|
|
|
|
|
|
EEG R&D » High Frequency EEG » 5.5
 

5.5  High frequency signal or noise?

In order to determine whether the high frequency EEG activity had any correlation with extreme drowsiness, a preliminary analysis was conducted on occipital channel EEG data (unless otherwise noted).  The analysis included 45 data segments for each of 3 typical (representative) test subjects.  The 45 data segments consisted of 25 hit and 20 missed events for each subject.  The 25 hit event data segments were arbitrarily selected from the set of data (based on) the following criteria:

1.  Successful response to visual stimulus
2.  No muscle/movement artifact detected in the record
3.  Data only taken from strings of at least 8 consecutive hits
4.  No data used from hit/miss or miss/hit transition boundaries

Similarly, data corresponding to missed events was arbitrarily chosen from the set of data meeting the following criteria:

1.  Response failure (to signal) not attributable to external causes
2.  No muscle/movement artifact detected in the record
3.  Data taken from strings of at least 8 consecutive misses
4.  No data used from hit/miss or miss/hit transition boundaries
5.  Initial missed events in strings of misses were not selected

Note:  The results discussed for the three tests subjects chosen for analysis in this section were indicative of the overall test population unless specifically noted otherwise.  The test subjects were chosen to demonstrate both the variability and consistence of various features which were relevant to this study.

Data from each of the three subjects was analyzed and examined individually.  Analysis was as follows:  A (windowed) power spectrum (PSD) was computed for each of the 45 data segments.  The PSD's corresponding to the 25 hit events were averaged and the PSD's for the 20 missed events were also averaged, resulting in two spectral plots.  These averaged spectrum provided a convenient mechanism to examine the changes in spectral components related to "typical" hit and miss events for each subject.

Note:  Because, at various scales in data segments, the 60Hz and subsequent harmonic components (especially the odd harmonics) tend to dominate the local scale of the power spectra, steps were taken to eliminate this activity from the PSD plots.  To eliminate the 60Hz components (and harmonics), Table 5.5.1 outlines the steps used.  Step 1, the center frequency to be eliminated (for example 60Hz) is determined.  Step 2, the location in the power spectrum of the undesirable frequencies and a small span of frequencies on either side are determined.  Step 3, components are replaced in the spectrum by the value of the nearest lower frequency component.

Step 1: freq={60,180,300,420Hz}
Step 2: loc={freq-2·res,freq-1·res,freq,freq+1·res,freq+2·res}
Step 3: PSD(loc)=PSD(freq-3·res)
with res = spectral resolution as computed in Section 5.2.

Table 5.5.1  Procedure to Eliminate 60Hz and Harmonic
Components from Power Spectrum


The first test subject to be examined was cia01.  The averaged spectra for hit and missed events is shown in Figure 5.5.1 for the classical frequency range of 0-40Hz.  In Figure 5.5.1(a), we see the average spectra of the selected 25 hit events.  Similarly, Figure 5.5.1(b) shows the average of the missed events (on the same scale as the hits).  The energy in the delta band remained relatively constant (decreases by a factor of 1.07), theta band energy increased by a factor of 2.57 from the hits to the misses, alpha power decreased by a factor of 2.39, and beta energy remained essentially unchanged.  These shifts in energy indicate (typically) relaxed wakefulness during the hits and possible stage 1 sleep for the missed events according to the R&K (1968) criteria (i.e. alpha activity being replaced by theta activity).

n
Figure 5.5.1 Average power spectrum from 0 to 40Hz: (a) Average
of 25 hits for cia01, (b) Average of 20 missed events.



A breakdown of the energy bands is given in Table 5.5.2.

Frequency Band
Energy (Hits)
Energy (Misses)
delta
2.001e+02      46.545% full
1.875e+02      41.601% full
theta
4.496e+01      10.462% full
1.155e+02      25.619% full
alpha
9.870e+01      22.965% full
4.138e+01       9.180% full
beta
2.303e+01       5.358% full
2.241e+01       4.972% full
0 to 30Hz
4.083e+02      94.991% full
4.439e+02      98.481% full
0 to 40Hz
4.135e+02      96.197% full
4.469e+02      99.141% full
0 to 475Hz (full spec.)
4.298e+02     100.000% full
4.508e+02     100.000% full

Table 5.5.2  Energy distribution of Fig. 5.5.1 shown as actual energy
and as a percent of total energy.


