Page 100 - The-5th-MCAIT2021-eProceeding

LSTM layers is analyzed.

2. Background
In most artificial intelligence applications involving sequential data or one-dimensional signals, RNN
architecture is widely used for prediction, detection, and recognition. This neural network is capable of
effectively retaining historical data and learning long-term text dependence information. The LSTM is a
variation of the RNN architecture that would handle input and output of different lengths. Because of this asset,
it has recently shown potential in sequence learning applications (Lay-Ekuakille et al., 2013)..

2.1 Bidirectional LSTM:
BiLSTM is a network extension of the LSTM. Figure 1 shows the general architecture of BiLSTM, which
entails replicating the network's first layer; the input sequence is then fed to the first layer as is, with a reversed
copy going to the replicated layer. A BiLSTM can be trained using all possible input data from the past and
future of a time step. The forward states (positive time direction) are controlled by a portion of the state neurons,
while the backward states are controlled by the other half (negative time direction). The BiLSTM layer's output
updates a collection of global features heavily affected by nearby inputs and localized (Cui et al., 2018).

Fig. 1. General Architecture of Bidirectional LSTM Network Framework.

2.2 Related Work:
There has been considerable work on LSTM models for arrhythmia classification. Work by Sharma et
al., 2020 utilizes deep LSTM networks to detect and identify four different types of abnormalities in ECG
signals. Another model was proposed in Yildirim (2018) for deep BiLSTM network-based wavelet sequences.
Here, for the generation of ECG signal sequences, a wavelet-based layer is utilized. Furthermore, the
performance of unidirectional LSTM and BiLSTM algorithms was compared. An automatic arrhythmia
classifier was developed in He et al., 2019 by combining residual convolutional neural networks (CNN) and
BiLSTM layers for the feature extraction from raw ECG signals. The feature vector that is obtained after feature
merging is trained for final classification. Finally, work in Xu et al., 2020 proposed a hybrid CNN-BiLSTM
architecture for ECG heart signal classification for diagnostic purposes. Specifically, two CNN layers, two
BiLSTM layers, and two fully connected layers were utilized in the proposed architecture.

E- Proceedings of The 5th International Multi-Conference on Artificial Intelligence Technology (MCAIT 2021) [87]
Artificial Intelligence in the 4th Industrial Revolution