The application of perception and edge computing technologies provides extensive data support for intersection conflict analysis. However, traditional threshold-based and semantic rule-based conflict analysis methods struggle to address the challenges posed by intersection heterogeneity and data diversity. This study integrates the kinematic features of trajectory data from eight different Holographic Intersections. Based on the proposed definitions for conflict severity and scenarios, 5,339 typical conflict events were labeled as the training dataset. Subsequently, four Transformer encoders with different combinations of heads and layers were trained, then integrated using the Weighted Voting method. The Ensemble Transformer was selected as the benchmark to construct the AI Conflict Observer (AICO) model. For the tasks of classifying conflict severity and scenarios, the Weighted F1 Scores of AICO reached 0.846 and 0.902, respectively. To validate the model's generalization performance, this study conducted case studies using conflict events from a 4-leg intersection and a 3-leg intersection sourced from different datasets as ground truth. A total of 560 and 136 conflict events were identified through threshold-based preliminary screening and manual verification, respectively. The conflict recognition results of the AICO model were then compared with the ground truth. The results indicate: 1) At the 4-leg intersection and the 3-leg intersection, AICO achieved classification accuracies of 89.97% and 86.13% for conflict severity, and 88.71% and 90.97% for conflict scenario, respectively. 2) Regarding spatial distribution, the kernel density values of conflicts identified by AICO were highly consistent with the ground truth. 3) For temporal distribution, AICO's results demonstrated a high degree of goodness-of-fit. The R2 values for common and serious conflicts were 0.927 and 0.987 at the 4-leg intersection, and 0.934 and 0.945 at the 3-leg intersection, respectively. 4) In terms of conflict severity distribution, there was no significant difference between AICO and the ground truth in identification of TTC. However, significant differences were observed in conflict duration identification (p < 0.05). AICO employed a more conservative strategy, tending to assign longer durations. The AICO model overcomes the limitations of traditional threshold and rule-based methods. It can generalize to different types of signalized and unsignalized intersections and achieve batch conflict identification. The model can be applied to near-real-time analysis of intersection operational risks, extraction of critical risk scenarios, providing decision support for intersection improvement, governance, and management.
Establishing the context-dependent relationship between abnormal driving events (ADEs) and crash risk is fundamental for developing next-generation proactive safety systems. However, current methodologies often oversimplify by neglecting that ADE risk significance is influenced by operating contexts, and by ignoring contextual heterogeneity in event evolution within fixed pre-crash windows-both crucial for understanding crash causation and timing interventions. This study proposes a two-stage analytical framework. First, a causal forest with debiased machine learning (CF-DML) approach is employed to quantify the effect of ADE exposure on crash risk and assess its effect heterogeneity across contexts. Second, a random parameters logit model with heterogeneity in means (RPLHM) is used to characterize pre-crash temporal patterns as either ADE-dense or ADE-sparse. The analysis utilizes crash and ADE data from 2023 to 2024 on two freeways in Shandong Province, China, integrated with matched weather, traffic, and roadway geometry. Results show that hard acceleration and braking are positively associated with crash risk under pronounced context-specific heterogeneity, while sharp turning consistently correlates with reduced risk. Several contextual variables-including temperature, traffic volume, truck proportion, and speed dispersion-significantly moderate the ADE-crash relationship. The pre-crash ADE distribution is closely linked to temperature, wind speed, daytime, traffic volume, speed dispersion, and crash type. Accordingly, three prototypical risk contexts are identified: ADE-informative (where ADE exposure strongly indicates crash risk), pre-crash ADE-dense, and pre-crash ADE-sparse. Targeted traffic management countermeasures proposed for each context advance the mechanistic understanding of crash risk and provide a foundation for developing targeted, context-sensitive, and effective safety interventions.
