Automated arrhythmia classification based on a pyramid dense connectivity layer and BiLSTM

Abstract

Background

Deep neural networks (DNNs) have recently been significantly applied to automatic arrhythmia classification. However, their classification accuracy still has room for improvement.

Objectives

The aim of this study is to address the existing limitations in current models by developing a more effective approach for automatic arrhythmia classification. The specific objectives include enhancing the receptive field sizes to capture more detailed information across various temporal scales, and incorporating inter-channel correlations to improve the feature extraction process.

Methods

This study proposes a pyramidal dense connectivity layer and bidirectional long short-term memory network (PDC-BiLSTM) to effectively extract waveform features across various temporal scales, which can capture the intricate details and the broader global information in the signals through a wide range of sensory fields. The efficient channel attention (ECA) is additionally introduced to dynamically allocate weights to each feature channel, assisting the model inefficiently prioritizing essential characteristics during the training process.

Results

The experimental results on the MIT-BIH arrhythmia database showed that the overall classification accuracy of the proposed method under the intra-patient paradigm reached 99.82%, and the positive predictive value, sensitivity and F1 Score were 99.64%, 97.61% and 98.60% respectively; under the inter-patient paradigm, the overall accuracy was 96.30%.

Conclusion

Compared with the latest research results in this field, the proposed model is also better than the existing models in terms of accuracy, which has the potential value of being applied to devices that assist in diagnosing cardiovascular diseases.

Keywords

arrhythmia classification pyramidal convolutional layer densely connected convolutional network efficient channel attention

1 Introduction

According to the World Health Organization statistics, more than 17 million people die yearly from cardiovascular diseases.¹ Arrhythmia, a pervasive cardiovascular disorder, is characterized by its nuanced nature, intricate mechanisms, and abrupt onset. This ailment manifests as irregular heartbeats or alterations in rhythm, and in severe instances, can lead to cardiac arrest, stroke, or other consequential complications. The timely detection of cardiac illnesses is of utmost importance, as it is crucial for individuals to promptly identify abnormal heart rhythms in order to gain insight into their cardiac health and maybe prevent fatalities among patients.^{2, 3}

The Electrocardiogram (ECG) is the most potent tool for non-invasive diagnosis of cardiovascular conditions. It carries a wealth of data, reflecting the intricacies of heart activity and providing practical, detailed insights. Conventionally, the interpretation and classification of heartbeats have depended on individual clinical expertise and theoretical understanding, a process that is labor-intensive and places a substantial burden on healthcare resources.⁴ Moreover, this approach can be influenced by the subjective judgment of cardiologists. Consequently, developing a precise, real-time, AI-powered method for heartbeat classification, capable of autonomous learning from historical data, has emerged as a crucial area of research.⁵

With the advancement of artificial intelligence technology, machine learning and deep learning algorithms in arrhythmia classification are becoming increasingly widespread. Machine learning-based arrhythmia classification typically involves two main steps: feature extraction and classifier training. For example, Mondéjar et al.⁶ presented an arrhythmia classification model combining multiple support vector machines to extract features from ECG signals. Jung et al.⁷ enhanced the representative features of the original ECG signal by combining the coefficients obtained from principal component analysis (PCA) and linear discriminant analysis (LDA). They applied the weighted k-nearest neighbor (WKNN) algorithm alongside a fitness rule to classify the ECG data. Rahul et al.⁸ proposed an improved arrhythmia classification method based on RR interval. Wan et al.⁹ devised a method for atrial fibrillation detection that combines ensemble learning and multi-feature discrimination. This approach integrates multi-domain features and constructs a robust BSK-Model classifier by optimizing multiple single machine learning classifiers. The method underwent testing on the AFDB dataset, achieving high specificity and accuracy rates of up to 99%. Automatic classification using machine learning enhances diagnostic accuracy compared to manual classification. However, these methods still rely on researchers manually selecting features and have limited self-learning capabilities, thus exhibiting a high dependence on expert knowledge.

Deep learning methods have recently been successfully applied in extracting electrocardiogram information for arrhythmia classification and have demonstrated strong capabilities. For example, Hannun A et al.¹⁰ constructed a DNN with 33 convolutional layers to categorize 12 rhythms. They used a dataset of 91,232 single-lead electrocardiograms from 53,549 individuals collected using a single-lead ambulatory ECG monitoring device. Yang et al.¹¹ combined convolutional neural networks (CNN) and long short-term memory (LSTM) networks to automatically identify normal beats, ventricular premature beats, atrial premature beats, and unidentified beats using a multi-input structure for feature extraction and merging of large and small-scale heartbeats. Gao et al.¹² introduced the attention mechanism to optimize the effect of neural networks in processing time-series signals and, through the comparative analysis of the experimental results, demonstrated that the attention mechanism could significantly enhance the classification performance of neural networks. Zhan et al.¹³ proposed an arrhythmia classification method based on a densely connected convolutional network (DenseNet) and an efficient channel attention (ECA) mechanism, which successfully classified six different types of beats. Zhao et al.¹⁴ merged ResNet's frequency feature extraction capability with TCN's time domain analysis to introduce a TCN-ResNet model for efficiently classifying atrial fibrillation in single-lead electrocardiograms. The experiments demonstrated a high accuracy of 97% and an F1 score of 87%.

The aforementioned deep learning methods have demonstrated promising results in arrhythmia detection, yet certain challenges persist. These include the need for intricate convolutional structures in algorithms and the limitations of standard CNNs with fixed receptive field sizes, which can hinder effective feature capture. Additionally, traditional convolution techniques often overlook inter-channel connections, leading to potential redundancy and inefficiencies in feature extraction. To address these challenges, this study presents a pyramidal dense connectivity layer and bidirectional long short-term memory network (PDC-BiLSTM) model for accurate classification of arrhythmias, which combines a densely connected network based on the pyramidal convolutional layer (PC-DenseNet) and a bidirectional long short-term memory network (BiLSTM). The key innovation lies in using PC-DenseNet to enhance the receptive field and extract multi-scale deep features from ECG data. Additionally, the model incorporates ECA for adaptive channel weighting, reducing redundant information transmission. Furthermore, by integrating the BiLSTM model, temporal characteristics of ECG data are effectively captured.

