As neurophysiologists, we require the use of some rather sophisticated electronic equipment to amplify and measure neural signals in the brain. Complex systems have been developed to efficiently record¬†10s to 100s of¬†channels of neural data from increasingly small sensors implanted in the brain. ¬†That being said, sometimes the equipment¬†is complicated… for basically no reason. ¬†At best, we might chalk this up to engineering laziness, at worst, a thinly veiled attempt to prevent customers from interfacing systems from different manufacturers. ¬†In this case, I think a manufacturer tricked even themselves.
The Silicon Probe
One type¬†of¬†hardware I use frequently is¬†the microelecrode. ¬†This is a device that measures¬†tiny voltage changes that happen outside cells (neurons) in the brain. ¬†These voltages reflect the activity of cells in the brain coding sensory and behavioral information. ¬†One particularly small and efficient microelectrode¬†is the multi-channel silicon¬†probe. ¬†These electrodes are¬†thinner than a human hair but fit up to 64¬†conductive¬†electrode pads on a single¬†shank. The more channels of neural data you have, the more activity (cells) you can measure¬†at once, allowing more complex¬†and efficient data collection. In addition, the smaller the electrode, the less damage done to the brain tissue while measuring neural signals. ¬† Due to these advantages and more,¬†I started using 32-channel silicon probes from NeuroNexus in 2014.
This is not going to be a highly technical post. The basic gist is that at the end of the day I just need to know which signals are recorded from which sites on the electrode¬†(the little white dots at the triangular tip of the silicon shank).
Unfortunately, the sites are not numbered 1,2,3,4… (etc) from one end,¬†like you might expect. ¬†In fact from the top, the sites of the A1x32 probe are numbered 17,16,18,15,19¬†… etc. While at first annoying,¬†this¬†is ultimately a¬†minor¬†inconvenience: the manufacturers provide detailed datasheets describing how the electrode pads are numbered and I remap the sites accordingly.
One added wrinkle, however, is that I use¬†an adapter to go from the silicon probe to the amplifier and data acquisition system. ¬†In this case, I bought the NeuroNexus A32-Z32 connector ($200) to go from the silicon probe to ZifClip headstage connectors that we already had in our lab. Of course, both the NeuroNexus A32-Z32 connector and the TDT system have their own arbitrary channel numbering¬†systems. ¬†This leaves me with a series of four (!) channel re-mappings:
For example, the channel I would call #9 if I counted from the top of the shank maps from 9(Counting) ->¬†21(NeuroNexus) -> 23 (A32-Z32, also NeuroNexus) -> 15 (ZifClip TDT). ¬†This is sort of insane, but whatever. ¬†As long as it works, it is solved: each manufacturer provides datasheets that allow me to map where the signal goes from one device to the next. Good. Done.
But… actually, not done.¬†Not done at all.
Don’t Trust the Strange Data
I started to notice something strange in my neural recordings:
Because we amplify the signals of individual cells in the brain, the amplitude of the signal should not “jump” channel-to-channel like the example above. ¬†Instead, it should vary¬†continuously across adjacent channels, with a peak signal on one channel:
I started to notice pretty consistent issues with a few of the channels.¬† My first thought was that I messed up the tedious mapping shown above. ¬†So I double-, triple- and quadruple- checked the datasheets. ¬†I started feeling a bit crazy because I did not see anything wrong. ¬†I next¬†decided to manually check the circuitry of the A32-Z32 adapter with a multimeter, which involved checking the resistance of¬†all 32 channels on the A32 side of the adapter to all 32 channels on the Z32 (the low-resistance “beep” setting on multimeters is¬†invaluable). ¬†If that doesn’t seem tedious enough on it’s own, consider that I needed to do the entire thing under a microscope because the adapter is about 1 x 2cm.
Number Me This
Low and behold, four¬†channels did not match the datasheet. ¬†The way they were switched affected a large swath of the probe. ¬†I was still sort of incredulous that this was possible, so I decided to e-mail NeuroNexus to see if I had¬†an outdated¬†datasheet for the A32-Z32. They replied:
“This adaptor design hasn’t changed since its first build , so this is the only [datasheet] we have. ¬†We will be contacting the other customers who have purchased this item to notify them of the error. ¬†I apologize for any trouble this has caused.”
Nope. NeuroNexus just¬†out-smarted themselves with the complex numbering scheme this time around. ¬†The datasheet was wrong.
And then they sent me a $25 gift certificate to Starbucks.
So all said and done I got paid about $2/hour to identify these issues and fix all¬†of the data I collected since 2014. ¬†I am a graduate student, so this is a good deal. ¬†No hard feelings, NeuroNexus. (I took the lab out to coffee and it was pretty nice).
But in all seriousness, this incident¬†left me with a few questions. Firstly, how many publications does something like this affect? ¬†I assume this adapter¬†is not commonly used, which is one reason why I was the only person to notice it was wrong. ¬†But I am likely not the only person with the adapter, and it has been on the market in a defective form since 2011. ¬†So somebody, somewhere, probably published some messed up data. ¬†Whether¬†the issue would¬†compromise the findings is another question.
But on a broader level, I think the problems I had with this adapter made me reconsider the¬†odds of simple mistakes hijacking data: these things happen.¬† To make matters worse, there are no clear mechanisms for identifying publications in the literature that are¬†affected once an issue is found.
So the lesson here¬†might be:¬†have caution with products that don’t¬†have many end users because they implicitly won’t be scrutinized as¬†closely. ¬†And trust your instincts¬†if something doesn’t make sense in your data!
I still have $2.47, if anyone wants¬†coffee.
The Corrected A32-Z32 mapping
For the¬†handful of (1? 2?) people who might ever have this problem, here’s my¬†corrected mapping (the channel labeled as 30 IN on the 2013 schematic (Molec¬†input) connects to channel 27 OUT (¬†ZifClip output), likewise 27 IN connects to 30 OUT, 28 IN connects to 29 OUT and 29 IN connects to 28 OUT.)