Self-Noise of OpenBCI with More Data

After posting about my short-duration measurement of the self-noise of the OpenBCI board, there was a comment that discussed EEG applications where very low frequency signals were being used.  To understand the self-noise of the OpenBCI board at these low frequencies, I repeated my self-noise measurements, but used a much longer recording...one hour instead of ten seconds.  I also to the opportunity to measure the noise on five channels, not just one like before.  Here are my results.

Measuring Self-Noise of Five Channels of OpenBCI by
Jumpering Them All to Analog Ground

Setup:  The setup that I used was the same as before except that I jumpered the first five inputs to analog ground, instead of jumpering just the first channel.  Also, yesterday, I de-activated all of the channels except for the first one.  Since today's test used five channels, I kept the first five channels and only de-activated the last three.

Data Collected:  I recorded about an hour of data.  Near the end of the data, there was a spike that I could see in the graphs...I must of bumped the setup with my hand while I was doing other things.  So, I trimmed the data to remove the spike.  Overall, I was left with 3180 seconds of data.  The lowest possible frequency that could be represented in this data is about 1/3180 = 0.00031 Hz.  That is a very slow signal.

Results for One Channel:  After removing the DC offset (ie, the mean) of the entire 3180 second recording, and after lowpass filtering the data with a cutoff at 65 Hz, I get the histogram shown below.  This is for Channel 1.  As can be seen, it is a nice Gaussian shape, which is very smooth because of the huge number of data points in my 3180 second recording.  The standard deviation is 0.15 uV.  Since the standard deviation is the same as the RMS value (when the mean is removed), the RMS noise value for this channel for this recording was 0.15 uVrms.  This is very close to the 0.16 uVrms value that I recorded yesterday for the shorter 10 second recording.  This agreement makes me feel very good.

Histogram of 3180 Seconds of Noise Recorded from
OpenBCI with its Inputs Jumpered to Analog Ground.

Noise Spectrum:  To see how the noise level varies with frequency, we can plot the spectrum of the signal.  Since the recording is so long, the spectrum reaches down to very low frequencies.  For channel 1, I got the spectrum shown below.  As can be seen, it has two regimes: (A) a flat "white noise" region above 0.07 Hz and (B) a sloped "1/f Noise" region where the noise density increases as the frequency gets lower and lower.  This "1/f" behavior (the "f" is for "frequency") is very commonly seen when evaluating the low-frequency noise of analog amplifiers.  It is impressive that the 1/f noise doesn't begin until about 0.05 Hz.  As a result, even at a frequency of 0.001 Hz, the noise level is only about 0.3 uV/sqrt(Hz).  That seems pretty good to me.

Spectrum of Self-Noise Recorded from OpenBCI
with its Inputs Jumpered to Analog Ground.

60 Hz Noise:  What is a bit surprising in this spectrum is the sudden appearance of the 60 Hz noise (there was none seen in my data yesterday) and of a spike at 0.080 Hz.  I'm thinking that, because I now have 5 channels active instead of one, and because the 5 channels share the single SRB2 input as the reference for the differential amplifier, that the common-mode rejection capability of the differential amplifier is degraded because of the 5x leakage current through the common SRB2 components.  Any imbalance on the two legs of a differential input will degrade the common-mode rejection, and the datasheet for the ADS1299 warns of this behavior.  It looks like we're seeing it.  While it is unfortunate, the level of the 60 hz (and 0.080 Hz) is still quite low...it reads about 0.3 uV/sqrt(Hz).  Since the "sqrt(Hz)" part is confusing for sinewave-like signals, I zoomed in on the graph around 60 Hz and I saw that the bandwidth of the 60 Hz signal is about 0.01 Hz.  From this, I estimate that the RMS value of the 60 Hz signal is 0.3 uV/sqrt(Hz) * sqrt(0.01Hz) = 0.03 uVrms.  So, it is a very small signal and probably not much to be concerned about.

Comparison Across Five Channels:  Unlike yesterday, where I measured just one EEG channel, today I measured five.  I would have done all eight, but I didn't have enough jumper wires.  For the five channels of data, I made the same plots.  They are very similar to each other.  The white noise is flat and the 1/f noise is similarly sloped.  They all contain the 0.080 Hz and the 60 Hz spikes.  The only real difference is that the amplitude of the two spikes varies a bit from channel to channel.  That does not change my overall conclusion that the five channels are sufficiently similar.  For the 3180 sec recording, I've tabulated the RMS noise (up to 65 Hz) below:

• Chan 1: Noise = 0.15 uVrms
• Chan 2: Noise = 0.16 uVrms
• Chan 3: Noise = 0.17 uVrms
• Chan 4: Noise = 0.18 uVrms
• Chan 5: Noise = 0.15 uVrms

Conclusion:  This longer, multi-channel recording that I performed today confirms the noise levels that I recorded yesterday.  Today's longer recording is also able to reveal the low-frequency noise behavior of the system, which transitions from white noise to 1/f Noise around 0.07 Hz.  At 0.001 Hz, the noise density is about 0.03 uV/sqrt(Hz).

Other EEG Systems:  While I think that these noise levels are pretty low, I still do not have any comparison data from other systems.  Does anyone know how they perform?

Self-Noise of OpenBCI

An important quality of any EEG system is its noise floor.  If an EEG system has a noise floor that is too high, the noise might mask the small EEG signals that you're trying to measure.  So, when working with an EEG system, it is important know that its noise floor is low enough.  I've been working with the OpenBCI system, but I don't know what it has for a self-noise level.  So, I decided to measure it.