In summary, the primary contributions of this study can be outlined as follows:

We innovatively constructed the PC-DenseNet, where the PC layer expands the scope of sensory fields. This architectural design enables the extraction of multi-scale features directly from raw ECG data while facilitating feature transfer and reuse through dense connectivity. Such characteristics endow our work with a unique advantage in ECG signal detection.

Another primary strength is our implementation of the ECA method, which dynamically assigns weights to each channel. This approach effectively suppresses the propagation of redundant information, thereby enhancing our model's discriminative capability.

Lastly, a notable aspect of our research is the superior performance exhibited by our model compared to state-of-the-art methods. The experimental results confirm the effectiveness and efficiency of the proposed method, which outperformed prior research findings across various metrics, including accuracy, sensitivity, positive predictive value, and F1 Score.

The paper is structured as follows: Section 2 explains the method, Section 3 presents the experiment and results, Section 4 describes the discussion, and Section 5 concludes the paper.

2 Method

The structure of the arrhythmia classification algorithm proposed in this paper is shown in Figure 1. The initial step involves performing wavelet denoising and heartbeat segmentation on the ECG signals affected by noise. Subsequently, the heartbeats are normalized, and the Borderline-SMOTE algorithm is employed to balance the data. Lastly, the PDC-BiLSTM network is developed to enable the automatic classification of arrhythmia using ECG signals.

Figure 1.

Flow chart of arrhythmia classification process.

2.1 ECG database

The MIT-BIH arrhythmia database, a collaboration between the Massachusetts Institute of Technology and Beth Israel Hospital,¹⁵ is currently recognized as the most extensively utilized in evaluating and validating arrhythmia classification algorithms. It comprises 48 dual-channel half-hour dynamic ECG recordings obtained from 47 patients(25 men aged 32 to 89 years and 22 women aged 23 to 89 years). The recordings comprised MLII (Modified Lead II) and V5 lead signals, both of which were sampled at a fixed frequency of 360 Hz. The data from the MIT-BIH arrhythmia database were categorized into five groups, as shown in Table 1, following the guidelines outlined in the ANSI/AAMI EC57: 2012 (AAMI) standard.¹⁶ These categories include normal beat (N), supraventricular premature beat (S or SVEB), ventricular ectopic beat (V or VEB), fusion beat (F), and unknown beat (Q). Notably, the four data points generated by pacemakers (102, 104, 107, 217) were excluded from the analysis, and category Q should have been excluded during the experimental phase in this work¹⁷ since it has no application value. This paper explores experiments conducted under two paradigms: intra-patient and inter-patient. Under the intra-patient paradigm, the dataset is randomly divided into training and test sets at a ratio of 7:3. Conversely, under the inter-patient paradigm, reflecting the variability between different patients, all records are partitioned into training set DS1 and test set DS2, as illustrated in Table 2.

Table 1.
Classification of heartbeats according to AAMI.

AAMI arrhythmia type Label Specific class

Non-ectopic beat(N) N Normal beat

L Left bundle branch block beat

R Right bundle branch block beat

e Atrial escape best

j Nodal (junctional) escape beat

Supraventricular ectopic beats(S) A Beat atrial premature contraction

a Aberrated atrial premature beat

J Nodal (Junctional) premature beat

S Premature or ectopic supraventricular beats

Ventricular ectopic beats(V) E Ventricular escape beat

V Premature ventricular contraction

Fusion beats(F) F Fusion of ventricular and normal beat

Unknown beats(Q) / Pace beat

f Fusion of paced and normal beat

Q Unclassified beats

AAMI arrhythmia type	Label	Specific class
Non-ectopic beat(N)	N	Normal beat
L	Left bundle branch block beat
R	Right bundle branch block beat
e	Atrial escape best
j	Nodal (junctional) escape beat
Supraventricular ectopic beats(S)	A	Beat atrial premature contraction
a	Aberrated atrial premature beat
J	Nodal (Junctional) premature beat
S	Premature or ectopic supraventricular beats
Ventricular ectopic beats(V)	E	Ventricular escape beat
V	Premature ventricular contraction
Fusion beats(F)	F	Fusion of ventricular and normal beat
Unknown beats(Q)	/	Pace beat
f	Fusion of paced and normal beat
Q	Unclassified beats

Table 2.

Segmentation of the dataset under the inter-patient.

Dataset	Patient recordings
DS1	101 106 108 109 112 114 115 116 118 119 122
(training)	124 201 203 205 207 208 209 215 220 223 230
DS2	100 103 105 111 113 117 121 123 200 202 210
(testing)	212 213 214 219 221 222 228 231 232 233 234

2.2 Data preprocessing

2.2.1 Wavelet denoising and heartbeat segmentation

In this paper, the discrete wavelet transform (DWT) is employed due to its effectiveness in analyzing non-stationary signals. The db6 wavelet basis function is utilized for denoising the data, involving the decomposition of ECG signals into nine levels. The denoised and smoothed ECG signal is obtained through the inverse wavelet transform from the third to the ninth level detail sub-bands.¹⁸ Figure 2(a) and Figure 2(b) represent ECG signals before and after denoising, respectively. Secondly, each consecutive ECG signal was segmented into heartbeats based on the annotated R peak position in the MIT-BIH arrhythmia database, and each processed heartbeat was standardized into a data segment consisting of 300 sample points to maintain input uniformity.¹⁹ Finally, the segmented heartbeats were Z-Score normalized to unify the signal scale and eliminate the offset effect. According to the AAMI standard, the corresponding label was annotated for each heartbeat, and Table 3 displays the number of heartbeats for each category in both intra-patient and inter-patient paradigms.

Figure 2.

ECG signals before and after denoising(109 record): (a) ECG signal before denoising, (b) ECG signal after denoising.

Table 3.

The number of samples in both intra-patient and inter-patient paradigms.

