Application of deep learning algorithms for automated detection of arrhythmias with ECG beats

Abstract

Arrhythmia is the anomalies of cardiac conduction system that is characterized by abnormal heart rythms. Prolong arrhythmias are life threatening and can often lead to other cardiac diseases. Abnormalities in the conduction system is reflected upon the morphology of the electrocardiographic (ECG) signal and the assessment of these signal can be extremely challenging and time-consuming. Morphological features of arrythmias ECG signals are low in amplitudes and the changes within can sometimes be very subtle. Therefore, the main aim of this study is to develop an automated computer aided diagnostic (CAD) system that can potentially expedite the process of arrhythmia diagnosis, which will allow the clinicians to provide better care and timely intervention for the patient.

In machine learning, the performance of classification largely depends on the quality of features extracted. Therefore, the process of obtaining useful information which effectively differentiate the specific classes into groups is crucial. Generally, there are two types of features used in machine learning, handcrafted features and learned features. Many of the techniques developed in earlier literature involved the use of handcrafted features. In order to engineer a handcrafted feature it typically requires one to have extensive domain knowledge and the latter experimentation cost in selecting the optimal features for a specific classification model can be costly as well. Learned features on the other hand is obtained though self-discovery by the artificial intelligence system, it obviates the process of manual engineering and the current state of the art technique used in obtaining learned feature is through deep learning.

In this research, two different deep learning architectures are tested for diagnosing arrhythmic ECG signals. The first proposed deep learning architecture is a hybrid neural network of convolutional layers and long short-term memory (LSTM) units capable of providing single class prediction for each variable-length data ECG segments. The second proposed model is U-net ,a fully convolutional auto encoder with skip connections, which provides a much detailed analysis for the ECGs as each of the detected beats can be marked with a specific heart conditions.

Both models are trained and tested against the MIT-BIT arrhythmia database. 5 cardiac conditions, normal sinus rhythm, atrial premature beats (APB), premature ventricular contraction (PVC), left bundle branch block (LBBB) and right bundle branch block (RBBB) are segmented from the recordings for evaluation. Additionally, the ten-fold cross validation strategy has been employed in the project to confirm the robustness of the proposed models.

Findings of this research will redound in benefiting the ECG screening procedures, considering that deep learning models are capable of achieving considerable accuracy and details in categorizing the individual arrhythmias beats with minimal preprocessing applied. The future work intends to acquire more ECG records to increase the variance of the current dataset, implementation of generative adversarial network (GAN) for ECG augmentation and to explore on other cardiac diseases.