Project Final Report

Project Title: Effects of speech-in-noise identification and question difficulty on cognitive load through EEG
analysis

Team Members: Jane Yu (yu.2164@[Link]) and Sheila Colbert (colbert.81@[Link])

Introduction:
The cognitive architecture consists of both a working memory, limited by capacity and time to process novel
information, and a long-term memory, unlimited capacity for storing learned information (Antonenko et al.,
2010). Within the working memory, there are various types of cognitive loads that can be applied which include
intrinsic load, germane load, and extrinsic load. Intrinsic load is the inherent complexity of information’s
content in relation to the specific learner’s level of expertise and germane load is the resources required to
handle this intrinsic load (Antonenko et al., 2010). Extraneous load is the unnecessary mental burden due to
information that does not contribute to learning (Antonenko et al., 2010). One method of increasing cognitive
load is to utilize intelligence tests and prompt subjects to answer progressively more difficult questions through
the duration of the test, as done in a study by Friedman et al (Friedman et al., 2019). In this study, the alpha and
beta power brain waves were recorded using EEG and the alpha and beta ratio was analyzed to evaluate
cognitive load (Friedman et al., 2019). In another study done by Marsella et al. cognitive load was manipulated
in participants with asymmetric hearing loss through speech-in-noise identification tasks where various noise
modifications were made (Marsella et al., 2017). The cognitive load was measured using EEG through
analyzing band density of the alpha and theta bands.

As displayed in these studies, cognitive load has been shown to be accurately measured by the physiological
index of EEG. EEG allows for noninvasive collection of continuous cognitive load data which provides more
detailed and accurate interpretation of cognitive load than other methods currently in use like objective rating
scales. Alpha, beta, and theta brain waves are the most relevant to cognitive load testing via EEG. Alpha brain
waves are the dominant wave in a normal human EEG with a frequency of 8-13 Hz (Antonenko et al., 2010).
Cognitive load can be manipulated in order to record its impact via EEG. This project is in collaboration with a
Nationwide client researching concussions in minors. The aim is to create an EEG acquisition circuit that is
portable, reliable, and robust in order to be able collect EEG data for more reliable detection of concussions in
patients. Sport-related concussions occur at an incidence rate of approximately 3.0 million per year and are
known to cause cognitive decrement so this EEG recording will improve ease of diagnosis of concussion
patients (Covassin & Elbin, 2010).

This project utilized a two-factor, two-level experiment in order to influence cognitive load based on the studies
mentioned previously. The first factor was question difficulty and was broken down into true/false and
descriptive questions. In this case, the descriptive questions cause the subject to utilize an increased cognitive
load as the questions require more mental capacity to anwer and therefore are more difficult. The second factor
was noise level and was broken down into no noise and high noise levels. The high noise causes the subject to
utilize increased cognitive load as the subject must remain attent and answer the question as the noise is applied.
The changes in cognitive load were recorded using an EEG acquisition circuit and then analyzed by the
alpha/beta wave ratios as in the study done by Friedman et al. The hypothesis of this experiment was that
increasing question difficulty and noise level will increase cognitive load as displayed by the alpha/beta EEG
wave ratio measured by the signal acquisition circuit.

Methods:
Subject and Experimental Preparation
Question Preparation
The factors of the experiment were established and prepared before testing. Both the true/false questions and the
description questions had a bank of 20 questions each in order to ensure that 20 total trials could be completed.
True/false questions were created based on physical appearance and common knowledge of the subject.
Description questions were created using a random noun generator. A random number generator was used to
determine the order and which questions would be asked during testing.

Trial Determination
A total of 5 replicates were completed with each replicate including four trials with all possible combinations of
the factor’s levels. The trial order within each replicate was determined using a random generator. The levels
were each assigned a value: description as 1, true/false as -1, high volume as 1, and no volume as -1. A table of
the trial order is included in Appendix A.

