EEG-based detection of mental workload level and stress: the effect of variation in each state on classification of the other

Abstract

Objective. A passive brain-computer interface (pBCI) is a system that continuously adapts human-computer interaction to the user’s state. Key to the efficacy of such a system is the reliable estimation of the user’s state via neural signals, acquired through non-invasive methods like electroencephalography (EEG) or near-infrared spectroscopy (fNIRS). Many studies to date have explored the detection of mental workload in particular, usually for the purpose of improving safety in high risk work environments. In these studies, mental workload is generally modulated through the manipulation of task difficulty, and no other aspect of the user’s state is taken into account. In real-life scenarios, however, different aspects of the user’s state are likely to be changing simultaneously—for example, their cognitive state (e.g. level of mental workload) and affective state (e.g. level of stress/anxiety). This inevitable confounding of different states needs to be accounted for in the development of state detection algorithms in order for them to remain effective when taken outside the lab. Approach. In this study we focussed on two different states that are of particular importance in high risk work environments, specifically mental workload and stress, and explored the effect of each on the ability to detect the other using EEG signals. We developed an experimental protocol in which participants performed a cognitive task under two different levels of workload (low workload and high workload) and at two levels of stress (relaxed and stressed) and then used a linear discriminant classifier to perform classification of workload level and stress level independently. Main results. We found that the detection of both mental workload level (e.g. low workload vs. high workload) and stress level (e.g. stressed vs. relaxed) were significantly diminished if the training and test data came from different as opposed to the same level of the other state (e.g. for mental workload classification, training on data from a relaxed condition and testing on data from a stressed condition, rather than both training and testing on the relaxed condition). The reduction in classification accuracy observed was as much as 15%. Significance. The results of this study indicate the importance of considering multiple aspects of a user’s state when developing detection algorithms for pBCI technologies.