Misusing Neuroscience Tools: Filters and Signal Artifacts.

Today I wanted to highlight an excellent primer Filters: When, Why, and How (Not) to Use Them in the Journal Neuron. Here, Dr. Cheveigne and Dr. Nelken, discuss what kinds of errors can be introduced by the filters neuroscientists apply to their raw data. Electronic filters are used to remove noise and artifacts from signals. Filters can also be used to analyze different frequency bands, as is common in EGG studies. Often, many neuroscientists use filters without much thought. For example, when recording any electrical signal its very common to use a 60Hz (or 50Hz) notch filter to remove artifacts from the power supply. This is because all electricity that comes from the power grid is alternating current or AC, and oscillates at 60Hz. If one doesn't remove this frequency, then you can see this oscillation in your data. Thus, most experiments use notch filters to remove this artifact. Of course, this means that you are removing information in the 60Hz frequency band, but often this is not a concern.

Figure 6 from Filters: When, Why, and How (Not) to Use Them. Notice all filters introduce false oscillations, and filters B and E introduce acausal signals, or activity before the input. The Left shows different filters introducing various artifacts. A is the original test signal called a box signal, or square pulse that lasts 50 ms. B-H different filters are applied to the boxcar signal. B is the signal ran through an acausal filter named filtfilt. C is a Butterworth lowpass filter with a 10Hz cutoff. D is a Butterworth lowpass filter with a 20Hz cutoff.  E is D filtered though the acaual  filtfilt filter. F is a Butterworth band pass filter order 8 with a pass gap of 10-20Hz. G is a Butterworth band pass filter order 2 with a pass gap of 10-20Hz. H is a Butterworth band pass filter order 2 with a pass gap of 10-12Hz. The right panel shows the error in decibels.

Even though there are lots of legitimate reasons to use filters in neuroscience, misusing filters still can occur.  Depending on the filters, one might create false oscillations, suppress certain oscillations, introduce signal delays, or even worse create an a-causal signal.  The primer does an excellent job describing the different artifacts various filters can introduce. There were two types of errors that stuck out to me. First was the false oscillatory ringing introduced by the filters. Because of my dynamical systems background, often I look at a voltage trace, and I am looking for sub-threshold ringing as a tell tale sign of complex eigenvalues, and resonating phenomena. This short-cut is very useful for tuning models, debugging models, and checking for numerical instability. Furthermore, often, for intracellular whole patch clamp recordings the sampling rate is high enough to be confident in the results. However, for slower sampling rates in calcium imaging, and EEG this assumption may longer valid. Filters may even add a false oscillatory ringing after a sudden impulse of activity. Thus when looking at filtered data, one may see these false oscillations. Thus, you should always be aware of your, or others, filter choices. What you, or the author assume is a resonance,or subthreshold oscillation, may in fact be nothing but the filter effect.

The second, and perhaps most surprising artifact, was signal delay and acausal signals. In particular, casual filters only include sample points in the past. However this means that your signal will have a phase offset, and will be delayed. To compensate for this often one will apply a zero-phase offset filter, which is acausal. It uses future data points to determine the current value. This is very useful if your trying compare the exact timing of peaks in your data. However, showing only the acausal filter response will give the impression that the system is anticipating the input response. This is course may not be the case.

This of course not to say that using a filter is unwise, as filters form the backbone of neuroscience. Without filters we wouldn't be able to analyze any signal produced by our brains. Furthermore, this doesn't mean resonance, prediction, and anticipation aren't real phenomena in the brain. But every tool can be misused. Dr. Cheveigne and Dr. Nelken give an excellent overview of some of the pitfalls in filter use. It is written at a very practical level, and almost no math background is necessary to understand. I suggest anyone interested in becoming a neurosciencist should consider giving this primer a read.


Author: Alexander White


Source: Filters: When, Why, and How (Not) to Use Them