Design of Driver Stress Prediction Model with CNN-LSTM: Exploration of Feature Space using Genetic Programming

Tingting Yang, Chenhao Xue, Jun Chen

Abstract: Road traffic accidents, primarily caused by driver-related issues such as stress, result in numerous fatalities and injuries. Effective prediction of driver stress within the driving phase is paramount for real-time accident intervention, while it requires prediction accuracy, stability, and enough predictive lead time. Physiological data contains rich information related to driver stress levels, which can be captured by machine learning models. However, those models are mainly developed for and perform well in static stress detection tasks, not addressing practical requirements of predictive lead time and performance stability. This study introduces a novel approach, the GP-CNN-LSTM model, Which employs a Convolutional Neural Network (CNN) with a Long Short-Term Memory (LSTM) network and Genetic Programming (GP), leverages GP to explore the feature space upon physiological sensor signals, and relies on CNN-LSTM to predict driver stress. Experiments show that this model achieves high accuracy for a 60-second forward stress prediction, with more stability compared to Fractional Fourier Transform-based benchmark models. We found explainable effective signals and features described by math functions via genetic programming that helped in improving the accuracy and stability of driver stress prediction.