Along with the classical findings, Table 5.5.2 begins to reveal the correlation between high frequency activity and drowsiness.  When the subject is responding to the visual signals (hits), the 0 to 40Hz frequency band accounts for 96.1967% of the total energy in the "alert" frequency spectrum.  But, when the subject was unable to respond to the visual signals due to excessive drowsiness, the 0 to 40Hz frequency band accounted for 99.1413% of the spectral energy.  In absolute terms, the spectral energy above 40Hz during the averaged hits was 16.3 and the energy above 40Hz during misses was 3.9, yielding an overall energy decrease by a factor of approximately 4.2.  This shows a change in the spectral power (and energy) distribution toward lower frequencies during failure.  This full-spectrum behavior is also consistent with the behavior of the 0 to 40Hz band in which shifts toward lower frequencies during the initial stages of sleep.

Figure 5.5.2 shows the remainder of the power spectrum (31 to 475Hz) with the hit and miss spectra again shown on the same scale.  Although there are no dominant peaks in this frequency range comparable to those seen in the 0-40Hz range (e.g. alpha peaks, etc.), we can see that the overall magnitude of the power in the high frequency band does decrease with drowsiness.  When the full spectrum is viewed on a single plot, the signals in the 0-40Hz band are much stronger than all the others and consequently, the high frequency activity is virtually invisible.  By viewing the various frequency bands on their own individual scales, we can see this otherwise obscured activity quite clearly.  Also, by comparing the hit and miss spectra, those changes in the high frequency components which might otherwise go unnoticed are detectable.  This situation is further improved because data for the hit/miss spectra are derived from data from a single test subject and useful trends are not obscured by possible person-to-person and/or day-to-day variations in activity.

n
Figure 5.5.2 Average power spectrum from 31 to 475Hz: (a) Average
of 25 hits for cia01, (b) Average of 20 missed events.


The energy distribution of Figure 5.5.2 is shown numerically in Table 5.5.3.  Among many findings, we can observe that the energy in the 31 to 100Hz band for this subject drops by a factor of 2.59 between the hit and miss power spectra, but the energy in the 100 to 475Hz band drops by a factor of 10.28 from hits to misses.  In fact, it appears that the frequency bands beyond 100Hz change by at least a full order of magnitude between good performance and failure!  Changes in energy of this magnitude are not generally observed in the classical frequency range.

Frequency Band
Energy (Hits)
Energy (Misses)
31 to 100Hz
1.547e+01      73.521% high
5.965e+00      91.672% high
100 to 200Hz
4.239e+00      20.146% high
4.701e-01       7.224% high
100 to 300Hz
5.249e+00      24.946% high
5.225e-01       8.029% high
100 to 475Hz
5.571e+00      26.479% high
5.419e-01       8.328% high
200 to 475Hz
1.333e+00       6.333% high
7.180e-02       1.103% high
300 to 475Hz
3.225e-01       1.533% high
1.941e-02       0.298% high
31 to 475Hz (high spec.)
2.104e+01     100.000% high
6.507e+00     100.000% high

Table 5.5.3  Energy distribution of Fig. 5.5.2 shown as actual energy
and as a percent of total high-frequency energy.


As Table 5.5.3 clearly indicates, not only does the total magnitude of the high frequency energy decrease with poor performance, but the energy distribution within the bands shift toward the lower frequencies (similar to the 0-40Hz band).  These are two identifiable changes which can be observed directly from the power spectrum and energy distribution and appear to be completely unreported in the sleep, performance, and EEG related literature.

Figure 5.5.3 shows a closer look at the 100-475Hz band where the ordinate has to be rescaled.  Because of the extreme difference in scale between the hit and miss spectra, the scale for the missed events is 10-times smaller than for the hits.


Figure 5.5.3 Average power spectrum from 100 to 475Hz: (a) Average
of 25 hits for cia01, (b) Average of 20 missed events.


A small localized peak at approximately 263.82Hz appears on the spectra for the missed events in Figure 5.5.3(b).  A peak at this frequency appears on similar plots for each of the three test subjects shown.  This peak appears for the averaged power spectra corresponding to missed events (as done here) because the averaging improves the signal-to-noise ratio.

In order to investigate this localized peak, Figure 5.5.4 shows the power spectrum computed from the ambient noise data which was simultaneously collected along with the neurological data.  The noise data was collected using a 10k dummy load at the EEG electrode input terminal.  The dummy load will certainly have dissimilar receiver characteristics as compared to the applied EEG electrodes attached to the test subjects.  But, in spite of the different characteristics, the noise channel can still be used to check for frequency content of non-neurologic origin when a questionable peak appears on an EEG power spectrum.  The data from the noise channel (specifically in the 31 to 475Hz range) remains unchanged from hits to misses, providing further evidence that the shifts in energy from hits to misses are not due to external causes..  Regarding the 263.82Hz peak: because neither noise spectra revealed the peak seen in the miss spectra given above, it was suspected that the peak could be of neurological origin.


