iCOR: End-to-End Electrocardiography Morphology Classification Combining Multi-Layer Filter and BiLSTM

Abstract

Accurate delineation of ECG signals is critical for effective cardiovascular diagnosis and treatment. However, previous studies indicate that models developed for specific datasets and environments perform poorly when used with varying ECG signal morphology characteristics. This paper presents a novel approach to ECG signal delineation using a multi-layer filter (MLF) combined with a bidirectional long short-term memory (BiLSTM) model, namely iCOR. The proposed iCOR architecture enhances noise removal and feature extraction, resulting in improved classification of the P-QRS-T-wave morphology with a simpler model. Our method is evaluated on a combination of two standard ECG databases, the Lobachevsky University Electrocardiography Database (LUDB) and QT Database (QTDB). It can be observed that the classification performance for unseen sets of LUDB datasets yields above 90.4% and 98% accuracy, for record-based and beat-based approaches, respectively. Beat-based approaches outperformed the record-based approach in overall performance metric results. Similar results were shown in an unseen set of the QTDB, in which beat-based approaches performed with accuracy above 97%. These results highlight the robustness and efficacy of the iCOR model across diverse ECG signal datasets. The proposed approach offers a significant advancement in ECG signal analysis, paving the way for more reliable and precise cardiac health monitoring.

1. Introduction

To accurately diagnose cardiac abnormalities, the morphology of the electrocardiographic (ECG) signal must be analyzed. However, the use of non-invasive ECG in clinical practice remains limited due to its low signal-to-noise ratio (SNR) [1]. Noises that complicate the extraction of non-invasive ECG include instrumental noise, baseline wander, and motion artifacts [2]. These noises can distort the ECG waveform, leading to the misinterpretation of cardiac information [2,3]. Hence, it is desirable to develop denoising methods that provide morphology ECG signals with an accurate representation of cardiac status.

Several methods have been proposed to remove noise from ECG signals, and they can be grouped into [4] model-based methods [5,6,7,8] and learning-based methods [3,9,10,11]. Deep-learning-based (DL) methods have recently achieved incredible success in various fields, including in ECG signal denoising, which is crucial for accurate diagnosis. Several DL methods have been proposed such as the adversarial DL approach [2]. This method involves training a deep neural network to learn a sequence of non-linear transformations that can take noisy ECG signals as input and produce clean signals as output. Another method is based on disentangled autoencoders [12]. This approach involves introducing a disentangled mechanism to separate the signal features from the noise features. The disentangled autoencoder network (DANet) uses attention to shunt latent variables, deconstructing the potential spatial separation between noise and the original ECG data. Supervised learning-based methods like the denoising convolutional autoencoder (FCN-DAE) [13], traditional methods like principal component analysis (PCA) [14], and discrete wavelet transformation (DWT) [15] are also used.

However, previous researchers primarily focused only on removing ECG noise and analyzing the results [16,17,18]. This study, however, goes further by delineating the denoised ECG signal to identify abnormalities in the ECG wave morphology through P-QRS-T-wave delineation [19,20,21,22]. ECG analysis involves understanding how to extract the morphology of ECG waves and segments (delineation). ECG signal delineation is a critical step that involves accurately identifying key feature points such as the onset and offset of P-waves, QRS-complexes, and T-waves [19,20,21,22]. However, accurately identifying the start, peak, and endpoint of the three primary ECG waveforms can be quite challenging. Moreover, diagnosing conditions through ECG signal analysis is a complex and time-consuming process that requires extensive training and specialized expertise. ECG delineation methods are broadly classified into two categories: digital signal processing [23,24] and intelligent processing [25,26,27,28]. Extensive research has been conducted on both approaches [23,24,25,26,27,28]. However, certain challenges persist, including the insufficient examination of noise effects in ECG signal delineation methods [20].

Numerous studies have proposed various ECG delineation models, demonstrating promising results [19,20,21,22,29,30]. However, accurately detecting and distinguishing between the P-, QRS-, and T-waves, handling noise and artifacts in the signal, and managing variations in wave morphology across different individuals and conditions remain challenges [11]. In addition, most previous studies only delineated normal signals as a baseline for interpreting abnormal morphology [21]. To address the limitations in recognizing abnormalities in ECG signals, this research makes the following contributions:

• Develop iCOR, an end-to-end DL model with a comprehensive ECG signal processing system incorporating noise removal, feature extraction, learning process, and morphology classification;
• Combine noise removal and feature extraction of ECG signals using multi-layer filters with DL-based signal delineation of normal and abnormal pathology to enhance generalization capabilities; and
• Evaluate with unseen data to accurately assess its performance and accuracy.

