Hybrid Risk Assessment Framework for Predicting Accident Severity from Driver Physiological Signals and Road Conditions

Abstract

The prediction of traffic accidents continues to provide a considerable problem, primarily because present models inadequately capture and integrate the dynamic condition of drivers with immediate environmental risks, resulting in weak severity predictions. This research proposes a Hybrid Risk Assessment Framework aimed at predicting accident severity through the integration of diverse data streams, including real-time driver physiological signals and current road conditions. The proposed system utilises a stacked ensemble architecture, incorporating a Bi-LSTM to model the temporal aspects of internal risk (e.g., Heart Rate Variability, RMSSD) and an XGBoost classifier for static exterior risk features (e.g., Road Surface, TIT). The forecasts from these specialised base-learners are integrated by a Meta-Classifier, allowing the framework to understand complex non-linear interactions. The findings indicate the enhanced effectiveness of the hybrid method, with a final accuracy of 92.1% and a macro-averaged F1-Score of 0.915. This performance markedly exceeds single-modal baseline models (e.g., XGBoost baseline F1-Score 0.829), validating the concept that decision-level data fusion is crucial for accurate accident severity prediction and facilitating the development of highly reliable proactive safety systems.

https://doi.org/10.65470/james.6