Cause-Based Unsupervised Risk Profiling of Injury-Involved Traffic Accidents with Post-Hoc Severity Validation

Understanding the heterogeneous nature of traffic accidents is significant for developing effective road safety policies and targeted preventive strategies. In this study, traffic accidents were examined independently of outcome variables, and an analytical framework based on accident-based risk profiling was proposed. The dataset used, consisting of official traffic accident records involving injuries, was converted into an accident-based structure using unique accident identifiers, yielding a more compact and interpretable representation for analysis. To enhance interpretability and reduce noise, the dataset was limited to summary variables related to road characteristics, environmental conditions, accident types, and vehicle and driver attributes. In addition, risk profiles were derived using the K-Means clustering method within an unsupervised learning framework. Only variables representing the structural conditions under which the accident occurred were included in the clustering process, while the number of deaths and injuries was excluded to prevent outcome leakage. Following clustering, a continuous severity score was defined, assigning a higher weight to accidents with fatalities, to compare the severity levels of the risk profiles. The results revealed the existence of four distinct and balanced risk profiles that were found to differ significantly in terms of accident severity, accident types, speed limits, and temporal distributions. The highest-risk profile was associated with run-off-road and rollover accidents occurring on roads with high speed limits, whereas lower-risk profiles were primarily characterized by contact-related accident types on roads with lower speed limits. These findings indicate that evaluating traffic accidents based on structural risk profiles, rather than solely on frequency or binary severity classes, provides more informative results. The proposed approach offers a practical methodological framework for developing decision support systems in traffic safety and for formulating targeted preventive policies.