An explainable machine learning framework for predicting driving states using electroencephalogram.
Academic Article
Overview
abstract
OBJECTIVES: Understanding drivers' cognitive load is essential for enhancing road safety, as cognitive demands fluctuate across different driving scenarios, potentially impacting performance, and safety, particularly for drivers with neurological disabilities. This study aims to predict driving states in healthy adult drivers using electroencephalogram (EEG) and machine learning (ML) models; and interpret the neural activity associated with each driving condition. METHODS: EEG data were collected from participants using a Cognionics Quick-20 EEG headset in Resting state, City-way, Highway, and Suburb-way driving states in 360 full-screen real car cabin inside the driving simulator. Participants drove while experiencing varied cognitive workloads due to various driving environments. EEG Features, including spectral band powers and power ratios, were extracted to extract neural activity patterns relevant to driving states. Gradient Boosting (GBoost) and Random Forest (RF) machine-learning classifiers were applied to classify the driving states based on EEG features. A model-agnostic explanation approach was implemented for model interpretability, which highlighted EEG spectral features contributing to each driving state. RESULTS: The GBoost model achieved the highest classification accuracy among tested models, with ROC-AUC values of 1.00 for Resting, 0.96 for Suburb-Way, 0.91 for Highway, and 0.90 for City-Way states, effectively distinguishing driving states using EEG features. Model agnostic feature attribution approach highlighted key EEG contributors, such as power in Theta and Delta bands, and power ratios (DAR, DTR), providing interpretability and aligning with established neurological indicators of cognitive workload. CONCLUSION: This study demonstrates the potential of EEG-based features combined with interpretable machine learning for driver states prediction. The approach offers a foundation for personalized, responsive driving assistance systems at improving road safety and preventing disability by monitoring to drivers' cognitive states.
