Patient-Specific ECG Beat Classification using Deep Learning Techniques

Abstract

The electrocardiogram (ECG) is a non-invasive medical tool used to capture the electrical activity of the heart. It records the electrical impulses generated by the heart muscle and displays them as a waveform on a screen or on paper. Manually, analyzing ambulatory ECG records can be a challenging task due to their duration and a high number of ECG beats. Therefore, an automated diagnosis tool is required for the automatic classification of ECG beats. This motivates to develop a better ECG beat classification system with efficient feature extraction methods for identifying different classes. In this regard, at first deep two-dimensional residual network (2D-ResNet) is developed, and S-transform (ST) based time-frequency ECG beat images are utilized as input to the network. In this context, ST-based ECG depiction in the time-frequency domain offers a response with consistent amplitude across frequencies and adaptable resolution. The resulting ST visuals serve as input for the suggested 2D-ResNet, which classifies five distinct ECG beat types in a manner tailored to each patient, following the guidelines set by the Association for the Advancement of Medical Instrumentation (AAMI). The first five minutes of ECG data from the test subjects are also included in the training data during the training phase of the proposed methods. Therefore, the proposed techniques can be treated as patient-specific in this work. The proposed 2D-ResNet, which does not utilize handcrafted features, leverages the benefits of ST and the 2D-ResNet model to detect the ECG beats automatically. It can capture both the signal’s time and frequency information and reveal patterns that are not apparent in the raw signal. Next, a combination of convolution neural network (CNN) and Bidirectional long short-term memory (Bi-LSTM) architecture is developed to utilize input data effectively during the training phase. Bi-LSTMs, in the proposed network architecture, traverse the input data in both forward and backward directions. Hence, the proposed model benefits from additional training information, contributing to improved performance and robustness. The Bi-LSTMs are better compared to the standard unidirectional LSTMs due to their fixed sequence-to-sequence prediction and increased training capacity. In this proposed model, the CNN component is usually employed as an initial feature extractor. It applies convolutional operations to capture local patterns in the input data. The Bi-LSTM layers in the model are responsible for capturing temporal dependencies and modeling long-range context information in both forward and backward directions. The major Abstract limitations of the above-discussed techniques are: first, the performance of the existing algorithms degrades comprehensively in the presence of similar morphological patterns with minor variations from different classes. Second, the generalization capability of the existing techniques with different datasets and higher-end graphical processing unit (GPU) requirements. To overcome the above-mentioned issues, a novel multi-stream deep learning algorithm with the random forest is proposed, which effectively extracts the features from similar morphological patterns with minor variations. Three different individual streams are utilized with CNN, residual, and bidirectional gated recurrent units (Bi-GRU) to extract more distributed representative, hierarchical and condensed, and long-term dependency features, respectively. These extracted features are utilized to form deep features with the help of concatenation and fusion techniques. The resulting features are able to capture both the morphology and temporal dynamics of the ECG signal. These features are more effective in identifying different types of arrhythmias, predicting future cardiac events, and filtering out noise and artifacts. The unique nature of the features obtained by combining CNN, residual blocks, and Bi-GRU enables a more comprehensive and accurate analysis of the ECG signal, which is particularly important for diagnosing and monitoring cardiac abnormalities. Finally, the extracted deep feature set is utilized to train and test the random forest algorithm. All the above discussed techniques performed well when the ECG signal quality is good and at an acceptable level, but detecting the acceptability of the ECG signal is a challenging and crucial task. To handle this issue, an automatic signal quality assessment (SQA) technique is also developed using the proposed deformable CNN architecture to verify the signal acceptability for further ECG beat classification. The performance of the proposed methods for ECG beat classification is tested with acceptable unacceptable ECG segments. The proposed algorithms are validated on three publicly available ECG arrhythmia datasets such as MIT-BIH arrhythmia, INCART, and MIT-BIH supraventricular. Furthermore, the qualitative and quantitative analysis of the proposed techniques outperforms the state-of-the-art methods on three different publicly available datasets in the literature.