Highway multilevel interchanges present unique challenges, characterized by complex road conditions, ambiguous routes and short intervals. Multiple ramps necessitate swift and continuous direction changes, significantly jeopardizing pathfinding safety. Guide signs play a crucial role in improving information recognition and driving safety. This study replaces accident conflicts with risk behaviors and constructs a risk prevention research framework: Firstly, through subjective demand analysis, the cognitive preferences and functional expectations of drivers were analyzed to clarify design orientation. Secondly, the influence of pathfinding process under traffic facility intervention and potential risk relationship was captured in a driving simulation experiment. Furthermore, focus on coupled collaborative analysis, the correlation between longitudinal stability and lateral maneuvering behavior was quantified, and collaborative effectiveness evaluation was conducted. Ultimately, the optimized sign served as a guide for improving behaviors. Results show that: 1) At different ramp decision points, setting "destination separation for different directions" and "lane guidance" information represent drivers'core demands; 2) Under the influence of 4 sign schemes, there are indeed differences in the pathfinding behavior of longitudinal and lateral dimensions, which should be improved in a coordinated manner. 3) Approaching the ramp exit, setting different directions information can help drivers make decisions more quickly, reach the target speed earlier, and reduce conflict time with mainstream traffic. Approaching the ramp for the first diversion, setting both information helps drivers tend to accelerate smoothly, and conduct lane change adjustments before making decisions. Approaching the ramp for the second diversion, simplify setting lane guide information can improve lateral maneuvering and enable faster path selection. 4) Providing information prompts in the form of navigation can offer continuous direction guidance, timely lane changing, with higher longitudinal safety margin and lateral stability. The proposed suggestions for risk prevention can achieve a coordinated improvement in traffic safety across multiple dimensions.
To describe tetanus vaccination practices and injury characteristics in Sandun Town, Hangzhou during February 2024 to January 2025. A retrospective cross-sectional study was conducted using data from trauma and animal-injury patients who received tetanus vaccination in Sandun Town, Hangzhou from February 2024 to January 2025. Demographic characteristics, injury profiles, vaccination status, and relationships with season/temperature were analyzed. Total of 3,174 patients were initially identified, and 2,825 were included in the final analysis after applying exclusion criteria. Among them, animal-induced injuries accounted for 1,029 cases (36.4%), traffic accident injuries for 943 cases (33.4%), cutting injuries for 374 cases (13.2%), blunt force injuries for 242 cases (8.6%), and other causes for 237 cases (8.4%). Males predominated in all trauma categories except animal-induced injuries, where females were the majority (53.3%). Employees/workers were the predominant occupational group across all categories (56.4-76.4%). The upper limb was the most common injury site across all categories (56.7-74.1%). High-risk wounds were observed in 88.2-98.9% of patients. The use of passive immunizing agents was generally low (2.1-10.4%), with the highest rate in animal-induced injuries. Following the implementation of China's 2024 non-neonatal tetanus guideline, the treatment of external injuries is becoming more standardized. However, gaps persist compared with developed countries, mainly reflected in the underutilization of passive immunizing agents for high-risk wounds and inconsistent application of guidelines. Targeted education for students, migrant workers, and pet owners-particularly during warmer months-and continued training for healthcare personnel are needed.
As electric scooter (e-scooter) use has expanded, understanding the factors associated with e-scooter rider injury severity has become increasingly important for road safety policy. This study analyses 2,128 crashes involving e-scooters and motor vehicles across England (2020-2023) to identify factors associated with severe and fatal injuries to e-scooter riders. Using the geographic coordinates of crashes, we developed a Bayesian spatial field model implemented via the Stochastic Partial Differential Equation (SPDE) approach for fast Bayesian estimation. Our approach accounts for spatial unobserved heterogeneity (area-level "context" effects) often overlooked in injury severity studies. Results indicate that severe or fatal injuries are more likely among older riders, male riders, and in crashes occurring in darkness, on single carriageways, on roads with speed limits of 40 mph or higher, involving heavy vehicles, at night or early morning, or with e-scooter skidding/overturning, frontal impacts, e-scooters entering main roads, or opponent vehicles moving straight. Conversely, motor vehicles performing moving-off manoeuvres are linked to lower odds of severe injuries. Importantly, the presence of authorised e-scooter trials was not found to be associated with rider injury severity outcomes. Our spatial analysis reveals higher odds of severe injury in parts of north-western and south-eastern England relative to the national average. Our research highlights the importance of vehicle kinematics, road environment, and spatial context in shaping injury severity and support targeted, evidence-based interventions, including infrastructure measures and vehicle-based safety technologies such as blind-spot detection.
