Nadia, This is a great question. Our group takes from the approach that you sample at an appropriate sampling frequency that would not require filtering. When you pilot your data, run a power spectrum analysis to see what frequency 99% of the data live under. Then set your sampling frequency based off of that. We come from the standpoint that variability in the time series is not noise but inherently interesting. Thus, we do not filter data because we do not want to lose that interesting data. It really depends on your research question you are asking. I would look at Nick Stergiou's book, Innovative Analysis of Human Movement, especially the final chapter. Hope that helps.

Jenna

Hi Jenna, thanks for sharing your perspective and for the book recommendation. The data set I'm working with is from adults with autism and intellectual disability, and we are using displacement of the COP and sway area as our dependent variables. Your point about the variability being inherently interesting is a good one - I'm going to try to find a copy of that book and see if there are additional variables we should also be considering. Thanks!

Using 99% (or even 95%) CPSD (cumulative power spectral density) is a reasonable logic. Loosing the informative variability inherent in the postural control system behavior, might be the most important concern about filtering the signal but, considering the physiological behavior of human sway during standing posture, data at frequencies higher than 5Hz, don't seem to be of physiological sources. We used the 99% CPSD method on data from Low Back Pain subjects and came to this frequency although the studied population might change it.

Please note that the traditional "Sway Velocity" used by many authors and implemented in many commercially available devices is defined as the SwayPath divided by task duration [Prieto 1996], without differentiation.

If you are interested in the instantaneous velocity, the numerical differentiation requires a LP filter before and after the filtering procedure to remove the electrical noise and its effect on differentiation. In motion analysis LP frequencies of about 6 Hz are used for trajectories, as in Winter's papers. Higher frequencies , up to 20, be necessary if you're working with data from patients with tremor. Orthostatic tremor is usually high frequency (14 Hz to 18 Hz).

The paper "The sensitivity of posturographic parameters to acquisition settings", from Schmid and colleagues, can be helpful.

In my experience with data from healthy subjects and kids, older people, stroke patients and PK patients, the effect of the LP cut off frequency is neglible with respect to the effect of pathology. COP area is not affected by LP freq. Path length and RMS are affected, but the clincally relevant information is always there, and can be enhaced by the task variation (e.g. eyes closed or soft surface)

Again, the instruction given to the subject (stay relaxed vs stay steady vs stay as firm as possible) may alter the COP much more than the LP frequency.

Hope that helps

Andrea

Hi Nadia,

Glad to see you here, I agree with Jenna and Amir. Filtering COP data depends on your research question. Usually COP data are oversampled. You should do a pilot study and using FFT method to determine the frequency below which 95% or cautiously 99% of data lives under. Then using the Nyquist sampling theorem, sample the data with twice that frequency. But if I were you I find the dominant frequency of fluctuations and sample with twice of that. Increasing sampling frequency induces noise in your captured data and more important, too close data points do not contain useful information; they just increase computation time. However if you have already captured the data, it does not matter. If you do not want to do nonlinear analyses you can low pass filter the data with above mentioned procedure, but with zero lag (like 2nd order butterworth filter) and I think that you find a cutoff frequency below 10 Hz. But if you are intended to do nonlinear analysis, filtering is not recommended; Instead of that you can do four point averaging technique. But when you are sampling the data for nonlinear analysis, it is good to find sampling frequency according to Poincare section of your pilot study data. I can devlope and send you matlab codes if you want.

Hamed

Dear Nadia, I don't understand why you are so surprised if some group doesn't filter the data. If your platform is a sufficently good platform, the signals it will give to you are "clean" enough and you will not need to filter them. If the noise is low, why to compulsory filter them? Bye

In general, EMG data band stops around 350-400 HZ, and therefor 350Hz is a good cut-off frequency.

Hi All,

as all of you expect to see, this is a question with lots of answers. filtering is a standard process in any instrumentation system, but filtering is always for a reason. In FFT it would be for anti-aliasing, in other studies it may be used to eliminate mains frequency. It is never a simple yes or no answer, however in all studies whether filtering or no filtering was used MUST be stated, as should the mode of filtering, low pass, band pass, high pass, chebychev etc. In your case you should determine WHY you think you need to filter the data, then select a suitable filter, and then compare pre and post-filtering data to see the outcome/effect.

I would suggest you look at a good instrumentation text book to learn about filtering.

good luck, P

I agree with Peter above, but would emphasise again that filtering should be used to help you answer a specific question. Generalised statements such as don't filer, or filter at X Hz are not that useful. For each situation you need to determine whether you are going to filter or not, and how.