These steps aim to build a robust and accurate system for ECG signal interpretation, leveraging advanced DL techniques for improved delineation and reliability.

2. Materials and Methods

2.1. Data Preparation

Data preparation precedes ECG pre-processing, involving three datasets: MIT-BIH Noise Stress Test Database v.1.0.0 (NSTDB) [31], QT Database v.1.0.0 (QTDB) [32], and Lobachevsky University Electrocardiography Database v.1.0.1 (LUDB) [33]. The metadata of the three databases are shown in

Table 1. The summary of NSTDB, QTDB, and LUDB.

Property	NSTDB	QTDB	LUDB
Frequency sampling	360 Hz	250 Hz	500 Hz
Number of channels	2	2	12
The length of the signal (number of nodes)	650,000	224,999	5000
Unit	mV	mV	mV
Channel name	Noise1 and Noise2	V2, mod.V1, ECG1, D4, CM4, CM5, V2-V3, V1-V2, V5, MLII, V1, V4, ECG2, V3, CM2, ML5, D3, V4-V5, and CC5	i, ii, iii, aVR, aVL, aVF, and V1–V6
Number of records	6	105	200
Pathology	ECG signal recording with baseline wander	Normal sinus Arrhythmia ST-change Supraventricular arrhythmia European ST-T Sudden death	Sinus rhythm Sinus tachycardia Sinus bradycardia Sinus arrhythmia Irregular sinus rhythm Abnormal rhythm
Function	Denoising ground truth	Denoising and delineation process	Delineation process

. In this study, the noise removal model utilizes the QTDB and NSTDB, while the delineation model uses both the QTDB and LUDB.

NSTDB contains ECG signals with three types of noise, i.e., baseline wandering (BW), electrode motion (EM), and muscle artifact (MA). BW is an undesired low-frequency fluctuation and interferes with ECG signal, as it can alter wave amplitudes and cause errors in the detection of important features [34]. EM is a disturbance in the ECG signal that typically appears as sudden fluctuations and can affect the accuracy of the information produced [35]. Meanwhile, MA appears as a high-frequency signal that can interfere with the identification of important ECG wave morphologies [36]. The sample of three noises from the NSTDB dataset is seen in Figure 1.

QTDB contains real ECG records representing a wide range of QRS and ST-T morphologies with considerable variability. It comprises 105 ECG signals, each 15 min long, with two channels sampled at 250 Hz. To incorporate real baseline drifts into the QTDB, real BW signals from the NSTDB were used. The NSTDB includes 30 min of 12 ECG recordings and three recordings of typical noise encountered in stress tests, sampled at 360 Hz. On the other hand, the LUDB is used exclusively to evaluate the delineation model. It contains real ECG records from 200 subjects, sampled at 500 Hz across 12 leads. The QTDB beats are up-sampled from 250 Hz to 360 Hz, conversely, for the delineation model with LUDB, beats are down-sampled from 500 Hz to 360 Hz, ensuring consistency with the NSTDB frequency of 360 Hz.

2.2. Data Pre-Processing

In the realm of ECG signal analysis, data pre-processing is pivotal to ensure the efficacy of subsequent models [37]. There are two primary models under consideration:

(i) The denoising model: the ECG beats are resampled to ensure uniformity in the data. We have up-sampled the QTDB to 360 Hz, the same as NSTDB’s frequency sampling. We have only experimented with the complete ECG beat, which consisted of P-waves, QRS-complexes, and T-waves. If the length of each beat is not fixed, padding is applied to standardize the length of each beat, accommodating the model’s input requirements. Padding is used to ensure that all model input sizes are consistent. Both the ECG signal noise removal and the ECG signal delineation model receive input in the form of vectors with a size of 512 nodes. Typically, the number of nodes in a beat varies between 200 and 300. To establish consistency, the size of 512 nodes was chosen for the input and output layers during the model training process. The reason 512 nodes were chosen is that the lift is not smaller than the range of variations in the number of nodes in one beat. Then, the dataset is then split into training, validation, and test sets, ensuring that the model can be appropriately trained and evaluated.

