Thank you for your response.  I haven't been able to find any articles that report this.  Can you provide a citation by any chance?

Hi Victor,

Many people have treated this problem in ways adapted to their needs and possibilties. A recent open access review in frontiers of physiology (2012) can be found here:

http://www.ncbi.nlm.nih.gov/pmc/articles/PMC3358711/pdf/fphys-03-00148.pdf

(Role of editing of R–R intervals in the analysis of heart rate variability by Mirja A. Peltola)

but a Google Scholar search using the keywords "data treatment HRV preprocessing artifacts" will show many (actually too many) solutions to the problem.

Beyond ectopic beats, missing data and erroneous R-wave detections also nonstationarities might introduce confounding factors for techniques requiring stationarity as  a prerequisite (see e.g. V. Magagnin et al, Physiol Meas, 32, 1775-1786, 2011).

Thank you all for the replies and references.  I have found many, many resources that  demonstrate the importance of editing and propose algorithms for identifying likely artifacts, but the problem I'm faced with now is not editing itself, or identifying RR intervals to be edited, but rather deciding at what point too much editing/excluding has to be done in a time slice and the slice should be thrown out (e.g., 20-25% of the slice excluded as artifact, as Saman Parvaneh suggested above).  We are dealing with ambulatory monitoring equipment and some files have segments with  technological artifact, as opposed to physical artifact.  Thanks again.

I think a part of the consideration needs to be the total time length of the recording.  From a functional point of view, if an individual is being monitored for, say, 15 minutes under lab conditions, I suggest that a criterion of 80% or more coherent or at least organized spectra would be defensible; on a 24 hour naturalistic recording, this criterion would not be appropriate.  We use the 80% rule for determing success in acquision of a self-regulation biofeedback technique based on resonance frequency breathing and positive emotional state.  I hope this helps.  We are also developing a custom software program (PEBL) to monitor (in the lab) the action cascade of HRV during 4 stages: vigilance, stimulus perception (orientaiton), appraisal, and response selection.     

Your 'errors' need to be differentiated between ectopy and simple equipment error, noise or EMG artifacts. Ectopy is frequently followed by heart rate turbulence.

http://www.ncbi.nlm.nih.gov/pmc/articles/PMC1555628/

Errors are much easier to handle than this because they are localised. Pay close attention to the subsequent beats to ectopy - if they have the characteristic pattern described, you have HRT.

In healthy participants, we viciously exclude people with frequent APC/VPCs... the subsequent 8-15 beats most definitely change the overall HRV, especially in frequency analysis.

If you are using ambulatory data, then I'm assuming you have some quite lengthy segments - this means you have a lot more freedom to throw some of them away.

Hi Victor,

I often use the following paper as reference when I need to work with ectopic beats.

Clifford G.D., Tarassenko L.: Quantifying errors in spectral estimates of HRV due to beat replacement and resampling, IEEE Transactions On Biomedical Engineering, April 2005, Vol. 52, No. 4. pp. 630-638.

I hope it can helps.

I have read the already existing answers and think that it is not necessary to add anything if answering only to the formulated question.

But for me exists also another very important problem, the actual nonstationarity of the RR-interval series in quiet state of the person under investigation. I mean an unpredictable appearance of considerable (up to hundreds of ms in height and of about 10-20 seconds in duration) singular waves ? 

How should I change the method of HRV analysis in this case ?

Is anything known concerning the reasons of such waves, which are very different from previously described by us swallowing tachyardic waves ?

Can you supply a screen shot of these waves?  I cannot understand if you are talking about the rhythm (erratic sinus rhythm) or the morphology.  If it is the morphology, it is possible that this is not sinus rhythm and HRV would have a very different meaning. 

As far as excluding intervals, once the interbeat interval series is accurately detected and labeled, which means each beat is labeled as normal, ventricular, supraventricular and unclassified (which could be misdetected or inserted or truly mystifying) and the universe of the highest and lowest true interbeat intervals is identified and the segments of pure noise also identified, it becomes a mathematical game and, as mentioned, a function of how much data there are and what the data length is.  Some scanners do not actually allow you to identify the detection point of the R-wave and even in those that do, it could be uneven or shift abruptly for a period of the recording. 

All you are ever getting is a statistical sample of the person's or animal's HRV over some time period and the questions are: is the sample representative of some "true"value and are the numbers you get usable?  The most reliable numbers for a 24-hour recording that has all of the hours, for example is average HR and SDANN which smoothes out scanning errors.  When recordings have not been obsessively annotated, we are left, pretty much with those values.  SDNN is probably okay too.  My rule of thumb is that if 50% of the NN intervals are there and accurate, that segment if usable for time domain HRV.  For frequency domain HRV analyzed at a 5-min level since the interval before and after each ectopic or misdetected beat is replaced by some sort of spline, in a general way, once you have 20% ectopic, you have 60% of intervals that are not "true,"  so you cannot do much with that interval. Our rule for 24 hour recordings is if >=75% of the 5-min segments are there and each of these segments contains at least 50% N-N intervals (for time domain) or 80% for frequency domain and non-linear analysis we can use that for 24-hour analysis.  Maybe just daytime is okay or nighttime, each period has the same criteria.  For HRT one needs at least 5 acceptable VPCs but it does not depend on anything besides the local recording.

I also recommend plotting an interbeat interval tachogram with the IBI replaced by an instantaneous heart rate so that it is easy to understand.  With experience, it is possible to identify the patterns that are unexpected and validate them on the original ECG.  I like to look at about 10 min at a time with about 1 inch per minute on the x-axis.  Once the recording is really validated, more subtle HRV patterns can be quantified, but I cannot tell you how many times I have seen extreme outliers, especially in beat to beat measures like rMSSD, reported without thinking twice.

I cannot emphasize enough that before you start calculating anything, you have to make sure that you have what you think you have.  Sometimes that is easy, sometimes that is maddening.