Hello,

you must be more specific on the purpose of your study. If you're interested in the arrhythmia analysis, then use the whole recording and report on types of arrhythmia per hour or minute or the whole recording (e.g. 24-hour ECG). If you want to analyze heart rate variability (HRV) then again there are different approaches, for physiological purposes and spectral methods, 5-minute segments are recommended. If you want to use time domain methods then 24-hour is standard, particularly when you use it for prediction purposes. For methods other than spectral or so called time domain, there are no specific recommendation. Another problem is how you want to select your segments, in beat space or time space. From physiological point of view time space is preferable and then the segments of the same duration should be chosen, e.g. 5-minute segments for all recordings. In beat space, some prefer to select the same number of beats, e.g. 256 or 516. However, please not that sinus node (i.e. physiological pacemaker) operates in time space as cardiac cells depolarize after some time and not after some beats. If, however, you want to analyse even something else than HRV, e.g. complexity, nonlinear dynamics, patterns etc, then there are no standards. As you wrote, you have also 2 other signals, probably blood pressure (PAP?) and respiration. Then with blood pressure you can also analyse blood pressure variability and baroreflex function. Depending on methods, authors/inventor recommend different lengths of recordings, however there are no standard approaches. My recommendation, as a general rule, please use always the same length of the recording in time space, not in beats.

If you have respiratory signal then you can also focus on separate breathing cycles. So, as you see, the answer is not easy, depends on what you're really interested nd thus please be more specific on your topic.

Please also remember that if there's too many artifacts and arrhythmic beats, most of the analyses will give you unreliable results. Extra supraventricular and ventricular beats should not be included into analyses of HRV, BPV, baroreflex function as they are not under physiological control as normal beats derived from sinus node.

Best wishes,

Przemyslaw Guzik

Hello,

you must be more specific on the purpose of your study. If you're interested in the arrhythmia analysis, then use the whole recording and report on types of arrhythmia per hour or minute or the whole recording (e.g. 24-hour ECG). If you want to analyze heart rate variability (HRV) then again there are different approaches, for physiological purposes and spectral methods, 5-minute segments are recommended. If you want to use time domain methods then 24-hour is standard, particularly when you use it for prediction purposes. For methods other than spectral or so called time domain, there are no specific recommendation. Another problem is how you want to select your segments, in beat space or time space. From physiological point of view time space is preferable and then the segments of the same duration should be chosen, e.g. 5-minute segments for all recordings. In beat space, some prefer to select the same number of beats, e.g. 256 or 516. However, please not that sinus node (i.e. physiological pacemaker) operates in time space as cardiac cells depolarize after some time and not after some beats. If, however, you want to analyse even something else than HRV, e.g. complexity, nonlinear dynamics, patterns etc, then there are no standards. As you wrote, you have also 2 other signals, probably blood pressure (PAP?) and respiration. Then with blood pressure you can also analyse blood pressure variability and baroreflex function. Depending on methods, authors/inventor recommend different lengths of recordings, however there are no standard approaches. My recommendation, as a general rule, please use always the same length of the recording in time space, not in beats.

If you have respiratory signal then you can also focus on separate breathing cycles. So, as you see, the answer is not easy, depends on what you're really interested nd thus please be more specific on your topic.

Please also remember that if there's too many artifacts and arrhythmic beats, most of the analyses will give you unreliable results. Extra supraventricular and ventricular beats should not be included into analyses of HRV, BPV, baroreflex function as they are not under physiological control as normal beats derived from sinus node.

Best wishes,

Przemyslaw Guzik

Hi Dear Guzik

Thanks for your response. This was complete and really helpful. But some questions come into my mind & i will be appreciate if you help me about them:

If i want to be more specific on my project,i am working on cardiac health diagnosis in ICUs and for this reason i want to get some features (such as wavelet or the same other pattern recognition methods) from the signals that i mentioned above, then put the features into one classifier such as neural network.

hence,if i use (for example) 5 min length segments, then my feature length will exceed and i will get poor results. and if i get HRV and BPV from this signals to have analysis on them, then i will lose the morphology of signal and i will not be able to diagnose more arrhythmia. what length do you suggest me for this reason?

the other problem that i have is that:

i use MGH/MF database from physionet. the type of arrhythmia for a signal is given for each beats of it. in other words, each beats of a signal are labelled with the type of arrhythmia that it has. if i use 5 min length segments, there will be some types of arrhythmia in a segment. then, how must i determine the type of arrhythmia for a segment?

Excuse me for more questions and poor English.

Thanks in advance

Akbar

For arrhythmia analysis typical length of 5 minutes is used, as a minimum can be only 10 seconds, but there isn't a good achievement with neural networks.

Hi, you may have a look at the method used in "Clinical decision support for diagnosing stress related disorders by applying psychophysiological medical knowledge" Nilsson et al., this seems to have some similarities with our work, finding patterns and connect patterns probability to a certain condition. In the beginning your confidence may be low, but with every new beat your confidence may increase. Best luck with your research.

Dear Esmaeelzadeh

Let me answer step by step to your questions:

"If i want to be more specific on my project,i am working on cardiac health diagnosis in ICUs and for this reason i want to get some features (such as wavelet or the same other pattern recognition methods) from the signals that i mentioned above, then put the features into one classifier such as neural network."

first, what is the reason of your studies, is it to explore some mathematical methods like wavelets analysis, looking for patterns (nonlinear dynamics, symbolic dynamics). If yes, then most probably you will spend a lot of time and find after all it was a nice (or not) exercise. However, if you want to do it for clinical reasons then I have another question - is it for diagnosis or prediction of risk. For the diagnosis using wavelet analysis or other methods applied for time series analysis is rather a waste of your time. If for prediction, then you may get some information. However, the best predictive value is usually for longer and not shorter ECG recordings, and besides you MUST analyse the normal rhythm of sinus origin, which means you should exclude all artifacts, all supraventricular, ventricular and paced beats. Still there is another "however... The presence of not-sinus and not paced beats may be also important in the risk assessment, particularly when these beats have specific patterns like tachycardias. But be aware that most of the monitoring systems have special algorithms which detects such beats, particularly those related sudden death like ventricular tachycardia or ventricular fibrillation. Sometimes it is a lack of any rhythm which means cardiac arrest, this is asystole. As you see there's even more and more questions. If you want to use the neural network (or other more or less sophisticated and fashionable method) then probably you want to build some predictive model but as you see I'm guessing.

