![](bg3.png)
3.1 The Muse Brainwave Reader
EEG data were collected using wireless, bluetooth-
enabled Muse headsets. Power from a headset’s
rechargeable battery typically lasted for about 45 min-
utes, which limited the length of a testing session.
The headsets were equipped with four sensors,
with two placed at the mastoids (on the ear clips) and
two at frontal regions Fp1 and Fp2. Muse headsets
initially oversample EEG and then downsample it
to yield a 220Hz signal with 2uV (RMS) noise.
(Kovacevic et al., 2015; Hashemi et al., 2016)
Participants. Fourteen undergraduate and graduate
level college students were recruited for this study.
They agreed to participate in four 20 minute sessions
in which they would solve GRE math and reading
problems as well as relax with open or closed eyes
while their brainwaves were being recorded using the
Muse headband.
In total, 12 subjects finished all the four surveys.
Two subjects had severe electrode connectivity issues
in at least one of their 4 sessions, and so their data
needed to be discarded. This left 40 valid record-
ings for 10 subjects. We had 80 minutes of recorded
data for each subject. The Muse headband generates
a wide variety of data, but we were only interested
in recording the relative power of the five standard
bands (alpha, beta, gamma, delta, theta) at a rate of
10Hz, which generated a total of 480,000 samples per
subject. Each sample consisted of five relative power
bands for each of four electrodes, yielding 20 floating
point numbers between 0 and 1. Since we were look-
ing at relative power, the sum of these five values for
each electrode was always equal to 1.0.
The average age of the subjects who completed
the experiment was 22.8. There were four Females
and six Males.
4 K-MEAN CLASSIFICATION OF
CORTICAL OSCILLATIONS
Over each 20-minute experimental session we used
the Muse band to collect ∼12,000 samples of cortical
oscillations data. Samples were taken over the course
of four five-minute conditions, presented one after an-
other. The four conditions were
• MATH (M) in which subjects subjects attempted
to solve problems from GRE Quantitative tests
• SHUT (S), in which subjects relaxed with eyes
shut
• READ (R), in which subjects read and tried to an-
swer questions from GRE Verbal tests
• OPEN (O), in which subjects relaxed with eyes
open
Our analysis used the alpha, beta, delta, gamma, and
theta band oscillatory power collected from each of
the four MuseBand sensors as output by the Muse-
band device itself.
In this paper we are only focusing on the READ
and OPEN data for each subject as we want to esti-
mate how effectively this data can be used to distin-
guish Reading activity from Open Eye relaxation, in
which subjects’ minds tend to wander.
After collecting all of the subjects’ data for all
four sessions, we extracted the READ and OPEN data
and combined it into a single dataset for each individ-
ual. Fig. 1 shows the raw data for subjects 1 and 2.
The horizontal axis is the time with tic marks every
5 minutes. The vertical axis is the relative power and
the four graphs from top to bottom correspond to the
four electrodes from left to right (left ear, left fore-
head, right forehead, right ear). The top plot shows
the results for Subject 1 and the bottom plot for Sub-
ject 2. The color of each line indicates the band.
Some of the boundaries between READ and OPEN
are clearly visible in these Figures, as they alter-
nate READ, OPEN, READ, OPEN, READ, OPEN,
READ, OPEN in 3000 sample blocks, but one sees
quite a bit of variance both between the two subjects
and within each subject.
Fig. 2 shows a magnified view of the data from
the right ear electrode of Subject 1 from minute 5 to
minute 7 as they are switching from reading to open
eye relaxation. We can see that the alpha band domi-
nates after minute 6. Some of the other transition are
harder to see visually, which is why machine learning
is needed to classify these activities.
Next, we applied a k-means clustering algorithm
(for various k, but k=12 was optimal for this data
and then used those clusters to form a classifier us-
ing the standard approach, as follows. For each clus-
ter, we calculated the number of samples for each
type (READ and OPEN) that were in that cluster, and
used that information to label each cluster as either a
READ or an OPEN cluster. Any new sample, would
then be compared to the cluster points, and assigned
the label of whichever cluster point was closest.
Fig. 3 shows an example of a k-means classifier
with k=12. This classifier was generated using the
first three sessions of Subject 1 (shown in top plot
in Fig. 1). The horizontal axis corresponds to the 12
clusters. Each cluster has two bars, one for the READ
samples and one for the OPEN samples that are clos-
est to that cluster point than any of the other 12 cluster
A Personalized Reading Coach using Wearable EEG Sensors
503