In accident analysis and prevention, passive vehicle safety technologies ensure the lower limit of driving safety, while active safety technologies determine the upper limit. This study aims to provide suggestions for the active safety management of commercial vehicles by identifying high-risk on-road scenarios. Firstly, taking regular-route passenger buses as the research object, based on multi-source data fusion technology, this study integrates driving alarm data, vehicle trajectory data within 5 min before the alarm, driver video data within 10 s before the alarm, and driving record video data to extract key features and construct a driving risk identification variable set. Secondly, Pearson correlation coefficient and variance inflation factor (VIF) are used sequentially to conduct collinearity tests and eliminate redundant variables. Considering that the variables include both continuous and discrete heterogeneous data, the K-prototype hybrid clustering method is adopted, and the optimal number of clusters (K = 4) is finally determined. Thirdly, an integrated method of 'multi-source heterogeneous data fusion-hybrid variable clustering-Ordered Logit modeling-SHAP interpretability analysis' is constructed. In an effort to explore active safety technologies, this study attempts to map the identified driving patterns to ordinal risk levels based on key vehicle kinematic parameters. Subsequently, the Ordered Logit model is applied to quantitatively analyze the marginal effects of significant variables. Finally, combined with the variable distribution characteristics of the clustering results and SHAP interpretability analysis, the core features and key incentives of the four risk levels are systematically characterized, and targeted active safety management suggestions are generated with the assistance of Large Language Models (LLMs). This study intends to provide certain insights for the research on vehicle active safety and offer references and suggestions for the dynamic monitoring and management of commercial vehicles.
An increased risk for an occupation-related SARS-CoV-2 infection has been linked to higher psychological distress. This online survey investigates the prevalence of depressive and anxiety symptoms in a large sample of 34,303 participants from the German cohort for digital health research (DigiHero) after the pandemic (late 2023, t1) and retrospectively from the Omicron wave (early 2022, t0), emphasizing variations across occupational groups and work-related risk factors. Participants reported their employment status (currently working; seeking employment; not working). Workers provided their primary occupation to assess occupational SARS-CoV-2 infection risk. Symptoms of depression and anxiety (assessed via PHQ-4) and additional occupational risk factors were solicited for each time point. Associations between occupational exposure and stressors with the four-level PHQ-4 outcome were analyzed separately for t1 and t0 using ordinal regression and expressed as odds ratios (OR) with 95% confidence intervals (CI). Over 60% of respondents were working at t1, and 1.4% classified themselves as seeking a job. Job seekers reported highest and non-working individuals lowest depressive and anxiety symptoms. Symptom severity varied by occupation with elevated odds in traffic/logistics professions exclusively at t1 (OR=1.24, 95% CI 1.04-1.48) and healthcare professions exclusively at t0 (OR=1.08, 95% CI 1.01-1.16). High occupational SARS-CoV-2 infection risk was linked to symptoms at t0. Overall, these associations were modest and partly attenuated after additional adjustment for individual work-related stressors (e.g., loneliness at work, chronic work-related stress, work-privacy conflicts). At both timepoints, individual stressors and sociodemographic factors showed stronger associations with severe symptoms than occupation (e.g., chronic work-related stress at t1 OR=2.87, 95% CI 2.70-3.04). Persistent post-pandemic depressive and anxiety symptoms among workers emphasize the importance of addressing individual psychosocial work-related stressors.
Bus crashes remain a significant public safety concern due to their potential to cause both passenger injuries and substantial vehicle damage. These two severity outcomes are interrelated, as the physical forces in high-impact crashes contribute simultaneously to occupant harm and structural damage. However, most existing studies model these outcomes independently, overlooking their statistical dependence and shared risk factors. To address this gap, this study employs a copula-based modeling framework to jointly analyze passenger injury and vehicle damage severity using bus crash data from Maryland (2015-2022). The joint estimation results reveal distinct sets of significant predictors for passenger injury severity and vehicle damage severity. For injury outcomes, the most influential variables include driver fault, airbag deployment, and safety equipment usage, underscoring the critical role of both human error and protective systems in shaping injury levels. In contrast, vehicle damage severity is more strongly associated with vehicle type, movement status at the time of impact, and mechanical condition. The inclusion of the vehicle type variable shows that school buses, despite being structurally safer, are frequently involved in crashes due to high exposure, contributing to notable damage risks. Additionally, adverse vehicle conditions and airbag deployment exhibit a strong association with disabling or destroyed damage, highlighting the role of structural integrity and energy dissipation mechanisms in post-impact outcomes. These patterns, revealed through the Copula-MNL model, reflect not only the different underlying risk structures of the two severity measures but also their shared dependence on critical safety-related factors.