Lastly, the ground truth signals are separated from the model inputs to facilitate accurate training and validation. The ground truth for the denoising model consists of the raw QTDB signal, which serves as the clean reference signal. To generate the noisy input signal for the denoising model, the amplitudes of the QTDB and NSTDB signals are combined. The NSTDB contributes baseline wandering noise, simulating the slow fluctuations caused by patient breathing. Additionally, to mimic electrode motion artifacts, a pseudo-random number generator is employed. To inject noise into the ECG signals, we adopted the approach used by the NSTDB as described in [11]. Noise is randomly introduced with amplitudes ranging from 0.2 to 2 times the maximum peak value of the ECG signal.

(ii) The delineation model: for the delineation process, we down-sampled the LUDB to 360 Hz, the same as the frequency sampling of the NSTDB. The delineation model shares some of these pre-processing steps, including resampling beats, padding beats, and splitting the dataset. However, a key additional step in the delineation model’s pre-processing is the inference of beats from the best denoising model. This step leverages the outputs of the denoising model to enhance the accuracy and reliability of the delineation process, ensuring that the model works with the cleanest possible signals. A sample of ECG in the QTDB and LUDB for the delineation process is presented in Figure 2.

2.3. iCOR Proposed Model

This study proposes a multi-layer filter (MLF) architecture used for noise removal and delineation processes on ECG signals. In the noise removal process, the MLF model with 1D CNN on the output layer applies the regression principle. Meanwhile, for the ECG delineation process, the MLF model with BiLSTM on the output layer applies the classification principle. This end-to-end learning process is called iCOR. The name of iCOR comprises intelligent and COR. COR comes from the Latin word for “heart”. COR is often used in medical terminology related to cardiology. Therefore, this study was developed the intelligent system for AI-powered systems for analyzing the heart using ECG signals. iCOR for ECG signals is presented in Figure 3. Figure 3 shows a two-step process involving noise removal and delineation that is described below:

The denoising process employs an MLF to reduce noise. The MLF model was adopted using principles from previous research [11]. In [11], they propose a novel algorithm for baseline wander noise filtering using deep learning techniques. From [11], the best-performing model identified is the deep filter model called Multibranch Linear and Non-Linear Dilated (MBLANLD), which achieved the highest performance in denoising. With the adopted MBLANLD, in this study, an MLF is simultaneously used for noise removal and feature extraction from the ECG signal. The ECG signals, with interference from some noise, undergo processing through several layers of filter modules, each with varying numbers of filters and dilation rates to control the sensitivity and scale of the filtering. The ECG noise removal includes an output layer with a 1D CNN for a regressor. In the denoising model that is presented in Figure 3, the output layer consists of a 1D CNN with 9 kernels, utilizing a linear activation function. The first two modules, (b) and (c), consist of a total of 64 convolutional filters and follow the internal structure. In the second multipath module (c), convolutional operations were applied with a dilation rate of (r = 3), enabling dilated convolutions. The same dilation rate was also used in the fourth and sixth modules. As the network progresses, the number of extracted features gradually decreases. The first two modules contain 64 features, followed by 32 features in the third and fourth modules. Finally, the fifth and sixth modules each extract 16 features. The final stage consists of a single convolutional filter with a kernel size of 9, which shapes the output signal.

The process of delineation involves the categorization of modules into two main groups: frozen-weight, also known as problem-independent, and trainable-weight, which are problem-dependent. Within the frozen-weight modules, the weights are pre-trained and fixed, facilitating the processing of clean input data by passing them through various filtering stages to establish a baseline for the denoising procedure. Concurrently, the noisy input data undergo processing by the trainable-weight modules, where the weights are adaptable throughout the training phase. These modules work to enhance the data, progressively converting the noisy input into a more refined version. The ultimate result is a visualization of the delineated data that accentuates specific characteristics, thus concluding the dual-stage process of purification and feature extraction. This design effectively integrates both problem-independent and problem-dependent modules to offer broad applicability across diverse datasets while still maintaining precision for particular objectives. The model’s outcome will deliver an ECG signal that has undergone denoising and has been categorized into four morphological classes: P-waves, QRS-complexes, T-waves, and isoelectric lines. To further refine the classification process, a combination of BiLSTM and dense layers is employed.

