Analysis of cardiovascular health is an urgent task, as the number of patients with cardiovascular diseases is growing in the world. For the early diagnosis of these diseases, various studies of the cardiovascular health of patients are used, among which the analysis of electrocardiograms (ECG) is widely used. Classifiers based on machine learning models have been developed for automatic ECG analysis. At the same time, in medical practice, the SCORE scale is often used to assess the probability of risk of cardiovascular disease, which analyzes gender, age, systolic blood pressure, total cholesterol and smoking. Therefore, this article proposes an approach and architecture for analyzing the cardiovascular health of patients that integrates automatic analysis of five types of heartbeats extracted from ECG with a linguistic assessment of the risk of cardiovascular disease on the SCORE scale. To classify heartbeats, a model based on RNN with LSTM neurons was developed, which showed an accuracy of 98.34% on the MIT-BIH Arrhythmia Database test set. A system of rules that integrates the heartbeats class and the patient's SCORE assessment was developed to provide a linguistic description of cardiovascular health, useful in forming a medical conclusion.
Diseases of the circulatory system occupy a leading place in the statistics on morbidity and mortality of the adult population in many countries, including Russia. At present, the reduction of mortality from cardiovascular diseases (CVD) is a strategic task of Russian healthcare and a key task of preventive medicine.
Informatization of health care and the use of digital technologies for data processing
are priority tasks in the development of a modern information society based on
knowledge in all countries. Joonseok Kim, Peter W. Groeneveld [
The analysis of the ECG is based on the study of the graphic patterns of its graphic
elements, such as teeth, segments and intervals, which reflect the work of the heart
muscle. Changes in the size of the teeth, uneven segments indicate violations in the
cardiovascular system, which requires further research and examinations, the
appointment of an effective course of treatment for this patient. Although the ECG
results are automatically briefly described by the electrocardiography device, they
require further interpretation and description by a cardiologist due to variation in
morphological and temporal characteristics of ECG waveforms. The ECG analysis
process to assess the heartbeat can include more than 50 subprocesses [
The implementation of such process using a trained classifier model will reduce the time spent on making a medical decision, as well as reduce the number of medical errors. To achieve this goal in this paper, we propose an architecture for a CVD risk analysis system that includes an automatic linguistic interpretation of cardiovascular health, based on classification of heartbeats using recurrent neural networks and assessment of cardiovascular risk on the SCORE scale. The proposed system architecture is developed to support medical decision making.
Among the studies in the field of methods for the classification of ECGs, two areas
can be distinguished: (1) algorithms based on morphological features and classical
signal processing techniques [
Recently, artificial neural networks (ANNs) have become increasingly popular for ECG classification due to their learning capabilities; they are more resistant to noise in the input data, which is important for ECG due to the presence of a large number of microalternations. Also, trained ANNs are able to produce results faster than other methods, which is important for emergency situations, for example, in cardiac resuscitation.
In the study by Mohammad Kachuee, Shayan Fazeli, Majid Sarrafzadeh [
In the paper U.R. Acharya, Shu Lih Oha, Yuki Hagiwara, Jen Hong Tan,
Muhammad Adam, Arkadiusz Gertych, Tan Ru San [
In the study by the authors R. J. Martis, U. R. Acharya, C. M. Lim, K. Mandana,
A. K. Ray, and C. Chakraborty [19] before QRS detection and segmentation,
denoising of ECG based on wavelet transform was performed [20]. After that,
principal components were calculated and used as features for further classification. In this
study, a feedforward artificial neural network (NN) [21] and a Least-squares
supportvector machine (LS-SVM) are compared in the classification accuracy of five types of
ECG heartbeats on the dataset [
When classifying complex data, a sequence of characters can enter the input and this sequence of characters must be recognized at the output. Often this is done using a classifier model based on a recurrent neural network (RNN) [22]. The presence of special blocks in the hidden layer NN, containing controlled elements of the «remembering / forgetting» state, is a special feature of RNN models. Currently, various configurations of recurrent neural networks based on different blocks in hidden layers are proposed, for example, BiRNN (bidirectional RNN), LSTM (RNN with long-term short-term memory), GRU (RNN with gates). The effectiveness of this kind of networks was shown in the works [23 -27].