Figure 5.5.4 Average power spectrum from 100 to 475Hz: (a) Average
of 25 hits for cia01, (b) Average of 20 missed events.  (ambient electrical room noise)


To further investigate the origin of this peak, extremely sensitive data (collected prior to testing during the laboratory and equipment shakedowns) which combined high sampling rates over extended time periods, unbalanced amplifier inputs, and high amplifier gain were averaged to improve the S/N ratio.  This noise data revealed a peak above background levels at the same frequency location as the peak being investigated.  We concluded that the peak at 263.82Hz is of non-neurologic origin and appeared on the missed-event spectra due to the improved S/N ratio of averaging the PSD's combined with the diminished activity in that frequency range during misses.  This peak will be ignored for the remaining examples.  Furthermore, by increasing the input impedance on the noise channel and by possibly adding longer leads between the test load and electrode input terminal to increase noise susceptibility, we can improve the effectiveness of the noise channel for future tests.

Although the averaged power spectral plots provide insight into the relationship between behavior and the EEG data, it is also useful to examine the non-averaged data segments.  To further illustrate the frequency shifts associated with the hit and missed events, Figures 5.5.a-f show the energy in various frequency bands for the individual hit and missed events.  The first 25 events (black) in each plot correspond to the hit events and the last 20 events (red) correspond to the misses.  The events chosen here are the same events that were used to compute the averaged power plots.

Figures 5.5.a-d show the energy for each of the hit and missed events in the classical frequency bands (delta, theta, alpha, and beta).  The changes in magnitude corresponding to hits and misses have already been discussed using the averaged spectral plots in these energy bands.  Although the aggregate behavior in each of the frequency bands is consistent, the behavior of individual events reveals a significant overlap in the energy magnitudes between the hits and misses.  That is, individual hit and missed events cannot be distinguished reliably by simply looking at the magnitude of the energy in a particular band.  This has also been demonstrated across subjects in this study.


Figure 5.5.a Hit and Missed Events:  Energy in DELTA band


Figure 5.5.b Hit and Missed Events:  Energy in THETA band


Figure 5.5.c Hit and Missed Events:  Energy in ALPHA band


Figure 5.5.d Hit and Missed Events:  Energy in BETA band

But, if we examine Figure 5.5.e (and Figure 5.5.f), the change in energy from hits to misses is approximately 10.15:1 (and 18.54:1 for Fig. 5.5.f) which appears sufficient to distinguish between hit and missed events by high frequency energy measurement alone for this selection of events.  This ability to distinguish hit and missed events based on changes in high frequency energy is consistent and universal among all test subjects under the criteria set forth for event data inclusion given previously in this section.


Figure 5.5.e Hit and Missed Events:  Energy in 100-475Hz band


Figure 5.5.f Hit and Missed Events:  Energy in 200-475Hz band

To further illustrate the high frequency activity, let's examine the another test subject (kal01) using the same procedure of creating spectral averages from 25 hit and 20 missed events.  In Figure 5.5.5, the 0 to 40Hz spectral plots show the differences between hit and missed events for this new test subject in the traditional frequency range, and Table 5.5.4 gives the accompanying energy distribution values.  Similar to the previous subject, an increase in theta energy accompanied the missed events by a factor of 3.83 (a value of 2.57 for the previous test subject).  But, whereas the previous subject had little change in delta energy and a decrease in alpha energy by a factor of 2.39, this subject decreased in delta energy by a factor of 2.14 and had virtually no change in their alpha energy levels between the hits to misses (increased energy by a factor of 1.01).  On the high frequency side, the 0 to 40Hz band during hits accounts for 94.1855% of the energy in the spectrum, but the 0 to 40Hz band during missed signals accounts for 99.5108% of the total energy.  This again shows a consistent change in the high frequency energy distribution which accompanies the behavioral failures.


Figure 5.5.5 Average power spectrum from 0 to 40Hz: (a) Average
of 25 hits for kal01, (b) Average of 20 missed events.

