7.4  Closing summary

This document investigated the role of fatigue and sleep in various accidents and the enormous costs (both monetary and human) associated with these events.  The basics of EEG signals and sleep were reviewed along with the literature on EEG's and drowsiness and the roles of the classical frequency bands of delta, theta, alpha, and beta.  After reviewing the literature regarding previous research on establishing correlations between characteristics of the EEG and drowsiness, various methods of EEG data analysis were reviewed.

The rationale and purpose behind the research study being conducted as part of the overall project was reviewed, and a detailed description of the experimental study was given along with a review of potential methods of experimental data analysis.  The outlined experiments and human-use protocols used in this study were voluntarily submitted and subsequently approved by the Case Western Reserve University Review Committee for Human Studies.  The experimental study was conducted in the laboratory according to the experimental description and physiologic data was collected for post-processing analysis.

Experimental data analysis followed along with a discussion of issues specific to the data collection techniques used in this experiment, including data extraction and handling methods.  Carefully selected segments of data were analyzed to find direct behavioral correlations with spectral content in the EEG signal.  This lead into a discussion of the differences in the signal content of the data collected using the traditional data collection methods (sampling and filtering) and the non-standard data collection methods used in this experiment.  Using a sample segment of data collected using the techniques employed in this study, it was shown that increased high frequency power existed in comparison to data collected using more traditional methods.  It was then necessary to determine if the high frequency components captured by the data collection techniques used in this study had any correlation with drowsiness.

Carefully selected data was analyzed for spectral content, using averaged power spectral density plots, from 25 hits and 20 missed events over three different test subjects.  A side-by-side comparison of data corresponding to hits and misses across several subjects was given.  The energy distribution in both the traditional and high frequency bands was examined and data correlating to hits and misses were compared.  A direct correlation was shown between spectral EEG components in the high frequency band and hits and misses in the visual test.  Changes in the EEG signal from hits to misses were usually accompanied by a decrease in high frequency energy by a factor of 10 to 20 in the frequency bands above 100Hz.

Analysis continued using larger data segments and the windowed FFT/spectrograms.  The high frequency shifts which were discovered during the initial data analysis were the focus of this stage of the analysis.  Various image and signal processing techniques were developed and applied to the spectrogram images to learn more about the behavior of the various frequency bands in relationship to performance.  It was shown that energy content in the high frequency bands was useful in distinguishing between hits and misses (which according to the strict data selection criteria was probably closer to distinguishing between alert behavior and sleep onset).  Also there is a continuous shifting of the energy (as opposed to discontinuous jump behavior) from the period of alert behavior (hits) through extreme sleepiness (misses) that is visible in the spectrogram images.  The spectrogram images also revealed that while the transition into the failures is a fairly smooth transition, the transition back to alert behavior can occur very rapidly.

Criteria for the development of a drowsiness tracking algorithm were outlined.  Building upon the information obtained during the analysis of power spectral plots and the spectrogram images, various frequency bands in both the traditional and high frequency range were examined and compared.  The high frequency range was quantized into equally sized energy bands and it was shown that all of the high frequency energy bands correlated extremely well with hits and misses and that the energy shifts in this range occur with very smooth transitions.  Low frequency energy bands also showed correlations with hits and misses, but did not have the same level of consistency as the higher frequencies.  When the low frequency bands (especially delta and theta) show distinctive changes associated with the hit and miss behavior, the high frequency bands always shows a very strong correlation too.  But, when the low frequency bands show little or no reaction, the high frequency bands show improved sensitivity by revealing a change in the high frequency energy levels.

In preparation for using the high frequency energy shifts as the cornerstone of an analytic measure, the inverse function was introduced as a useful enhancement of the high frequency information.  An initial off-line drowsiness tracking algorithm was developed using the correlations that were identified between energy in the high frequency bands and test performance.

A block diagram of the tracking algorithm was given and the basic theory behind the processing of raw EEG signal to obtain the drowsiness measure was given.  The frequency information used by the drowsiness measure was purposely chosen as the new high frequency range, no traditional frequency band information was used.  The general behavior of the measure was described, threshold values for all subjects were established and explained, and the measure was computed on several segments of EEG data representing various states of alertness for various subjects.  The measure showed stability during alert segments of data and the measure was not effected or influenced by the status of the eyes.  The measure also showed large magnitude changes that were highly correlated temporally with failure segments from the visual test.  The measure was finally used on segments of data corresponding to sleep where the algorithm showed behavior that was similar to the behavior seen during the test failures.  The measure during the sleep and sleep-onset phases also proved to be very stable.  The measure clearly tracked the period of alertness through sleep with a very distinct trend and allowed easy recognition of periods during the sleep phase where the subject experienced an arousal from sleep (as verified by video of the sleep segments).

<PREVIOUS

NEXT>