Confounding Variables
Before testing, the subject reported current physical state including consumption of one cup of coffee three
hours prior to testing, eight hours of sleep the night before testing, and minimal stress [Link]
confounding variables were also recorded. Ambient room noise was recorded to be approximately 50-60 dB
throughout testing. The subject’s heart rate remained around 85 BPM throughout testing. The subject was
placed in a preparation position with legs and arms uncrossed, eyes open, and facing a blank wall and remained
in this position throughout testing.

Procedure
Calibration
The equipment was calibrated before testing to ensure proper data collection. After connecting the Arduino to
the computer, the Arduino was calibrated using the code file: Arduino_Calibration.ino. MATLAB was also
calibrated to ensure proper connection to the Arduino with the code file: Arduino_Read.m. Once calibration was
established, the EEG acquisition circuit was connected to the Arduino. The circuit was also covered with a
piece of dark fabric to act as a shield to prevent light from interfering with data collection.

Sensor Placement
The sensors were then placed on the subject’s scalp in a manner similar to Figure 1 below. The hair was parted,
the scalp was wiped with an alcohol pad, and electrode gel was added to obtain the best possible connection.
Earbuds were also placed in the subject’s ears at this point and remained in the ears for the duration of the
testing. After electrode placement, a swim cap was added over top of the electrodes, and then the head was
wrapped with tied saran wrap to establish proper contact. The subject then remained in the preparation position
as described above to ensure the subject was in a relaxed state before data was acquired.
 Figure 1: Electrode placement on subject

Data Acquisition
For each trial shown in Appendix A, the code: Arduino_Read.m was run for 15 seconds to acquire data. During
the 15 second recording, if a trial involved high volume, the noise began playing at 80 dB during MATLAB
selection of recording settings. No noise trials simply involved leaving the earbuds within the ear, but playing
no noise. During true/false trials, two true/false questions were asked spaced evenly in the 15 second recording
window using the questions determined by the random number generator from the previously established bank.
During description questions, one question was asked at the beginning of the 15 second window and the subject
was given the remaining time to answer the question. Each trial was saved with a descriptive name to the
computer to later be analyzed.

Data Analysis
Data Filtering
After data had been collected for all 20 trials, each trial’s data was run through various MATLAB codes for a
notch filter, and bandpass filtering. Then, the data was run through the code: windowedEEGAnalysis.m to
determine the mean and maximum peaks for the EEG frequency bands. Approximately 13-14 data points of
mean alpha and beta bands from each trial were used to determine the alpha/beta band ratio.

Statistical Analysis
Then, statistical analysis of the data was performed by running a linear regression in the program SPSS. Along
with the EEG wave values, the values used to differentiate the levels, and the repetitions were included in the
regression.

Results:
After filtering the data, 13 data points remained for each of the 20 trials performed. Each data point
corresponded to approximately one second of the EEG recording for the trial. A linear regression for alpha/beta
ratio was performed in SPSS first. The dependent variable was alpha/beta ratio, and the independent variables
used were the two factors along with repetition number. This analysis resulted in a p-value greater than the
alpha value of 0.05 for all independent variables. No significance was found for question difficulty (p-value =
0.765), noise level (p-value = 0.916), or repetition (p-value = 0.151). A histogram and normal probability plot
for this regression can be found below in Figure 2. A box plot of the ratios can be found below in Figure 3.
From this it was apparent that none of the factors had an effect on the alpha/beta ratio for measuring cognitive
load.
 Figure 2: Normal probability plot (left) and histogram plot (right) for alpha/beta ratio linear regression

Figure 3: Box plot for alpha/beta ratios

A second linear regression was then performed on the data for just the mean alpha band power. The dependent
variable was mean alpha band, and the independent variables used were the two factors along with repetition
number. This analysis resulted in a p-value less than the alpha value of 0.05 for repetition (p-value < 0.001) and
question difficulty (p-value < 0.001). No significance was found for noise level (p-value = 0.277). A histogram
and normal probability plot for this regression can be found below in Figure 4. A box plot of the ratios can be
found below in Figure 4. From this, it showed that question difficulty and repetition did have an effect on
cognitive load as measured through mean alpha band power.

