Estimating cognitive or affective states from brain signals is a key but challenging step in the creation of passive brain-computer interface (BCI) applications. So far, estimating mental workload or emotions from EEG signals is only feasible with modest classification accuracies, thus leading to unreliable neuroadaptive applications. However, recent machine learning algorithms, notably Riemannian geometry based classifiers (RGC) and convolutional neural networks (CNN), have shown to be promising for other BCI systems, e.g., motor imagery-BCIs. However, they have not been formally studied and compared together for cognitive or affective states classification. This paper thus explores such machine learning algorithms, proposes new variants of them, and benchmarks them with classical methods to estimate both mental workload and affective states (Valence/Arousal) from EEG signals. We study these approaches with both subject-specific and subject-independent calibration, to go towards calibration-free systems. Our results suggested that a CNN obtained the highest mean accuracy, although not significantly so, in both conditions for the mental workload study, followed by RGCs. However, this same CNN underperformed in both conditions for the emotion data set, a data set with small training data. On the contrary, RGCs proved to have the highest mean accuracy with the Filter Bank Tangent Space classifier (FBTSC) we introduced in this paper. Our results thus contributed to improve the reliability of cognitive and affective states classification from EEG. They also provide guidelines about when to use which machine learning algorithm.

While often presented as promising assistive technologies for motor-impaired users, electroencephalography (EEG)-based Brain-Computer Interfaces (BCIs) remain barely used outside laboratories due to low reliability in real-life conditions. There is thus a need to design long-term reliable BCIs that can be used outside-of-the-lab by end-users, e.g., severely motor-impaired ones. Therefore, we propose and evaluate the design of a multi-class Mental Task (MT)-based BCI for longitudinal training (20 sessions over 3 months) of a tetraplegic user for the CYBATHLON BCI series 2019. In this BCI championship, tetraplegic pilots are mentally driving a virtual car in a racing video game. We aimed at combining a progressive user MT-BCI training with a newly designed machine learning pipeline based on adaptive Riemannian classifiers shown to be promising for real-life applications. We followed a two step training process: the first 11 sessions served to train the user to control a 2-class MT-BCI by performing either two cognitive tasks (REST and MENTAL SUBTRACTION) or two motor-imagery tasks (LEFT-HAND and RIGHT-HAND). The second training step (9 remaining sessions) applied an adaptive, session-independent Riemannian classifier that combined all 4 MT classes used before. Moreover, as our Riemannian classifier was incrementally updated in an unsupervised way it would capture both within and between-session non-stationarity. Experimental evidences confirm the effectiveness of this approach. Namely, the classification accuracy improved by about 30% at the end of the training compared to initial sessions. We also studied the neural correlates of this performance improvement. Using a newly proposed BCI user learning metric, we could show our user learned to improve his BCI control by producing EEG signals matching increasingly more the BCI classifier training data distribution, rather than by improving his EEG class discrimination. However, the resulting improvement was effective only on synchronous (cue-based) BCI and it did not translate into improved CYBATHLON BCI game performances. For the sake of overcoming this in the future, we unveil possible reasons for these limited gaming performances and identify a number of promising future research directions. Importantly, we also report on the evolution of the user's neurophysiological patterns and user experience throughout the BCI training and competition.