6.2  Tracking frequency bands

Low frequencies

As a first example, we will examine the traditional low frequency energy bands from the spectrogram image shown in Figure 6.2.1L (cia01).  For consistency, the term "spectrogram data" will be used to refer to the sequence of windowed Fourier transforms with the associated signal processing (filter compensation, noise removal, etc.), and the term "spectrogram image" will be used to refer to the image created from the spectrogram data for display (re-scaling, gridding, smoothing, etc.).  Note:  Spectrogram images are used strictly as a visual representation of the spectrogram data on which the analysis is being performed.

Figure 6.2.1L  Low frequency spectrogram

By computing the energy in each of the four classical frequency bands as defined in Table 2.2.1 (delta, theta, alpha, beta) , we are able to trace the evolution of each band over time for comparison against test performance.  Figure 6.2.2 shows the time history of the four marked energy bands of Figure 6.2.1L with event markers superimposed over the energy to show the corresponding test performance.

Figure 6.2.2  Low frequency band tracking

In Figure 6.2.2 we can see that there is an increase in the amount of beta and theta energy only during the middle portion of the failure, a small decrease in beta activity, and a sustained loss of alpha activity across the entire failure.  The segment of data during the actual failure contains some characteristics that may be consistent with a brief episode of S1 sleep in the middle portion of the failure.  For this example, the alpha activity correlates well with the failure, this activity appears mostly linked with the status of the eyes (as soon as the eyes close, the alpha activity disappears) as shown in Figure 6.2.3.  This figure shows the vertical EOG over the duration of the data segment with the event marks superimposed over the plot.  Since the data segment is 9.7 minutes in duration, the amplitude envelope is the only discernable feature of the plot, but this is sufficient to determine the presence or absence of eye activity.  As we have shown previously in chapter 5, alpha activity behaves differently for various people and is not consistent enough to be used as a detection measure.  Also, we can see from this plot that the alpha activity in not even completely consistent with the state of the eyes, the subject is able to respond to visual signals even during periods of low alpha activity.

Figure 6.2.3  Vertical EOG

From the standard frequency plots of Figure 6.2.2, it would be difficult to closely determine the time of failure onset or conclusion.  These plots reveal why the largest variance (disagreement) among sleep scorer's scoring the same portion of data occur during the stage W to stage 1 transition.  We will now proceed directly to analysis of the high frequency data.

High frequencies

No predefined high frequency bands have been defined, as far as we know, so the 60-475Hz range is divided into bands of approximately 41.5Hz each (the actual value depends on the spectral resolution).  Although different bandwidths have very similar behavior (as can be observed from the spectrograms), a 41.5Hz bandwidth appears to provide sufficient resolution to distinguish between the activities in various frequency ranges, does not produce excessive amounts of data, and divides the frequency range of interest into ten parts of equal bandwidth to allow for direct comparisons.  Prior analysis of the data has shown that there are no bands of energy in the range 60-475Hz that behave inconsistently with the overall band trends, the only differences between frequency bands appears to be the magnitude (and possibly the time) of the response.  Therefore, we have freedom to discretize the high frequency range as long as the bands are small enough to capture the individual performance characteristics of various frequencies, but not too small so as to produce inconsistent results.

Using the data shown in Figure 6.2.1H, we are now able to track each of the 10 energy bins (the energy contained in the frequency bands) comprising of the 60-475Hz range.  Figure 6.2.4 tracks the evolution of the energy in each of the 41.5Hz frequency bins over time.  Once again, the blue event markers are superimposed over the graph to show the behavioral information.  As always, the event marks and the scaling of the abscissa (time bins) are directly comparable with those shown on the associated spectrogram images.

Figure 6.2.1H  High frequency spectrogram

Little or no useful information can be obtained directly from of the graphs in Figure 6.2.4 due to the large differences in amplitude between the small amplitude rhythmic energy which we are tracking and the large energy associated with the spikes which dominate the scale.  This plot was included to illustrate (quantitatively) the actual band energy and the large magnitude fluctuations which almost completely obscure the data we are interested in observing.  This figure also helps highlight the effectiveness of the various image processing techniques which captured all of these interesting features on a single image.  Figure 6.2.5 is a rescaled version of Figure 6.2.4 and provides a much clearer view of the energy changes that we are interested in.

Figure 6.2.4  (unscaled) High frequency band tracking

---

Figure 6.2.5  (scaled) High frequency band tracking

All bands show a very distinct downward trend as failure approaches (incipient behavior) and a continuous drop in energy content during the course of the entire failure, which is consistent with the previous analysis.  Some bands show more natural fluctuations that others (which may be a visual function of scale), but all react very similarly by losing energy during the failure.

