Enhancing single-trial mental workload estimation through xDAWN spatial filtering

Abstract

Mental state monitoring is a topical issue in neuroengineering, more particularly for passive brain-computer interface (pBCI) applications. One of the mental states that are currently under focus is mental workload. The level of workload can be estimated from electroencephalographic activity (EEG) and markers derived from this signal. In active BCI applications, a well-known neurophysiological marker, the event-related potential (ERP), is commonly enhanced using a spatial filtering step. In this study, we evaluated how a spatial filtering method such as the xDAWN algorithm could improve mental workload classification performance. Twenty participants performed a Sternberg memory task for 18 minutes with pseudorandomized trials of low vs. high workload (2/6 digits to memorize). Three signal processing chains were compared on their performance to estimate mental workload from the single-trial ERPs of the test item (i.e. present/absent in the memorized list). All 3 included an FLDA classifier with a shrinkage covariance estimation and a 10-fold cross-validation. One chain used the ERPs of a relevant electrode for workload estimation (Cz) and the 2 others used the ERPs of the 32 electrodes and an xDAWN spatial filtering step with either 1 or 2 virtual electrodes kept for classification. Statistical analyses revealed that spatial filtering significantly improved mental workload estimation, with up to 98% of correct classification using the xDAWN algorithm and 2 virtual electrodes.