Theoretical and experimental study of P300 ERP in the context of Brain-computer interfaces. Part II: An experimental study of inter-and intra-subject variability based on EEG recordings

Abstract

The P300 event-related potential (ERP) is closely related to cognitive processes such as: attention, working memory, consciousness, among others. This signal has been shown to have a large variability that occurs independently of the subject or the cognitive process being assessed. Detecting and characterizing this variability is important for understanding cognitive changes and using this knowledge to improve the performance of P300-based brain-machine interfaces (BCIs). There are several studies that have shown that the cognitive processes associated with P300 signal generation are highly variable between subjects, as this variability depends on factors such as: attention level, memory, age, experimental characteristics, and so on, but also that variability exists within the same subject when the same task is performed in different time periods. For this project, a study and analysis of inter- and intra-subject variability is proposed by characterizing the P300 signal according to the specificities of each subject. For this purpose, an experimental study is proposed to be performed in the laboratory of the Biological Neurocomputation Group (GNB) of Autonomous University of Madrid on 12 subjects, using the oddball paradigm to generate visual P300 ERP by electroencephalography (EEG). The proposed experiment was designed to evaluate variability in different circumstances: i) variability between the different subjects, ii) variability within the same subject during the 3 proposed experimental days, and iii) variability determined by the difference between the two helmets considered for the experiment due to their design and repositioning on the different experimental days. Two EEG helmets were used for this study: Enobio 8 and g-Tec g.USBAmp with dry electrodes. For the analysis, the characterization of the signals was first performed, and the cosine distance was calculated using the coefficient of determination (r2) results for each subject to show the difference between the P300 and non-P300 signals. In addition, a selection of electrodes using the Bayesian Linear Discriminant Analysis with an exhaustive search of electrodes by ”forward selection” (BLDAFS) is proposed to determine how the selected electrodes vary for each subject or experimental day. After characterizing the behavior observed in the different subjects, it can be concluded that variability is widespread and undeniable both between subjects and within the same subject when performing the same task on different days. This variability is highly related to the underlying neural activity of each individual as well as to the experimental characteristics. Moreover, the study of intra-subject variability has allowed us to observe how the results differ between the two helmets used for the experiment, highlighting that the difference between the accuracy values in the detection of P300 calculated for both helmets in the analyses performed is not representative. The exploratory variability analysis performed allowed the attainment of a configuration that reached accuracy values up to 76.10 % for the Enobio helmet and 79.34 % for the g-Tec helmet, both with passive and active dry electrodes. It was also found that not accounting for variability can degrade the performance of a BCI system.