In the paper [28], RNN Encoder-Decoder [29] is used in conjunction with CNN
which extracts a sequence of heartbeat features from an ECG, the latter are then fed to
the RNN Encoder-Decoder input. The Encoder encodes the heartbeat features using
LSTM cells [30], while the Decoder classifies the heartbeat type. The authors of [28]
use BiRNN instead of the usual RNN [31], which can process data in both forward
and backward directions, in normal time order, t = 1, ..., T for the forward network,
and in reverse time order, t = T, ..., 1 for the backward network. Finally, the weighted
sum of the outputs of the two networks is computed as the output of the BiRNN. In
this work, ECG classification is carried out for 4 classes MIT-BIH arrhythmia
database [
To better capture the patterns in the ECG waveform in paper [32], an algorithm
employs both LSTM recurrent neural networks and classical features, i.e., wavelet, at
the same time, is proposed. Then the results are blended by multi-level perceptron
(MLP) with two hidden layers, which generates the final type of beat. Authors
consider and study 3 approaches of ECG classification on MIT-BIH arrhythmia database
[
In this study, the open-source MIT-BIH arrhythmia database [
This paper proposes an architecture of system “Cardio+” for linguistic summarization of a cardiovascular health that includes an automatic ECG classification based on LSTM based recurrent neural networks. Since, when forming medical conclusions, in addition to the ECG, the cardiologist analyzes various other health indicators, the system includes a module for combining the results of ECG classification with an assessment of cardiovascular risk by SCORE-scale [33]. The proposed system architecture can serve as an experimental model of support tool for medical decision making. The architecture of the CVD risk analysis system is shown in Fig. 1
There are two main components in the system. The first component is designed for ECG classifier training and testing, the second one uses the trained model to classify new ECG and produces linguistic summarization of cardiovascular health taking into account the assessment of cardiovascular risk by SCORE-scale. Both components of the system “Cardio+” preprocess ECG signal to obtain heartbeat representation of equal length.
The input of the system receives data from the MIT-BIH arrhythmia database [
The developed system makes it possible to analyze cardiovascular health based on the classification of heart rhythm pathologies in the ECG, integrating the heartbeat class obtained using the RNN model and the CVD risk indicator calculated using the assessment of CVD risk by SCORE-scale [33]. SCORE-scale assesses risk of fatal CVD within the next 10 years for patients in regions of Europe by gender, age, systolic blood pressure, total cholesterol and smoking. As a result of evaluating these indicators, an integer X is formed in the range from 0 to 47. The larger this number, the higher the risk of fatal cardiovascular events in a given patient. This paper examines four degrees of CVD risk according to the scale of assessment of CVD risk by SCORE [33]: 2=< Х < 5: The 1st degree of risk (low risk) means that the probability of cardiovascular complications in this patient is less than 5%; 5 ≤ Х < 10: The 2nd degree of risk (medium risk) assumes the probability of complications of 5-10%; 10 ≤ Х ≤ 15: The 3rd degree of risk (high risk) implies the probability of complications of 10-15%; Х > 15: The 4th degree of risk (very high risk) implies a probability of complications of more than 15%.
Table 1 shows the developed rules that combine the heartbeat class recorded on the ECG and the degree of assessment of CVD risk by SCORE-scale. These rules made it possible to obtain a more informative comprehensive linguistic assessment of the risk of developing cardiovascular disease in a patient and formulate recommendations for a doctor.
The description of some details of the system implementation is given further. Each ECG is read using the pandas library. The continuous ECG signal is split into windows using slices of arrays. A random number i from is selected and a slice of the array is taken from i to i + 3600. The amplitude values are normalized in the range from 0 to 1 using sklearn, which implements the normalization by the minimax criterion: (1) Next, the set of local maxima function of the signal module of the scipy library (see 28). From the set of local maxima, a set of R-waves() function of the signal module of the scipy library: From the set of local maxima, a set of R-waves is found by applying a threshold value of 0.83. From a practical point of view, this is achieved by passing the appropriate condition to the where () function of the numpy library: (2) Time intervals (I) are calculated as the difference between the indices of adjacent Rwaves in the segmented ECG signal: (3) Possible risk of developing CVD. There are no deviations in the work of the heart. Recommend a healthy lifestyle to the patient.
