Analyzing EEG signals to find association between student's cognitive load and learning profiles

Description

This thesis has two main goals. First, it aims to evaluate the ability of the
Curvex headset to measure cognitive load. Secondly, it analyzes if there is a
relationship between the learning profile of students and their EEG data.
The reliability of commercial headsets with single-channel devices has been investigated
in multiple research. However, they did not investigate its reliability
in a learning context. One of these commercial single-channel headsets is Neurosky
which has been involved in numerous research, some of which showed
promising results. This study uses the Curvex device, which includes the Neurosky
sensor in a new headset design along with an application originally made
for meditation. The paper is the first to use this device, and it provides an
evaluation of its performance.
The methods used to evaluate the capability of the device are inspired by previous
similar research. The experiment was conducted with 15 students who were
asked to complete the Raven’s Advanced Progressive Matrices (RAPM) test,
and their learning profiles were determined using the Index of Learning Styles
Questionnaire (ILS).
Four different machine learning models were used to predict task difficulty and
student learning profiles: Naive Bayes, Support Vector Machines (SVM), Random
Forest, and eXtreme Gradient Boosting (XGBoost). Nominal task difficulty
(easy/difficult) was predicted with an accuracy of 71.11% and an F-score
of 73.74% using the SVM model with the inputs of theta, high alpha, low beta,
and flow measures of the device. Students’ preference for active/reflective learning
could be predicted with an accuracy of 68.17% and an F-score of 60.15% using XGBoost and the previously listed inputs. These results are similar to
the outcomes of relevant studies.
It was shown that the cognitive load measure provided by the device does not
have a good predictive power of task difficulty or learning profiles, and theta
and alpha values have better distinguishing power. Thus, in the future development of the learning profile recognition, multiple inputs should be considered to increase the accuracy. Several concerns are raised about the validity of the measurements, and improvement points are suggested to increase the validity of the device.

Period	24 Jan 2022 → 24 Jun 2022
Examinee	Szoboszlai Anna
Degree of Recognition	National