EEG-based workload estimation using ignored probes : A comparison of three tasks

Mental workload (MW) is frequently defined as task difficulty and the associated mental effort (Gevins & Smith, 2007). This mental state is under a lot of focus in the neuroergonomics community, as most researchers aim at better assessing this state in order to develop smart technologiesthat enhance operator’s safety and performance (Parasuraman et al., 2012). It can be estimated using markers derived from electroencephalography (EEG) such as event-related potentials (ERPs). A few authors propose a classification of these ERPs for MW estimation. Brouwer et al. used 7electrodes and achieved 64% of correct binary classifications (2012). Recently, it was proved that ERP spatial filtering could significantly enhance workload classification (Mühl et al., 2014; Roy et al., 2015). The authors respectively achieved 72% and 98% of correct classifications. Yet, this literaturefocused only on one type of task -memory task- and used task-dependent probes. This preliminary study investigates the use of ERPs for MW estimation for 3 tasks that modulate MW based on different cognitive processes: the MATB-II, an arithmetic task, and the Sternberg memory task.In order to get closer to a real life implementation with minimally intrusive probes, we used infrequent ignored task-independent auditory probes (1000 Hz; 6-30s interval). Three volunteers performed each task for 10 minutes (2 MW levels). This research was approved by the French health safety agency (B140052-31). The participants’ behavioral responses were measured, in addition to their answer to the RSME questionnaire and their EEG activity (32 electrodes; 500 Hz). A single-trial subject-dependent processing chain was applied to the data. It included preprocessingsteps (e.g. automatic artifact removal), an ERP extraction step (600 ms; 200 ms baseline), a spatial filtering step (2 CCA filters) and a Fisher Linear Discriminant Analysis classification with a 5-fold cross-validation. The behavioral and subjective results confirmed MW modulation for all tasks. As regards MW estimation, promising results were obtained for the MATB-II with as high as 93% of correct classifications, whereas significantly lower results were obtained for the 2 other tasks (respectively 53% and 66%; p<0.01). All results were significantly higher than random (p<0.05). Thisstudy demonstrates that EEG-based MW estimation is feasible using the spatially filtered ERPs elicited by ignored task-independent probes. Yet its performance seems to depend on the type of task. Indeed, higher accuracies were reached for the task that was the most stimulating (MATB-II).The others tasks, that gave rise to alpha-corrupted ERPs, might not have been stimulating enough and were therefore prone to time-on-task induced noise.