Preprocessing of ECGs for classification of ventricular arrhythmia

I have coded preprocessing mentioned in this research paper . Is my preprocessing code correct? Am I going in right direction?

Please bother to run my code on random recordings of NSR, CU Ventricular Tacharrhythmia, MIT BIH SupraVentricular Arrhythmia and MIT BIH Malignant Ventricular Ectopy. These recording are available to download from PhysioNet ECG Data-set. I am referring external link because data-set is more than 5MB.

If code is correct and I am going in right direction, how should I proceed?

Should I concatenate 500 milli sec windows of signals and start Feature Extraction?

My pre-processioning code is named as code5.m By default sampling frequency of NSR and Supraventricular ECGs is 128 Hz and for others it's 250 Hz.

[Star Strider]

I do not access files on sites external to MathWorks (for MATLAB Answers posts). Please upload all of them here using the ‘paperclip’ icon, and be sure to complete both the ‘Choose file’ and ‘Attach file’ steps.

[Explorer]

@ Star Strider

I have attached research paper but PhysioNet ECG Data-set size is more than 5 MB. How can I upload files that are greater than 5 MB?

Answer 1

[Star Strider]

I would not use the Symbolic Math Toolbox at all for signal processing, unless it involved a one-off derivation that I would then transform to a numeric calculation.

Here:

[q, r] = quorem(sym(nu), sym(ss)) ;% quotient

you apparently want to divide the length of the signal by 4*Fs and get the quotient and remainder. It’s much faster to do this numerically:

q = fix(nu/ss);
r = rem(nu,ss);

Here:

for k= 1:size(splitedParts,2);
    y_nsr16265(:,k)= filtfilt(b,a,splitedParts (:,k) ) ; 
end

I would filter the entire signal first, then split it. That reduces the effect of end transients that can be part of any filtering procedure. I would also use the second-order-section (‘sos,g’) implementation for stability, rather than using the transfer function coefficients.

Beyond that, while I do not agree with the preprocessing procedure in the paper (and the reason I advised you to do it differently), your code looks acceptable. You need to test it to see if it produces the results you want, and change it if it does not.

[Explorer]

My above code is just displaying one wave after pre-processing. I want to view and compare ECG before and after pre-processing.

I also want to concatenate 500 milli second windows in order to process further. Each window contains 128 samples. Windows are saved win_for variable. How should I do it?

[Star Strider]

It’s difficult to follow what you’re doing with the ‘splittedParts’ matrix, but apparently you’re dividing your signal into small windows of about 4 beats (or 4 seconds) per window, although apparently you later changed it. Depending on the heart rate, a 500 ms window could be one complete beat (at 120 bpm).

I would start by not windowing it before filtering it. I would just do the preprocessing on the complete EKG record, then plot it as a function of time. You don’t have to do any concatenation then, and you can do the windowing and any other processing you need to do later.

[Explorer]

splittedparts is just 4 second segment of first trace. This is what I have done:

Can you please now tell the way to view and compare the preprocessed signal with raw ECG?

By the way, I have done filtering of 4 sec ECG because I am following paper.

[Star Strider]

First, you’re attempting to replicate the results of a conference paper that is not refereed (only read to eliminate obvious fraud) and does not have space to completely describe the methods they used. I would look up a subsequent full journal article by the same authors on the same subject to see if it eliminates the ambiguities in the conference paper.

Second, the preprocessed EKG is the raw EKG, at least as I understand it. Splitting your data into four second segments is not a problem if you do all the filtering and other preprocessing on the full record first.

[Explorer]

Okay. So I have to look another paper for feature extraction of ECG dataset.

What is difference between full paper and conference paper? And where am I supposed to find it?

Is the attached paper full paper or conference paper?f

[Star Strider]

The difference is usually in the amount of detail, particularly with respect to the materials and methods. A conference paper is intended as a summary, with emphasis on the results. Since the paper discusses EKG processing, it would probably be easiest for you to search the Internet using the authors’ names. You can also probably also find it in PubMed. Read the PubMed search instructions, since it’s a database of all published papers of medical interest, and has some search options that can make your search easier.

The attached paper looks like a refereed paper, so it’s not a conference paper. It also seems to go into some detail with respect to the methods. You likely need to do some reading on wavelets first, before you implement its methods. There are a number of good references, of which the documentation for the MATLAB Wavelet Toolbox is likely the best place to begin, especially if you have it to work with.

[Explorer]

Thank you Star Strider :)

[Star Strider]

As always, my pleasure!