	Intra-patient		Inter-patient
AAMI arrhythmia type	Training set	Testing set	Training set	Testing set
N	63,026	27,011	45,799	44,238
V	4862	2084	3726	3220
S	1904	816	883	1837
F	542	232	386	388
Total	70,334	30,143	50,794	49,683

2.2.2 Borderline-SMOTE data balancing

The traditional SMOTE²⁰ algorithm enhances the intra-class aggregation of minority class samples by generating them based on nearest neighbor characteristics. However, it overlooks the distribution characteristics of majority class samples, potentially leading to duplication between classes and decreased recognition accuracy. In contrast, the Borderline-SMOTE²¹ algorithm addresses these limitations by selecting the most representative boundary samples to generate new samples. This approach helps overcome overfitting and intra-class confusion, thereby improving the accuracy of the classification results.²² To amplify the rare S and F classes in the database, the new samples are synthesized by first calculating the Euclidean distance among the K samples closest to the minority class sample X_i:

d (X, Y) = \sqrt{\sum_{i}^{m} {(x_{i} - y_{i})}^{2}}

(1)

In the formula, x_i and y_i represent two sample points in the two-dimensional space, and d(X,Y) denotes the Euclidean distance between these two sample points. To identify whether each minority class sample qualifies as a danger sample, we consider if over half of its K-nearest neighbor samples are from the majority class. If this condition is met, the sample is classified as a danger sample. Subsequently, Equation 2 demonstrates the application of linear processing on the identified danger samples.

X_{n e w} = X_{i} + δ \times (X_{z i} - X_{i}), 0 < δ < 1

(2)

Where X_new is the newly synthesized sample set, X_zi is the minority class sample set among the K-nearest neighbors of X_i.

2.3 Model construction

The proposed model PDC-BiLSTM in this study is divided into three main parts: the feature extractor PC-DenseNet, the ECA, and the temporal information extractor BiLSTM. Initially, it extracts and fuses features from diverse perspectives, enhancing the learning process with multiple DenseNet layers. Subsequently, the ECA mechanism is utilized to reduce redundancy and refine the feature selection. Furthermore, the temporal characteristics of the ECG data are comprehensively captured using the BiLSTM network. Finally, Softmax is employed for classification, facilitating more straightforward output interpretation. The specific structure is shown in Figure 3.

Figure 3.

Arrhythmia classification model based on PDC-BiLSTM.

2.3.1 PC-DenseNet network architecture

Traditional convolutional neural networks are limited by the inherent constraint of the receptive field in each convolutional layer, thereby restricting them to only a tiny portion of the input sequence and yielding limited feature information. To address this limitation and enhance the comprehensive extraction of features from ECG data, this paper introduces a novel architectural design, PC-DenseNet, as illustrated in Figure 4. The architecture comprises two distinct types of convolutional layers: the pyramidal convolutional layer (PC) and the DenseNet layer. The raw ECG signals undergo data preprocessing to obtain heartbeat segments of size 1 × 300, which are then fed as input to the PC layer. This layer consists of several parallel convolution branches with different kernel sizes, specifically set to 1 × 16, 1 × 8, and 1 × 4. Subsequently, features of different scales are concatenated to obtain cascaded features. Finally, a convolution kernel of size 1 × 1 is used to realize cross-channel feature fusion, thereby capturing global information. The structure of the PC layer is shown in Figure 4, and the mathematical expression is provided by Equation (3).

{\begin{cases} X = C o n v (x) \\ Y = C o n c a t (X_{1}, X_{2}, \dots X_{n}) \\ R = C o n v (Y) \end{cases}

(3)

Figure 4.

The structure of PC.

Where x represents the input ECG signal, Conv() is the convolution operation, X is the feature map obtained from one convolution branch, X_n represents the feature map obtained from the nth convolution branch, Concat() is the splicing operation, Y represents the multiscale feature after the n convolution branch feature maps have been spliced. R is the fusion feature obtained after passing through the convolution kernel of size 1 × 1.

In order to augment the network's capabilities of local feature extraction and feature reuse, the features fused post the PC layer are directed to two DenseNet²³ layers for further processing. As depicted in Figure 5, DenseNet primarily comprises of the dense block and the transition layer. The dense block, composed of a batch normalization layer (BN), Relu function, and Conv, facilitates improved information interchange between layers and appends dimensions to the output of each layer, thereby bolstering the transmission of ECG signal waveform features. Located between two densely connected blocks, the transition layer amalgamates the high-dimensional signals produced by the preceding dense block. This layer addresses the issue of network width and resolves the inability of the densely connected block to decrease the size of the feature map via pooling operations.

Figure 5.

The structure of DenseNet.

2.3.2 The efficient channel attention

In the process of arrhythmia classification, accurately capturing key features is crucial for correct classification. In the feature extraction process, the feature maps generated after PC-DenseNet often contain a large amount of redundant or similar information across different channels, which affects the performance of the model. Therefore, this paper considers embedding the efficient channel attention (ECA)²⁴ module after the PC-DenseNet network to adaptively allocate feature attention and assign differentiated weights to each extracted feature, enhancing the network's ability to represent important features in ECG signals. The structure of ECA, as illustrated in Figure 6, involves a series of steps to process the input matrix F0 with dimensions (C, H, W). Initially, the feature of unreduced dimensionality denoted as (C, 1, 1) is acquired through global average pooling (GAP). Subsequently, 1D convolution is applied to capture information specific to the channel domain. Lastly, the Sigmoid activation function normalized the feature weights to a range of [0, 1]. These weights are then utilized to perform a weighted combination with the input matrix F0, generating the feature matrix F1.

Figure 6.

The structure of ECA.

2.3.3 Bidirectional long short-term memory network

A long short-term memory network (LSTM) is a special RNN that can effectively tackle the problem of gradient disappearance and gradient explosion and record long-term dependencies, improving the network's performance while dealing with sequential data. However, LSTM models are limited to capturing solely past contextual information, hence needing more ability to incorporate future contextual information. In order to tackle this issue, a bidirectional long short-term memory network (BiLSTM) is implemented. As shown in Figure 7, this network consists of two LSTM units operating in opposite directions, which enhances the capacity to capture time series features. This paper uses the bidirectional temporal information processing capability of BiLSTM to fuse and process features extracted from PC-DenseNet at various scales. This enables the model to identify valuable patterns and regularities within a sequence of continuous ECG waveforms, thereby enhancing the classification accuracy of arrhythmias.

Figure 7.

The network structure of BiLSTM.

