4.2 Interrelation of the performed studies
In Figure 4.1 the underlying research question, as well as the dependencies of the per- formed studies, reported in the following chapters, is shown.
Designing the optimal learning material for an EEG-based learning system is the starting point for each study. Therefore, research question 1: “How does suitable learning material for an adaptive learning environment look like?” is treated in several chapters (5, 7, 10 and 11), while the obtained knowledge of each study influences the task design for the subsequent studies.
“What kind of features lead to a precise workload detection in EEG signals?” is research question 2. Since suitable neurophysiological features are key prerequisites for successful classification results, different feature selection and extraction methods are comparatively shown in chapter 6. These findings were considered for data analysis in the following studies, reported in chapter 7ff.
An additional essential research question for developing an efficient and realizable learning environment is question 3: “Can generalizable classifiers be developed for an accurate workload prediction?” This research question is extensively examined in chapter 7, 8, 9 and 10, where workload prediction is realized by using cross-task classification or cross- subject regression. Further analysis as subjective cognitive workload labeling and task order effects are analyzed while using cross-task classification.
“Is a precise workload prediction across subjects possible in real-time?” is research question 4 and the goal of this thesis. Thus, the findings from all prior studies regarding task design, suitable neurophysiological features, as well as workload prediction methods were considered for the last study, reported in chapter 11.
The innovation and novelty of this thesis is: using an objective, non-obtrusive adaptation measurement, which can ensure that a learner remains in their optimal workload capacity range. To the best of my knowledge, there are no studies dealing with online workload de- tection based on EEG data and complex learning material adaptation, keeping a learner in his/her optimal workload capacity range. The results from previous studies are promising for using EEG data to adapt tasks. Thus, it seems advisable to develop an EEG-based adap- tive learning environment, providing individual technological support for learners. There- fore, this learning environment adapts learning materials to learners’ levels of expertise and workload capacity, to support them in their learning process successfully.
4 Basic ideas, hypotheses and objectives of this thesis Research Question 1 Learning Material (Chapter 5, 7, 10, 11) Neurophysiological Features (Chapter 6) Research Question 2 Generalized Classifier (Chapter 7ff.) Research Question 3 Cross-Task Classification (Chapter 7) Subjective Labeling (Chapter 8) Task Order (Chapter 9) Cross-Subject Regression (Chapter 10) Research Question 4 Online Application (Chapter 11)
5 Workload classification using complex
learning material
In the following sections, it will be explored, if complex learning material imposes different levels of cognitive workload which could be measured by using EEG data. This chapter deals with classifying different cognitive workload levels during learning and non-learning tasks with the help of EEG data from students. The research question was, if single-subject single-trial EEG data allows for classification and which features are used. The results presented in this chapter were partially published in [127].
5.1 Study design
An introduction in the participants data, the EEG recordings, the task design as well as data analysis will be given in the following sections.
5.1.1 Participants and EEG recordings
Subjects were ten German students (12 – 15 years old, mean age: 13.6, 6 female, 4 male) without prior knowledge of angular geometry who participated voluntarily in the experi- ment. None of the subjects had participated in an EEG study before. A set of 16 passive electrodes, attached to the scalp, were used to record EEG signals. The electrodes were placed according to the International Electrode (10-20) Placement System. An electroocu- logram (EOG) was recorded through three additional electrodes: two placed horizontally at the outer canthi of both eyes and one placed in the middle of the forehead between the eyes. EEG and EOG signals were amplified by two 16-channel biosignal amplifier sys- tems (g.USBamp Generation 3.0). The sampling rate was 256 Hz and the impedance of each electrode was < 10 kΩ. The EEG data was high-pass filtered at 0.1 Hz and low-pass filtered at 100 Hz during the recording. Furthermore, a notch-filter was utilized between 48 Hz − 52 Hz to filter the noise of the power line.
5.1.2 Paradigm
The experiment is a within-subject design and comprises of four phases. First, the subjects had to solve a pre-test to assure they had no prior knowledge of angular geometry. The pre-test is composed of four main tasks with 30 sub-tasks in total. The participants had to solve these tasks with no time constraints. Phase 2 consisted of three learning cycles, 11 minutes each. In each cycle, the subjects were asked to study five angle theorems and watch five comic-strips (see Figure 5.1). Each theorem and comic-strip was presented for
5 Workload classification using complex learning material Pre- Test Post- Test Comic Strip 45 Time (sec) Learning Session 10 Feed- back 45 Angle Theorem 10 10 Feed- back 10
Figure 5.1:Chronological stimulus presentation during the learning session.
45 sec in an alternating order. By studying difficult material like angle theorems, high levels of workloads were induced. On the other hand, studying easy materials as reading comic- strips caused low levels of workload. Furthermore, the comic-strips were used, since the stimuli despite from the learning stimuli should also induce continuous eye-movements and require cognitive processes. All participants studied 15 angle theorems (3 × 5) and 15 comic-strips. In phase 3, students applied the theorems to geometrical exercises with a German version of the Carnegie Learning’s Cognitive Tutor provided by Schwonke et al. [128]. Finally, a post-test had to be accomplished, where participants had to solve the same problems as in the pre-test, to achieve a direct comparison of their knowledge before and after the learning cycles. The whole experiment took approximately three hours for each subject.