Studying the generalisability of cognitive load measured with EEG

• Published in: Biomedical signal processing and control 2021-09, Vol.70, p.103032, Article 103032
• Main Authors: Cabañero Gómez, Luis, Hervás, Ramón, González, Iván, Villarreal, Vladimir
• Format: Article
• Language: English
• Subjects: Cognitive load; Electroencephalography; Generalisability; N-back test; Stroop test
• ISSN: 1746-8094; 1746-8108
• DOI: 10.1016/j.bspc.2021.103032

•Study of generalisability for estimating cognitive load along with contexts and subjects.•Proved methods and features for studying cognitive load determined with EEG.•Low generalisability while comparing models from different people (inter-subject).•High generalisability while comparing models from different cognitive load tests (inter-context). Cognitive load is defined as the information the working memory is holding at specific moment. It has been studied not only in education but also in medicine, particularly regarding the diagnosis of disorders such as mild cognitive impairment. In our study, we examined the generalisability of cognitive load metrics obtained with electroencephalography (EEG) across various participants and contexts in four combinations: (a) inter-subject and intra-context, (b) inter-subject and inter-context, (c) intra-subject and intra-context, and (d) intra-subject and inter-context. EEG signals were recorded from 19 participants as they completed two cognitive assessment tests with differentiable levels of cognitive load: the n-back test and Stroop test. The data obtained were processed to extract numerous features that were later reduced following a forward feature selection and used to train different models for the various combinations of generalisability. Analysing the performance of trained models revealed that classification in (a) showed results close to random classification, in (b) showed significative differences between Stroop levels, in (c) revealed a classifier able to find patterns associated with the various levels of cognitive load for both tests, and in (d) indicated that Stroop levels were differentiable. Although the methods analysed can be used to determine patterns for each participant in a particular context and applied to different contexts, they cannot establish generalisable models among participants in a single context. Our analysis of the feature selection captured a group of powerful algorithms and parameters potentially usable for extracting features in cognitive load analysis.