, p &lt; 0.001). Perceived ease of use indirectly influenced behavioral intention through perceived usefulness (0.306, p &lt; 0.001), attitude (0.102, p &lt; 0.001), trust (0.047, p &lt; 0.001), and their combinations. Trust indirectly affected behavioral intention via attitude (0.130, p &lt; 0.001). Perceived security and privacy risks had indirect negative effects through trust and attitude (−0.035, p &lt; 0.001; −0.005, p &lt; 0.05). <em>Conclusion:</em> These results suggest that fostering acceptance among less tech-savvy individuals may help promote positive attitudes, increase conditionally automated vehicle adoption, and potentially enhance road safety. <em>Practical implications:</em> These findings suggest a need for targeted programs to enhance perceived usefulness and trust while addressing security and privacy concerns, ultimately contributing to safer road systems through the adoption of conditionally automated vehicles.</div></div>","PeriodicalId":48224,"journal":{"name":"Journal of Safety Research","volume":"91 ","pages":"Pages 423-430"},"PeriodicalIF":3.9,"publicationDate":"2024-10-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142532197","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Recent Latino immigrants to Miami-Dade County, Florida: Impaired driving behaviors during the initial years after immigration and the pandemic lockdown","authors":"Eduardo Romano ,&nbsp;Mariana Sanchez","doi":"10.1016/j.jsr.2024.09.009","DOIUrl":"10.1016/j.jsr.2024.09.009","url":null,"abstract":"<div><div><em>Introduction</em>: Typically, recent Latino immigrants (RLIs) experience a decline in driving while impaired (DWI) rates soon after immigration, largely due to limited access to vehicles. Such a transitional period offers a window of opportunity for intervention for RLIs at risk of engaging in DWI and riding with an impaired driver (RWID). This manuscript examines the rates of DWI, RWID, and driving while impaired by drugs (DWID) among RLIs upon arrival to Miami/Dade County (MDC), Florida. <em>Methods:</em> Collected between 2018 and 2021, data originates from a longitudinal study examining self-reported drinking and driving trajectories among 540 RLIs to MDC. At baseline retrospective pre-immigration data were obtained simultaneously with first-year post-immigration data. Two follow-up surveys conducted one year apart (N=531 and N=522), collect data on RLIs initial 3 years in the United States. <em>Results:</em> Pre- to post-immigration trajectories for mean number of drinks per month (d/m) revealed a “U-shaped” curve: 18.3 d/m, 13.9 d/m, 10.4 d/m, 12.9 d/m, and 16.4 d/m, from pre-immigration (T0), first year (T1), second year before COVID (T2-BC) and during the pandemic lockdown (T2-DC), and third year in the United States (T3). The use of illicit drugs showed a constant decline, from 14.6% at T0 to 2.1% at T3. The prevalence of DWI at T1 was significantly lower compared to rates in the country of origin (T0) and continued declining through T3. DWID rates remained low across the assessment period. RWID was significantly more prevalent than DWI across all study time points. C<em>onclusions:</em> Although the relatively low prevalence of DWI, drug use, and DWID among the RLIs during their initial years in the United States is encouraging, the surge in alcohol use at T3 warns about the need for interventions to prevent increases in DWI. <em>Practical applications:</em> Findings from the present study point to an opportunity to develop early interventions to prevent the escalation of impaired driving among RLIs to MDC.</div></div>","PeriodicalId":48224,"journal":{"name":"Journal of Safety Research","volume":"91 ","pages":"Pages 401-409"},"PeriodicalIF":3.9,"publicationDate":"2024-10-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142532195","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A vehicle occupant injury prediction algorithm based on road crash and emergency medical data","authors":"Tetsuya Nishimoto ,&nbsp;Kazuhiro Kubota ,&nbsp;Giulio Ponte","doi":"10.1016/j.jsr.2024.09.015","DOIUrl":"10.1016/j.jsr.2024.09.015","url":null,"abstract":"<div><div><em>Introduction</em>: Advanced Automatic Collision Notification (AACN) systems are an automobile safety technology designed to reduce the number of fatalities in traffic accidents by optimizing early treatment methods. AACN systems rely on robust injury prediction algorithms, however, despite the importance of time to treatment, current injury prediction algorithms used in AACN systems do not take this critical time period time into consideration. <em>Method</em>: This study developed a vehicle occupant injury prediction algorithm by using emergency transport time in addition to mass crash data, to determine the risk of serious injury for vehicle occupants in a road crash. Two sources of de-identified data were used: The South Australian Traffic Accident Reporting System (TARS) database and the highly detailed South Australian Serious Injury Database (SID). Firstly, the TARS data, a large statistical crash dataset, was imputed into