How to Classify ECG Signals Using Continuous Wavelet Transform and AlexNet
ECG signals represent the electrical activity of the heart observed from a strategic point of the human body, characterized by Quasi-periodic voltage. AlexNet is a convolutional neural network that has eight different layers. This method is commonly used in the science sector and for image studies. <!--more--> Transfer learning (TL) is a research problem in machine learning that focuses on storing knowledge gained while solving one problem and applying it to a different but related problem.
For example, knowledge gained while learning to recognize cars could be used when trying to recognize trucks.
This tutorial will classify the ECG signals using pre-trained deep CNN (AlexNet) via transfer learning in Matlab. For this purpose, we utilized the strength of a container wavelet transforms to represent the one dimension ECG signals as images. It makes it possible to be used as an input in the AlexNet.
Prerequisites
To follow along with this tutorial, you'll need to have:
Types of ECG signal for classification
Primarily, we are taking three types of ECG signals:
- ARR: Arrhythmias
- CHF: Congestive heart failure, and
- NSR: Normal Sinus Rhythm
They are shown below:
The main objective here is to train a CNN to distinguish between ARR, CHF, and NSR. These three signals (162 ecg recording) are obtained from ecg signal databases from the physionet. The databases include:
- MIT-BIH Arrhythmia database (96recordings)[ARR signals]
- MIT-BIH Normal Sinus Rhythm Database (30recordings)[NSR signals]
- BIDMC Congestive Heart failure Database (36recordings)[CHF signals]
The first step is downloading the data from the github repository. Next, click on the code tab and select download the zip to download the database.
In this .zip file, we only need the ECGData.mat. It means that we unzip this file to get it. Afterwards, we place it in Matlab's current directory. Then, executing the load function with the database as the argument to download it to Matlab:
load(ECGData.mat); %load ecg
To get the signal in your data, execute the code below:
data=ECGData.Data;
Also, if you want to store the labels, we use the ECGData.Labels. This command reads all the signals and stores all the labels available here:
labels=ECGData.Labels; %getting labels
Now, this signal is a matrix of size 162 x 65536. So it means that it carries 162 ECG signals of size 65536 samples each.
ECG signal database preparation
For our problem, we pre-process the database. Each recording is 65536 samples. It means we will break the signals into small signals of length 500 samples. This is done to increase the size of the database to make it appropriate to train CNN.
For this purpose:
- We take 30 recordings of each type to have equal distribution.
- Each record is broken down to 10 pieces, each of length 500 samples.
- Out of the 900 recordings, we are using 750 recordings for training and 150 for testing.
ECG signal to image conversion using cwt
Now, we want to convert all the 1-D signals into images cwt to be used as inputs to CNN for classification.
The deep CNN that we are using is Alex, and it takes images as the input. For this purpose, we take the CWT of each one dimension signal, and all the coefficients are arranged to form a cwt scalogram.
Each scalogram is represented in the color map of the type jet of 128 colors. Scalograms are then converted into images and saved into folders corresponding to each class since we create folders for each signal type. Each image is of the size 227x227 and in the RGB color format.
For cwt, the wavelet used is analytic Morlet(amor). These are wavelets with one-sided spectra and are complex-valued in the time domain. These wavelets are suitable for obtaining a time-frequency analysis using the cwt. The wavelet has equal variance in time and frequency.
Transfer learning via AlexNet
Transfer learning is called fine-tuning a pre-trained CNN to perform classification on a new collection of images. Transfer learning is quick and easier rather than training a CNN from scratch, which requires millions of inputs, lots of training time, and high-speed, efficient hardware.
For ecg signal classification, we use a pre-trained deep CNN. AlexNet has been trained on over one million images and can classify images into 1000 objects categories.
Matlab code for database creation (image signal to scalogram image conversion)
We first load the ecg signal using the load command and the labels giving the data a .Label extension. This extension extracts the data labels:
% program to create CWT image database from ecg signal
load('ECGData.mat'); %loading ECG database
data = ECGData.Data; %getting database
labels = ECGData.Labels; %getting labels
Now we will take 30 recordings of each signal:
ARR = data(1:30,:); %Taken first 30 recordings
CHF = data(97:126,:);
NSR = data(127:156,:);
signallength = 500;
The signal length is specified in the database and is equal for all the signals.
Define the cwt filter cwtfilterbank, used to find the cwt coefficients. This function takes in the signal length, wavelet type amor, and the wavelet bandpass filter VoicesPerOctave:
%Defining filters for CWT with Amor wavelet and 12 filtering per octave
fb = cwtfilterbank('SignalLength', signal length, 'Wavelet', 'amor', 'VoicesPerOctave', 12);
Make a folder for each label in the current directory. To make a folder using matlab command window, we use the mkdir function. This function creates the folder automatically. The function uses the folder's name as the argument.