It is also interesting to note the difference in the energy content between the low and high frequency bands.  In the low frequency bands, the energy appears in bursts of activity (sawtooth) that fluctuate down to zero, whereas the high frequency energy has a much smoother appearance and even though there are natural fluctuations, the energy reveals much smoother transitions.

As further illustration of the high frequency activity shown against the classical frequency bands, data from the remaining four spectrograms first shown in chapter 5 will be examined further.

Figure 6.2.6L shows the low frequency spectrogram with Figure 6.2.7 detailing the temporal evolution of the classical frequency bands.  There is an increase in both delta and theta band activity during the course of the failure episode and a slight decrease in both alpha and beta activity as the failure progresses.  This activity persists through the remainder of the failure, but does not manifest until approximately one-third of the way into the failure.  The decrease in alpha activity is noticeable when shown against the event markers, but the alpha activity has many periods where the activity is at the same minimal level as during the failure.  Alpha is certainly not a good indicator in this instance.  Bands such as delta and theta which do show changes in amplitude would benefit from an integrator (accumulator) which would smooth the instantaneous changes in energy levels and provide a more stable measure.  Because of the way the energy increases in bursts, rapid tracking of behavioral state would be difficult using these frequencies.

Figure 6.2.6L  Low frequency spectrogram

Figure 6.2.6H  High frequency spectrogram

Figure 6.2.7  Low frequency band tracking

 

If we now look at the high frequency activity of Figure 6.2.6H and Figure 6.2.8, we can observe the evolution of the high frequency energy bins over time.  Again, we can observe the same characteristics as observed previously in Figure 6.2.5.  The high frequency bands all show the same shifts during the failure.  The high frequency bands appear to decrease enough in energy as to be distinguished from the previous levels at approximately the same time as the delta and theta changes became noticeable.  However, the standard bands showed no distinct trends revealing the failure prior to manifestation and did not appear sensitive to the changes which occurred during the course of the failure.  The high frequency bands, on the other hand, again showed a distinct downward trend prior to the failure and through the course of the failure.  This demonstrates the potential of using the high frequency information to develop a tracking algorithm (measure) which will have improved sensitivity over the traditional bands.

Figure 6.2.8  High frequency band tracking

In Figure 6.2.9L and Figure 6.2.10 there is good correlation between an increase in delta and theta activity when a failure occurs.  The alpha activity shows no consistent correlation with the failure and the beta activity shows virtually no change.

Figure 6.2.9L  Low frequency spectrogram

 

Figure 6.2.9H  High frequency spectrogram

 

Figure 6.2.10  Low frequency band tracking

But again in Figures 6.2.9H and 6.2.11, the same downward shift in energy can be observed prior-to and during the course of the failure.  The high frequency begins to show a downward trend prior to the failure episode and immediately returns to "normal" levels after the failure has passed.  The downward trend then begins, correlated with further failure episodes and continuing to the end of the segment.

Figure 6.2.11  High frequency band tracking

Finally, Figures 6.2.12L and 6.2.13 show the last example in this section.  For this example, the delta, theta, and beta frequency bands show little useful correlation with the large failure episode.  The alpha activity shows a very distinct drop in energy across the failure period, but the energy does not re-appear until after the subject had already recovered from the failure and responded to a couple of signals.  The large (and small) failure(s) in this segment may have gone undetected using a standard low frequency measure.

Figure 6.2.12L  Low frequency spectrogram

Figure 6.2.12H  High frequency spectrogram

Figure 6.2.13  Low frequency band tracking

Using the high frequency information, we can see a distinct downward trend in Figure 6.2.13 (corresponding to data from Figure 6.2.12H) leading into the failure in the middle of the segment and a smaller trend which did not drop as much energy as in the larger failure corresponding to the four missed signals.

Figure 6.2.14  High frequency band tracking

In summary, from the analysis of the low frequency (traditional) bands, the delta and theta bands show the most consistent correlation with test failure.  These bands show increased energy as their correlation to failures.  The activity in the alpha and beta bands appears to correlate with failures with a decrease in overall energy when correlations are observed.  These findings are consistent with many researchers, including Beatty and Greenberg (1974), who find that occipital theta frequency band activity appears as a reliable correlate of vigilant behavior; and Belyavin and Wright (1987) who also correlated increased delta and theta band activity with worsening performance.  The high frequency bands are the most consistent in that they correlate with drowsiness with a decrease in energy across all of the frequency bands during failures.  Furthermore, tracking the energy in the traditional (low frequency) bands appears to give information indicating the presence or absence of failure conditions at best (if correlations exist), energy in the high frequency bands shows a remarkable sensitivity to detecting the onset as well as the presence or absence of a failure event or condition, and is also sensitive to the "extent" of the failure (behavioral state) through trends in the magnitude of the energy.

<PREVIOUS

NEXT>