Mental fatigue impacts workload level classification performances

International audienceBrain-computer interfaces (BCIs) are information transfer systems between a machine and one's brain that were initially designed to provide surrogate communication and motor abilities to handicapped people [1]. Lately, BCIs have been applied on other neural correlates than sensorimotor rhythms and visual potentials, and have been employed to monitor an operator's cognitive and emotional state, and to implicitly adapt systems in real-time using only brain waves as an input [2]. Those interfaces, now known as 'passive' BCIs, are the new means to answer neuroergonomics and mental state monitoring (MSM) issues. Most of those systems include measures of workload [1]. Workload can be defined as the load in working memory (i.e. number of items), the number of tasks to be performed simultaneously, and more generally as a measure of the amount of mental resources engaged in a task. Thus, it is considered a measure of task difficulty. Mental workload has been extensively documented in the MSM literature. Yet its interaction with mental fatigue has not. In real life situations, systems are mainly calibrated before being used for a long period of time (e.g. for driving). Therefore the calibration occurs only once, when the participant is awake and fully responsive. With increasing time-on-task, mental fatigue and reduced alertness arise. Classification algorithms have been designed to detect both mental fatigue and workload levels from electroencephalographic (EEG) signals. However, the impact of mental fatigue on workload classification performance has never been studied or taken into account. With this study, we intend to fill that void by presenting preliminary results. We manipulated workload using a modified Sternberg paradigm. Participants (n=4) had to memorize a list of sequential digits presented visually, and perform a recognition task on a probe item. Two levels of workload (low/high; 2/6 digits) and 2 levels of mental fatigue (low/high alertness) were considered. They were statistically assessed thanks to behavioral measures (reaction time and accuracy). Mental fatigue, assumed to increase with time-on-task, was also evaluated by subjective self-ratings. In addition to those measures, we recorded participants' EEG signal from 32 electrodes with a sampling frequency of 1000Hz. The signal was band-pass filtered between 1 and 40Hz and corrected for ocular artifacts using the signal recorded from 4 electro-oculographic electrodes. EEG data were then epoched to only analyze the segments where participants were loaded and had not yet performed the recognition task (segments of 800ms). This way, we avoided analyzing memory-encoding and memory-scanning neural correlates. For each participant, 160 segments were considered, with 40 segments per mental fatigue and workload level. Several processing steps were performed on the signal in the 5 frequency bands delta [1-4Hz], theta [4-8Hz], alpha [8-12Hz], beta [12-30Hz], and gamma [30-40Hz]. Those steps are: channel selection (selection of 15 channels per frequency band using a method based on Riemannian geometry [3]) and spatial filtering (with 6 CSP filters). A binary classification was then performed across experimental conditions using Fisher's linear discriminant analysis (FLDA). Using a 10-fold cross-validation procedure, we obtained results in classification performance for mental fatigue and workload level independently, as well as for workload level per mental fatigue state. Participants reported feeling increasingly tired with time-on-task (p<0.05). Their mental fatigue was also reflected by longer reaction times to the probe in the high workload condition (p=0.055). As regards the workload effect, participants were slower to respond in high workload conditions than in low ones (p<0.01). These behavioral results confirm the capability of the experimental setup to modulate both workload and mental fatigue. Regarding the EEG signal classification results, both mental fatigue and workload levels had a classification performance significantly different from random (p<0.01 and p<0.05 respectively), with a mean of 98.45% of good classification for mental fatigue and 69.25% for workload. When workload classification was performed for a specific mental fatigue state, classification performance reached 65.65% in the high alertness condition, which is significantly different from random (p<0.05), whereas it only reached 57.85% in the low alertness condition, which is not significantly different from random (p = 0.33). Moreover, to place ourselves in a real life situation, we trained our classifier for workload level discrimination on the data from the high alertness condition, and then tested it on the data from the low alertness condition. Again, classification performance was significantly degraded (p