Is the attached paper okay for Feature Extraction of ECG dataset?

If it's okay, which wave should I consider in filtered plot of cu01m as current beat?

Plot of filtered cu01m.mat

[Star Strider]

It looks like it could work. It uses a radial basis function neural network, and the diagram seems to use the radbas function. (In fact, it looks like it was taken from the Radial Basis Neural Network MATLAB documentation, or the MATLAB diagram using the network they created from it.) You will need to use the Neural Network Toolbox to do that analysis most efficiently. Programming complicated backpropagation neural nets by hand is not fun, even in MATLAB.

[Explorer]

Your comment on paper is about feature matching part. What do you think about its feature extraction part?

[Star Strider]

They don’t go into significant detail on how they developed their (4x14) matrix, among other things. (You will also need the Wavelet Toolbox to use the db6 wavelet decomposition if you want to reproduce their paper.) They’re not doing very sophisticated rate and rhythm classification — normal sinus rhythm (NSR), ventricular tachycardia (VT), and ventricular fibrillation (VF) are significantly different enough to differentiate with much simpler methods. Besides, the QRS complex does not exist in classical VT or VF, so I have no idea how they would go about detecting it. I would be tempted to use the RBF network and a classifier on the filtered signal without all the preprocessing if I only needed to differentiate NSR, VT, and VF.

[Explorer]

Okay thanks. Is it possible to apply classifier without extracting features if I want to differentiate NSR, VT and VF? If so, how to do that?

[Star Strider]

Well, you have to extract some features, but the RBF network should be able to do that for you as the input to the classifier. (I actually did EKG classification with a RBF NN classifier once, but so long ago I don’t remember the details.)

[Explorer]

Can you please refer me the paper you followed for EKG classification? I haven't studied Neural Networks in my graduation.

[Star Strider]

I did not follow a paper. I used Haykin’s book on neural networks and programmed it in FORTRAN myself. (This was before the Neural Network Toolbox first appeared.)

I would search your university library for Haykin’s book and others on RBF networks. If you have the Neural Network Toolbox, start with its documentation (for your MATLAB release) for a comprehensive introduction, then go on from there.

[Explorer]

Thanks.

[Star Strider]

My pleasure.

EKG analysis is not trivial. I have not ever seen a paper that could do reliable EKG analysis with any degree of reliability or reproducability. Wavelets are likely the best initial analysis tool. Processing the wavelet features to correctly analyse the EKG remains a significant problem, even for the relatively simple task of doing heart-rate variability (HRV) studies, which was my interest at the time. (For that, it’s necessary to detect a normal P-wave preceding the QRS complex in order for that beat to be considered in the HRV analysis. That’s much more challenging than it at first appears.)

[Explorer]

What do you think about the method in attached paper? Should I start to code up this?

[Star Strider]

That one seems a quite plausible approach, with enough detail and results to convince me that it would work. The MATLAB Statistics Toolbox has the classify and fitcdiscr functions that can do the quadratic discriminant analysis once you have calculated the features.

What book is it from? I’ll look into adding it to my library.

[Explorer]

Thank you for the comments.

Book Name: Advances in Cardiac Signal Processing

Author: Acharya, U. Rajendra

[Star Strider]

My pleasure!

Thank you for the book! I’ll take a look at it later.