There is a risk of developing CVD. Abnormalities in the work of the heart according to ECG type F or V. Recommend a healthy lifestyle to the patient.
Conduct additional research, prescribe treatment.
There is a risk of developing CVD. Abnormalities in the work of the heart according to ECG type Q, S Recommend a healthy lifestyle to the patient.
Conduct additional research, prescribe treatment.
Average CVD risk. There are no deviations in the work of the heart. Recommend a healthy lifestyle to the patient, monitor the SCORE scale indicators.
Average CVD risk. Abnormalities in the work of the heart according to ECG type F or V. Recommend a healthy lifestyle to the patient, monitor the SCORE scale indicators. Conduct additional research, prescribe treatment.
Average CVD risk. Abnormalities in the work of the heart according to ECG type Q, S. Recommend a healthy lifestyle to the patient, monitor the SCORE scale indicators. Conduct additional research, prescribe treatment.
High risk of developing CVD. There are no deviations in the work of the heart. Recommend to the patient a healthy lifestyle, to carry out daily monitoring of the SCORE scale indicators and to perform electrography once a year.
High risk of developing CVD. Abnormalities in the work of the heart according to ECG type F or V. Recommend to the patient a healthy lifestyle, to carry out daily monitoring of the SCORE scale indicators and to perform electrography twice a year. Conduct additional research, prescribe treatment.
High risk of developing CVD. Abnormalities in the work of the heart according to the ECG type Q, S. Recommend the patient to a healthy lifestyle, carry out daily monitoring of the SCORE scale indicators and perform electrography twice a year. Conduct additional research, prescribe treatment.
Very high risk of developing CVD. There are no deviations in the work of the heart. Recommend to the patient a healthy lifestyle, daily monitoring of SCORE scale indicators and electrography twice a year Very high risk of developing CVD. Abnormalities in the work of the heart according to ECG type F or V. Recommend the patient to a healthy lifestyle, to carry out daily monitoring of the SCORE scale indicators and to perform electrography three times a year. Conduct additional research, prescribe treatment.
Very high risk of developing CVD. Abnormalities in the work of the heart according to ECG type F or V. Recommend to carry out daily monitoring of the SCORE scale indicators and to perform electrography three times a year.
Conduct additional research, prescribe treatment. 1 1 2 2 2 3 3 3 4 4 4
F, V Q, S N F, V Q, S N F,V Q, S N F, V
Q, S Then the period of the heartbeat of this segment is found. To do this, first the median value (m) of the time intervals between adjacent R-waves using the median () function of the numpy library takes. And for the period of the heartbeat (T), the value is taken equal to 1.2 * m: The maximum number of ECG time series points after the described changes is 187, but depending on T, there may be less of them. Since for the ANN needs to submit a standardized tensor to the input, the remaining part of the ECG time series, if it had less than 187 values, was supplemented with 0. The addition of exactly zero values was made due to the probability of accepting nonzero values for other types of teeth when training the classifier.
In this paper, data augmentation was performed using displacement and stretching of existing ECGs, as well as by adding a little noise.
Then the data goes to the Module ECG pre-processing. As a result of passing this module, a set of ECG time series is formed, which is later entered into the Google Colaboratory system record (ECG_classification_RNN.ipynb), where, based on the preprocessed data, a recurrent classifier based on LSTM neurons is trained and tested using machine learning tools. After training and testing, the best model (model.hdf5) is saved for later classification of the ECG signals. Now, for the analysis, it is enough to download the ECG signal (ECG_signal.csv). This signal enters the Module heartbeat classification, which uses pre-processing components, machine learning components, a Trained RNN model to classify and provide a linguistic assessment of the severity of CVD. Further, the data obtained during the analysis are sent to the Module linguistic summarization, which builds ECG and ECG time series graphs using the Matplotlib component.