Driving behavior and interactions with bicyclists on rural roads have not been quantified and modeled extensively. Naturalistic bicycling data for 1,991 passing events were collected on a rural two-lane roadway (55 mph, 88 kph speed limit) to quantify how opposing traffic and vehicle platooning influence passing lateral distance, speed, and aerodynamic forces. Results indicate that opposing traffic significantly reduces passing lateral distance by an average of 2.0 ft (61 cm) and decreases speed by an average of 2.3 mph (3.7 kph). Platooning leads to progressively reduced passing distance and speed among following vehicles. The reductions reflect limited available space and increased risk for bicyclists when opposing vehicles are present. The estimated aerodynamic lateral forces created by passenger vehicles were well below tolerable safety limits for bicyclists. To surpass tolerable limits, passenger vehicles would have to pass at a lateral distance of 0.9 ft (27 cm) at a speed of 55 mph (88 kph). Lateral distance and speed were found to be independent at a disaggregate level. Leading vehicles' lateral distance followed a Log-normal distribution and speed followed a Weibull distribution. Theoretical joint probability density functions were developed for leading and following vehicles with and without opposing traffic. Pairwise differences among lead and follower vehicles were similar and resembled a Normal distribution. The developed joint probability density functions can be used for calibration and validation of driving simulators, or development of autonomous and artificial intelligence driving models. Results contribute to developing safer design guidance and risk mitigating strategies for bicyclists.
The rapid advancement of Battery Electric Vehicle (BEV) in automotive technology and market adoption present safety challenges due to their distinct design and operational characteristics compared to Gasoline Vehicles (GVs). A proactive approach to evaluating the safety performance of Advanced Driver Assistance Systems (ADAS) among BEVs and GVs is necessary to address these emerging risks. A total of 8,118 Florida crash reports with ADAS engaged were analyzed in this study, consisting of 6,052 from GV and 2,066 from BEV. Comparability between BEVs and GVs regarding ADAS safety evaluation across diverse traffic conditions was established using Propensity Score Matching (PSM), examining injury severity, environmental conditions, driver maneuvers, and vehicle types. And then a partially constrained Random Parameter Logit (RPL) model was employed to identify both the unique and shared factors influencing injury severity of BEVs and GVs. PSM showed that ADAS-engaged BEV crashes had lower injury severity (75.13% no-injury vs. 69.82% for GVs) but higher involvement in Vulnerable Road User (VRU) crashes (3.2% vs. 1.8%). The RPL model revealed ADAS-engaged BEV crash risk in work zones (+0.0057 severe injury probability), GV crash risk at processing straight and backing (+0.0042 minor injury probability), and shared risk at speeds above 61 mph (+0.0084 severe injury probability). This study offers an exploratory analysis of ADAS safety performance between BEVs and GVs using real-world crash data. These findings provide valuable insights for manufacturers and other stakeholders to inform decisions regarding the deployment and application of these technologies.
Pedestrian road safety represents a critical challenge in rapid urbanization processes, with land use and point of interest (POIs) configurations playing a central role in shaping crash risks. Existing studies mainly rely on traditional quantitative models, which often struggle to capture nonlinear spatial relationships and lack real-time, visual feedback for planning. This study develops a Generative Adversarial Network (GAN)-based framework to predict pedestrian crashes, integrating land use, POIs, and commuting distance to capture both direct and indirect associations. Using 504 spatial analysis units in Auckland, New Zealand, and 2870 pedestrian crash records (2016-2024), the model performs image-to-image translation to simulate crash risks under different urban configurations. Findings show that pedestrian crash risks vary nonlinearly across space and are linked to the combined effects of land use, POIs, and commuting distance. In the CBD, reducing commuting attraction areas emerges as a key factor associated with improved safety outcomes, with public open space expansion reducing pedestrian crashes by up to 14.7%. Sub-centre areas demonstrate sharp risk amplification (up to 80.4%) when educational, commercial, and transportation POIs simultaneously reach high densities. Residential areas exhibit significant threshold effects, where high-density combinations are associated with extreme risk scenarios (up to 35.5%). Moreover, direct associations significantly outweigh indirect pathways. This study offers a scenario-based visual tool for exploring the safety implications of objection and POI planning, providing empirical evidence for data-driven urban design.
