Section: New Results
Mental state decoding from EEG signals using robust machine learning
Participants: Aurélien Appriou, Smeethy Pramij, Khadijeh Sadatnejad, Aline Roc, Léa Pillette, Thibaut Monseigne, Fabien Lotte
External collaborators: Andrzej Cichocki, Pierre-Yves Oudeyer, Edith Law, Jessie Ceha, Frédéric Dehais, Alban Duprès, Sarah Blum, Nicolas Drougard, Sébastien Scannella, Raphaëlle N. Roy
Modern machine learning algorithms to classify cognitive and affective states from electroencephalography signals: 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. We have thus explored such machine learning algorithms, proposed new variants of them, and benchmarked them with classical methods to estimate both mental workload and affective states (Valence/Arousal) from EEG signals. We studied 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 little 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. This work was just accepted for publication in the IEEE System Man and Cybernetics magazine.
Towards decoding curiosity from Brain and physiological signals: The neurophysiological mechanisms underlying curiosity and intrinsic motivation are currently not well understood. However, being able to identify objectively, from neurophysiological signals, the curiosity level of a user, would bring a very useful tool both to neuroscientists and psychologists, to understand curiosity deeper, as well as to designers of human-computer interaction, in order to trigger curiosity or to adapt an interaction to the curiosity levels of its users. A first step to do that, is to collect neurophysiological signals during known states of curiosity, in order to develop signal processing/machine learning tools to recognize those states from such signals. We designed and ran an experimental protocol to measure both brain activity through Electroencephalography (EEG) and physiological responses (heart rate, skin conductance, Electrocardiogram) when subjects were induced into different states of curiosity. During the experiment, fun facts were presented to subjects to induce different levels of curiosity. We obtained those fun facts using the Google functionality "I’m feeling curious" as well as crowdsourcing. A subject could choose a fun fact that made him curious, and push forward with a 4-to-10 questions chain on this theme. For each question on a given theme, a subject could choose to reveal the answer (interpreted as a curious state) or to skip it (interpreted as a non-curious state). Skipping an answer will automatically break the chain and will point the subject to the next fun fact. Neurophysiological signals were collected from 28 subjects, between a question and the choice of revealing the answer. Then those subjects graded the question on a 1-to-7 curiosity level scale. We are currently working on finding biological markers of curiosity by analyzing the collected signals using machine learning.
Channel Selection over Riemannian manifold with non-stationarity consideration for Brain-Computer interface applications: EEG signals are essentially non-stationary. Such non-stationarities, including cross-trial, cross-session, and cross-subject non-stationarities, are the result of various neurophysiological and extra-physiological causes. Such non-stationarities lead to variations in BCI users' performance. To handle this problem, we designed and compared multiple criteria for selecting EEG channels over the Riemannian manifold, for EEG classification. These criteria aim to promote EEG covariance matrix classifiers to generalize well by considering EEG data non-stationarity. Our approach consists of both increasing the discriminative information between classes over the manifold and reducing the dispersion within classes. We also reduce the influence of outliers in both discriminative and dispersion measures. The criteria were evaluated on EEG signals recorded from a tetraplegic subject and dataset IVa from BCI competition III. Experimental evidences confirm that considering the dispersion within each class as a measure for quantifying the effects of non-stationarity and removing the most affected channels can improve BCI performance (see Figure 9). This work was submitted to ICASSP 2020.
Monitoring Pilot's Mental Workload Using ERPs and Spectral Power with a Six-Dry-Electrode EEG System in Real Flight Conditions: Recent technological progress has allowed the development of low-cost and highly portable brain sensors such as pre-amplified dry-electrodes to measure cognitive activity out of the laboratory. This technology opens promising perspectives to monitor the “brain at work” in complex real-life situations such as while operating aircraft. However, there is a need to benchmark these sensors in real operational conditions. We therefore designed a scenario in which twenty-two pilots equipped with a six-dry-electrode EEG system had to perform one low load and one high load traffic pattern along with a passive auditory oddball. In the low load condition, the participants were monitoring the flight handled by a flight instructor, whereas they were flying the aircraft in the high load condition. At the group level, statistical analyses disclosed higher P300 amplitude for the auditory target (Pz, P4 and Oz electrodes) along with higher alpha band power (Pz electrode), and higher theta band power (Oz electrode) in the low load condition as compared to the high load one. Single trial classification accuracy using both event-related potentials and event-related frequency features at the same time did not exceed chance level to discriminate the two load conditions. However, when considering only the frequency features computed over the continuous signal, classification accuracy reached around 70% on average. This study demonstrates the potential of dry-EEG to monitor cognition in a highly ecological and noisy environment, but also reveals that hardware improvement is still needed before it can be used for everyday flight operations. This work was published in the journal Sensors in .