a logistic regression analysis to produce a base injury prediction algorithm. The important effect of emergency transport time on the risk of death and serious injury was then independently quantified as an odds ratio (OR) from the SID. The ORs were converted into regression coefficients and subsequently introduced into the base injury prediction algorithm to produce an enhanced injury prediction algorithm. <em>Results</em>: The ORs calculated from the SID showed that the risk of death and serious injury increased with increasing transport time: 61–90 min (OR = 1.6), 91–120 min (OR = 3.3), and &gt; 120 min (OR = 4.9), compared to a transport time of 60 min or less. An assessment of the base algorithm compared to the enhanced injury prediction algorithm through Receiver Operating Characteristic (ROC) analysis, demonstrated a prediction accuracy improvement from AUC 0.70 to AUC 0.73 when evaluating the respective algorithms. The injury prediction calculations indicate that the impact of two risk factors, transport time and age-related decline in human injury tolerance, are significant, and both have a strong influence on the increased risk of serious injury. <em>Conclusions</em>: The impact of emergency transport time on the risk of fatal and serious injuries was determined from a relatively small, but data rich SID. Subsequently this was incorporated into an injury prediction algorithm constructed from the large (TARS) statistical crash data set to produce an enhanced injury prediction algorithm. <em>Practical Application</em>: By adding the effect of transport time to enhance the basic injury prediction algorithm, an AACN that incorporates such an algorithm can be used to determine the probability of death or serious injury due to delayed treatment. Further, such a system can be used to improve policies and procedures to optimize emergency transport time.</div></div>","PeriodicalId":48224,"journal":{"name":"Journal of Safety Research","volume":"91 ","pages":"Pages 410-422"},"PeriodicalIF":3.9,"publicationDate":"2024-10-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142532196","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Fatal injuries among landscaping and tree care workers: Insights from NIOSH and state-based FACE reports","authors":"Gregory D. Kearney ,&nbsp;Nancy Romano ,&nbsp;Anna Doub","doi":"10.1016/j.jsr.2024.10.005","DOIUrl":"10.1016/j.jsr.2024.10.005","url":null,"abstract":"<div><div><em>Context:</em> A comprehensive assessment of the National Institute for Occupational Safety and Health (NIOSH) and State-based Fatal Assessment and Control Evaluation (FACE) investigative reports involving landscaping and tree worker fatalities have not been fully examined. <em>Methods:</em> Narrative text from 93 FACE reports from 1987 to 2023 involving landscaping and tree care workers were reviewed, manually coded and analyzed on major variables. Univariate analyses was conducted to summarize results of decedent workers and workplace characteristics. <em>Results:</em> Among the total number of worker fatalities (n = 95), the most commonly reported incidents were, electrocutions from power lines (18.3%), falls from trees (16.1%), and incidents involving a worker being either caught, pulled, or dragged into wood-chipping machine (12.9%). More than 66.0% of fatal incidents occurred among tree care workers that had been on the job for one year or less. Among reports, 60.2% of employers lacked a written safety plan, and 34.4% did not provide job training to their workers. <em>Conclusions:</em> FACE case reports alone are not a valid measure of workplace fatalities. Nevertheless, the codification and descriptive summary of more than three decades of case reports increases understanding of circumstances and contributing risk factors associated with these tragic, and yet largely preventable incidents. A comprehensive approach is urgently needed that includes: (a) taking immediate action to reduce occupational risks while cultivating a robust safety culture across the industry, and (b) increasing research to evaluate the effectiveness of interventions and prevention measures. <em>Practical Application:</em> The interconnectedness of safety challenges requires a multi-faceted approach that includes addressing issues related to new and diverse workers, employer commitments to the implementation of safety plans, and comprehensive training and mentorship programs. Intervention strategies and implementation measures are essential to diminishing fatalities in these high-risk jobs.