Maritime Autonomous Surface Ships (MASS) with multiple modes of operation are increasingly adopted to improve efficiency and address crew shortages. Safe mode transitions, especially dependable human takeovers in urgent scenarios, are essential for the effective prevention of navigation accidents. However, existing human reliability analysis methods often overlook how information degradation and cognitive processes jointly affect takeover performance. To address this challenge, the information, decision, and action in crew context model (IDAC) is extended by introducing an external filtering stage that explicitly accounts for information loss or distortion before reaching the operator's cognition. Building on this, an integrated system control and cognitive perspective is proposed. The model combines System-Theoretic Process Analysis with the enhanced IDAC model, and embeds Bayesian Networks to identify causal chains of takeover failures, quantify Human Error Probability, identify critical factors. Through scenario-based reasoning, it further derives the key causal paths of takeover failure in urgent scenarios. A case study involving both remote and onboard takeover scenarios demonstrates the framework's applicability. Results indicate that the diagnosis and decision-making stage is the most critical, with fatigue, attention, available time and trust level emerging as dominant factors. Based on these findings, this study proposes a three-level pre-alert mechanism and targeted intervention strategies to enhance human reliability during MASS mode transitions. This study provides a scalable and behaviourally grounded framework that supports the development of takeover guidelines and safety standards, aiming to prevent navigation accidents caused by human takeover failures during the development of ship autonomy.
In China, there are usually no exclusive right-turn phases or right-turn-on-red at signalized intersections. Right-turn crashes account for over 30% of intersection-related crashes, highlighting the critical role of intersection design in traffic safety. However, the relationship between specific design features and crash risk remains unclear. This study identifies key design variables and quantifies their causal effects and heterogeneity on right-turn crashes at signalized intersections. A total of 271 signalized intersections in Suzhou, China, were analyzed from 2022 to 2024. The SHapley Additive exPlanations method was applied to rank variable importance, followed by Generalized Random Forest modeling to estimate both Average and Heterogeneous Intervention Effects (HIEs) while controlling for confounding bias. The minimum right-turn radius and intersection skewness were identified as the most influential factors for total and fatal crashes, respectively. A turning radius of approximately 15 m was associated with the lowest total crash risk, and right-angle intersections were linked to reduced fatal crash rates. A bicycle lane barrier width on the minor road between 1.01 and 3 m also contributed to crash reduction. HIEs showed that smaller turning radii improved safety at high-volume, multi-lane intersections, and skewed intersections with low traffic volumes required special attention. Facilities such as bicycle lanes, physical barriers, and channelization islands enhanced safety performance at skewed intersections. As turning radii increase, expanding the bicycle lane barrier width may be beneficial, although wider barriers should be applied cautiously under low traffic conditions. This study provides a causal inference framework for evaluating right-turn design and offers evidence-based guidance for improving intersection safety in urban environments.
