Treffer: Enhancing traffic safety by forecasting the severity of road accident injuries using pyramidal dilation attention convolutional networks designed by the reptile search algorithm.

Objective: This work aims to give a method that is both efficient and comprehensible for forecasting the extent of injuries sustained in traffic accidents. This addresses the limitations of existing GNN-based frameworks, which often struggle with complexity, limited interpretability, scalability issues, and the need for extensive data pre-processing and advanced graph representation learning.
Methods: In this manuscript, Predicting Road Crash Injury Severity utilizing Pyramidal Dilation Attention Convolutional Network optimized with Reptile Search Algorithm (PRCIS-PDACN-RSA) is proposed. Firstly, the input data is gathered from the UK road accident dataset. The data is then sent to pre-processing, where the Robust Maximum Correntropy Kalman Filter (RMCKF) is applied to eliminate null, noisy, or incomplete entries. The pre-processed data is fed into Adaptive SV-Borderline SMOTE (ASV-SMOTE) to balance the imbalanced dataset. Then the balanced dataset is given to the Pyramidal Dilation Attention Convolutional Network (PDACN) to predict road crash injury severity and classify it as either severe or non-severe. The Reptile Search Algorithm (RSA) is used to optimize the PDACN parameters, enhancing its predictive performance.
Results: The proposed PRCIS-PDACN-RSA technique is implemented in Python and evaluated using performance metrics, including accuracy, F1-score, recall, precision, Receiver Operating Characteristic (ROC), and Matthews's correlation coefficient (MCC), to assess its efficiency. The proposed PRCIS-PDACN-RSA approach attains 97.2% accuracy, 0.90% MCC, and 98.11% recall compared with existing methods, including Road Crash Injury Severity Prediction Utilizing a Grey Wolf Optimization-driven Artificial Neural Network for Predicting Road Crash Severity (GWO-ANN-PRCS), Graph Neural Network Framework (RCI-SP-GNN), and Multi-View Graph Convolutional Networks for Traffic Accident Risk Prediction (MGCN-TARP).
Conclusions: The results demonstrate that the proposed PRCIS-PDACN-RSA framework outperforms existing methods in predicting road crash injury severity. Its high accuracy, robustness, and efficient handling of pre-processing and optimization highlight its suitability for real-world intelligent traffic safety systems.