To create a subfolder inside the main folder, you re-run the command with the mainfolder\subfolder as shown below:
mkdir('ecgdataset'); %main folder
mkdir('ecgdataset\arr'); %sub folder
mkdir('ecgdataset\chf');
mkdir('ecgdataset\nsr');
ecgtype = {'ARR', 'CHF', 'NSR'};
ecgtype is a cell array that carries the names of the signal by string.
Now we are converting these ecg signals to images. This is done using the ecg2cwtscg function:
%Function to convert ECG to image
ecg2cwtscg(ARR, fb, ecgtype{1});
ecg2cwtscg(CHF, fb, ecgtype{2});
ecg2cwtscg(NSR, fb, ecgtype{3});
This function takes the signal type and the filterbank fb as inputs. It takes the signal, gets its coefficient using fb, and then converts it to an image.
ecgtype are the string types. ecgtype{1} represents ARR, ecgtype{2} represents CHF, and ecgtype{3} represents NSR.
It is how we create our databases. We have the main folder dataset created with the subfolder arr, nsr, and chf as shown below when we execute this program:
mainfolder
subfolders
Inside the subfolders
The ecg2cwtscg is not a matlab function. It is a custom function that we created for the conversion purpose. The function is as shown:
function ecg2cwtscg(ecgdata, cwtfb, ecgtype)
nos = 10; %number of signals
no1 = 500; %signal length
colormap = jet(128);
if ecgtype == 'ARR'
folderpath = strcat('mainpathfolder/arr/');
findx = 0;
for i = 1:30
indx = 0;
for k = 1:nos
ecgsignal = ecgdata(i, indx+1: indx+no1);
cfs = abs(cwtfb.wt(ecgsignal));
im = ind2rgb(im2uint8(rescale(cfs)), colormap);
filenameindex = findx + k;
filename = strcat(folderpath, sprintf('%d.jpg', filenameindex));
imwrite(imresize(im, [227 227]), filename);
indx = indx + no1;
end
findx = findx + nos;
end
elseif ecgtype == 'CHF'
folderpath = strcat('mainfolderpath/chf/');
findx = 0;
for i = 1:30
indx = 0;
for k = 1:nos
ecgsignal = ecgdata(i, indx+1: indx+no1);
cfs = abs(cwtfb.wt(ecgsignal));
im = ind2rgb(im2uint8(rescale(cfs)), colormap);
filenameindex = findx + k;
filename = strcat(folderpath, sprintf('%d.jpg', filenameindex));
imwrite(imresize(im, [227 227]), filename);
indx = indx + no1;
end
findx = findx + nos;
end
else ecgtype == 'NSR'
folderpath = strcat('mainfolderpath/nsr/');
findx = 0;
for i = 1:30
indx = 0;
for k = 1:nos
ecgsignal = ecgdata(i, indx+1: indx+no1);
cfs = abs(cwtfb.wt(ecgsignal));
im = ind2rgb(im2uint8(rescale(cfs)), colormap);
filenameindex = findx + k;
filename = strcat(folderpath, sprintf('%d.jpg', filenameindex));
imwrite(imresize(im, [227 227]), filename);
indx = indx + no1;
end
findx = findx + nos;
end
end
end
The number of signal nos is 10. Since we have three signals, 10 by 3 gives you the number of recordings we said to be 30. The colormap we are using is jet(128), and the signal length remains the same.
There is a section of this function that is repeated. The else and the elseif statements are repeated with only the signal type changed.
Lets look at a section of that:
if ecgtype == 'ARR'
folderpath = strcat('mainfolderpath/arr/');
findx = 0;
for i = 1:30
indx = 0;
for k = 1:nos
ecgsignal = ecgdata(i, indx+1: indx+no1); %extracting signals
cfs = abs(cwtfb.wt(ecgsignal));
im = ind2rgb(im2uint8(rescale(cfs)), colormap);
filenameindex = findx + k;
filename = strcat(folderpath, sprintf('%d.jpg', filenameindex));
imwrite(imresize(im, [227 227]), filename);
indx = indx + no1;
end
findx = findx + nos;
end
This program means that if the signal is of ARR type, create a folder arr and store the images there. The for loop defines the number of iterations which is 30 for the first 30 recordings.
Inside the for loop, we extract the signals using ecgdata(i, indx+1: indx+no1). We then find the wavelet coefficients with the filter fb.
The abs(cwtfb.wt(ecgsignal)) command gives the wavelet coefficients stored in the variable cfs. The abs means taking the absolute value since the coefficients are complex:
im = ind2rgb(im2uint8(rescale(cfs)), colormap);
Convert the coefficients to images using ind2rgb, but first rescale these coefficients using rescale function. im2uint8 converts the signals to the uint8 type so that it becomes an image. The colormap defines the color since the images are in rgb form:
filenameindex = findx + k;
filename = strcat(folderpath, sprintf('%d.jpg', filenameindex));
imwrite(imresize(im, [227 227]), filename);
indx = indx + no1;
The image is saved using the imwrite function with the proper name after resizing it to 227 x 227 pixels. The images are saved in the indexed form, i.e. 1.jpg, 2.jpg etc. using the indx function.
