Beat-to-Beat Electrocardiogram Waveform Classification Based on a Stacked Convolutional and Bidirectional Long Short-Term Memory

Abstract

Delineating the electrocardiogram (ECG) waveform is an important step with high significance in cardiology diagnosis. It refers to extract the ECG morphology in start, peak, end points of waveform. Due to various shapes and abnormalities presented in ECG signals, several conventional computer algorithms always fail to extract the essential feature of heart information. Thus, it is critical to investigate an automated ECG signal delineation with its result accuracy. In this study, we propose the delineation process by using bidirectional long short-term memory (BiLSTM) classifier. Such process was conducted as one beat to the next (beat-to-beat), that means the ECG waveform classification is start of P-wave1 to start of P-wave2. However, such classifier lack of feature extraction process, reducing the classification accuracy result. To improve the classifier performance, convolutional layers as facture extraction are stacked with BiLSTM named ConvBiLSTM. We conducted the experimental based on seven-class ECG waveform using a publicly available QT Database with annotation of the main waveforms to produce high accurate classifier, i.e., P start -Pend, Pend -QRSstart, QRSstart -Rpeak, Rpeak -QRSend, QRSend -Tstart, Tstart -Tend, and Tend -Pstart. It was found that the proposed model showed remarkable results with overall average performances of 99.83% accuracy, 98.82% sensitivity, 99.90% specificity, 98.86% precision, and 98.84% F1 score. Based on these promising results, the efficacy of the proposed stacked ConvBiLSTM model in classifying ECG waveform provides a great opportunity to help cardiologists in diagnosis decision-making for faster assessment.

DAE-ConvBiLSTM: End-to-end learning single-lead electrocardiogram signal for heart abnormalities detection

“…The DAE, a variant of the auto-encoder (AE), is composed of encoding and decoding layers; the encoding layer keeps the lower dimensional representation in the hidden layer, and the decoding layer extracts features to reconstruct the input [25]. With the excellent performance of the DAE to enhance the ECG signal conditions from noise and artifacts, this study aims to combine the DAE as the ECG denoising technique with our previous model, ConvBiLSTM, for detecting heart abnormalities [26]. The convolution layer as the feature extraction, part of convolutional neural networks (CNN) [27,28], focused solely on one-dimensional ECG signal data.…”

Short Single-Lead ECG Signal Delineation-Based Deep Learning: Implementation in Automatic Atrial Fibrillation Identification

“…The baseline wander noise must be removed from the raw signal to produce satisfactory results from the delineation process. These factors were removed by applying a discrete wavelet transformation (DWT) with low-pass and high-pass filters [ 14 , 21 ]. This strategy can separate the ECG signal into various frequency bands and maintain a good representation of a nonstationary signal [ 7 , 22 ].…”

“…A previous study achieved best performance when the ECG signal was delineated automatically with a CNNs and bidirectional LSTM (Bi-LSTM) architecture [ 21 ]. A unidirectional phase also has limits because future input information is impossible to get from the current state.…”

“…The feature extraction of CNNs process used four convolutional layers without max pooling. The convolution process can be expressed as follows [ 7 , 21 ]: where is the activation of the neuron of the filter for the convolutional layer, is the kernel size [ 8 ], is the neural activation function, is the shared bias of the filter, are the shared weights of the filter, and are the corresponding inputs.…”

“…The Bi-LSTM architecture processes a data sequence in both directions by two hidden layers, which are fed forward to the output layer and backward along with their corresponding weights ( ) and biases ( ). Two parallel directions of forwarding and backward passes are provided below [ 21 ]. …”