5.4  Preliminary Data Analysis

Choice of data segments

The goal of this analysis phase was to identify changes in the EEG signal which correlate directly with extreme behavioral sleepiness.  Specifically, this behavior will be manifect as either the successful response to a visual stimulus (awake/alert) or a failure to respond (extreme drowsiness).

Data was carefully selected during this phase to include only those events which were classified as fast-hits and those classified as missed signals (as described in section 4.2.1) across various test subjects.  In addition, data corresponding to missed signals was only taken from the middle and end of strings of missed signals.  This should help ensure that the test subject is in the midst of a true failure which should make the differences in the two datasets more noticeable.  Any data with a questionable behavioral origin was not included for selection.  Also, each data segment was processed to detect the presence of muscle or movement artifact.  Any segment of data which showed artifact-contamination was excluded from further analysis at this stage.  Methods used for artifact detection will be discussed later.

Using the above selection criteria, a representative set of data that corresponds to 25 hits and 20 missed events for each of three different test subjects (1 Male/2 Female) was chosen.

Data windowing

The data segments selected above which correlate to visual events were extracted by truncation (rectangular window) from the main dataset by the extraction software.  Because rectangular windowing is generally undesirable when performing spectral analysis (high side-lobe leakage), each segment of data was multiplied by a standard Hanning windowing function which indices smaller side-lobes than the rectangular function (at the expense of a broader main-lobe).  If analysis warrants, more complex windowing functions can be used (such as the Dolph-Chebyshev windowing function) with the potential for improved characteristics over the Hanning function at the expense of slightly increased computational complexity.

Channel selection

Analysis in this report will use the occipital EEG channel, since this channel is most often used clinically for scoring the wake to stage 1 sleep transition (typically known as the sleep onset period - although some researchers feel that the sleep onset period may be better defined as the appearance of the first sleep spindle in stage 2 sleep [Merica, 1992; Ogilvie, 1989; Ogilvie, 1984; Bonnett, 1982]).

Differences introduced from our data collection methods

In order to understand how the data collection techniques used in this experiment differ from the traditional data collection techniques, a single EEG data segment has been chosen for closer analysis.  The data segment chosen is taken from the occipital EEG channel and corresponds to the successful response (hit) to a visual stimulus with a reaction time of 0.6147s (within the fastest 10% of this subject's responses).  Please note, this data segment is considered typical of the types of data segments that have been collected and all of the specific discussion on this segment can and will be extended to the larger dataset.

Figure 5.4.1 shows how this example EEG channel appears in the time domain using traditional EEG data collection techniques (i.e. sample rate of 256Hz per channel and an anti-aliasing lowpass filter at 30Hz).

Figure 5.4.1  Typical EEG data segment

The frequency responses of the anti-aliasing filters on most modern EEG machines are typically of low order (i.e. 1st or 2nd-order Butterworth).  The frequency response of the 30Hz lowpass filter used in this study (Grass Instruments Model 12C Neurodata Acquisition System) has been experimentally derived and is shown in Figure 5.4.2.  The filter breaks at 21.5Hz with a shallow rolloff.  Grass Instruments names their lowpass filters for their 50% amplitude points.  For example, the 30Hz lowpass filter reaches its 50% output amplitude point at approximately 30Hz.  The response shown in Figure 5.4.2 is typical of the filters used routinely in EEG data collection.  Obviously this filter would be sufficient to keep pen recorders from chattering, but may not be ideal for certain applications from a signal processing perspective because of the shallow rolloff.  Of the available filters on the Grass amplifiers, the 30Hz lowpass filter is most appropriate for the typical 256Hz sampling rate because the next higher filter (100Hz for our equipment) does not attenuate sufficiently at the Nyquist rate to be effective.

Figure 5.4.2  Grass Instruments 30Hz lowpass filter

Figure 5.4.3 shows the resulting power spectrum of the signal shown in Figure 5.4.1 which was filtered according to Figure 5.4.2.  It is interesting to note the prominent peak of energy in the alpha band (8-12Hz) shown in the power spectrum in this particular example.  The total energy within the 0-30Hz frequency range accounts for 99.479% of the total energy in the spectrum and the alpha band accounts for 50.459% of that total.

Figure 5.4.3  PSD of data segment shown

This particular example is interesting if we recall (briefly) the literature regarding the relationship between spectral parameters and performance.  The literature states that alpha intrusion while the eyes are open is usually an indication of extreme drowsiness and often associated with poor task performance.  In this case, the alpha activity was present, but performance was not only acceptable, but corresponds to one of the faster reaction times this subject achieved.  This test subject did have their eyes open during this segment as verified by the vertical EOG and the fact that the subject responded to the visual stimulus.  This type of activity can be seen in other EEG segments (shown later) and may help explain why many proposed drowsiness detection systems relying heavily on this type of activity have proven unsuccessful.

In contrast to the typical EEG data collection methods, the data collection used throughout this study used a sampling rate of 950Hz per channel with a 100Hz lowpass filter (Grass Instruments).  Figure 5.4.4 shows the same EEG data segment in the time domain under these new conditions.  The frequency response of the 100Hz lowpass filter used for anti-aliasing was derived experimentally and is shown in Figure 5.4.5.