This procedure remains the same for other signals. The only thing that we change is the signal type, hence the long function above.
Matlab code for training AlexNet (transfer learning)
For this, we need to download a toolbox (deep learning toolbox for AlexNet network). Click on the adds on and select get add on on the home page to get the toolbox. When you click this, you get a window requesting your Mathworks account. Log in and search for deep learning toolbox for alexnet network and download it.
Afterwards, start the training process. The first step is reading the images in the database folder (images that we converted):
% Training and validation using AlexNet
DatasetPath = 'folderpath/ecgdataset';
% reading images from the image database folder
images = imageDatastore(DatasetPath, 'IncludeSubfolders', true, 'LabelSource', 'foldernames');
The DatasetPath stores the path of our datasets. imageDatastore is used to read multiple images from a database. It includes images in the folder and IncludeSubfolders' subfolder.
We need to distribute the images into training and testing. We will take 250 images from each folder, and we will have a total of 750 images for training. Since each subfolder has 300 images, the number of images used for testing will be 50 from each folder giving a total of 150 images:
% Distributing images in the set of training and testing
numTrainFiles = 250;
[TrainImages, TestImages] = splitEachLabel(images, numTrainFiles, 'randomsize');
All the training images are stored in the TrainImages variable, and test images are stored in the TestImages. The splitting is done by the splitEachLabel function, which is randomized.
Load the pre-trained network alexnet. To load this network, we execute the command alexnet. This command is not functional until you download the toolbox:
net = alexnet; %importing pre-trained alexnet(requires support package)
layersTransfer = net. Layers(1:end-3); %preserving all layers except last three layers
net.Layers loads all the layers of the network. We preserve all the layers in the variable layersTransfer and modify the last three, thus 1:end-3.
Define the number of output classes:
numClasses = 3; %Number of output classes: ARR, CHF, NSR
Now, lets define the three layers that we are modifying, that is fullyConnectedLayer, softmaxLayer and the classificationLayer:
%defining layers of alexnet
layers = [layersTransfer
fullyConnectedLayer(numClasses, 'WeightLearnRateFactor', 20, 'BiasLearnRateFactor', 20)
softmaxLayer
classificationLayer];
For for more information on the layers, check here.
Now we want to define the training options. It gives the direction in which the training should be carried. We define the solver, sgdm, the learning rate, 1e-4, validation frequency training progress in plot form:
%training options
options = trainingOptions('sgdm', 'MiniBatchSize', 20, 'MaxEpochs', 8, ...
'InitialLearnRate', 1e-4, 'Shuffle', 'every-epoch', 'ValidationData', TestImages, ...
'ValidationFrequency', 10, 'Verbose', false, 'Plots', 'training-progress');
After defining all the layers and the transfer options, we start the training. To begin, we use the command trainNetwork function. The function takes the images and trains them according to the options. It means the arguments of this function are the training images, layers, and options:
% training the Alexnet
netTransfer = trainNetwork(TrainImages, layers, options);
Now let us classify the images using the classify command:
%Classifying images
YPred = classify(netTransfer, TestImages);
YValidation = TestImages.Labels;
accuracy = sum(YPred == YValidation)/numel(YValidation);
The classify command takes the trained network and the test images and stores them in YPred variable for the classified. They are the predicted classifications.
Yvalidation will store the actual validation or labels of the images, thus TestImages.Labels. When comparing the predicted classification YPred and the actual validation YValidation, we get the accuracy.
For proper visualization, plot the confusion matrix. This plot shows the number of correctly classified data and how many are misclassified:
%plotting Confusion Matrix
plotconfusion(YValidation, YPred)
Let's now execute our program.
When we run the program, we see the training progress as shown below:
We have the two graphs here. The upper one shows the accuracy, while the lower one shows the loss. As the accuracy increases, the losses decreases. The number of epochs is 8. Also, we have the plots for training and that for validation.
The plot for the final training is as shown below:
In the final training, we see that the accuracy achieved is 96.67%, which is good.
Let's look at the confusion matrix:
The confusion matrix shows the relation between the output class and the target class regarding classification and misclassifications. For example, in arr, we see that out of 50 testing images, 48 are correctly classified, 2 are misclassified as chf. On the other hand, for chf, 49 images are correctly classified, and one is misclassified as nsr.
Conclusion
This is how we can train our deep CNN, a pretrained network, via transfer learning. It is also how we can use the cwt to represent a dimensional signal in the form of two-dimensional images.
The funny part is that the output or the accuracy of this method depends on the signal length and the colormap you choose. Although we use the jet(128), if you choose some other colormap, you receive a different result.
I hope you found this tutorial helpful.
Happy Coding!
Peer Review Contributions by: Monica Masae