2.3.4 Model structure and parameters

To optimize the algorithm in this study, the Adam optimization method was utilized. During model training, the step size for one-dimensional maximum pooling (MaxPool1D) was set to 2, and the one-dimensional convolution (Conv1D) was configured as 1. The batch size was 128, the number of epochs was 50, and the learning rate was specified as 0.0005. The network model was trained using a focal loss function²⁵ (FL) to enhance classification performance. By adjusting the weights of the classification samples, the FL function guided the model's focus towards the challenging data from the S and F classes. The FL function is formally defined as follows:

f o c a l l o s s (p_{i}) = - α_{t} (1 - p_{t})^{γ} \log (p_{t})

(4)

Where p_t denotes the predicted probability of the model for that sample. γ is the focus factor, which is used to adjust the weights of the samples. α is the balance factor, which is used to balance the proportionality between positive and negative samples. In this paper, we refer to the literature,²⁶ and set the values of γ and α of the focal loss function to 2.0 and 0.25, respectively. The specific structural parameters of the model are shown in Table 4.

Table 4.

The parameters of the proposed model.

Number	Type	Name of the layer	Filters	Kernel size	Output
0	——	Input layer	——	——	300 × 1
1	Conv1	BN + Relu + Conv1D	32	16 × 1	300 × 32
2	Conv2	BN + Relu + Conv1D	32	8 × 1	300 × 32
3	Conv3	BN + Relu + Conv1D	32	4 × 1	300 × 32
4	——	Concatenate	——	——	300 × 96
5	Conv4	BN + Relu + Conv1D	32	1 × 1	300 × 32
6	DenseNet1	BN + Relu + Conv1D + Concatenate	16	5 × 1	300 × 32
7	DenseNet1	BN + Relu + Conv1D + Concatenate	16	5 × 1	300 × 48
8	DenseNet1	BN + Relu + Conv1D + Maxpool1D	32	5 × 1	150 × 32
9	DenseNet2	BN + Relu + Conv1D + Concatenate	32	5 × 1	150 × 64
10	DenseNet2	BN + Relu + Conv1D + Concatenate	32	5 × 1	150 × 96
11	DenseNet2	BN + Relu + Conv1D + Maxpool1D	48	5 × 1	75 × 48
12	ECA	——	——	——	75 × 48
13	BiLSTM	——	——	——	75 × 256
14	——	FC(Fully connected layer)	——	——	19,200
15	——	Softmax	——	——	4

3 Experiment and results

3.1 Experimental platform and evaluation indicators

The hardware configuration utilized to train the model in this study includes an Intel Xeon E5-2620 processor, 32GB of operating memory, an NVIDIA GeForce GTX 1080Ti graphics card, Keras for deep learning, and TensorFlow for the back-end. Windows Server 2016 Datacenter 64 was the operating system used to train the model. The four indicators of accuracy (Acc), sensitivity (Sen), specificity (Spe), positive predictive value (Ppv), and F1 Score (F1) are used to evaluate the performance of this model, and the calculation formulas are provided below:

\begin{aligned} A c c (%) & = \frac{T P + T N}{T P + F P + T N + F N} \times 100 % \end{aligned}

(5)

\begin{aligned} S e n (%) & = \frac{T P}{T P + F N} \times 100 % \end{aligned}

(6)

\begin{aligned} S p e (%) & = \frac{T N}{T N + F P} \times 100 % \end{aligned}

(7)

\begin{aligned} P p v (%) & = \frac{T P}{T P + F P} \times 100 % \end{aligned}

(8)

\begin{aligned} F 1 & = \frac{2 \times S e n \times P p v}{S e n + P p v} \times 100 % \end{aligned}

(9)

3.2 Data balance validity experiments

The experimental steps are as follows:

Under the intra-patient paradigm, the total number of sparse heartbeats, categorized as either S or F, is increased to 3810.

Under the inter-patient paradigm, the quantity of sparse cardiac beats, S and F, inside the training set DS1 is expanded to 7400.

The datasets under the two paradigms are inputted into the network model for training.

The confusion matrices of the classification results of the model before and after data balancing are shown in Figure 8. Under the intra-patient paradigm, the neural network demonstrates significant accuracy in correctly classifying heartbeats N, V, and S when trained with the original dataset. However, it fails to identify heartbeats belonging to class F, misclassifying them as class N. After using data balancing techniques, the identification rate for class F significantly improves, reaching an impressive 93%. Under the inter-patient paradigm, the application of Borderline-SMOTE for data balancing resulted in significant enhancements in the recognition accuracies of classes V, S, and F. Prior to the use of data balancing techniques, a considerable quantity of S and F heartbeats were erroneously categorized as N, rendering them nearly indistinguishable. However, after applying data balancing methods, the accurate recognition rates for S and F heartbeats showed notable improvement, surpassing 60%. Integrating the above analysis into the two paradigms highlights the necessity of achieving a balanced dataset.

Figure 8.

Confusion matrix before and after balancing the dataset: (a) before data balancing under the intra-patient, (b) after data balancing under the intra-patient, (c) before data balancing under the inter-patient, (d) after data balancing under the inter-patient.

3.3 Intra-patient paradigm

3.3.1 The performance of the model

By employing the experimental procedures outlined in Section 3.2 under the intra-patient paradigm, Table 5 presents the specific experimental results. Under the intra-patient paradigm, the model achieves an overall accuracy of 99.82%, Ppv of 99.64%, Sen of 97.61%, F1 of 98.60%, and Spe of 99.99%. Figure 9 and Figure 10 describe the trends of accuracy and loss values over time under the intra-patient paradigm, and the model did not suffer from severe overfitting.

Figure 9.

Change curves of accuracy under the intra-patient paradigm.

Figure 10.

Change curves of loss values under the intra-patient paradigm.

Table 5.

Experimental results under the intra-patient paradigm.

