^{1}. Generally, manual elucidation may take time for the physician to recognize the symptoms of MI, and also it varies for time to time and patient to patient. Thus, the rapid growth of wearable computerized devices and their conceivable techniques help to save the lives of millions of people whom are getting affected by MI. Here, electrocardiogram (ECG) signals are used to identify and examine myocardial infarction in time. Three different wave forms of each cardiac cycle in ECG signals are: P wave, QRS complex, and T wave in normal rate. Nowadays, deep learning methods like convolutional neural network (CNN), recurrent neural network (RNN), auto encoder and attention mechanism have accomplished great success in various domains, that include natural language processing, biomedical signal and image processing. Hence, the proposed algorithm based on combined CNN and LSTM (Enhanced Deep Neural Network) is used here to classify and predict whether the patient has MI or healthy.

Myocardial Infarction characteristics include ST-segment elevation, abnormal Q wave appearance, and T-wave inversion. These are commonly used for classification of feature vectors. The ECG signals are varying in ST Elevation interval length, and T waveform. In ^{2}, a deep CNN model has been proposed to provide an effective framework for an automated detection of MI. PTB diagnostic database has been used in their experiment. In^{ 3, }CNN based multi-class MI classification model has been implemented for detecting MI by applying all 12 lead signals. This model makes the decisions through the features that are extracted from the signals. The classification model proposed in for MI detection^{ 4 i}s based on Long Short Term Memory. In this system the source 8-lead ECG signals are preprocessed and partitioned into heartbeat sequences. Then these sequences are fed to LSTM network to make it learn. The features are extracted with the help of the deep learning networks in a hasty manner and that substitutes the manual method of fetching features.

Deep neural network presented in^{ 5 }can classify the 12 rhythm ECG classes. This network possesses 33 convolutional hidden layers followed by a linear output layer. Input to this network is the unprocessed ECG signal data. With this data, it learns predicts and outputs the classified 12 rhythms of each of the 256 samples. In^{ }^{ }

RNN and GRU architecture compared with LSTM architecture is presented in ^{8} to obtain the best sequence model for ECG signal processing. This proposal proved that LSTM architecture has the better performance. In^{ 9 }an artificial neural network is used. In that, the parameters are modified based on the changes of ECG signals. Block-based Neural Network has been trained using PSO algorithm. The BBNNs use morphological and temporal features which have been extracted from ECG signals, and create the input vector of the BBNN. MIT-BIH arrhythmia database is used in their experiment. In ^{ 11 } to provide an automatic recognition of MI.

A novel framework has been developed^{ 12 }for automatic MI detection and location. The experiment provides a new insight into the application of attention mechanism and parallel feature extraction structure based on deep learning. In^{ }

The dataset is obtained from PTB diagnostic database. It holds 290 subjects. Each subject is denoted by three records. Each records consists of 15 signals. They are 12 conventional leads and 3 frank leads^{ }

CNN is a Filtered Back Projection based artificial neural network that shares the weight. It has the connectivity resembling the biological network. CNN has pooling layer, convolutional layer, and fully connected layer^{ }

From the equation (1), the input feature map is denoted by M_{j}; total number of layer is denoted by l; k denotes the convolution kernel; and the network bias vectors is denoted by b. During practical application, Max –pooling is frequently used. Its mathematical model is shown in the Equation (2):

Here, the value of t neuron of i feature map in layer l is denoted by

The workflow of an LSTM model is shown in

According to the above workflow diagram, the following calculation can be done in the equations (3 to 7).

where σ indicates sigmoid function; the input gate is represented by_{; }forget gate is denoted by _{t}_{ }in the hidden layer; and W_{ci }, W_{cf}, and W_{co}_{ }represents the weight matrix of opening connections.

The LSTM unit has a memory cell to keep its state value for a long while and a gating system consisting of three non-linear gates, to point out, an input gate, a forget gate, and an output gate. The intended role of the gate’s, is to regulate the flow of signals into and out of the cell, in order to be effective in regulating long-range dependencies and achieve successful RNN training. Since the inception of the LSTM unit, many modifications have been introduced to improve performance. Adding more components in the LSTM architecture may produce better performance. It is exposed in the architecture of the proposed model in

Based on that, this study proposed an algorithm EDN based on the CNN and LSTM algorithm. In the proposed methodology, the Bias of h were added to the Cell state vector to improve the performance. As the output gate was less important than the Input gate and Forget gate. The proposed algorithm modified the Hidden state vector by adding Point wise Hadamard Multiplication among the previous Output gate parameter and previous Cell state vector. The Equation (8 to 13) represents the mathematical model for EDN.

Where,

Input :

D – Represents the No. of memory Blocks.

S_{j }– Represents the No. of Cells in Block j.

Process:

Step 1: Read the data, then find out standard Deviation and separate the data.

Step 2: To obtain Total independent variable numbers.

Step 3: Set CNN layers with Input Layers and sub sampling layer.

Step 4: Pass the CNN output layers units as Input vectors to the Input gate, Forget Gate and Output Gate of LSTM unit.

Step 5: For each and every block in the Memory, compute the Input, Forget and Output gate for j=1 to D do

Evaluate the Input Gate:

Evaluate the Forget Gate:

Evaluate the Forget Gate:

for Ѵ =1 to S_{j }do

Finally update the hidden state by computing

Evaluate the Hidden State:

End for

Step 6: Return EDN Layers.

The proposed EDN can process the data in a sequential manner, so that each vector in Hidden state can implicitly dependent on previous Cell State unit. EDN uses the convolutional Neural Networks to extend the effective neighborhood identification process.

By using Convolutional Neural Network the output is shown.

Predicted Class 0 | Predicted Class1 | |

Class 0 | 175 | 511 |

Class 1 | 93 | 3235 |

Accuracy | 84.95% |

(b) Details of each layer’s parameters of the CNN Model

By applying LSTM, the output is shown below.

Predicted Class 0 | Predicted Class 1 | |

Class 0 | 362 | 328 |

Class 1 | 212 | 3685 |

Accuracy | 85.23% |

(b) Details of eachlayer’s parameters of the LSTM Model

In this study, a deep learning model is created to provide high recognition performance on ECG signals based on the combination of CNN and LSTM. The 7-layer EDN model with a block representation is shown in

(b) Details of each layer’s parameters of the EDN Model

Predicted Class 0
Predicted Class1
Class 0
244
384
Class 1
118
3772
Accuracy
88.89%

Metrics
CNN model
LSTM model
EDN model
Precision
0.8635
0.9182
0.9776
Recall
0.7720
0.8455
0.8696
F1 Measure
0.9146
0.9217
0.9376
Cohen Kappa Coefficient
0.2996
0.5051
0.2355
Accuracy
84.95%
85.23%
88.89 %

Going along with the direction of forward progress of the deep learning algorithms, this study proposes the EDN algorithm for classifying the ECG signals of normal and MI affected people. For demonstrating the efficiency of the proposed deep learning EDN algorithm, its performance is compared with the two most prominent algorithms of the deep learning realm namely CNN and LSTM. It is ten times faster than the LSTM due to its speed of convergence in training. Through the confusion matrices of the respective algorithms it is obvious that the EDN model achieved 88.89% accuracy; which is 3.66% and 4.04% superior than the LSTM and CNN algorithms respectively. Same way, the proposed model shows performance improvisation through the other performance metrics such as Precision, Recall, F1 measure, and Cohen Kappa Coefficient. Although the difference is small, in healthcare sectors this difference plays crucial role in saving the life of human being.