-This paper explores the possibility of application of machine learning methods for detecting the signs of cardiovascular diseases. The ECG Heartbeat Categorization Dataset that contains electrocardiogram data of various heart rhythms is taken for the study. The paper describes classification of five various arrhythmia types by the use of machine learning methods. The classification has used such methods as the random forest algorithm, the decision tree and the convolutional neural network. The results of the study show the highest accuracy of the neural network that equals to 0.9347.
Cardiovascular diseases (CVD) are one of the most
significant mortality causes in the modern world. The
forecasting of cardiovascular diseases is a major concern in
the area of analyzing of clinical data [
Neural network technologies are aimed at resolving
various complicated tasks that include a number of issues in
medicine.First of all, the use of neural networks relates to the
fact that a researcher receives a large amount of different
factual materials that do not already have any mathematical
models. Application of the machine learning can improve the
accuracy of medical visualization and diagnostics of
pathologies. With the use of various algorithms, machine
learning methods are suited to address the tasks with unknown
mechanisms of situation-based developments and
dependencies between inputs and outputs [
The objective of the paper is to analyze the application of machine learning methods for detecting the electrocardiograms with indicators of some abnormal heart rhythms. The scientific novelty of the research lies with the use of the convolutional neural network for the analysis of constantly incoming data and the testing of machine learning methods to compare the accuracy of algorithms. The experiment shows that deep neural networks are better suited for solving the problem of classification of heart rhythms.
In order to deal with the issues arising from the manual
analysis of electrocardiogram signals, many scientists
considered the application of machine learning methods for
accurate detection of abnormalities in the signal. For example,
the study [
Our study consists of two stages. Data pre-processing at the first stage solves the problem of unbalanced classes. We have processed each electrocardiogram signal by Gaussian noise overlaying.
The second stage represents the resolution of the multiclass classification task with the use of various machine learning methods such as random forest, the decision tree and the convolutional neural network.
For the analysis of the electrocardiogram, we have
selected the MIT-BIH Arrhythmia database with 48 half-hour
excerpts of two-channel ambulatory electrocardiogram
recordings obtained from 47 subjects as the data source. It is
important to note that this study uses pre-processed data from
the MIT-BIH Arrhythmia database presented in the Kaggle
platform (ECG Heartbeat Categorization Dataset) [
The training dataset contains information on 87 thousand heartbeats. Each heartbeat falls within one of five following categories: 1.
In all our experiments, we have used electrocardiogram lead II re-sampled with the sampling frequency of 125 Hz as the input signal.
For further interpretation of the data, we show an example of input data for each class.
At the next step, we have added Gaussian noise with the
mathematical expectation of 0 and the dispersion of 0.05 [
Figure 4 and Figure 5 show data before and after transformation of data respectively. Such transformation is necessary to summarize the data and remove duplicates.
The major part of the dataset (72471) falls under the regular rhythm. Figure 2 shows the bar chart of number of examples depending from classes (0 - Normal beat, 1 Supraventricular premature beat, 2 - Premature ventricular contraction, 3 - Fusion of ventricular and normal beat, 4 Unclassifiable beat).
Figure 2 demonstrates the big difference between the
numbers of input data for each of the classes. We have used
the method described in [
As the result, after data pre-processing, we have received the sampling consisting of 100 thousand heartbeats uniformly ranged into classes.
III. DEVELOPMENT AND TRAINING OF MACHINE LEARNING
METHODS
We have built the models based on such methods as the random forest algorithm, the decision tree and the convolutional neural network for detecting the signs of cardiovascular diseases. The paper describes in detail the operation principle of each of the methods:
Decision Trees are a non-parametric supervised learning
method used for classification and regression. The goal is to
create a model that predicts the value of a target variable by
learning simple decision rules inferred from the data features
[
The random forest consists of many decision trees. The
random forest is used in statistics, data analysis and machine
learning. Figure 6 demonstrates the model of this algorithm.
