Automated interpretation of fundamental physiological signals like Electrocardiogram (ECG) plays an important role as part of smart healthcare ecosystem to build reliable and on-demand cardio-vascular diseases (CVD) screening system. Myocardial infarction is a life-threatening condition. Acute myocardial infarction leads to fatal condition commonly known as heart attack. It is medically established that myocardial infarction is often ignored initially, and treatment starts late. In this paper, we present an automated method of myocardial infarction detection from off-the-shelf single lead ECG signals so that early warning can be generated, and timely diagnosis can take place. We propose regularized deep neural network based model that is capable of classifying myocardial infarction condition from normal heart rhythm in single lead ECG signals. More precisely, we propose intensely regularized deep residual networks (ResNet) where both L2 (also known as Tikhonov regularization) and L1 (commonly known as Lasso) regularizations are used to construct a compact residual learning model. We demonstrate through empirical study on publicly available relevant ECG dataset from UCR timeseries archive that the proposed method demonstrates considerably superior performance over baseline methods and current state-of-the-art algorithms. We have also performed ablation study to depict the efficacy of the proposed intense regularization over only L2 or L1 regularizations.
Electrocardiogram (ECG) is one of the
fundamental markers for preliminary
investigation of Cardio-Vascular Diseases
(CVDs). Owing to the affordable availability of
personal ECG Sensors like Alivecor [
Future smart cities with IoT based applications and eco-system demand such automated clinical management system that ensures ondemand diagnosis and quick response to the critical treatment requirements. In a general setup, patients interact with their smartphone applications to get associated in the smart healthcare systems and derive the benefit of remote assistance of healthcare facility with digital therapeutics. is inferred based on the rule of minimum DTW distance of the training signals set. In fact, similarity measure is often considered a good approach for classification particularly in the area distance of data
mining. DTW-based
measurement
and
1-NN
elastic
based
classification is the conventional choice of
baseline time series classification algorithm [
and shapelet domains [
Another important time series classification algorithm is
Bag of
Approximation (SFA) symbols (BOSS) [
Recently Residual Network (ResNet) have
been emerged to solve time series classification
problem. In fact, ResNet based approach seems
to be a promising one [
Myocardial infarction detection from ECG signal considers number of input ECG signals as part of the overall training space. We can represent each of them as: = [ 1, 2, 3, … , ] ∈ ℝ is an ordered set of real values
number of time steps at each of the training samples. Each of the training samples is associated with a with class label ∈ [ ,
], where rhythm and indicates normal sinus indicates myocardial infarction. Thus, the training dataset consists of labeled ECG signals as Ω = { , }, for = 1,2,3, … , examples.
with total
number of training
Typically, normal sinus rhythm consists of
regular ECG morphology while ECG signal of
myocardial infarction patients have different
morphology [
In order to extract the distinctive features
from ECG signal to distinguish normal rhythm
and myocardial infarction, we use deep residual
network based representation learning [
The immense success of deep residual
networks for computer vision applications
motivates us. In deep residual networks, the
underlying layers fit a residual mapping instead
of directly stacking layers to overcome learning
degradation problem through layer-wise
recursive learning using skip-connection [
ℛ = ℛ −1 + ℋ (ℛ −1)
The main intuition of a residual network is that the skip connection or short-circuit connection provides direct path of input signal
We can conceptualize residual networks as micro-structures of residual blocks connected sequentially along with skip connection as shown in Figure 3.
We construct a pretty deep residual neural network with total fifteen convolution layers. There are three residual blocks. First residual block consists of five convolution layers. There are eight convolution layers in second residual block, while third residual block consists of six convolution layers. Including the three skip connection layers, total twenty layers are present in the proposed deep learning model, with nineteen effective convolution layers.
We have used Batch Normalization (BN)
and Rectified Linear Unit (ReLU) as the
activation layer. After the last convolution
layer, we use Global Average Pooling, which
acts a structural regularizer. In fact, it natively
prevents overfitting for the overall structure
with added advantage of no need for
hyperparameter optimization [
It is understood that ECG signal annotation
is an expensive process and it is likely to have
scarce number of training examples. While the
proposed deep neural network has a
sophisticated capability of representation
learning over the training space, there exists
higher chance of overfitted model generation.
In order to eliminate the risk of overfitting to
the training examples, we regularize the model
using 2 and 1 regularizations [
∗( ) = ( ) + ( )
Where, ( ) is the penalty function and is the regularization co-efficient, with ∈ [0, ∞]. The 2 or Tikhonov regularization is expressed as: ∗2( ) = ( ) + 2 Τ 2 Where, ( ) = Τ2 .
Where, the network parameter gradient is: Similarly, the Lasso or 1 regularization is defined as:
∗1( ) = ( ) + 1‖ ‖1
Where, 2 and 1 are 2 regularization and 1 regularization factors respectively.
∗
Τ ( ) = ( ) + 2 2 + 1‖ ‖1
We intend to emphasize that regularization impacts on network parameters for 2 and 1 regularizations are different. With or 1 regularization, we have sparser parameter matrix while 2 regularization clips the values of the parameters. Thus, a sparser yet controlled network parameter space model is constructed.
We consider relevant dataset ECG200 from
publicly available UCR time series archive
[
For this experimentation purpose, we
consider the hyperparameters as described in
Table 1.
[
as: i.e.
when ( ).
Ablation study to depict the efficacy of the proposed method, showing that L1 and L2 regularization is more effective than L1 or L2 regularizations.
Regularization
L1
L2
Our method (L1 and L2)
Test accuracy 0.80 0.86 0.91
We
illustrate
a
plausible
deployment
architecture of remote diagnosis of myocardial
infarction (CVDs in larger scope) in Figure 5.
The
proposed ECG
analysis
algorithm is
supposed to be hosted
over cloud and a
smartphone-based application interacts with the
user to send the captured ECG signal to the
cloud. For ECG signal capturing purpose,
offthe-shelf sensors like Alivecor can be used. The
cloud server responds the inference result to the
user through the same smartphone application
Residual block #3
state-of-the-art algorithms. Currently, BOSS
as well as alerts nearby medical facilities if
abnormal cardiac activity is detected and
myocardial infarction is suspected. Thus, a
complete diagnosis eco-system can be built
using the proposed method, which will have
immense benefit in providing early-warning
alarms to initiate necessary medical attention.
However, it is imperative to mention that
physiological signals like ECG consists of
potential sensitive information and appropriate
trust [
In this paper, our focus is to develop a classification algorithm to reliably detect myocardial infarction or heart attack condition using ECG signal. The performance of the proposed model is superior than the currently available relevant methods. We are confident that ECG analytics as part of a smart healthcare ecosystem will ensure the necessary impetus for human-centered purpose in contributing better quality of life through the development of technology-driven early detection and treatment for life-threatening CVDs like myocardial infarction.
Leandro Marin is partially funded by Project PID2020-112675RB-C44 funded by MCIN/AEI/10.13039/501100011033.
Antonio J. Jara is funded by the European Union’s Horizon 2020 research and innovation programme under grant agreement No 732679, ACTIVAGE project https://www.activageproject.eu/.