Class	Acc%	Ppv%	Sen%	F1%	Spe%
N	99.82	99.83	99.97	99.90	99.99
V	99.95	99.76	99.52	99.64	99.99
S	99.93	99.87	97.43	98.64	99.99
F	99.94	99.09	93.53	96.23	99.98
Average value	99.91	99.64	97.61	98.60	99.99
Overall accuracy	99.82

3.3.2 Ablation experiments

Under the intra-patient paradigm, ablation experiments were used to analyze the impact of the PC layer and ECA strategies on arrhythmia classification performance. As shown in Table 6, the Acc, Ppv, Sen, and F1 scores of model A2 increase by 0.94%, 7.17%, 8.63%, and 8.51%, respectively, compared to model A1. Notably, there is an improvement in the ability to classify limited samples into S and F categories. Both Sen and F1 scores for Class S and Class F increased by at least 15%. These results indicate that incorporating the PC layer into the model can improve its classification performance. Upon comparing A2 and A3 in the provided table, it is evident that the implementation of ECA results in an overall improvement of 0.53% in classification accuracy. More specifically, Sen values for the difficult-to-identify S and F categories exhibit 13.97% and 1.72% enhancements. The results above highlight the effectiveness of integrating ECA into the system.

Table 6.
Comparison of model classification performance results under the intra-patient.

Model Clsss Acc% Ppv% Sen% F1% Overall acc%

A1 N 98.42 98.60 99.65 99.12

V 99.74 97.79 98.50 98.14

S 98.89 94.46 62.75 75.41

F 99.64 75.92 77.59 76.75

Average value 99.17 91.69 84.62 87.36 98.35

A2 N 99.33 99.38 99.88 99.63

V 99.83 99.85 98.66 98.75

S 99.49 97.15 83.46 89.78

F 99.93 99.07 91.81 95.30

Average value 99.65 98.86 93.45 95.87 99.29

A3 N 99.82 99.83 99.97 99.90

V 99.95 99.76 99.52 99.64

S 99.93 99.87 97.43 98.64

F 99.94 99.09 93.53 96.23

Average value 99.91 99.64 97.61 98.60 99.82

Model	Clsss	Acc%	Ppv%	Sen%	F1%
A1	N	98.42	98.60	99.65	99.12
V	99.74	97.79	98.50	98.14
S	98.89	94.46	62.75	75.41
F	99.64	75.92	77.59	76.75
Average value	99.17	91.69	84.62	87.36	98.35
A2	N	99.33	99.38	99.88	99.63
V	99.83	99.85	98.66	98.75
S	99.49	97.15	83.46	89.78
F	99.93	99.07	91.81	95.30
Average value	99.65	98.86	93.45	95.87	99.29
A3	N	99.82	99.83	99.97	99.90
V	99.95	99.76	99.52	99.64
S	99.93	99.87	97.43	98.64
F	99.94	99.09	93.53	96.23
Average value	99.91	99.64	97.61	98.60	99.82

A1: DenseNet + BiLSTM, A2: A1 + PC, A3: A2 + ECA.

3.4 Inter-patient paradigm

3.4.1 The performance of the model

By applying the experimental procedure described in Section 3.2 to the inter-patient paradigm, Table 7 shows the specific performance of the proposed model. he model is effective in classifying categories N and V, with Sen reaching 97.95% and 95.75%, respectively. The relatively low Sen values for categories S and F are mainly due to the small number of samples and the significant differences in the shape of the beats of different patients when arrhythmia classification is performed under the inter-patient paradigm, which can lead to a reduction in classification performance. However, The overall accuracy of the model reaches 96.30%, which can still accurately classify important arrhythmia-related diseases.

Table 7.
Experimental results under the inter-patient paradigm.

Class Acc% Ppv% Sen% F1% Spe%

N 96.57 98.20 97.95 98.08 98.79

V 98.46 84.41 95.75 89.16 97.06

S 98.74 99.28 67.17 80.13 99.76

F 98.91 38.71 61.86 47.62 94.96

Average value 98.17 80.00 80.68 78.75 97.64

Overall accuracy 96.30

Class	Acc%	Ppv%	Sen%	F1%	Spe%
N	96.57	98.20	97.95	98.08	98.79
V	98.46	84.41	95.75	89.16	97.06
S	98.74	99.28	67.17	80.13	99.76
F	98.91	38.71	61.86	47.62	94.96
Average value	98.17	80.00	80.68	78.75	97.64
Overall accuracy	96.30

3.4.2 Ablation experiments

As shown in Table 8, an inter-patient performance comparison experiment is devised to examine the effect of PC layer and ECA strategies on the classification performance of arrhythmia across various classification models. Upon comparing A1 and A2 in the provided table, it is observed that including the PC layer substantially improves overall accuracy, increasing it from 92.60% to 94.68%. Furthermore, significant improvements are observed in two key metrics, the Sen and the F1 Score of the S class. Specifically, the Sen of the S class demonstrates a notable increase of 32.01%, while the F1 Score exhibits a substantial increase of 48.37%. These results presented illustrate the effectiveness of integrating the PC layer. After comparing A2 and A3 in the provided table, it is observed that the implementation of ECA led to an overall accuracy improvement of 1.62%. Furthermore, the Sen, Ppv, and F1 Score for the sparse sample categories S and F increased. This improvement signifies an improved capability to recognize the S and F categories to a certain extent, thereby indicating the effectiveness of incorporating ECA.

Table 8.
Comparison of model classification performance results under the inter-patient.

Model Clsss Acc% Ppv% Sen% F1% Overall acc%

A1 N 95.10 94.86 99.89 97.31

V 99.46 97.67 93.91 95.75

S 96.32 100 0.05 0.11

F 99.22 50 0.26 0.51

Average value 97.53 85.63 48.53 48.42 92.60

A2 N 94.85 96.43 97.84 97.13

V 97.82 75.51 98.32 85.42

S 97.48 99.33 32.06 48.48

F 99.20 16.67 0.52 1.00

Average value 97.34 71.99 57.19 58.01 94.65

A3 N 96.57 98.20 97.95 98.08

V 98.46 83.41 95.75 89.16

S 98.74 99.28 67.17 80.13

F 98.91 38.71 61.86 47.62

Average value 98.17 80.00 80.68 78.75 96.30

Model	Clsss	Acc%	Ppv%	Sen%	F1%
A1	N	95.10	94.86	99.89	97.31
V	99.46	97.67	93.91	95.75
S	96.32	100	0.05	0.11
F	99.22	50	0.26	0.51
Average value	97.53	85.63	48.53	48.42	92.60
A2	N	94.85	96.43	97.84	97.13
V	97.82	75.51	98.32	85.42
S	97.48	99.33	32.06	48.48
F	99.20	16.67	0.52	1.00
Average value	97.34	71.99	57.19	58.01	94.65
A3	N	96.57	98.20	97.95	98.08
V	98.46	83.41	95.75	89.16
S	98.74	99.28	67.17	80.13
F	98.91	38.71	61.86	47.62
Average value	98.17	80.00	80.68	78.75	96.30

A1: DenseNet + BiLSTM, A2: A1 + PC, A3: A2 + ECA.