Each individual tree is a rather simple model with branches,
nodes and leaves. Attributes are recorded in the nodes and
objective function depends on the attribute values. Further,
values of the objective function get into the leaves through the
branches. As part of the classification process for a new case
we have to go down the tree through the branсhes all the way
until a leaf, moving through all attribute values based on the
logical principle “if-then” In accordance with these
conditions, a target variable receives one or another value or
class (the target variable gets into the specific list). The
objective of the decision tree construction is to develop a
model that predicts the value of the target variable depending
on several variables at the input [
We apply decision tree and random forest algorithms with the use of the Scikit-learn library for the Python programming language. The paper considers the n_estimators parameter (the number of trees in the forest) for random forest. We have selected this parameter in experimental studies and it equals to 50.
In recent years, convolutional neural networks are gaining popularity. In this paper, the model of convolutional neural network contain 1 convolution consisting of 32 feature maps. Then the batch-normalization layer comes that normalizes previously obtained data. The normalized data have mathematical expectation of 0 and dispersion of 1. Sub Nyquist sampling layer (MaxPooling1D) compresses the obtained data twice and the Flatten layer shrinks the compressed data into 1D dimension array. Three cascade fully connected layers containing 64, 32 and 5 neurons respectively follow these layers. At the output, we have five neurons that show the probability of belonging to each class. Figure 7 shows the described neural network architecture.
The neural network is introduced through the Keras library in Python programming language and we have trained it with the use of Adam (Adaptive Moment Estimation) optimizer.
Adam is an update to the RMSProp optimizer. This
optimization algorithm uses running averages of both the
gradients and the second moments of the gradients. Given
parameters of the weight - w ( t ) , where t indexes current
training iteration are updated according to the following
algorithm [
m ( t 1) mˆ w 1 w t 1 ;
1 v ( t 1) vˆ w w 1 t 1 ;
2
w ( t 1) w ( t )
mˆ w
vˆw
where ϵ is a small scalar used to prevent division by 0 , and
1 and 2 are the forgetting factors for gradients and second
moments of gradients, respectively. Squaring and
squarerooting is done elementwise [
We select ReLU function as the activation function on each layer except the last one. For the last layer we use Softmax function that is often applied in neural networks to show unnormalized network outputs into a probability distribution over the predicted output classes. Softmax function is as follows: S o ft m a x ( z ) i , e zi 5 e z j j 1 where z i is an unnormalized input vector.
We use logloss as an error function:
1 q l lo g lo s s y ij lo g a ij , q i 1 j 1 where q is a number of elements in sampling, l is a number of classes, a ij is a response (probability) of the algorithm on the object i to the question on its belonging to class j , y ij 1 if object i belongs to class j , if not y ij 0 .
Each of the above-mentioned methods has its own advantages and disadvantages. In this paper, we present the training of all models mentioned above to identify which model solves the classification problem more accurately.
For experimental studies, we have launched the program based on test sampling consisting of 21,892 heartbeat records.
We have started up the training process of the neural network in 15 epochs. Figure 8 shows a graph of the accuracy of neural network training depending upon an epoch for each training and validation set.
For each model, we consider Accuracy characteristic that is formed through the ratio of the total number of correctly predicted objects ( P ) to the total number of objects ( N ) : A c c u r a c y
P
.
N
Table 1 shows the training results (Accuracy for validation data).
Table 1 demonstrates that the convolutional neural network has showed the best result.
As the neural network results further show, we have calculated the normalized confusion matrix that presents the proportion of correct and incorrect predictions for each class. Table 2 demonstrates the values of the confusion matrix.
Table 2 indicates that the neural network classifies Supraventricular premature beat much worse. This is simply ause of the small data amount of this class. The neural network identifies Unclassifiable beat more accurately. This appears to be due to the specificity of the electrocardiogram data in this class.
In this paper, we compare various machine learning algorithms for detecting the signs of cardiovascular diseases. We have pre-processed all signals in accordance with Section 2 before applying the algorithms. The study shows that the convolutional neural network classifies heart rhythms better (0.9347) than other algorithms. The random forest algorithm also demonstrates excellent results (0.9090) that are a bit worse than the results of the neural networks. The decision tree presents the worst results (0.7870). The significance of the study lies in the feasibility of application of the achieved results in predicting the possible development of cardiovascular diseases. The findings obtained in the present study can be used to create medical or other expert systems based on artificial neural networks under conditions of low volume of statistical material.