Development and Validation of a Human-Inspired Risk Prediction Model for Traffic Accidents

Visualizing In-Cabin Risk Using a Multi-layered Supplemental Union Model

Abstract

In this study, we developed a novel traffic accident risk prediction model inspired by human cognitive processes, in contrast to conventional black-box-type machine learning approaches. The proposed model targets risks within the vehicle cabin at the moment of a crash and employs a layered "multi-layered supplemental union model," which integrates individual regression models built for each explanatory variable (e.g., vehicle damage severity, occupant characteristics, presence of safety devices).

This approach mimics the structure of human reasoning, which relies on the accumulation of relative impressions such as “larger,” “smaller,” “higher,” or “lower.” The model incorporates the ambiguity and overlap inherent in each factor without reduction, emphasizing intuitive understandability and practical interpretability over pure predictive accuracy—thereby offering value distinct from traditional methods.

The evaluation introduces flexible and realistic performance metrics, including ±1 class tolerance and average consistency across classes, to emphasize trend capture and model explainability. As a result, the model demonstrated a high level of agreement with actual injury trends, supporting its potential use in real-world decision-making and preventive safety applications.

This study is based on the premise that traffic accidents are not random, but rather occur as a consequence of the accumulation of multiple factors—a “deterministic structure.” It represents an effort to construct a model that reproduces human-like decision-making patterns.