Measuring the Self-Noise of the OpenBCI Board by
Shorting the Inputs to Analog Ground

Expected Noise Level:  The heart of the OpenBCI board is the Texas Instruments ADS1299, which is a "low-noise, 8-channel, 24-bit analog front-end for biopotential measurements". Its datasheet says that it has a self-noise of 1 uV over a bandwidth of 0.01 Hz to 70 Hz.  That 0.01 Hz value is a very low frequency and would require at least 100 seconds of data, if not several hundred seconds of data (to get a good average) to properly evaluate.  Since I don't care about signals down to 0.01 Hz, and since I want a faster test, I looked at the datasheet for other statements regarding its noise level.  The best value that I found was in the Table 4 in the data sheet, which I copied below for convenience.  A footnote on the table says the values are based on a recording of 1000 data points.  When running at 250 Hz, this means that their sample had 4 seconds of data and not the 100s of seconds of data needed to recreate the value reported for 0.01 Hz.  This is a shorter test.  Great.

Table 4 from the ADS1299 Data Sheet Showing Self-Noise for a Gain of 24.
For a sample rate of 250 Hz, the noise level is 0.14 uVrms for a 65 Hz bandwidth.

We run the ADS1299 at a sample rate of 250 Hz, so looking near the bottom of the table, we see that we should expect a noise level of 0.14 uVrms for a bandwidth of 65 Hz (presumably 0.25 Hz up to about 65 Hz).  To make sure that I'm interpretting this table correctly, I found Figure 3 in the datasheet (copied below), which shows a sample of noise recorded with a gain of 24 and a sample rate of 250 Hz.  Based on the amplitude of the noisy signal, I would say that this graph is consistent with the 0.14 uVrms value that we took from the table above.  That's good.  Now, let's do our own experiment to see if the self-noise of the OpenBCI board is at this low level.

Figure 3 from the ADS1299 Data Sheet Showing a Sample of Self Noise.
This picture is consistent with the 0.14 uVrms level given in the table.

How to Measure "Self Noise":  When evaluating the self-noise of a sensing device, what you want to measure is the amplitude of the signal that the device thinks that it sees, even when there is not "real" signal present.  If there is no real signal present, then anything that is present is noise.  One easy way to ensure that no real signal is present is to simply short the inputs to ground.  That's what we'll do.

Shorting the Input to Ground:  For the OpenBCI board, each input is actually a differential measurement between the input and the common reference labeled "SRB2".  So, as you can see in the picture above, I shorted "input 1" to "SRB2" with the yellow jumper wire, and then I shorted SRB2 to analog ground with the white jumper wire.

Other Hardware:  The OpenBCI board was mounted on an Arduino Uno.  It was plugged into a PC via USB.  The Arduino (and, therefore, the OpenBCI board) were being powered from USB.

Software Setup:  I configured the OpenBCI board for normal EEG data collection.  This means that it was using a sample rate of 250 Hz with a gain of 24.  All channels except for channel 1 were turned off.

Results:  I recorded 10 seconds of the digital values that were output by OpenBCI board via the Arduino.  After scaling the raw counts into volts (1 count = 0.022 uV), and after filtering the noise to a bandwidth of 0.1-65 Hz, I made the time-domain plot below, which mimics the plot above from the ADS1299 datasheet.  As can be seen, it looks very similar.  The RMS value is 0.16 uV, which is very close to the 0.14 uV reported in the data sheet.  Given that datasheet values are usually very hard to achieve in practice, I find the 0.16 uVrms value to be a remarkable result.

Noise Recorded from OpenBCI with its
Inputs Jumpered to Analog Ground.

Characteristics of the Noise:  In addition to the RMS value, it can also interesting to look at the histogram and spectral properties of the noise.  The histogram below is for the same sample of OpenBCI data shown above.  It appears to be Gaussian shaped, which is good because that is what is expected.  The spectrum shown below  is also good because the spectrum is generally flat all the way up to the Nyquist frequency for this sample rate (Nyquist = half the sample rate = 250 Hz / 2 = 125 Hz).  Looking in detail, the magnitude of the noise density appears to average about 0.02 uV/sqrt(Hz) for frequencies up to 65 Hz, which is also good because it is consistent with our RMS value being 0.16 uV (this is consistent because 0.16 uV / sqrt(65 Hz)  = 0.02 uV/sqrt(Hz)).  Good news all-around!

Histogram of 10 Seconds of Noise Recorded from OpenBCI with its
Inputs Jumpered to Analog Ground.

Spectrum of Self-Noise Recorded from OpenBCI with its
Inputs Jumpered to Analog Ground.

Conclusion:  Based on this single 10 second recording, the self-noise of channel 1 of my OpenBCI board is 0.16 uVrms over the bandwidth of 0.1 to 65 Hz.  This is consistent with the value reported in the datasheet for the ADS1299 (which was 0.14 uVrms), so I have confidence in the value that I measured.

Other EEG Systems:  This 0.16 uVrms value sounds pretty good to me, but I do not have much experience with other EEG systems.  How about you?  Do you know the noise level for other systems?  Can you share the values?  I'd love to build a table comparing the different systems!

Follow-UP:  Based on reader's comments, I've extended my measurements to go lower in frequency and to look across multiple channels.  Check out the results here!