A RNN based on LSTM neurons was implemented using Keras, a machine learning library that runs under the control of the Tensorflow library. It was decided to use the Tensorflow library, since it is an open development by Google and is included by default in the Google Colaboratory [34], in which the training of the created recurrent ANN was carried out.
The RNN contains four sequential layers: the first consists of 50 LSTM neurons, the second 25 LSTM neurons, the third contains 5 LSTM neurons, the last layer generates the network output and contains 5 multilayer perceptron, each with one hidden layer. For LSTM layers, it is obligatory to indicate the dimension of the structure supplied to the input - the input_shape () parameter, which is passed: the number of ECG for training (in our case, it is a little over 450,000); the number of points in the time series (n_features) (in our case, it is 187).
A parameter function is passed to all recurrent layers defining the activation function. In this work, the ReLU activation function was chosen, since recurrent networks with this activation function showed the best results in several applications. Softmax is used as the activation function for the last layer of the network, since it represents the output of the network as a probability distribution of belonging to each of the classes. After each recurrent layer, the probability of deactivation of each neuron is also set. Between the first and second, and the second and third layers, it is 0.2 for each neuron; between the third and fourth - 0.1 for each neuron.
The model of the network considered above was selected empirically. During the experiments, various structures of RNNs were considered: a different number of hidden layers (from 1 to 5); different types of memory cells (LSTM and GRU); different number of memory cells in each layer (ranging from 250 to 25 with a step of 25 neurons); different probability of deactivation of each neuron in the layer (probably varied from 0.3 to 0 with a step of 0.05). In this study, categorical cross entropy was chosen as the loss function, the adaptive method Adam was chosen as the gradient descent optimizer. 5
Four main measures usually were considered in the literature to evaluate the performance of heartbeat classification models including the sensitivity (SEN), positive predictive value (PPV), accuracy (Acc) and F1-score:
SEN = TP/(TP + FN)
PPV = TP/(TP + FP) Acc = (TP + TN)/(TN + FP + FP + FN)
F1=2*(PPV*SEN)/( PPV+SEN), (6) (7) (8) (9) where TP (True Positive), TN (True Negative), FP (False Positive) and FN (False Negative) indicate the number of heartbeats correctly labeled, number of heartbeats correctly identified as not correspond to the heartbeats, number of heartbeats that incorrectly labeled, and number of heartbeats which were not identified as the heartbeats that they should have been, respectively.
Estimates of the accuracy of the ECG classifier, implemented on the basis of RNN for five classes of heartbeats (N (normal and bundle branch block beats), S (supraventricular ectopic beats), V (ventricular ectopic beats), F (fusion of N and V beats) and Q (paced beat or unclassifiable beat) are given in Table 2.
The total number of ECG records on which the RNN based classifier was tested was 4000 samples, and 800 samples were selected for each of 5 classes for testing.
In this contribution we utilize LSTM based RNN model which contains four layers. It should be noted that the authors in [28] classified only 4 classes of heartbeats, while in this work the ECG was classified into 5 classes.
Comparison of the classification results according to the criteria of average accuracy is shown in Table 3. According to Table 3, the average percentage of correct answers for the proposed RNN model for ECG classification is 98.34%, so the accuracy achieved in this work is competitive with modern models.
Unlike other studies a system of rules that integrates the heartbeats class and the patient's SCORE assessment was developed to provide a linguistic description of cardiovascular health, useful in forming a medical conclusion. 6
Operational analysis and automatic diagnosis of cardiovascular diseases is an important and urgent task in medical practice. The development and application of machine learning methods that allow solving this problem will improve the quality of medical services provided, identify diseases at an early stage and inform patients about the decline in cardiovascular health. In this work, a modular structure of a system for analysing cardiovascular health is proposed and implemented, in which, on the basis of the developed RNN model, an automatic classification of the patient's ECG is performed, the results of which are integrated with the degree of development risk to generate useful information for the doctor in a linguistic form. The developed LSTM based RNN has a simple architecture and, on the MIT-BIH arrhythmia database, has been shown to be effective in classifying ECGs into five classes. In this study, we obtained the highest average accuracy of 98.34%. In addition, a feature of this study is the possibility of integrating the assessment of CVD risk by SCORE with the results of ECG classification. Further research is planned towards the development of the Cardio+ system to expand the integration of CVD indicators characterizing the behavioral, anthropological and psychological characteristics of patients.
