Study of Multi-class Arrhythmia Detection and Classification from ECG data using Deep learning

Cardiovascular diseases are responsible for a significant number of fatalities worldwide each year and the death rate from these diseases has increased in recent years. ECG signals are one of the most important sources of information about the heart’s electrical activity which can be used to detect abnormalities in the heart. Detection of heart diseases using ECG signals still largely depends on the physicians. A well-trained automated system can go a long way to give assistance to the physician and reduce the time and cost of the detection process. Deep learning models can be trained on large datasets of ECG signals to learn to identify patterns that are associated with different heart conditions. In this paper, we have presented a comparative study of different 1D convolutional deep-learning model architectures with attention mechanisms for electrocardiogram (ECG) signal classification. The models are trained on five large publicly available datasets of ECG signals, which is a diverse collection of ECG signals from patients suffering from various cardiac diseases. In our work, we have shown how different deep learning techniques perform in a large collection of ECG datasets from various sources. We have also proposed an ensemble technique that achieved a significantly higher score in terms of both AUC-ROC(98%) and F1-score(66.9%) than previously suggested techniques.