</div></div>","PeriodicalId":48224,"journal":{"name":"Journal of Safety Research","volume":"91 ","pages":"Pages 393-400"},"PeriodicalIF":3.9,"publicationDate":"2024-10-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142532254","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"The aggressive driving performance caused by congestion based on behavior and EEG analysis","authors":"Shuo Zhao,&nbsp;Geqi Qi,&nbsp;Peihao Li,&nbsp;Wei Guan","doi":"10.1016/j.jsr.2024.10.004","DOIUrl":"10.1016/j.jsr.2024.10.004","url":null,"abstract":"<div><div><em>Introduction:</em> Traffic congestion is closely related to traffic accidents, as prolonged traffic congestion often results in frustration and aggressive behavior. Moreover, in daily commuting, drivers often have to pass through multiple congested road sections, and aggressive driving performance due to exiting or re-entering traffic jams has rarely been analyzed. <em>Method:</em> To fill this research gap, we designed an intermittent traffic congestion scenario using a driving simulator and employed unsupervised learning algorithms to extract high-level driving patterns gathered with EEG data to investigate the continuous effects of traffic jams, particularly when drivers exit and re-enter traffic jam conditions. <em>Results:</em> We discovered that drivers, upon exiting congested areas, engage in abrupt braking with a decrease in braking time of approximately 0.47 s and smooth lane changes with an increase in lane change time of approximately 0.5 s to maintain high-speed driving conditions. When drivers re-enter a traffic jam, they exhibit more abrupt stop-and-go behaviors to escape the traffic jam. The results of the risk assessment of driving behavior indicated that after leaving congested areas, free-flow segments have greater risk factors than other segments. Electroencephalogram (EEG) data were analyzed to identify instances of mind-wandering when a driver transitions into free-flowing segments, followed by a substantial increase in brain activity upon re-entry into congested traffic conditions. <em>Practical Applications:</em> The research outcomes suggest that optimizing the road segments after congestion, using appropriate entertainment systems to reduce driver stress, and implementing adaptive traffic signals to achieve smooth transitions during intermittent congestion can reduce aggressive driving behavior and enhance traffic safety.</div></div>","PeriodicalId":48224,"journal":{"name":"Journal of Safety Research","volume":"91 ","pages":"Pages 381-392"},"PeriodicalIF":3.9,"publicationDate":"2024-10-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142445490","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Rates and ratios of fatal and nonfatal drowning attended by ambulance in New South Wales, Australia between 2010 and 2021","authors":"Edwina Mead ,&nbsp;Chen-Chun Shu ,&nbsp;Pooria Sarrami ,&nbsp;Rona Macniven ,&nbsp;Michael Dinh ,&nbsp;Hatem Alkhouri ,&nbsp;Lovana Daniel ,&nbsp;Amy E. Peden","doi":"10.1016/j.jsr.2024.09.019","DOIUrl":"10.1016/j.jsr.2024.09.019","url":null,"abstract":"<div><div><em>Introduction</em>: Drowning is a preventable cause of mortality, with 279 unintentional drowning deaths per year in Australia. Despite larger estimated numbers, less is known about nonfatal drowning compared to fatalities. This study aimed to examine the burden of fatal and nonfatal drowning in the Australian state of New South Wales using pre-hospital case capture. <em>Methods:</em> A cross-sectional analysis of individuals attended by an ambulance in NSW for drowning between 2010 and 2021 was conducted. Ambulance data (paper-based and electronic medical records) were linked to emergency department and death registry. Ratios of fatal to nonfatal drowning were constructed overall, by sex, age, and remoteness of incident and residential locations. <em>Results:</em> 3,973 ambulance-attended drowning patients were identified (an annual rate of 4.16/100,000 persons). Six percent (6.1%; n = 243) died within 30 days, 82.7% (n = 201) of which died on the day of incident, including at the scene. Mean survival time for those who died between 2 and 30 days was 4.6 days. The overall ratio of fatal to nonfatal incidents was 1:15. Ratios were highest for 10–19 year-olds (1:77), females (1:22), and in metropolitan incident (1:20) and residential (1:23) locations. Across the study drowning declined by 14 incidents and 0.18 fatalities per year. <em>Discussion:</em> Temporal trends indicate declining drowning incidents and fatalities. However, this study highlights significant numbers of nonfatal incidents among those traditionally seen as lower risk, such as adolescents and females, necessitating a widened focus on improving water safety among these groups. <em>Conclusions:</em> Nonfatal drowning results in significant, yet preventable health system burden in New South Wales. <em>Practical Applications:</em> This study highlights the importance of documenting the full burden of drowning, including health system impacts of a preventable cause of injury and death. Such data may be used to encourage further investment in primary prevention efforts.