The Fatality Analysis Reporting System (FARS) and Crash Report Sampling System (CRSS) report a unique "transit bus stop-related" pedestrian crash type. This study merged fatal and non-fatal transit bus stop-related pedestrian crashes from FARS and CRSS (2016-2023) to examine injury severity and identify risk factors for pedestrians at bus stops. To address imbalance in severity (KA = 183, BCO = 86), synthetic minority oversampling for nominal and continuous features (SMOTE-NC) was applied. Then, a severity analysis using elastic net penalized logistic regression was conducted on the original and SMOTE-NC sample with 17 predictors, as regularization handles high-dimensional data and multicollinearity better than traditional parametric approaches. The SMOTE-NC model showed severe transit bus stop-related pedestrian crashes were associated with poor lighting, midblock locations, and higher-speed, wider roadways. Results showed amplified effects for predictors in the SMOTE-NC models, including several that were zero in the original model but were selected after oversampling, suggesting oversampling potentially mitigated bias due to class imbalance. The results provided a foundation for identifying tangible solutions to address dangerous bus stops. Improving street lighting or installing bus stop lighting may be effective, as dark, unlighted conditions were associated with higher odds ratios for KA injuries (OR = 3.18) than dark, lighted conditions (OR = 1.69), relative to daylight. Evaluating pedestrian behavior near midblock bus stops, like unmarked crossings, can further mitigate risks exacerbated by darkness and high-speed multilane roads. These findings support data-driven policy and design decisions by transit agencies and transportation planners to address higher-risk bus stops.
Ocular injury is a significant public health concern, which may threaten vision and impose a heavy burden on healthcare systems. In South Korea, understanding the epidemiology of ocular injury is critical for developing effective prevention strategies. This study utilized data from the National Emergency Department Information System from 2016 to 2022. Patients presenting with ocular injury to emergency departments (EDs) nationwide were analyzed. We categorized ED visits based on the types of injury and the risk of vision loss by a modified Delphi process. Multivariate logistic regression was performed to identify factors associated with a high risk of vision loss. A total of 581,264 cases of ocular injury were analyzed. Majority of cases occurred in males (72.3%) and individuals aged 40-64 years (37.02%). The elderly (≥ 65 years) exhibited an increasing trend in ocular injury, rising from 9.41% in 2016 to 17.82% in 2022. Individuals covered by industrial accident compensation insurance showed a significantly higher risk of vision loss, particularly among males aged 20-39 years (adjusted odds ratio [aOR], 5.55; 95% confidence interval [CI], 4.25-7.26) and 40-64 years (aOR, 2.94; 95% CI, 2.41-3.59). Machinery-related injuries were also identified as a major risk factor for severe vision loss, with an aOR of 4.83 (95% CI, 3.27-7.14) in males aged 40-64 years. Residents in low-populated areas showed significantly greater risks of severe vision loss compared to those in high-populated areas. This study provides a comprehensive epidemiological overview of ocular injury in South Korea, suggesting the need for a more detailed understanding of the mechanisms underlying these injuries, promoting the use of appropriate eye protective gear, and improving access to emergency care in low population density areas.
Commercial fishing remains one of the most hazardous occupations globally, with smallscale fleets exhibiting persistent safety challenges. In Türkiye, fishing operations are characterized by low regulatory compliance, insufficient training, and fatigue-related risks. This study assessed occupational health and safety (OHS) compliance levels among Turkish fishing vessel crews and identified key predictors of safety outcomes across vessel size categories. A cross-sectional study was conducted (June-August 2018) across Türkiye's Aegean, Marmara, and Black Sea regions, involving 356 crew members from 180 vessels. Data collection included structured questionnaires, observational checklists, and interviews. Analyses employed descriptive statistics, χ² tests, independent t-tests, and multiple linear regression (SPSS v26). Of all participants, 38.8% reported at least one occupational accident in the past year. The most frequent injuries were cuts (12.9%), falls (9.3%), and equipment-related trauma (5.9%). The main contributing factors were the hasty work pace in the workplace (52.2%), inadequate training (28.9%), and fatigue due to long working hours (19.0%). PPE compliance was low at 18%, and only 27% of participants had received formal safety training. A significant association was found between vessel size and accident occurrence (χ² = 12.45, p = 0.002), with smaller vessels having a significantly higher accident risk than larger vessels. Workers involved in accidents reported longer working hours (M = 14.3, SD = 1.8) than their counterparts (M = 13.1, SD = 2.2; p < 0.001). Regression analysis identified formal training (β = 0.35, p < 0.001), education level (β = 0.21, p < 0.001), and vessel size (β = 0.14, p = 0.01) as significant predictors of OHS compliance (R² = 0.29). Occupational health and safety compliance in Türkiye's fishing sector remains inadequate, particularly for small-scale vessels. Prioritizing training expansion, work-hour regulations, and targeted support for high-risk fleets is essential.
