Page 102 - The-5th-MCAIT2021-eProceeding

4. Results

Table I highlights the performance of the proposed BiLSTM model for cardiac arrhythmia detection versus
varying BiLSTM layers. The number of filters for each layer is set to 32. In addition, Adam optimizer and
binary cross-entropy loss are considered. From Table I, it can be observed that accuracy and specificity
performance improve with increasing BiLSTM layers. In contrast, the recall performance slightly decreases
with increasing BiLSTM layers. Another thing to notice is that the training performance in a majority of the
cases is much higher than the validation performance. A possible option to improve the validation performance
is to use a higher number of filters for each layer. Another option is to make use of nonlinear activation functions
and batch normalization techniques. However, these changes come at the cost of increased complexity and
latency.
Table 1: Performance of the proposed BiLSTM model for cardiac arrhythmia detection versus varying BiLSTM layers

BiLSTM Module Accuracy Recall (%) Precision (%) Specificity (%)
layers (Training/ (%)
Validation)
1 Layer Training 81.0 91.7 75.1 70.9
Validation 56.9 40.5 39.9 66.0
2 Layers Training 85.5 90.7 81.6 80.5
Validation 55.7 30.1 35.8 69.9
3 Layers Training 76.0 67.6 80.2 84.1
Validation 63.6 35.3 48.8 79.4
4 Layers Training 80.3 88.9 75.3 72.0
Validation 56.0 11.6 25.2 80.8
5 Layers Training 85.5 87.5 83.5 83.5
Validation 56.1 67.0 18.7 83.6

5. Conclusion:
Prediction of cardiac arrhythmia through ECG signal analysis is gaining considerable importance to current
researchers and clinicians. The BiLSTM model is one of the prominent deep learning models being used for
arrhythmia detection. However, a study analyzing the impact of BiLSTM variation on the performance of
cardiac arrhythmia detection is still lacking. In this paper, the performance of the BiLSTM model for arrhythmia
classification with different numbers of BiLSTM layers is analyzed. It can be said that the accuracy and
specificity performance of the model increases with increasing BiLSTM layers. As part of the future work, the
impact of the number of filters, nonlinear activation functions, and dropout rate will be studied to achieve further
performance improvements.

Acknowledgments:
The authors acknowledge the support of this research by the Yayasan Universiti Teknologi PETRONAS
Fundamental Research Grant (YUTP-FRG) under Grant 015LC0-244.

E- Proceedings of The 5th International Multi-Conference on Artificial Intelligence Technology (MCAIT 2021) [89]
Artificial Intelligence in the 4th Industrial Revolution