Figure 5.4.4  EEG data segment with our collection techniques

Figure 5.4.5  Grass Instruments 100Hz lowpass filter

Side-by-side comparison of the two time-domain signals shows that additional high frequency components have been captured using the higher bandwidth data collection techniques than were captured using the standard procedures.  The performance characteristics of the 30Hz and 100Hz lowpass filters are qualitatively similar, with the main difference being the location of the cutoff.

Figure 5.4.6 shows the resulting power spectrum associated with the higher bandwidth signal shown in Figure 5.4.4.  The spectrum now has frequency content from the EEG signal up to 475Hz, but the majority of the spectral energy is still contained in the 0-30Hz band as would be expected from the EEG literature.  Using the higher filter cutoff and faster sampling rate, the 0-30Hz frequency band now accounts for 95.156% of the total spectral energy.  This leaves 4.844% of the spectral energy outside of the 0-30Hz band, whereas before, only 0.521% was outside of the 0-30Hz band.  If we also consider the energy in the 60, 180, 300, and 420Hz frequency bands which could be introduced into the signal from the electrical system, this can account for an additional 0.204% of the total energy, resulting in 95.360% of the total energy within the 0-30Hz band, still leaving 4.640% unaccounted for.

Figure 5.4.6  PSD of higher frequency data segment

To reveal more detail of the low frequency portion of the spectrum, Figure 5.4.7 displays only the 0-128Hz portion of the spectrum which can then be compared directly with the spectrum shown in Figure 5.4.3.

Figure 5.4.7  PSD displayed from 0-128Hz

A visual comparison shows that both power spectrum plots have the same general shape and power distribution within the 0-30Hz band as we would expect.  Beyond 30Hz, the high frequency power spectrum of Figure 5.4.7 reveals a very small amount of activity through 100Hz and a small 60Hz peak.

To improve viewing of the high frequency activity (i.e. above 30Hz), Figure 5.4.8 shows the power spectrum without the 0-30Hz frequency band.  This allows the higher frequency components to be shown on their own scale without being obscured by the much higher power low-frequency components.

Figure 5.4.8  Power Spectrum 31-475Hz

To gain a perspective of the change in scale, the 60Hz peak in Figure 5.4.7 which was barely visible when displayed on the 0-30Hz scale is now the largest amplitude component and fills the full height of the new spectrum (1/100th original scale).  We can now plainly see that there are high frequency components beyond 300Hz which were previously obscured.  If we further rescale the power spectrum by limiting the display range to the 100-475Hz frequency band as in Figure 5.4.9, we can see that there are broadband high frequency components of various magnitudes throughout the frequency spectrum.

Figure 5.4.9  Power Spectrum 100-475Hz

At this point, we must determine if these high frequency components are merely broadband high frequency noise introduced by the increased bandwidth of the data collection or "useful" neurologic signals.

Justification for high frequency analysis

The rhythms present in the EEG signal from approximately 0-30Hz have been studied and examined extensively for decades and are well understood and researched, and frequencies in the 40Hz range have also been examined, but, to a lesser degree [Loring and Sheer, 1984].  There is also a tremendous body of literature correlating this traditional frequency band activity for staging of human sleep.  Unfortunately, researchers using characteristics of this same frequency band for the incipient detection of drowsiness have not been extremely successful.

Discourage search for high-freq. components

To date, we are not aware of any research in the literature whose goal was the exploration for the presence of sustained high frequency rhythms.  In fact, the language used throughout the current literature tends to discourage exploration for "useful" high frequency activity.  For example, Gaillard (1987  pp.9-11) refers to frequencies above the beta band as "high frequency noise"; Pritchard (1995  pp.378) refers to the higher frequency activity as white or near-white noise; Carskadon and Rechtschaffen (1987  pp.668) state that a high frequency filter setting in the range of 30 to 35Hz will generally pass through the essential wave forms, while minimizing high frequency interference; and O'Hanlon and Beatty (1977  pp.195) also refer to "noise" greater than 30Hz.

Other sources of high frequency

There are some physiological situations where higher frequency components (above 30Hz) are present in the EEG signal, but these should not be confused with the continuous appearance of the high-frequency energy which we are attempting to classify.  Common sources of high frequency activity are most often associated with burst activity such as neurological spike activity (such as that observed during an Epileptic seizure) and muscle/movement artifact which is an extremely common source of high-frequency contamination found in nearly every EEG record.  These sources of high frequency activity come from highly localized time domain features having sharp transitions and therefore, have very broad representations in the frequency domain.

Therefore, researchers trying to detect spike activity in the EEG (common in Epileptic research) tend to increase the bandwidth of their anti-aliasing filters and use faster sampling rates during digitization in order to capture more of the frequency content of the spike for improved identification, analysis, and reconstruction.

A final common source of high-frequency components previously mentioned is electrical contamination.  The most common source of this electrical noise is from the 60Hz power line which intrudes in the EEG signal during collection.  For this reason, modern EEG equipment is equipped with a sharp 60Hz notch filter (50Hz in Europe) to help remove this contamination from the EEG.

Burst activity (spikes), muscle/movement artifact, and electrical contamination appear to be the only sources of high-frequency EEG activity that we have found that has been reported in the current medical literature.

<PREVIOUS

NEXT>