Hi everyone,
I’m actually writing Matlab code to perform simple frequency analysis for HRV.
I have a 5-min recording of hear rate from PPG (rest condition on a men, age 30). It is pre-processed to extract the RR tachogram, linearly interpolated and resampled at 8-Hz. So my tachogram is a 2400 points vector, and each point represent the instantaneous RR interval in ms (values varied around 1200-ms).
Now I’m performing a frequency analysis from FFT. The problem is that, when I compared the power spectrum from my analysis to others from the literature, it seems that my power values are way too large (for exemple, values up to 10 x 10^5 ms^2 for low frequency band 0.04-0.15 Hz, see the chart inclosed).
Can anybody have an idea about where could it come from ? The thing is that, when I performed a log-transform, it seems better but the unit are in dB instead of ms^2, and the papers from which I compared my data show their charts in ms^2 (see for example Singh et al. 2004, enclosed). Do I miss something about ms^2 unit / dB unit / any step in my processing ?
Thank a lot,
Regards,
Yannick
http://www.ncbi.nlm.nih.gov/pubmed/15513744
I suggest to use Power Spectral density Estimation instead of FFT.
In MatLab you can use Pwelch function, read the following document:
http://it.mathworks.com/help/signal/ref/pwelch.html
PWelch in MATLAB is in fact a periodogram-based method which also uses FFT, but applies smoothing by averaging and windowing.
Dear Yannick,
Can you display the time domain signal before and after the interpolation.
From my point of view, the fft spectrum must be normalized such that the power content of the waveform and the spectrum of the fft must be the same. That is P= Integral v^2(t) dt/ T = intergral V^2(f) df. This will scale down the V^2 (f) to normal values. Where t is the integration period.
please tell us after normalization.
wish you success
Hi everybody,
Thank a lot for your three contributions.
Dr Zekry, here is enclosed two charts depicting the "time domain" signal before and after the interpolation (don't be afraid for the 3 or 4 artefacts that appears visually, I didn't perform a manual correction on my signal, this are just pilots data to develop the matlab routine).
I have already performed your suggested normalization, but thank a lot for having suggested it.
I was thinking, maybe my scale problem could come from an initial issue of scale conversion : I realized that when I plot the FFT from my power transformation, it express power in ms^2 function of frequency in Hz --> the power values then depend of whether the initial data for RR-interval are expressed in ms or in s, when compare to frequency in Hz (that is, expressed in 'events per second'). So I have scaled down my time vector to ms for the FFT transform, and it looks like power values are coherent with the literature.
I hope this is the great way to fix the issue. If someone have a look at this discussion and identify a mistake, please tell us.
Thank a lot,
Dear Yannick,
After inspecting your curves in time domain and in frequency domain, it seems me theta the values appearing in the spectral power density agrees well with the sample values in time domain. It is so if the signal amplitudes in the order of 1000 ms the square of the amplitude will be 1000000 ms^2.
My question is why do use ms to represent the amplitude of the signal. Why not using seconds. If you use seconds, then your values will be normalized to small values. That is use seconds the order of the magnitude will be 1 second and its square will be 1 second squared.
The other question why not expressing the signal amplitudes in volt?
wish you success
I recommend you to use STFT instead of FFT, in MATLAB you can use the "spectrogram" function, I used to have same problem and I solved it by using Spectrogram
Thank a lot Dr Zekry and Dr Bazgir.
For your first question Dr Zekry, I use ms just because RR-intervals range are found between 350-ms and 1500-ms (see method in Jiang et al. 2015). So I preferred to express it in ms in the time domain, but as you underlined and as I suggested before, this might lead to huge scale when in the frequency domain.
For the second question : the signal is not expressed in volt simply because this is not volt ! At first, when you acquire your PPG signal, each sample are in Volt --> the signal looks like a sort of sinusoidal curve, in which every main peak represent a heart beat event --> you compute time interval between peaks, so the result is expressed in ms ! (or second, juste depending on the scale you want to).
Many regards for your help and suggestions,
Yannick
- Badges
- Science topic