Frequency Band
Energy (Hits)
Energy (Misses)
delta
2.814e+02      64.801% full
1.317e+02      33.743% full
theta
3.199e+01       7.367% full
1.224e+02      31.360% full
alpha
3.717e+01       8.558% full
3.750e+01       9.608% full
beta
2.537e+01       5.841% full
2.466e+01       6.319% full
0 to 30Hz
4.025e+02      92.687% full
3.870e+02      99.166% full
0 to 40Hz
4.091e+02      94.185% full
3.884e+02      99.511% full
0 to 475Hz (full spec.)
4.343e+02     100.000% full
3.903e+02     100.000% full

Table 5.5.4  Energy distribution of Fig. 5.5.5 shown as actual energy
and as a percent of total energy.

Figure 5.5.6 and accompanying Table 5.5.5 shows the remaining 31 to 475Hz frequency range.  As observed in the previous subject, order of magnitude changes in the various high frequency bands decrease consistently across the high frequencies throughout the spectrum during failures.


Figure 5.5.6 Average power spectrum from 31 to 475Hz: (a) Average
of 25 hits for kal01, (b) Average of 20 missed events.

Frequency Band
Energy (Hits)
Energy (Misses)
31 to 100Hz
2.243e+01      72.904% high
2.652e+00      85.950% high
100 to 200Hz
6.760e+00      21.972% high
3.479e-01      11.275% high
100 to 300Hz
7.985e+00      25.955% high
4.085e-01      13.238% high
100 to 475Hz
8.337e+00      27.096% high
4.335e-01      14.050% high
200 to 475Hz
1.576e+00       5.124% high
8.560e-02       2.774% high
300 to 475Hz
3.510e-01       1.141% high
2.503e-02       0.811% high
31 to 475Hz (high spec.)
3.077e+01     100.000% high
3.086e+00     100.000% high

Table 5.5.5  Energy distribution of Fig. 5.5.6 shown as actual energy
and as a percent of total high-frequency energy.

Figures 5.5.g-j show the energy in the traditional frequency bands for the individual hit and missed events.  The delta and theta energy bands appear to be the best indicators of the failure, but again, individual hits and missed events could not be distinguished in all instances by simply looking at the magnitude of the energy in a particular band.


Figure 5.5.g Hit and Missed Events:  Energy in DELTA band

 


Figure 5.5.h Hit and Missed Events:  Energy in THETA band


Figure 5.5.i Hit and Missed Events:  Energy in ALPHA band


Figure 5.5.j Hit and Missed Events:  Energy in BETA band

Figures 5.5.k-l correspond to the high frequency energy bands.  These examples demonstrate that hits and misses can be reliably distinguished simply by examining changes in the high frequency energy content.


Figure 5.5.k Hit and Missed Events:  Energy in 100-475Hz band

 


Figure 5.5.l Hit and Missed Events:  Energy in 200-475Hz band

Finally, the third test subject in the series (cvj01) is shown in Figures 5.5.7 and 5.5.8 with accompanying Tables 5.5.6 and 5.5.7, respectively.  On the low frequency side, delta energy increased by a factor of 5.29, theta energy increased with missed signals by a factor of 9.46, but, this time alpha energy increased during misses by a factor of 1.70.  On the high frequency side, we see that the 0 to 40Hz band accounted for 93.2165% of the total energy during hits and increases to 99.7772% during the missed signals.


Figure 5.5.7 Average power spectrum from 0 to 40Hz: (a) Average
of 25 hits for cvj01, (b) Average of 20 missed events.

Frequency Band
Energy (Hits)
Energy (Misses)
delta
6.464e+01      43.605% full
3.419e+02      54.959% full
theta
1.528e+01      10.310% full
1.446e+02      23.243% full
alpha
2.172e+01      14.655% full
3.695e+01       5.940% full
beta
1.667e+01      11.248% full
2.243e+01       3.605% full
0 to 30Hz
1.350e+02      91.042% full
6.196e+02      99.604% full
0 to 40Hz
1.382e+02      93.217% full
6.207e+02      99.777% full
0 to 475Hz (full spec.)
1.482e+02     100.000% full
6.221e+02     100.000% full

Table 5.5.6  Energy distribution of Fig. 5.5.7 shown as actual energy
and as a percent of total energy.


Figure 5.5.8 Average power spectrum from 31 to 475Hz: (a) Average
of 25 hits for cvj01, (b) Average of 20 missed events.

