27
Identification from Wearable Device Brain Signals
3. Watching video: A user watched a video for 1–3 minutes.
4. Game playing: A user played a computer game for 1–3 minutes.
5. Listening to music: A user listened to music for 1–3 minutes.
The above activities were repeated 10 times for five individuals resulting in a total of 50
datasets.
2.6 Knowledge Representation
One of the key aspects of any data mining activity is representing real-world entities
using the pertinent numeric data available for them. In our case, we want to capture
individuals and their activities using the signals emanating from their brain. This
chapter focuses on summarizing, and not manipulating, the signals collected from the
wearable into a fixed-length representation. As mentioned before, the Muse headband
collects data from four positions around the head. Each position provides five types
of waves: alpha, beta, gamma, delta, and theta. That means at any given point in time,
we receive a record with 20 values. There will be a stream of these 20-valued records
that will be recorded at a discrete time interval of 0.1 seconds. That means if we record
activity for one person for one minute we will have a total of 600 records. Realistically,
if we are recording an activity for a person, we cannot put an exact time limit on each
person. In our experiment, the recording time per activity ranged anywhere from 60 to
180 seconds.
Table 2.1 shows the summary statistics for all the five waves: alpha, beta, gamma, delta,
and theta for each of the four locations for all the participants. It can be seen that the values for each type of wave have similar ranges in all locations. However, the frequency
range for different waves varies considerably. For example, the absolute frequency range
of alpha values is from 7.5 Hz to 13 Hz, while beta values range from 13 Hz to 30 Hz [27].
In our dataset, we used Muse’s relative band powers, which normalize the absolute band
powers as a percentage of the total absolute band powers. This resulted in alpha values
ranging from 0.00 to 0.98, while beta values range from 0.00 to 0.95, as shown in Table 2.1.
Using the relative band powers decreased the variability of the value ranges; however,
our collected data still had slight variations in the range for each wave. Therefore, we
normalized the values for each wave using 90% of the maximum value for that wave, as
shown in Equation 2.10. That made sure that the values for all the waves were in the same
range. For instance, the value of wave maximum would be 0:98 for an alpha value, denoted by
value relative , determined using relative band powers.
value
value
wave
normalized
relative
0.90 *
maximum
=
(2.10)
Another issue with the data collection was the variable length of time for different
activities as well as the length of each record. The length of the record could vary anywhere from 60 to 180 seconds. Figure 2.2 shows an example of one of the waves, alpha,
from position 3 while player 1 was playing a game for 60 seconds. In order to fix the
length of the record to a fixed and more manageable value, we studied the frequency distribution of the records. After experimenting with different number of bins, we decided
Identification from Wearable Device Brain Signals
3. Watching video: A user watched a video for 1–3 minutes.
4. Game playing: A user played a computer game for 1–3 minutes.
5. Listening to music: A user listened to music for 1–3 minutes.
The above activities were repeated 10 times for five individuals resulting in a total of 50
datasets.
2.6 Knowledge Representation
One of the key aspects of any data mining activity is representing real-world entities
using the pertinent numeric data available for them. In our case, we want to capture
individuals and their activities using the signals emanating from their brain. This
chapter focuses on summarizing, and not manipulating, the signals collected from the
wearable into a fixed-length representation. As mentioned before, the Muse headband
collects data from four positions around the head. Each position provides five types
of waves: alpha, beta, gamma, delta, and theta. That means at any given point in time,
we receive a record with 20 values. There will be a stream of these 20-valued records
that will be recorded at a discrete time interval of 0.1 seconds. That means if we record
activity for one person for one minute we will have a total of 600 records. Realistically,
if we are recording an activity for a person, we cannot put an exact time limit on each
person. In our experiment, the recording time per activity ranged anywhere from 60 to
180 seconds.
Table 2.1 shows the summary statistics for all the five waves: alpha, beta, gamma, delta,
and theta for each of the four locations for all the participants. It can be seen that the values for each type of wave have similar ranges in all locations. However, the frequency
range for different waves varies considerably. For example, the absolute frequency range
of alpha values is from 7.5 Hz to 13 Hz, while beta values range from 13 Hz to 30 Hz [27].
In our dataset, we used Muse’s relative band powers, which normalize the absolute band
powers as a percentage of the total absolute band powers. This resulted in alpha values
ranging from 0.00 to 0.98, while beta values range from 0.00 to 0.95, as shown in Table 2.1.
Using the relative band powers decreased the variability of the value ranges; however,
our collected data still had slight variations in the range for each wave. Therefore, we
normalized the values for each wave using 90% of the maximum value for that wave, as
shown in Equation 2.10. That made sure that the values for all the waves were in the same
range. For instance, the value of wave maximum would be 0:98 for an alpha value, denoted by
value relative , determined using relative band powers.
value
value
wave
normalized
relative
0.90 *
maximum
=
(2.10)
Another issue with the data collection was the variable length of time for different
activities as well as the length of each record. The length of the record could vary anywhere from 60 to 180 seconds. Figure 2.2 shows an example of one of the waves, alpha,
from position 3 while player 1 was playing a game for 60 seconds. In order to fix the
length of the record to a fixed and more manageable value, we studied the frequency distribution of the records. After experimenting with different number of bins, we decided