</div></div>","PeriodicalId":48224,"journal":{"name":"Journal of Safety Research","volume":"91 ","pages":"Pages 373-380"},"PeriodicalIF":3.9,"publicationDate":"2024-10-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142419174","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"The maximum potential benefits of safety systems on light van crashes in the United States","authors":"Aimee E. Cox,&nbsp;Jessica B. Cicchino","doi":"10.1016/j.jsr.2024.09.021","DOIUrl":"10.1016/j.jsr.2024.09.021","url":null,"abstract":"<div><div><em>Introduction:</em> The retail landscape has shifted from brick-and-mortar sales to e-commerce, which surged during the COVID-19 pandemic. Light vans are popular vehicles to meet the rising home delivery demands. Two New Car Assessment Programs developed van ratings programs based on their equipment of safety features. This study was designed to estimate the maximum potential benefits that safety technologies could provide light vans based on their historical involvement in relevant crash scenarios. <em>Methods:</em> We used U.S. crash data from 2016—2021 to estimate the average annual total (police-reported), injury, and fatal crashes involving light vans. We determined the proportion of total crashes where front crash prevention, lane departure prevention, blind spot detection, and intelligent speed assistance systems might help the driver prevent crashes or mitigate their severity. We determined the proportions of injury and fatal crashes that resulted in an injury to someone not traveling in the light van. <em>Results:</em> Of the systems studied, front crash prevention that detects vehicles, pedestrians, and cyclists was relevant to largest percentage of light van crashes and could prevent as many as 17% of their involvements, 14% of their injury crashes, and 19% of their fatal crashes. Combined, the four systems have the potential to reduce up to 26% of light van crashes, 22% of their injury crashes, and 36% of their fatal crashes. Sixty-two percent of injury crashes and 56% of fatal crashes relevant to these technologies resulted in injuries or fatalities to occupants of other vehicles or other road users. <em>Conclusions:</em> Light vans are a growing market that can benefit from safety technology, especially when considering their impact on others with whom they share the road. <em>Practical Applications:</em> People and businesses in the market for a light van should seek these systems. Aftermarket products can be installed on light vans not equipped with them.</div></div>","PeriodicalId":48224,"journal":{"name":"Journal of Safety Research","volume":"91 ","pages":"Pages 366-372"},"PeriodicalIF":3.9,"publicationDate":"2024-10-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142419173","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Multi-label material and human risk factors recognition model for construction site safety management","authors":"Jeongeun Park ,&nbsp;Sojeong Seong ,&nbsp;Soyeon Park ,&nbsp;Minchae Kim ,&nbsp;Ha Young Kim","doi":"10.1016/j.jsr.2024.10.002","DOIUrl":"10.1016/j.jsr.2024.10.002","url":null,"abstract":"<div><div><em>Introduction:</em> Construction sites are prone to numerous safety risk factors, but safety managers have difficulty managing these risk factors for practical reasons. Moreover, manually identifying multiple risk factors visually is challenging. Therefore, this study aims to propose a deep learning model–based multi-label risk factor recognition (MRFR) framework that automatically recognizes multiple potential material and human risk factors at construction sites. The research answers the following questions: How can a deep learning model be developed and optimized to recognize and classify multiple material and human risk factors automatically and concurrently at construction sites, and how can the decision-making process of the model be understood and improved for practical application in preemptive safety management? <em>Methods:</em> Data comprising 14,605 instances of eight types of material and human risk factors were collected from construction sites. Multiple risk factors can occur concurrently; thus, an optimal model for multi-label recognition of possible risk factors was developed. <em>Results:</em> The MRFR framework combines material and human risk factors into a single label while achieving satisfactory performance with an F1 score of 0.9981 and a Hamming loss of 0.0008. The causes of mispredictions by MRFR were analyzed by interpreting the decision basis of the model using visualization. Conclusion: This study found that the model must have sufficient capacity to detect multiple risk factors. Performance degradation in MRFR is primarily due to difficulties recognizing visual ambiguities and a tendency to focus on nearby objects when perspective is involved. <em>Practical applications:</em> This study contributes to safety management knowledge by developing a model to recognize multi-label material and human risk factors. Furthermore, the results can be used as guidelines for data collection methods and model improvement in the future. The MRFR framework can be used as an algorithm to recognize risk factors preemptively and automatically at real-world construction sites.</div></div>","PeriodicalId":48224,"journal":{"name":"Journal of Safety Research","volume":"91 ","pages":"Pages 354-365"},"PeriodicalIF":3.9,"publicationDate":"2024-10-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142419172","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}