4.2 Data Analysis


4.2.1 Initial Data Analysis Procedures

The data analysis procedures outlined in this section are intended to increase the general level of knowledge about the signals being collected under the laboratory conditions of the proposed study and should not be interpreted as a complete list of the techniques used to solve the problem being investigated in this research.  Subsequent chapters will describe more specific signal processing techniques that have been developed to differentiate between alertness and sleep, and ultimately those features that reveal incipient sleep onset.

Unlike other studies which rely heavily on statistics and data averaging during the initial data analysis phase, our objective is to identify those moment-by-moment (temporal) changes which we believe exists in the data and can only be uncovered through tight behavioral correlation, data synchronization, and analysis over very small time increments (initially).  The research goal is to identify changes in the EEG signal which correlate directly with behavior, where in this work, behavior refers to the act of responding to the visual stimulus or a failure to respond.

Because we are collecting behavioral data at the same rate as the physiologic information, data synchronization is natural and straightforward.  All data is processed and test performance reports are generated automatically.  These reports show a complete breakdown of each SVS test event including: hit/miss/reaction time, time since last visual event, time and sample number of when each SVS event starts and stops, time and sample number of when each response starts and stops, etc.  From this information each event is studied and classified according to the supposed behavior involved, i.e. fast response hit, slow response hit, missed response, response with no stimulus, etc.

Behavioral event information is then analyzed further using the video and vertical EOG measurement.  The video recording and EOG measurements are used to examine the behavioral and physiological characteristics of each subject to determine the cause of any missed responses or slow response times (e.g. if the subject was not looking at the display board when the signal first appeared, or if the subject's eyes were closed for a much longer duration than normal eye-blinking, or if the subject had nodded off during a microsleep, etc.).  If external factors can not be ruled out, missed signals will be classified as a potential lapse or microsleep behavior.

All hits and misses are analyzed before attempting correlation with processed EEG data.  Attempting to correlate EEG data with a failed response due to an external cause can lead to incorrect results.  On the other hand, changes in a subject's response time or characteristics not attributed to external causes could indicate a transient change in alertness and therefore require analysis and correlation.

The first and most important step is to differentiate changes in the EEG during both hit and missed signals.  Events identified as fast-hits (Rt=mean response time ± 1sd) and missed signals without apparent external causes (Rt=inf) are the only data segments included in this initial analysis.  Physiologic data corresponding to the presentation of each event is extracted.  The data that contains the most meaningful behavior information is that which is taken during an SVS event.

The data analysis plan is to use signal processing methods to clearly understand and define the characteristics in the EEG signal that are correlated with transitions between alertness and extreme sleepiness (hits and misses).

To begin with, the physiologic data is analyzed using the Fourier transform for spectral analysis.  Physiologic data (events) are grouped as hits and misses and the power spectra are compared for individual and grouped events (using averaged spectra) in order to discover various qualitative changes in the EEG signal correlating with hits and misses.  If an accompanying frequency shift in the EEG signal occurs as the result of a transition from alertness to extreme drowsiness, it should be apparent from the straightforward test conducted in this research and this initial data analysis.  Note, we are only analyzing the endpoints of the behavioral transitions seen in the SVS experiment and not the potentially subtler transitional periods between these endpoints.  Search and comparison of the power spectra is not limited to the standard frequency bands, it may be that the frequency shifts we are searching for do not occur in these traditionally defined bands.

Once relationships between changes in response characteristics and changes in the EEG signal characteristics are identified, this information is used to develop and test particular correlations.  For instance, certain variables are selected (e.g. representing energy shifts, energy within particular bands, ratios of energy within bands, etc.) and used to estimate correlations with test performance (behavior).


4.2.2 Subsequent Data Analysis Procedures

After the initial data analysis is completed, the set of qualitative features in the data correlating to hits and misses is available for further study.  The analysis is now expanded to time periods prior to, and after each of these SVS events.  Specifically, the segments of data that are known to contain a behavioral transition are isolated for further evaluation.  For example, segments of data that contain a series of hits and then a transition to a series of misses, or visa-verse require more analysis.

Initially, the specific signal processing techniques to be used in this phase of the analysis was unknown and had to be determined based on the results from the previous analysis phase.  As will be described in detail in subsequent chapters, the interesting features uncovered in chapter 5 are frequency shifts located in nontraditional frequency bands of the EEG signal.  This suggests focusing on spectral data methods for continued analysis.

Using data segments containing behavioral transitions, various window lengths and different amounts of overlap between segments are initially used in the computation of spectrograms.  Consistent with the preliminary data analysis, the resulting spectrograms revealed frequency shifts identified in the first phase of analysis.  The temporal nature of the frequency shifts are studied in greater detail and this allowed us to track both the speed and nature (slow shifts, jumps, etc.) of the transitions.  A 2048 sample data window and Hanning windowing function are chosen for spectrogram analysis and various techniques for spectral enhancement of those regions identified during the first phase of data analysis are used.

The computation of windowed Fourier transforms is not exclusively limited to viewing spectrograms.  Data from a windowed Fourier transform is also used to track the evolution of a single frequency and combinations of frequencies (bands) over time.  Using frequency points and bands, temporal changes (behavior) of the signal are tracked.  The idea is to identify those frequency points or bands with the best tracking characteristics.  In this way, various frequency shifts are identified, isolated, and studied.  The behavior of these shifts can be tracked over an extended period of time to look for any anomalies that may have been missed during the previous data analysis phases.  Data from the alert day test and drowsy day tests are both analyzed.  The sleep data is also analyzed for frequency shifts similar to those present in the test data.  If our hypothesis about the frequencies and behavior involved are correct, these shifts should also be apparent as the subject is allowed to fall asleep in bed.  This sleep data constitutes a baseline representing the lowest state of CNS arousal in the study.

Finally, data from the alert, drowsy, and sleep tests are used to create a baseline measure which takes these qualitative findings and transforms them via an algorithm into quantitative data that correlates directly with the physiological state of the individual.

<PREVIOUS

NEXT>