7.3  Specific contributions of this work

The results of this research represent a new and highly innovative approach to drowsiness detection.  The main objective of this research was to correlate the signal characteristics from a single channel of EEG data with extreme sleepiness.

This research has resulted in the discovery of an entirely new range of frequencies in the EEG signal that correlate with states of consciousness from alertness through extreme sleepiness and various stages of sleep.  These new signals have a much high frequency than the traditional EEG bands and these new frequencies were previously considered broadband noise and as such, were typically filtered out of the EEG signal.  In fact, laboratory tests and data analysis conducted in this work has established for the first time that the high frequency range of the EEG signal contains useful information for the drowsiness tracking application.  In the course of this research, we have be able to explore some of the characteristics of these new frequencies and compare them to the behavior of the standard frequency bands before moving on to the construction of a tracking algorithm.

In addition to discovering an entirely new range of useful frequencies in the EEG signal, a method has been given which, through effective signal analysis and processing, can allow these frequencies to be used directly in a drowsiness tracking and detection system.  In fact, the algorithm outlined in this document is constructed exclusively from those frequencies that are routinely eliminated from typical EEG records.

In addition to developing a drowsiness tracking algorithm and analyzing its behavior under various conditions, a set of thresholds has been determined to detect behavioral failures on the visual test.  Because the thresholds appear to work over a number of different test subjects, the algorithm as constructed may be invariant to the differences between individuals.  Certainly, the threshold values determined in this work are task specific.

Also, because the drowsiness tracking algorithm relies on tracking changes in the activity of these naturally occurring frequencies, in contrast to complicated algorithms such as waveform analysis which are used to process EEG signals, the algorithm developed in this work can use standard signal processing methods to compute the drowsiness measure and is suitable for real-time implementation without substantial modifications.

Certainly the contributions from this research will span many disciplines.  First, the discovery of a "new" frequency range will advance the general field of EEG research by giving researchers access to more of the information content of the EEG signal.  Next, since no system is currently available which can effectively use the EEG signal for continuous drowsiness tracking and detection, this work represents a significant contribution to this general technology area.  Even a recent report by Wierwille, et al. (1994) to the US DOT surveying methods of drowsiness detection states:

"Planque et al. state that automatic processing of the EEG signal has proved very difficult to implement.  Presently, various phases of sleep (stage 1, stage 2, REM, etc.) are identifiable via automated methods, however an examination of drowsiness and sleep onset is distinguishable by much less distinctive physiological events.  Therefore, Planque et al. suggest the manual method for analysis of EEG as well as EOG which was discussed previously."

Also, those signal processing techniques which can successfully differentiate between alertness and the initial stages of sleep would greatly improve the state-of-the-art in EEG analysis for medical and diagnostic purposes.  Improved ability to track and differentiate between waking and sleep states will help advance the field of sleep analysis and potentially improve sleep scoring techniques.  The analysis of early sleep stages and the ability to recognize even brief periods of arousal from sleep can be critical in helping to identify many sleep disorders, including a very common complaint, the diagnosis of clinical insomnia.  And finally, the potential spin-offs of this technology cannot be underestimated once other researchers in various EEG and sleep related fields build upon this initial work.

In clinical sleep scoring, it is typically necessary to use two channels of EEG (occipital for sleep onset and central for the remainder), chin EMG, and horizontal EOG to score the basic stages of sleep.  Scoring a sleep session requires a high level of skill and experience and results are always susceptible to variations in interpretation by different scorers.  The algorithm developed in this work  provides a very straightforward method of determining levels of alertness and sleep and may prove worthwhile in simplifying and automating the sleep scoring process.

Using the drowsiness tracking algorithm developed in this work, there is potential for an improved understanding of the sleep onset phase.  We can observe from our limited dataset that the transition from wakefulness into sleep appears to be a smooth transition as demonstrated by the measure.  Merica and Gaillard (1992) reinforce our findings by stating that the "... continuity of the process of falling asleep is not adequately accounted for by the sleep staging process.  A step function is sought there where, in fact, a smooth function exists."  Also from the test results, the exit transitions from sleep to wakefulness appear more as jumps as opposed to smooth transitions.  This may simply be due to the conditions under which the subjects are woken abruptly either as a result of their head dropping or intervention by the test administrator.  Also, the sleep portions of the data seem to reveal transitions and stable plateaus which can now be investigated and compared to the more traditional sleep stages.  This may reveal a correlation between the plateaus and stages of sleep or may reveal a new characteristic or sleep such as an additional stage, etc.

We have also verified, within the scope of the experimental study, that the behavioral failures due to extreme sleepiness appear to be caused by the initial stages of sleep as we would expect and not some other mechanism.

Finally, the laboratory experiments conducted as a part of the research have verified some of the correlations between energy in the traditional frequency bands and extreme drowsiness.  For this experiment and under these testing conditions, we have determined that among the traditional frequency bands, increases in delta and theta activity were the most reliable EEG correlates/indicators with test failures.  In addition to these frequencies, a disappearance of alpha activity and reduction of beta energy also show correlations with test failures but with less consistency than the delta and theta bands.

<PREVIOUS

NEXT>