Figure 4: Normal probability plot (left) and histogram plot (right) for mean alpha band power linear regression
 Figure 5: Box plot for mean alpha power

Discussion and Conclusions:

Based on the results of this study, repetition (p-value < 0.001) and question difficulty (p-value < 0.001) did have
a significant effect on cognitive load when measured through mean alpha power. However, the same did not
hold true for when alpha/beta ratio was used to measure cognitive load. No significance was found for any
factor (p-value > 0.05) when alpha/beta ratio was used.

The initial predictions of the experiment were that an increase in cognitive load by increasing question difficulty
and noise levels could be observed by alpha/beta wave ratios in EEG acquisition. Based on the p-values higher
than the significance value α = 0.05 from the regression analysis of the alpha/beta wave ratios considering both
factors and repetition of the experiment, it was determined that the results failed to meet the initial predictions.
These results were consistent with prior literature when mean alpha power was used to analyze cognitive load.
However there were inconsistencies when using alpha/beta ratio, as previous literature did find significant
differences with speech-in-noise identification affecting cognitive load.

Although it cannot be determined whether or not this EEG was recording cognitive load by the collected data,
this EEG acquisition circuit was able to record EEG waves. The accuracy of these recordings is unable to be
concluded from the collected data.

For future experiments, a major goal will be to better understand what factors would be the best measure of
cognitive load. As seen with the difference in results for when alpha/beta ratio was used, to when mean alpha
power was used, understanding which is a better measure of cognitive load will be important. Further testing
should also increase the number of levels used for each factor, to get a more precise understanding of the trends
for how each factor affects cognitive load. Beyond this a more controlled testing environment would help to
create more accurate data. With the data collected in this experiment, the ambient noise levels fluctuated near 60
dB and could have acted as a confounding variable, which would have led to skewed results. As such, a closed
environment with little to no ambient noise and minimal visual stimulus (moving people, flashing lights,
computer screens, etc.) would help reduce noise in the collected EEG signal.

Acknowledgements:
JY and SC both contributed equally towards conceptualization, methodology, experimentation, and drafting the
report. We would like to thank Dr. OR, Devleena Das, Mukul Govande, and Genevieve Thanh for technical
support, feedback, and guidance throughout this project.
Appendix A

In the table below for Factor 1 (Question Difficulty): 1 = Description, -1 = True/false

Total Run Repetition Run Factor 1 Factor 2

1 1 1 -1 1

2 1 2 1 1

3 1 3 -1 -1

4 1 4 1 -1

5 2 1 1 -1

6 2 2 1 1

7 2 3 -1 1

8 2 4 -1 -1

9 3 1 -1 1

10 3 2 1 -1

11 3 3 -1 -1

12 3 4 1 1

13 4 1 1 -1

14 4 2 -1 -1

15 4 3 1 1

16 4 4 -1 1

17 5 1 1 1

18 5 2 1 -1

19 5 3 -1 -1

20 5 4 -1 1
References:

1. Antonenko, P., Paas, F., Grabner, R., & van Gog, T. (2010). Using Electroencephalography to Measure
Cognitive Load. Educational Psychology Review, 22(4), 425–438.

2. Friedman, N., Fekete, T., Gal, K., & Shriki, O. (2019). EEG-Based Prediction of Cognitive Load in
Intelligence Tests. Frontiers in human neuroscience, 13, 191.

3. Marsella, P., Scorpecci, A., Cartocci, G., Giannantonio, S., Maglione, A.G., Venuti, I., Brizi, A.,
Babiloni, F. (2017). EEG activity as an objective measure of cognitive load during effortful listening: A
study on pediatric subjects with bilateral, asymmetric sensorineural hearing loss. International Journal
of Pediatric Otorhinolaryngology, 99, 1–7.

4. Covassin, T., & Elbin, R. J. (2010). The cognitive effects and decrements following concussion. Open
access journal of sports medicine, 1, 55–61.