Road traffic crashes pose a serious public safety challenge, particularly due to fatal and serious injuries. Although machine learning (ML) has been widely used for crash severity prediction, many studies remain accuracy-oriented and insufficiently address class imbalance, decision thresholds, and probabilistic reliability. This study proposes a safety-oriented and explainable ML framework for predicting killed or seriously injured (KSI) crashes using nationwide United Kingdom traffic accident data from 2020-2024. Crash severity is reformulated as a binary classification task distinguishing slight injury crashes from KSI outcomes, aligning model objectives with road safety priorities. A Light Gradient Boosting Machine (LightGBM) model is developed with imbalance handling using SMOTE, safety-oriented decision threshold optimization, and probability calibration. Model performance is evaluated using ROC-AUC, precision-recall analysis, confusion matrices, the Brier score, and a utility-based evaluation metric, while interpretability is ensured through SHapley Additive exPlanations (SHAP). Results show that default threshold settings fail to adequately detect severe crashes. At an optimized threshold of 0.35, the model achieves a Recall(KSI) of 0.605 - representing a substantial 73% improvement compared to conventional configurations - while maintaining acceptable precision. In addition, probability calibration confirms reliable risk estimation (Brier score = 0.190), supporting risk-based interpretation. Comparative analysis demonstrates that the SMOTE-based model provides a more balanced and operationally effective trade-off than class-weighted learning. SHAP analysis identifies speed limit, road class, lighting conditions, and urban context as key variables associated with KSI risk. The findings highlight the importance of safety-oriented learning design and context-aware performance interpretation for effective, risk-based traffic safety management.
The effectiveness of in-vehicle Connected Information (CI) is often limited by uniform warning strategies that overlook the interaction among warning design, traffic context, and driver state. This study establishes a causal machine learning framework to quantify how CI modality and lead time influence driver workload (WL) and perceived risk (PR) across multiple traffic scenes. A 3 × 3 × 5 factorial driving-simulator experiment collected multimodal physiological, behavioral, and self-reported data from 52 drivers. Double Machine Learning with Causal Forests (DML-CF) was applied to identify both average and heterogeneous causal effects. The results show strong context dependence and clear nonlinear patterns in CI effectiveness. Moderately early warnings in the range of 5.5 to 6.5 s combined with congruent dual-modality cues consistently reduced WL and PR. Delayed or single-modality cues frequently increased cognitive demand and perceived hazard. Conditional treatment effect (CATE) analyses revealed substantial heterogeneity and identified driver subgroups with distinct physiological and behavioral characteristics. For instance, certain individuals exhibited elevated WL due to heightened arousal even when exposed to multimodal early warnings that were beneficial for most drivers. These findings indicate that optimal CI design requires continuous alignment between warning parameters, situational uncertainty, and the driver's state. The evidence provides a quantitative causal basis for developing adaptive CI systems capable of tailoring information delivery to enhance safety and human-machine interaction.
The TraumaRegister DGU® of the German Society for Trauma Surgery (TR-DGU) has been collecting data on seriously injured persons for decades. Based on a few key data points about the accident, far-reaching insights into accident epidemiology can be derived. The present evaluation refers to traffic accident victims, who account for about half of all seriously injured persons in the register. For a 10-year period (2015-2024) information on the accident victims (age, gender), the circumstances of the accident and the time of admission to hospital (date, time) was extracted from the TR-DGU. Only patients who were primarily admitted to and treated in a German trauma center and who met the criteria of the registry's population base were considered. The analysis is based on 121,534 accident victims with an average age of 47.4 years and 71% were male. Car occupants were the largest subgroup (39%), followed by motorcyclists (25%), cyclists (22%) and pedestrians (10%). The