Noise in ECG data

hi, I am doing project of ECG acquiring on C5505. I got results successfully on the PC Application. My ECG results are good and their shape matches those available on the real ECG machines.

   But the ECGs are still suffering from some humming noise. So when I plot the data on Matlab, I got all forms of zig-zags.

   I want to know is it possible that fft of the signal is taken and noise is removed, that is those frequency components can be removed from them in Matlab??
   I got the fft spectrum, but the plot does not show the frequency information related to samples...How I will extract the frequency info??

         Thanks.

[Sean de Wolski]

Can you show us what you've done and supply a small set of sample data?

Answer 1

[Rick Rosson]

Hi Talha,

1. Do you know the sampling rate of the signal?
2. Also, what is the size of the time-domain representation (how many rows and columns)?
3. Can you please show us your MATLAB code?

Thanks.

Rick

[Talha]

Answers to your questions:

1. 500 HZ

2. N x 12(N depends on duration of acquiring of ECG data)

3. I was able to get the frequency spectrum and now I want to remove the noise between 45-55 Hz and then 90-115 Hz. I have applied low pass and band stop filters and All that I have done has made the waveforms with more and more ripples.

The code is written is below:

Fs = 500;

xdft = fft(d);

% below works for even length x

psdest = 1/(length(d)*500).*abs(xdft(1:length(d)/2+1)).^2;

psdest(2:end-1) = 2*psdest(2:end-1);

df = 500/length(d);

freq = 0:df:500/2;

plot(10*log10(psdest),freq);

Answer 2

[Rick Rosson]

Hi Talha,

1. Why are you using low-pass and band stop filters? Why not use two band-stop filters?
2. Is the noise you are trying to remove related to 50 Hz electrical power? Is it the fundamental (50 Hz) and the second harmonic (100 Hz) of the electrical power supply? If yes, have you considered using two notch filters, one at 50 Hz and the other at 100 Hz?
3. How did you design the low-pass and band-stop filters? Did you use the FDATool, MATLAB filter design function calls, or some other method? If you designed by function calls, can you please post your filter design code?
4. Can you please post the filter coefficients for both filters?
5. How did you apply the two filters? Can you please post your filter application code?

Thanks!

Rick

Answer 3

[Rick Rosson]

Hi Talha,

Here is some improved code:

   %%Computations:

   x = d;
   [N,P] = size(x);

   Fs = 500;

   dt = 1/Fs;
   t = dt*(0:N-1)';

   dF = Fs/N;
   f = (-Fs/2:dF:Fs/2-dF)';

   X = fftshift(fft(x))/N;
   Xpsd = 20*log10(abs(X));

   %%Plotting:

   for k = 1:P

       figure;

       subplot(2,1,1);
       plot(t,x(:,k));

       subplot(2,1,2);
       plot(f,Xpsd(:,k));

       title(['Time Domain and PSD for Channel #' num2str(k)]);

   end

   %%Filtering:

      % Specify:
      spec(1) = fdesign ...
      spec(2) = fdesign ...

      % Design:
      filt(1) = design(...); 
      filt(2) = design(...);

      % Filter:
      y = filter(filt(1),x);
      z = filter(filt(2),y);

[Talha]

hi Rick,

Again the answers to your questions:

1. Does not matter, Passing through the filters cause ripples.

2. Two frequencies are not related and I am in a process of implementing a two notch filters but in the DSP stage not after acquiring the data due to the ripples.

3. I used a .m file I made to have the notch filters' coefficients, when I use the fdatool the coefficients were in decimals and rounding them caused distortion.

5. Yes, I have applied the two filters. The code is:

dt=fdesign.bandstop('Fp1,Fst1,Fst2,Fp2,Ap1,Ast,Ap2',40,45,55,56,0.5,80,0.5,500);

de=fdesign.bandstop('Fp1,Fst1,Fst2,Fp2,Ap1,Ast,Ap2',85,86,115,116,0.5,80,0.5,500);

Hd = design(dt,'butter');

Hd1= design(de,'butter');

%view filter magnitude response

fvtool(Hd);

output = filter(Hd,d(:,1));

output1 = filter(Hd1,output);

(I used the code you sent, and yes it has made it much more clear. Thanks for that.)

