6.1  Tracking single frequencies

In chapter 5, several examples of drowsiness induced failures were examined.  An objective of this work is to develop algorithms that can detect such failures using the information discovered in chapter 5.  Specifically, the energy changes that occur in the high frequency bands (i.e. shifts toward lower frequencies) during failures.  The algorithm developed in this chapter is capable of tracking the temporal changes in the EEG signal which correlate with increased drowsiness and using this information to detect distinct behavioral states, i.e. discriminate one behavioral state from another.

The functional requirements of the drowsiness tracking algorithm are defined in terms of the set of performance specifications and criteria that follows:

1. Provide a continuous tracking of the behavioral state
2. Use only a single channel of EEG data
3. Unaffected by the status of the eyes
4. Tunable for different individuals
5. Algorithm can be implemented in real-time
6. Sensitive to extreme drowsiness and sleep with a low false alarm rate
7. Small time lag and fast update rate
8. Thresholding adjustable for the detection of a change in behavioral state and
9. Performance level for a variety of tasks

Visual inspection of the spectrograms given in chapter 5 clearly indicates the changes in energy content that occurs in the high frequency band as a result of extreme drowsiness and behavioral failures.  This information along with additional features highlighted by image processing techniques are used as the starting point for algorithm development.

The computational framework of the spectrogram is used to track the temporal evolution of single frequency points as the subject experiences the behavioral transitions which we are interested in detecting (from good performance, to task specific failures, and back again to good performance).  The analysis in this chapter will focus on those frequencies from 60Hz to 475Hz, knowing (from chapter 5 section 5) that we can expect the most significant changes in high frequency energy to occur beyond 100Hz.

Tracking the evolution of single frequency points has proven to be unprofitable, primarily for two reasons.  The spectrogram is computed using a discrete Fourier transform via the FFT, consequently the frequency range is discretized and each point on the frequency axis is influenced by neighboring frequencies.  This influence by neighboring frequencies is a function of the spectral resolution.  Therefore, single frequency analysis is dependent upon the specific spectral resolution of the transform being used and would not be consistent across various transforms involving different spectral resolutions.  Secondly, there appears to be a natural point-to-point variability both across time and frequency which makes the extremely narrow view of single frequencies variable and unreliable as a predictive measure.  For example, one particular frequency may behave exactly as we would expect during the transition into a failure, but an adjacent frequency point may not show the same sensitivity.  If we then examined these same frequency points on a different test subject, the frequency behavior is often not consistent.  Therefore, analysis involving single frequencies will not be explored further.

Figures 6.1.1L and 6.1.1H show both the low and high frequency spectrograms respectively without the smoothing function applied.  From these examples, the variability in the frequency activity discussed above is quite noticeable.

Figure 6.1.1L  (unsmoothed) Low frequency spectrogram

Figure 6.1.1H  (unsmoothed) High frequency spectrogram

A solution to the limitation of tracking single frequencies is to analyze the temporal evolution of energy in narrow frequency bands.  When we examine the evolution of energy bands as opposed to single frequencies, the output is more consistent and stable from person to person due to the averaging effect of integrating over multiple frequencies.  Tracking the behavior of a range of frequencies reduces the impact of small changes in spectral resolution on output behavior.

<PREVIOUS

NEXT>