"" hence,if i use (for example) 5 min length segments, then my feature length will exceed and i will get poor results. and if i get HRV and BPV from this signals to have analysis on them, then i will lose the morphology of signal and i will not be able to diagnose more arrhythmia. what length do you suggest me for this reason?"

I do not understand what you mean by a feature length and why it should exceed? From the ECG signal we can get loads of information, there is different information in the analysis of the structure or morphology of ECG, like P wave, PQ, QRS, QT, ST and T wave. By analysis of the time series of RR intervals (or PP intervals, however QRS complexes are more discrete and easier to identify then P waves and thus PP intervals), QT intervals or T wave amplitude you can get variability or complexity of these signals, or in case of T wave you can also get T wave alternans. You will easily find references on these topics. I do not want to advertise what we do but one of the methods which we have invented for the description of heart rate asymmetry (i.e. physiological phenomenon related to unequal contribution of heart rate decelerations and accelerations to heart rate variability, structure and complexity), which describes the structure of heart rate (based on the monotonic runs analysis) has shown predictive value in post-infarction patients (you can find it on my profile or:

1. Piskorski J, Guzik P. The structure of heart rate asymmetry: deceleration and acceleration runs. Physiol Meas. 2011 Aug;32(8):1011-23.

2. Guzik P, Piskorski J, Barthel P, Bauer A, Müller A, Junk N, Ulm K, Malik M, Schmidt G. Heart rate deceleration runs for postinfarction risk prediction. J Electrocardiol. 2012 Jan-Feb;45(1):70-6.).

Recently there has been a short paper in Computing in Cardiology (http://cinc.mit.edu/archives/2012/pdf/0633.pdf) on the use of our method in ECG monitoring in preterm infants to help diagnosis of sepsis. We have also some data (manuscripts in preparation) which show that this extremely simple method have a lot of predicting potential.

If you can record blood pressure on a beat to beat manner, together with ECG signal then you can have not only blood pressure variability (or other derivative methods as in RR intervals, i.e. complexity, structure or asymmetry) but also you can use different methods for baroreflex function. Still with the pressure waveform you can apply different methods (e.g. Modelflow) for estimation a number of hemodynamic parameters like stroke volume, vascular resistance, cardiac output etc, and some other methods for arterial stiffness. I would not worry about too much of information, if you can manage them automatically then this is not your problem but the software and your PC (or servers). You can get much more information. I recommend you to check on the number of papers by my friend Prof. Georg Schmidt - they have collected thousands of Holter ECG recordings and > 1000 recordings of 30-minute duration with ECG, continuous BP and respiratory curves and they do a lot of signal processing and completely new discoveries. One of the most important discoveries of his team is heart rate turbulence (www.h-r-t.org), i.e. the dynamic change of sinus rhythm after a premature ventricular beat. Please be, however, aware, that the ventricular beats are not taken into the analysis, they are used as triggers of some changes in the normal heart rhythm. Please also note that HRT is patented so for research purposes you should ask him for agreement (his email is on the website and he should agree with no problems).

"the other problem that i have is that:

i use MGH/MF database from physionet. the type of arrhythmia for a signal is given for each beats of it. in other words, each beats of a signal are labelled with the type of arrhythmia that it has. if i use 5 min length segments, there will be some types of arrhythmia in a segment. then, how must i determine the type of arrhythmia for a segment?"

again I may have dozens of questions. First, we have tested some of the recordings from PhysioNet and the quality is far from optimal for some of the recordings. But this is my private opinion, I want you to be just cautious with what you take from there and what you get. Second, this is good that every beat or RR interval is labelled with a proper name for the type of beat, I cannot imagine any reliable analysis of cardiac time series without such annotations. Again, for HRV you should analyse only normal RR intervals or if you want to analyse the variability of some tachycardias, e.g. VT, then please use only RR intervals from such tachycardia. But please, please, please, never mix them because you will kill physiology. One of the most common errors which I have witnessed so many times and which is done by numbers of engineers, physicists and mathematicians (which I admire) is that they take all RR intervals irrespective of their origin as they say that all the beats are in the system and all should count. they use mathematical methods which already exist and were invented for completely different purposes, e.g. astronomy, meteorology, economy name whatever you want, and try to apply them in heart rate. Most frequently they do not respect most of the assumptions for these methods, the good example is the use of Fast Fourier Transform for spectral analysis of heart rate, then they get some results and try to build different theories in what is there or not. There is more wishful thinking with spectral analysis of HRV than proofs. Heart rate is not stationary, even if there are some oscillation in the cardiovascular system they are not symmetric like the classic sinus wave, they start and end rapidly, have different duration of period, there is a lot of other non-oscillatory influences etc.

Coming back to another part of your question about determining the type of arrhythmia - it should be done by the system, if you;re not an expert in arrhythmia analysis do not do this alone by yourself. In Holter analysis we use numbers showing how many different types of arrhythmias are during 24-hour or per 1 hour. I've never seen any type of report showing it for 5 minute segments.

Best regards,

Przemyslaw Guzik