5.1  Traditional EEG Data Collection Overview

Typical EEG literature states that the information content of an EEG signal is bandlimited between roughly 0.5 and 30Hz.  Within this range, the frequency spectrum is quantized into bands with standardized names such as delta, theta, alpha, and beta.  This combination of frequency bands spans the 30Hz range of frequencies.

Sampling rate criteria

The Nyquist sampling-rate criteria requires sampling at a rate of at least two times the maximum frequency in a (bandlimited) signal to prevent aliasing.  To ensure that the signals are bandlimited, lowpass (anti-aliasing) filters (LPF) are placed in series with the input signal prior to digitization.  Typically, the lowpass filters for EEG data are set at approximately 30Hz to bandlimit the signal.  Therefore, to digitize a 30Hz bandlimited signal, the sampling rate should be at least 60Hz.  Because the lowpass filters built-in to most EEG amplifiers are typically of low order (a 1st-order Butterworth filter is common), sampling rates are usually set at 128Hz (sometimes 256Hz) to allow for the shallow filter rolloff (-20dB/dec for Butterworth filters) and prevent aliasing.

Line Noise Filter

Also, because of the extreme sensitivity (high gain) of the amplifiers used in EEG data collection, 60Hz line noise is a typical contaminant of EEG records.  So, in addition to the anti-aliasing lowpass filters standard on most amplifiers, a sharp 60Hz notch filter is also typically available to attenuate this very common source of interference.


5.1.1  Experimental Data Collection

Sampling rate

In our experiment, we extended the data capture bandwidth beyond 30Hz to thoroughly investigate the hypothesis that higher frequency components exist and are useful in drowsiness detection.  It seemed entirely possible that a system as fast and complex as the human brain could have components working faster than the typical rates observed.  Therefore, sampling and A/D conversion was performed at a rate of 7.6kHz (950Hz/channel over 8 channels) which gives a maximum possible bandwidth of 475Hz (Nyquist Thm.).  The anti-aliasing lowpass filters on the EEG amplifiers were set to 100Hz (discussed later) and the 60Hz notch filter was not used to prevent any additional data loss.  Because the data will be post-processed, 60Hz line noise (and harmonics) can be eliminated during the data analysis.  Use of faster sampling rates also allow for a tighter temporal correlation between the EEG data and test performance.

The main disadvantage of a higher sampling rate is the large amount of data which must be collected and continuously streamed to disk over extended periods of time (typically several hours) without loss of data.  Also, there is an increased computational burden during the post-processing data analysis when the higher data rates are used.  Certainly, if no additional information was revealed from the higher sampling rate, the data could be subsequently lowpass filtered and decimated to reduce the amount of data which must be handled and processed.  Because the EEG data as collected in this experiment does not conform to standard practice (lowpass filtering and sampling rates), data needs to be lowpass filtered (around 30-50Hz) and decimated to standard rates (128 or 256Hz) for comparison with standard EEG data.

Reasons others may sample fast

There are some occasions, however, when EEG researchers choose faster sampling rates than those typically used in the collection of spontaneous EEG signals.  Typically, these researchers are analyzing EEG signals for their time-domain features.  Certain types of analysis, such as evoked potential (EP), use very short segments of the time-domain EEG signals (usually on the order of milliseconds) for waveform analysis (peak detection, etc.), usually in response to an externally generated (evoked) stimulus (visual, audible, etc.).  For this reason, faster sampling rates allow for a more precise view of the input signal (analogous to a high-speed camera being used to slow down an event for closer analysis) with improved temporal resolution and correlation.

Also, researchers analyzing spike activity in the EEG typically use faster sampling rates because the faster data rates reduce the distortion of spikes (high bandwidth events) which occurs because the data collection bandwidth was too small.

<PREVIOUS

NEXT>