I still like David Winter's classic book.

Hi everyone, thank you all for your input.

After looking through the literature, I posted the question because I was curious to hear from the community about "why we filter" or "why we don't filter" from a theoretical perspective. For example, Jenna's statement "We come from the standpoint that variability in the time series is not noise but inherently interesting." I understand the mechanics of filtering, residual analysis, finding an appropriate cutoff frequency, etc., but up to now I had never come across a situation in my research where you would not filter. I have worked with EMG, accelerometers, electrogoniometers, and motion capture up to now, and we have always filtered the data. This was my first venture with the force platform, and based on my past experience with instrumentation it was surprising to see this type of data go unfiltered. I simply wanted to get some informal accounts of the theory behind why you would not want to filter the COP data.

Thanks again!

Nadia.

During balance measurements for quite stance you have two variables that are typically used to estimate the average sway velocity in any axis. Typically you want to use the center of mass (gravity) in these calculations as opposed to the center of pressure. Low pass filtering on the center of pressure time series is the preferred method of estimating the center of mass.

In general all exeprimental measurments should be filtered. Unfiltered data can be usefull to indicate the noise level from the measuring instrument,or when the signal level is similar to the noise level (but then probably the instument is not chosen properly for the application). I can't think of any other really usefull cases. Force plate measurments as any experimental measurments contain noise. Therefore the instrumental noise should be filtered out, to leave you just the measured quantity. Depending on how accurate sensor there is in the force plate and how much does it pick up noise from environment you should adjust your filter. Good rule of tumb is to indicate first what is the highest frequency that you expect to appear from the phoneomea you are measuring and then check the noise level in the data. Noise frequency should be much higher then the highest frequency you expect to be important, then you can simply low-pass filter the data.

My general opinion is to use filters *only* when you can give a good reason for it.

Typical examples are:

- you see e.g. measurement vibrations in the power spectrum (-> use e.g. bandstop filter)

- you want to use the derivative of the signal (-> use lowpass, cut-off based e.g. on cummulative power density)

Because filters can introduce "artifacts", you should use them sparingly. For example, if you use filters in inverse dynamics, your results will be physically inconsistent (although I'm not aware of a good solution for this).

The best is to look at "robust" quantities, which do not depend on filtering (e.g. "max CoP sway amplitude / time used for that motion" would be robust).

Regarding the cut-off frequency: it's always good to make a visual check on the spectrum. For example, you might have strong vibrations in your data which you don't want to have, but they contribute to your cummulative power density.

ps: I like to sample at high frequencies to get an increased resolution. You might need this in later analyses of the same data, if you look at different quantities. Resampling to lower sampling rate is always possible.

The sampling rate depends on what you are doing with the data. Low pass filtering is only necessary if you plan to double-differentiate the data, which you almost never do with force data. If you are using the force data to permit inverse dynamics, however, you may want to filter the forces so as not to contaminate the moments of force. Moments of force are used to reflect the contractions of muscles. Since muscle contractions only can produce frequencies as high as 8 Hz, the higher frequencies in the force data will create artificially high frquencies in the moments of force that cannot be explained as coming from muscle contractions. On the other hand, if you want to know the forces experienced by the bones, etc., than you may want to keep the higher frquencies experienced during an impact. In your case, you probably don't need to filter since you are only differentiating once to get COP velocity. Nevertheless, since COP sway has a very low frequency spectrum and is thus totally controlled by muscle contractions, something like a 10 Hz cutoff, low-pass filter will remove non-muscle perturbances to the force signals (possibly caused by motors or building vibrations).

IMHO it depends a lot. I have had a few surprises wih force platforms. I think that deciding "to filter or not to filter" should be based on the answers to some questions (the list is not complete) :

i) what is the sampling frquency ?

ii) what about the eletromagnetic experimental environment ? may this induce a parasitic signal ?

iii) what about the floor vibration ? are there any vibration dampers mounted under the platform ?

iv) what type of AD converters are used on the force platform (16 bit, 24 bit, etc) ?

v) how are going to process the data ? will you differentiate the values ?

Answers to these questions are already given above.

To get an idea, one thing I have done is a "dead body" test on the platform. I took a 45 kg concrete slab and put it on the platform. Once recorded, the signal proved to be quite noisy, and I had to filter. Give it a try and come back to tell us.

Filtering is not a sure matter, but something naif very close to an art (2The art of filtering"). You need to filter depending on how you recorded the signals and relative noises which interferred with them. The typical characteristic of noise is its extreme enviromental variability: so, the filter must adapt to it: therefore, it is clear why different authors use different filters.