The authors acknowledge the reported study was funded by RFBR, project number 20-07-00672. 17. Pan J., Tompkins W.J.: A simple real-time QRS detection algorithm, https://www.researchgate.net/publication/3728672_A_simple_realtime_QRS_detection_algorithm, last accessed 2020/10/10. 18. Z-Score: Definition, Formula and Calculation, https://www.statisticshowto.com/probability-and-statistics/z-score, last accessed 2020/10/10. 19. Martis R. J., Acharya U. R., Lim C. M., Mandana K., Ray A. K. and Chakraborty C.: Application of higher order cumulant features for cardiac health diagnosis using ECG signals.
International Journal of Neural Systems, 23(4), (2012). 20. Wavelet Transform, https://www.sciencedirect.com/topics/computer-science/wavelettransforms, last accessed 2020/10/10. 21. Haykin S. S.: Neural networks: a comprehensive foundation, 2nd Edition, Prentice Hall, (1999). 22. Graves, A.: Supervised Sequence Labelling with Recurrent Neural Networks. Studies in
Computational Intelligence. Springer. (2012). 23. Afanasieva, T., Platov P.: The Study of Recurrent Neuron Networks based on GRU and LSTM in Time Series. In: 6th International conference on Time Series and Forecasting, pp. 190-202, Granada (2019). 24. Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio: Empirical evaluation of gated recurrent neural networks on sequence modeling, https://arxiv.org/pdf/1412.3555.pdf, last accessed 2020/10/10. 25. Laptev N., Yosinski J., Li Erran, and Smyl S.: Time-series extreme event forecasting with neural networks at uber. In International Conference on Machine Learning 34 (2017). 26. Hojjat Salehinejad, Sharan Sankar, Joseph Barfett, Errol Colak, and Shahrokh Valaee: Recent Advances in Recurrent Neural Networks, https://arxiv.org/pdf/1801.01078.pdf, last accessed 2020/10/10. 27. Zachary C. Lipton, John Berkowitz, Charles Elkan : A Critical Review of Recurrent Neural Networks for Sequence Learning, https://arxiv.org/abs/1506.00019, last accessed 2020/10/10. 28. Mousavi S., Afghah F., Acharya U.R.: Inter- and Intra-Patient ECG Heartbeat Classification For Arrhythmia Detection: A Sequence to Sequence Deep Learning Approach, https://arxiv.org/pdf/1812.07421.pdf, last accessed 2020/10/10. 29. Kyunghyun Cho, Bart van Merriënboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, Yoshua Bengio: Learning Phrase Representations using RNN Encoder–Decoder for Statistical Machine Translation, https://arxiv.org/pdf/1406.1078.pdf, last accessed 2020/10/10. 30. Hochreiter, S. and Schmidhuber, J.: Long short-term memory. Neural Computation, 9(8):1735–1780 (1997). 31. Schuster M., and Paliwal K.K. : Bidirectional recurrent neural networks. IEEE Transactions on Signal Processing 45(11), 2673–2681 (1997). 32. Saadatnejad S., Oveisi M., Hashemi M: LSTM-Based ECG Classification for Continuous Monitoring on Personal Wearable Devices. IEEE Journal of Biomedical and Health Informatics (JBHI) (2019). https://www.researchgate.net/publication/329608349_LSTMBased_ECG_Classification_for_Continuous_Monitoring_on_Personal_Wearable_Devices, last accessed 2020/10/10. 33. SCORE, https://www.escardio.org/static_file/Escardio/Subspecialty/EACPR/Docume nts/score-charts.pdf, last accessed 2020/10/10. 34. Google Colaboratory, https://colab.research.google.com, last accessed 2020/10/10.