Answer 4

[Rick Rosson]

Everything that you are doing looks correct, as far as I can tell. When you say that "Passing through the filters causes ripples", what exactly do you mean? Are you seeing "ripples" in the time domain signal, or in the frequency domain representation? Can you please post one or more screen shots showing the signal before and after passing through the two filters?

[Talha]

By ripples I mean waveforms start to show more peaks than before.

and I did not find any option to post screen shots here.

By the way can I use Moving Average Filter here?? Or Adaptive Filters??

[Rick Rosson]

1. Are you seeing the ripples in the time domain or the frequency domain?

2. People have been posting screen shots to other web locations, and then providing links to the images.

3. Yes, you can use moving average or adaptive filters, but I do not know if they will help or not.

[Talha]

1. Its in time domain

2. Links are here. I have passed through the bandstop filters

http://tinypic.com/view.php?pic=264t0yh&s=7 (Before)

http://tinypic.com/view.php?pic=21f08wn&s=7 (After)

3. How to smooth out or remove baseline wandering?? In the original signals, which looks much cleaner now,Due to bandstop filter I implemented in the DSP stage.

[Rick Rosson]

Are you sure that the sampling rate Fs really is 500 samples per second? The only explanation I can think of is that the sampling rate you are using to design the filters is different from the sampling rate of the source signal.

[Talha]

Its given in the DSP documents that the sampling rate is 500 Hz. I want to mention here that the Sampling rate of PC application which is used to record the signals form DSP to PC is 250. is it relevant here??

[Rick Rosson]

Yes, that could very much affect the results. Please try designing your filters for Fs of 250 samples per second, and then reapplying them to the signal. That might do the trick.

[Talha]

Unfortunately, results were not as expected. Can I use wavelet transform here, cause ultimately I have to extract different locations of peaks. The only problem I am facing is the PQRST waves have random locations i.e. I can't follow simple zero crossings algorithm or findpeaks, on the waveforms.

Answer 5

[Bjorn Gustavsson]

Talha, what is the noise you're after in your original image? Is it the point-to-point variation? If so you should be fine with a very simple low-pass filter, something you'd need to use only matlab's filtfilt or filter functions. Something like this:

I_filtered = filtfilt([.5,1,.5]/2,1,I_original);

HTH

[Talha]

The noise riding on the signal is due to environment and its over all distributed. The waveforms are also suffering from baseline wandering.

Thank you Bjorn,I tried the code you sent me. It did smooth the waveform further. Now I state again, how can I extract different waves locations? Smoothing of data is been done to satisfaction level.

[Bjorn Gustavsson]

I don't know enough about ECG-signals to know what is "true" ECG signal and what constitutes base-line variations. But considering that you only show us ~2 periods it might be completely impossible to determine. I suggest you could do something similar to conditional integration - but here rather "conditional resampling" (I just made that one up.):

Identify the peak of the spikes (the ones at ~2690 and 2950 and so on) then resample/interpolate something like this:

iPeaks = find-local-peaks(ECG);

for i1 = 1:(length(iPeaks)-1),

iCurr = iPeaks(i1):iPeaks(i1+1);

ECGresampled(i1,:) = interp1(t(iCurr),ECG(iCurr),linspace(t(iCurr(1)),t(iCurr(end)),nS);

end

That should give you a 2-D matrix with the individual ECG-periods along the rows and might give you some view of variations in base-line variations that are way slower than one ECG-period. How to determine what is ECG-signal and what is base-line variations in a single period I would not like to tell! Maybe this flutter is a heart condition but not that one over there...

Answer 6

[haneen]

hi , i am haneen my broject (removes the baseline wander in ecg signal by using wavelet transform) so, i need code in matlab to do this, thanks