4 Discussion

Classification of arrhythmias typically involves the utilization of machine learning techniques, two-dimensional images, and one-dimensional ECG signals. However, traditional machine learning methods occasionally require the manual generation and selection of features, which can be laborious and time-consuming. Furthermore, most approaches fail to adequately capture the temporal information and fully exploit the time-dependent characteristics of the ECG signal.^27–30 Numerous existing approaches to arrhythmia classification that employ 1D CNN concentrate predominantly on local feature extraction from the ECG signal. Nevertheless, these algorithms fail to sufficiently capture the feature information across various frequency ranges and scales.^31–34 Spatial information is extracted from ECG signals by classifying arrhythmias based on two-dimensional images. These techniques employ image operations, including scaling, translation, and rotation, to improve the data quality and the classification's performance. However, it is critical to acknowledge that this methodology is computationally intensive, necessitating substantial computational resources. Additionally, it is susceptible to information loss and requires more storage space to convert 1D signals to 2D images. This study presents an innovative approach by combining PC-DenseNet and BiLSTM as feature extractors for ECG data and temporal features, respectively. The initial step involves using the PC layer to extract the signals at various scales. This enables the capture of various aspects of ECG signals across multiple frequency ranges. As shown in Tables 6 and 8, the overall accuracy of ECG signals passing through the PC layer increased by 0.94% under the intra-patient paradigm and 2.05% under the inter-patient paradigm. Subsequently, dense connectivity is employed to minimize the loss of information transmission. Finally, the temporal dependency of ECG signals is modeled using BiLSTM.

Additionally, different channels contain different ECG signal features in the arrhythmia classification task. These features demonstrate a certain degree of connection among themselves. Meanwhile, ECG signals often encompass a substantial amount of redundant information, including noise and interference. In order to enhance the model's ability to build channel relationships and mitigate information redundancy, this study proposes integrating an efficient channel attention mechanism into the PC-DenseNet module. The channel attention mechanism is a method that effectively trains and directs attention toward the most pertinent channels or features within an input sequence. Tables 6 and 8 show that with the inclusion of ECA, the model collects key feature information more efficiently and performs better in both paradigms.

Furthermore, a notable disparity exists in the distribution of arrhythmia samples among various kinds in the arrhythmia classification task. This study employs the Borderline-SMOTE algorithm for data balancing to mitigate the issue of network training disproportionately emphasizing a more significant number of samples while inadequately learning from a few categories. Figure 8 shows the effectiveness of the Borderline-SMOTE, under the intra-patient paradigm, the number of correct classifications of difficult-to-categorize samples with algorithmic treatment grew from nearly unrecognizable to 217 samples for class F. Under the inter-patient paradigm, the number of correct classifications of difficult-to-categorize samples with classes S and F was also much higher than before the treatment, with a correct classification rate of more than 93%.

As shown in Table 9 and Figure 11, this paper compares the classical arrhythmia classification models in recent years. The focus of the literature^35,36 aligns with the aim of this paper in that both consider the model's performance in both intra-patient and inter-patient paradigms. Literature³⁵ and this study utilize DenseNet as the backbone network to extract features from ECG signals. However, it overlooks the correlation between each signal channel, resulting in insufficient recognition accuracy for a limited number of challenging samples that are difficult to identify. Literature³⁶ transforms one-dimensional signals into two-dimensional images for arrhythmia classification, effectively enhancing the accuracy of arrhythmia classification. However, the use of image-based input in the network significantly lengthens the processing time, posing a challenge to real-time arrhythmia classification. Literature³⁷ successfully combined the excellent feature extraction capability of CNN with the high sensitivity of gated recurrent units (GRU) to temporal signals, which performed well on the MIT-BIH arrhythmia database. However, this method needed to fully account for feature changes at different time scales; thus, its feature extraction capability could have been improved. Literature³⁸ uses the AdaBoost method to classify arrhythmias, but this method requires manual feature selection during the data preprocessing stage, which is time-consuming and labor-intensive. Literatures³⁹ and⁴⁰ both build networks from the perspective of extracting multi-scale features. Literature³⁹ proposed a multi-scale ResNet-based heartbeat classification method. This method first extracts shallow features from the convolutional layer. Then, it sends the feature map to three ResNet branches with different kernel sizes to combine receptive fields of different sizes. The method in this paper aims to expand the receptive field through the PC-DenseNet layer. The DenseNet layer is more concise and effective than three parallel complex ResNet branches. Its dense connection method allows the network to directly access the feature maps of all previous layers, reducing information loss and improving feature utilization. Literature⁴⁰ proposed a method for arrhythmia classification based on multi-scale convolution. This method adds convolution modules of different sizes to the front layer of CNN to extract richer feature information. This scheme considers the idea of multiple receptive fields, but its classification performance is not as good as the method proposed in this paper. The above analysis shows that the model proposed in this paper obtains high Acc, Spe, F1 Score, Sen, Spe and Ppv values for arrhythmia classification.

Figure 11.

Comparison with existing methods: (a) histogram comparing multiple methods under intra-patient paradigm, (b) histogram comparing multiple methods under inter-patient paradigm.

Table 9.

Comparison with existing methods.

