Recognizing and explaining driving stress using a Shapley additive explanation model by fusing EEG and behavior signals

Abstract

Driving stress is a critical factor leading to road traffic accidents. Despite numerous studies that have been conducted on driving stress recognition, most of them only focus on accuracy improvement without taking model interpretability into account. In this study, an explainable driving stress recognition framework was presented to quantify stress based on electroencephalography (EEG) and behavior data. Based on the extraction of key EEG and behavior features and feature selection, low, medium, and high levels of driving stress were identified using seven machine learning algorithms. The recognition results when only using EEG or behavior features were compared with the result when fusing EEG together with behavior features. Then, the dependency effects between brain activity, driving behavior, and stress were analyzed using the SHapley Additive exPlanation (SHAP) method, and fuzzy rules were obtained by decision tree method. Results indicated that after feature selection, the accuracy of the combined EEG and behavior feature set improved by 8.56% and 26.51% compared to the single EEG and behavior feature sets respectively, and the accuracy rate of 84.93% was achieved. Furthermore, the variations in driver behavior and physiology under stress were identified by the visualization results of SHAP and the quantitative analysis method of decision tree. The changes of different brain regions in the same frequency band showed higher synchronicity under driving stress stimulation. The changes caused by increased stress can be explained by lower speed, smaller maximum lateral lane deviation, smaller accelerator pedal depth and larger brake depth, along with the power changes of the θ and β-band of the brain.