Frequency Band
Energy (Hits)
Energy (Misses)
31 to 100Hz
1.001e+01      77.103% high
2.118e+00      91.408% high
100 to 200Hz
2.484e+00      19.123% high
1.621e-01       6.997% high
100 to 300Hz
2.876e+00      22.141% high
1.884e-01       8.129% high
100 to 475Hz
2.974e+00      22.897% high
1.991e-01       8.592% high
200 to 475Hz
4.901e-01       3.774% high
3.696e-02       1.595% high
300 to 475Hz
9.819e-02       0.756% high
1.074e-02       0.463% high
31 to 475Hz (high spec.)
1.299e+01     100.000% high
2.317e+00     100.000% high

Table 5.5.7  Energy distribution of Fig. 5.5.8 shown as actual energy
and as a percent of total high-frequency energy.

Figures 5.5m-r show the energy in some of the various frequency bands (traditional and high frequency) for the individual hit and missed events used in the averaged spectra.


Figure 5.5.m Hit and Missed Events:  Energy in DELTA band

 


Figure 5.5.n Hit and Missed Events:  Energy in THETA band

 


Figure 5.5.o Hit and Missed Events:  Energy in ALPHA band

 


Figure 5.5.p Hit and Missed Events:  Energy in BETA band

 


Figure 5.5.q Hit and Missed Events:  Energy in 100-475Hz band

 


Figure 5.5.r Hit and Missed Events:  Energy in 200-475Hz band



Concluding remarks for this section

Table 5.5.8 summarizes the changes in energy in specified bands for each of the three example subjects discussed.  The changes are listed in the form of multipliers from hits to misses.  For example, a value of -9.02 is interpreted as a decrease in energy by a factor of 9.02 from hits to misses.  Similarly, a value of +9.02 would indicate an increase in energy from hits to misses.

ID
delta
theta
alpha
beta
0-40
0-475
31-100
100-475
200-475
300-475
31-75
cia
-1.07
+2.57
-2.39
-1.03
+1.08
+1.05
-2.59
-10.28
-18.56
-16.61
-3.23
kal
-2.14
+3.83
+1.01
-1.03
-1.05
-1.11
-8.46
-19.23
-18.42
-14.03
-9.97
cvj
+5.29
+9.46
+1.70
+1.35
+4.49
+4.20
-4.73
-14.94
-13.26
-9.15
-5.61

Table 5.5.8  Summary of the Changes in Energy Distribution (multipliers)


Because the criteria that were used for including data to be analyzed in this section were so strict, we expect that all subjects are in essentially the same physiological state during the hits and during the misses.  This should provide a stable means of comparing frequency changes across various subjects.  It is surprising that only the theta band proved consistent within the 0-40Hz frequency range.  This band consistently increased in energy from hits to missed events for these subjects.

Even though the theta energy behaved consistently for the three test subjects shown, the alpha energy during missed events was erratic with: no changes for one subject, decreased for another, and increased for yet a third.  Certainly any detection system which relies on alpha energy for drowsiness detection would be confounded by the potentially inconsistent behavior of this frequency band.  We have observed, in general, that alpha energy levels may correlate well with performance for some test subjects and not others, and that alpha levels may correlate well with performance during some portions of a test and not others.  Delta energy levels were also inconsistent with failures (missed events) because there were decreases for some subjects and increases for others.  Beta energy showed relatively little sensitivity during failures in the test subject's for this study.  The inconsistencies discovered in the various traditional frequency bands in this work may also help explain the possible inconsistencies reported in scoring stage W to stage 1 sleep.  Of course, when scoring sleep segments, the data to be scored is in 15 or 30 second epochs and not 2 second clips.  This may certainly account for some of the variations which could be averaged out in the larger data segments.  Small segments are used in this work because the ultimate goal is the development of a detection device and small data segments are preferred to increase the responsiveness of a drowsiness measure observed from the data.  Response times on the order of a few seconds are desirable for such a system.

Looking at the high frequency bands, we can see that all of the high frequency bands shown consistent decreases during failures.  The largest changes occur in the higher frequencies as shown.  The high frequency band appears to be much more sensitive than the lower frequency bands with relative changes on the order of 10-times and greater from hits to misses.  Our analysis has clearly demonstrated that the high frequency range has a definite correlation with extreme drowsiness and has the potential for improved sensitivity over the traditional frequency bands.  For these reasons, the remaining research will focus on the high frequency activity.


As a quick recap:

1. Theta energy has consistently increased during the missed events.
2. Alpha, Beta, and Delta energy has behaved inconsistently between the hits and misses.
3. High frequency energy has consistently decreased during the missed events.

We now have the necessary justification to fully pursue high frequency analysis.



 
Copyright © 2010 Consolidated Research of Richmond, Inc. All rights reserved.