Methods	Database	Type	Overall Acc/%	Ppv/%	Sen%	F1%
DenseNet³⁵	MIT-BIH	intra-patient	99.44	96.11	95.69	95.89
GASF + ResNet³⁶	MIT-BIH	intra-patient	99.52	96.41	96.13	96.27
CNN + GRU³⁷	MIT-BIH	intra-patient	99.56	98.76	97.92	98.33
RseNet³⁹	MIT-BIH	intra-patient	98.94	90.81	93.88	92.31
AdaBoost³⁸	MIT-BIH	inter-patient	95.60	62.20	81.20	66.20
DenseNet³⁵	MIT-BIH	inter-patient	92.37	60.35	68.29	63.49
GASF + ResNet³⁶	MIT-BIH	inter-patient	95.48	69.88	82.31	76.34
MCNN⁴⁰	MIT-BIH	inter-patient	95.60	61.23	69.88	66.63
This paper	MIT-BIH	intra-patient	99.82	99.64	97.61	98.60
This paper	MIT-BIH	inter-patient	96.30	80.00	80.68	78.75

5 Conclusion

This paper proposes a PDC-BiLSTM model to categorize four types of arrhythmias, N, V, S, and F, in both intra-patient and inter-patient paradigms. Multi-scale features are created using distinct convolutional branches in different feature channel dimensions, based on which the transfer and reuse of features are promoted using a dense connection. Meanwhile, ECA is introduced to focus on crucial aspects and eliminate information duplication. The model developed in this research boosts the capacity to describe multi-scale feature information and substantially improves classification resilience and accuracy. Subsequent studies will attempt to apply the proposed model on portable devices to enable doctors to conduct electrocardiogram analysis promptly and offer improved, personalized medical care for enhancing arrhythmia diagnosis and ensuring timely treatment.

Footnotes

Acknowledgments

The authors have no acknowledgments.

ORCID iDs

Xiangkui Wan

Yunfan Chen

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the Wuhan Knowledge Innovation Project (No. 2022020801010258) and Natural Science Foundation of Hubei Province (No. 2022CFA007).

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

References

Roth

Mensah

Johnson

, et al. Global burden of cardiovascular diseases and risk factors, 1990–2019: update from the GBD 2019 study. J Am Coll Cardiol. 2020; 76: 2982–3021. doi:https://doi.org/10.1016/j.jacc.2020.11.010

Quadri

. Analysing and evaluating the performance of deep-learning-based arrhythmia detection using electrocardiogram signals. Int J Adv Res Comput Sci. 2024; 15: 74. doi:https://doi.org/10.26483/ijarcs.v15i2.7065

Ansari

Mourad

Qaraqe

, et al. Deep learning for ECG arrhythmia detection and classification: an overview of progress for period 2017–2023. Front Physiol. 2023; 14: 1246746. doi:https://doi.org/10.3389/fphys.2023.1246746

Jha

Kolekar

. Electrocardiogram data compression techniques for cardiac healthcare systems: a methodological review. IRBM. 2020; 43: 217–228. doi:https://doi.org/10.1016/j.irbm.2021.06.007

Liu

Jin

, et al. A review of arrhythmia detection based on electrocardiogram with artificial intelligence. Expert Rev Med Devices. 2022; 19: 549–560. doi:https://doi.org/10.1080/17434440.2022.2115887

Mondéjar-Guerra

Novo

Rouco

, et al. Heartbeat classification fusing temporal and morphological information of ECGs via ensemble of classifiers. Biomed Signal Process Control. 2019; 47: 41–48. doi:https://doi.org/10.1016/j.bspc.2018.08.007

Jung

Lee

. An arrhythmia classification method in utilizing the weighted KNN and the fitness rule. IRBM. 2020; 38: 138–148. doi:https://doi.org/10.1016/j.irbm.2017.04.002

Rahul

Sora

Sharma

, et al. An improved cardiac arrhythmia classification using an RR interval-based approach. Biocybern Biomed Eng. 2021; 41: 656–666. doi:https://doi.org/10.1016/j.bbe.2021.04.004

Wan

Liu

Mei

, et al. A novel atrial fibrillation automatic detection algorithm based on ensemble learning and multi-feature discrimination. Med Biol Eng Comput. 2024; 62: 1809–1820. doi:https://doi.org/10.1007/s11517-024-03046-7

10.

Hannun

Rajpurkar

Haghpanahi

, et al. Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network. Nat Med. 2019; 25: 65–69. doi:https://doi.org/10.1038/s41591-018-0268-3

11.

Yang

Huang

Cai

, et al. Arrhythmia beat classification model based on CNN and BiLSTM. Chin J Biomed Eng. 2020; 39: 719–726. doi:https://doi.org/10.1109/ICSAI48974.2019.9010393

12.

Gao

Wang

Liu

. A novel approach for atrial fibrillation signal identification based on temporal attention mechanism. IEEE Eng Med Biol Soc. 2020; 42: 316–319. doi:https://doi.org/10.1109/EMBC44109.2020.9175823

13.

Zhan

, et al. A novel CNN model with dense connectivity and attention mechanism for arrhythmia classification. 2022 IEEE 35th International Symposium on Computer-Based Medical Systems(CBMS). 2022; 35: 50–55. doi:https://doi.org/10.1109/CBMS55023.2022.00016

14.

Zhao

Zhou

Ning

, et al. Atrial fibrillation detection with single-lead electrocardiogram based on temporal convolutional network-ResNet. Sensors. 2024; 24: 98. doi:https://doi.org/10.3390/s24020398

15.

Moody

Mark

. The impact of the MIT-BIH arrhythmia database. IEEE Eng Med Biol Mag. 2001; 20: 45–50. doi:https://doi.org/10.1109/51.932724

16.

ANSI/AAMI EC57:2012/(R)2020. Testing and reporting performance results of cardiac rhythm and ST segment measurement algorithms. 2013. doi: https://doi.org/10.2345/9781570204784.ch1.

17.

Essa

Xie

. An ensemble of deep learning-based multi-model for ECG heartbeats arrhythmia classification. IEEE Access. 2021; 9: 103452–103464. doi:https://doi.org/10.1109/ACCESS.2021.3098986

18.

Chen

Liu

, et al. An atrial fibrillation detection algorithm based on lightweight design architecture and feature fusion strategy. Biomed Signal Process Control. 2024; 91: 106016. doi:https://doi.org/10.1016/j.bspc.2024.106016

19.

Wang

Yang

, et al. Arrhythmia classification algorithm based on multi-head self attention mechanism. Biomed Signal Process Control. 2023; 79: 104206. doi:https://doi.org/10.1016/j.bspc.2022.104206

20.

Chawla

Bowyer

, et al. SMOTE: synthetic minority over-sampling technique. J Artif Intell Res. 2002; 16: 321–357. doi:https://doi.org/10.1613/jair.953

21.