proportion of male accident victims was almost always greater than the female proportion and in the 16-71 years age group it was more than twice as high. In relation to the German population 18-year-olds have the highest incidence with 40 per 100,000 per year. On weekends, 10-20% more people are involved in accidents and during the day the rate is highest in the late afternoon (4-6 p.m.). The number of accidents is about twice as high in the summer months as in winter, which is mainly due to 2‑wheeled vehicles. Other external factors such as a full moon or Friday the 13th have no effect on the number of serious injuries. With just a few details about the person, the time and the mechanism of traffic accident victims, it is possible to identify areas that lead to more accidents. These findings can also be effectively used for prevention purposes. EINLEITUNG: Das TraumaRegister DGU® der Deutschen Gesellschaft für Unfallchirurgie (TR-DGU) sammelt seit Jahrzehnten Daten schwer verletzter Personen. Anhand weniger Eckdaten zum Unfall lassen sich weitreichende Erkenntnisse zur Unfallepidemiologie ableiten. Die vorliegende Auswertung bezieht sich auf Verkehrsunfallopfer; diese machen etwa die Hälfte aller Schwerverletzten im Register aus. Für einen 10-Jahres-Zeitraum (2015–2024) wurden Angaben zum Unfallverletzten (Alter, Geschlecht), zum Unfallhergang und zum Zeitpunkt der Krankenhausaufnahme (Datum, Uhrzeit) aus dem TR-DGU extrahiert. Es wurden nur Patienten betrachtet, die primär in einem deutschen Traumazentrum aufgenommen und behandelt wurden und den Kriterien des Basiskollektivs des Registers entsprachen. Die Auswertung basiert auf 121.534 Unfallopfern, das Durchschnittsalter betrug 47,4 Jahre, und 71 % waren männlich. Autoinsassen waren die größte Untergruppe (39 %), gefolgt von Motorradfahrenden (25 %), Fahrradfahrenden (22 %) und Fußgängern (10 %). Der Anteil männlicher Unfallopfer war nahezu immer größer als der weibliche Anteil, und im Alter von 16 bis 71 Jahren sogar mehr als doppelt so hoch. Bezogen auf die deutsche Bevölkerung zeigen die 18-Jährigen die höchste Inzidenz mit 40/100.000 und Jahr. Am Wochenende verunfallen 10–20 % mehr Menschen, und im Tagesverlauf ist die Rate am späten Nachmittag (16–18 Uhr) am höchsten. Die Unfallzahlen sind in den Sommermonaten etwa doppelt so hoch wie im Winter, was insbesondere an den Zweirädern liegt. Andere externe Faktoren wie der Vollmond oder Freitag, der 13., zeigen keine Auswirkung auf die Zahl Schwerverletzter. Durch wenige Angaben zur Person, zum Zeitpunkt und zum Mechanismus von Verkehrsunfallopfern lassen sich Bereiche erkennen, die vermehrt zu Unfällen führen. Diese Ergebnisse lassen sich auch für die Prävention sinnvoll nutzen.
Real-time corridor-wide crash-occurrence risk (COR) prediction is challenging, since existing near-miss EVT models oversimplify collision geometry, neglect vehicle-infrastructure (V-I) interactions, and fail to adequately account for spatial heterogeneity in traffic and roadway conditions. To do so, this study develops a geometry-aware 2D-TTC near-miss extraction and integrates it with a hierarchical Bayesian structure grouped random parameters (HBSGRP-UGEV) to estimate short-term COR in urban corridors. Building on prior grouped EVT formulations while explicitly accommodating both V-V and V-I near-miss processes within a unified corridor-wide modeling framework. High-resolution trajectories from the Argoverse-2 dataset were analyzed across 28 sites on Miami's Biscayne Boulevard to extract extreme near-miss events. The model incorporates vehicle dynamics and roadway features as covariates, with partial pooling across segments and intersections to capture corridor-wide heterogeneity. Results show that the HBSGRP-UGEV framework outperforms fixed-parameter HBSFP-UGEV models, reducing DIC by up to 7.5% (V-V) and 3.1% (V-I). Predictive validation using ROC-AUC confirms strong accuracy (0.89 for V-V segments, 0.82 for intersections, 0.79 for V-I segments, and 0.75 for intersections). Grouped random-parameters (HBSGRP) framework indicate that relative (speed, distance, and deceleration) dominate V-V near-miss risk on segments, whereas V-I segment risk is primarily associated with relative distance; at intersections, V-V risk is driven by relative (speed and distance), while V-I dynamics exhibit no statistically significant effects. These findings demonstrate the value of a geometry-aware, spatially adaptive framework for proactive corridor safety management, supporting both real-time interventions and long-term Vision Zero goals.