Predictive Markers in EEG data forSusceptibility: A Data Driven Approach

Abstract

In this thesis, I test whether data-driven approaches can predict cognitive load based on EEG data. Previous research has tested in an oddball paradigm how the magnitude of an Event-Related Potential relates to cognitive load. However, an intervention is not always desirable in everyday scenarios. I therefore test, whether it is possible to predict the different cognitive load conditions before the intervention (i.e., oddball stimulus) is presented in three data-driven experiments. In experiment 1, I used machine learning to train a model that classifies the data for different conditions. In experiment 2, I test which characteristic/features of the data have the highest predictive power for each condition. In experiment 3, I tested how a Fourier transformation and data-driven approach can be used to complement each other. The combined results show that using a data-driven approach; I can predict cognitive load for the experiment at hand. However, the machine learning approach (experiment 1 and 2) require a long processing time. The combined approach (experiment 3) provides a consistent pattern that can differentiate between the cognitive load conditions. Further research is needed to test the generality of these findings for different datasets.