Han

Wang

Mao

. Borderline-SMOTE: A new over-sampling method in imbalanced data sets learning. In: International Conference on Intelligent Computing , 2005, pp. 878–887. doi: https://doi.org/10.1007/11538059_91

22.

Vitianingsih

Othman

Kama

, et al. Application of the synthetic over-sampling method to increase the sensitivity of algorithm classification for class imbalance in small spatial datasets. Intell Eng Syst. 2022; 15: 676-690. doi:https://doi.org/10.22266/ijies2022.1031.58

23.

Huang

Liu

Van Der Maaten

, et al. Densely connected convolutional networks. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2017, pp. 4700–4708. doi: https://doi.org/10.1109/CVPR.2017.243

24.

Wang

Zhu

, et al. ECA-Net: efficient channel attention for deep convolutional neural networks. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2020, pp. 11531–11539. doi:https://doi.org/10.1109/CVPR42600.2020.01155

25.

Lin

Goyal

Girshick

, et al. Focal loss for dense object detection. In: IEEE International Conference on Computer Vision (ICCV) , 2017, pp. 2980–2988. doi:https://doi.org/10.48550/ar-Xiv.1708.02002

26.

Qian

. Inter-patient arrhythmia classification with improved deep residual convolutional neural network. Comput Methods Programs Biomed. 2022; 214: 106582. doi:https://doi.org/10.1016/j.cmpb.2021.106582

27.

Zhou

. ECG Classification using wavelet packet entropy and random forests. Entropy. 2016; 18: 285–294. doi:https://doi.org/10.3390/e18080285

28.

Gupta

Mittal

. A novel FrWT based arrhythmia detection in ECG signal using YWARA and PCA. Wirel Pers Commun. 2022; 124: 1229–1246. doi:https://doi.org/10.1007/s11277-021-09403-1

29.

Sahoo

Dash

Behera

, et al. Machine learning approach to detect cardiac arrhythmias in ECG signals: a survey. IRBM. 2020; 41: 185–194. doi:https://doi.org/10.1016/j.irbm.2019.12.001

30.

Gupta

Mittal

. QRS complex detection using STFT, chaos analysis, and PCA in standard and real-time ECG databases. J Inst Eng (India): Series B. 2019; 100: 489–497. doi:https://doi.org/10.1007/s40031-019-00398-9

31.

Degirmenci

Ozdemir

Izci

, et al. Arrhythmic heartbeat classification using 2d convolutional neural networks. IRBM. 2022; 43: 422–433. doi:https://doi.org/10.1016/j.irbm.2021.04.002

32.

Wang

Zang

, et al. A pooling convolution model for multi-classification of ECG and PCG signals. Comput Methods Biomech Biomed Engin 2024; 1–14. doi:https://doi.org/10.1080/10255842.2023.2299697

33.

Balamurugan

Meera

. Hybrid optimized temporal convolutional networks with long short-term memory for heart disease prediction with deep features. Comput Methods Biomech Biomed Engin. 2024; 1–25. doi:https://doi.org/10.1080/10255842.2024.2310075

34.

Çınar

Tuncer

. Classification of normal sinus rhythm, abnormal arrhythmia and congestive heart failure ECG signals using LSTM and hybrid CNN-SVM deep neural networks. Comput Methods Biomech Biomed Engin. 2021; 24: 203–214. doi:https://doi.org/10.1080/10255842.2020.1821192

35.

Gan

Shi

, et al. Parallel classification model of arrhythmia based on DenseNet-BiLSTM. Biocybern Biomed Eng. 2021; 41: 1548–1560. doi:https://doi.org/10.1016/j.bbe.2021.09.001

36.

Wan

Luo

Liu

, et al. An image classification method for arrhythmias based on Gramian angular summation field and improved Inception-ResNet-v2. J Biomed Eng. 2023; 40: 465–473. doi:https://doi.org/10.7507/1001-5515.202207049

37.

Zhu

. ECG Classification exercise health analysis algorithm based on GRU and convolutional neural network. IEEE Access. 2024; 12: 59842–59850. doi:https://doi.org/10.1109/ACCESS.2024.3392965

38.

Wang

, et al. Imbalanced heartbeat classification using EasyEnsemble technique and global heartbeat information. Biomed Signal Process Control. 2022; 71: 103105. doi:https://doi.org/10.1016/j.bspc.2021.103105

39.

Zhang

Yang

Huang

, et al. Multi-scale and attention based ResNet for heartbeat classification. In: 2020 25th International Conference on Pattern Recognition (ICPR) , 2021, pp. 1529–1535. doi:https://doi.org/10.1109/ICPR48806.2021.9413052

40.

Zhou

Sun

Wang

. Inter-patient ECG arrhythmia heartbeat classification network based on multiscale convolution and FCBA. Biomed Signal Process Control. 2024; 90: 105789. doi:https://doi.org/10.1016/j.bspc.2023.105789

Automated arrhythmia classification based on a pyramid dense connectivity layer and BiLSTM

Abstract

Background

Objectives

Methods

Results

Conclusion

Keywords

1 Introduction

2 Method

2.2.1 Wavelet denoising and heartbeat segmentation

3.1 Experimental platform and evaluation indicators

3.3.1 The performance of the model

3.4.1 The performance of the model

Table 7. Experimental results under the inter-patient paradigm. Class Acc% Ppv% Sen% F1% Spe% N 96.57 98.20 97.95 98.08 98.79 V 98.46 84.41 95.75 89.16 97.06 S 98.74 99.28 67.17 80.13 99.76 F 98.91 38.71 61.86 47.62 94.96 Average value 98.17 80.00 80.68 78.75 97.64 Overall accuracy 96.30

Footnotes

Acknowledgments

ORCID iDs

Funding

Declaration of conflicting interests

References

Table 7.
Experimental results under the inter-patient paradigm.

Class Acc% Ppv% Sen% F1% Spe%

N 96.57 98.20 97.95 98.08 98.79

V 98.46 84.41 95.75 89.16 97.06

S 98.74 99.28 67.17 80.13 99.76

F 98.91 38.71 61.86 47.62 94.96

Average value 98.17 80.00 80.68 78.75 97.64

Overall accuracy 96.30