By generating the cohesive pipeline, the methodology of this study has the ability to enhance the ECG signal quality through noise removal, then extracts essential features using the MLF, and finally applies a BiLSTM model to accurately delineate ECG wave components. This combination leverages the strengths of both DL models to improve the accuracy and reliability of ECG morphology recognition and analysis (Figure 3). Figure 3 shows the input signal data (Figure 3a), the multi-layer filter modules (Figure 3b,c), the convolution layers and activation function (Figure 3d,e), and the output (Figure 3f).

2.4. Performance Metric Evaluation

The model will produce an ECG signal that has been denoised and classified into four categories: P-waves, QRS-complexes waves, T-waves, and isoelectric lines. Its performance will be evaluated using both regression and classification metrics to provide a comprehensive assessment of accuracy and effectiveness [38]. This dual approach ensures the model’s predictions are accurate, distinguishes between different classes reliably, and effectively removes noise from ECG signals. Regression metrics include maximum absolute distance (MAD) to measure the largest absolute difference between the predicted values and the actual values,

$$MAD\left(s_1,s_2\right)=\mathrm{m}\mathrm{a}\mathrm{x}\left|s_1\left(n\right)-s_2\left(n\right)\right| 1\le n\le N$$

(1)

The sum of squared distances (SSD) between the predicted values and the actual values is

$$SSD\left(s_1,s_2\right)={\sum }_{n=1}^N{(s_2\left(n\right)-s_1\left(n\right))}^2 1\le n\le N$$

(2)

Percentage root-mean-square difference (PRD) represents the root-mean-square difference between the predicted values and the actual values as a percentage,

$$PRD\left(s_1,s_2\right)=100\mathrm{\%}·\sqrt{{\displaystyle \frac{{\sum }_{n=1}^N{(s_2\left(n\right)-s_1\left(n\right))}^2}{\sum _{n=1}^N{(s_2\left(n\right)-\overline{s_1\left(n\right)})}^2}}} 1\le n\le N$$

(3)

Cosine similarity (CosSim) measures the cosine of the angle between two non-zero vectors,

$$CosSim\left(s_1,s_2\right)={\displaystyle \frac{\sum _{n=1}^N(s_1\left(n\right)·s_2\left(n\right))}{\sqrt{\sum _{n=1}^N{s}_1^2\left(n\right)}·\sqrt{\sum _{n=1}^N{s}_2^2\left(n\right)}}}={\displaystyle \frac{⟨s_1,s_2⟩}{\left|\right|s_1\left|\right|\left|\right|s_2\left|\right|}} 1\le n\le N$$

(4)

Signal-to-noise ratio (SNR) measures the ratio of the power of the signal to the power of the noise,

$$SNR\left(s_1,s_2\right)=10{\mathit{log}}_{10}\left[{\displaystyle \frac{\sum _{n=1}^N{s}_2^2\left(n\right)}{\sum _{n=1}^N{\left(s_2\left(n\right)-s_1\left(n\right)\right)}^2}}\right] 1\le n\le N$$

(5)

where $s_1$ is the original signal, $s_2$are the filtered signals, $n$ is the index of the current sample, and $N$ is the length of the signals.

The metrics provide various perspectives on the accuracy and reliability of regression models, helping to identify both average performance and specific areas of strength or weakness in the predictions. The classification metrics are assessed using a confusion matrix, including accuracy, sensitivity, precision, and F1 score [39]. However, in this study, we highlight the F1 score. In the ECG signal delineation study, the F1 score allows us to evaluate how well our model distinguishes between the P-wave, QRS-complex, T-wave, and isoelectric line classes across unseen data. By using the F1 score, we ensure that the proposed model maintains a good balance between identifying true positives and minimizing false positives and false negatives, leading to more reliable and accurate ECG signal classification.

2.5. Platform Implementation

In this study, we deployed the proposed model on a computer with the following specifications: an Intel^® Core™ i9-9920X CPU @ 3.50 GHz (Santa Clara, CA, USA) as the processor, 490,191 MB RAM, and a GeForce 2080 RTX Ti GPU manufactured by NVIDIA Corporation (GV102, rev a1) (Santa Clara, CA, USA). The operating system used was Ubuntu 18.04.5 LTS. All experiments and algorithms were implemented using Python and the PyTorch 1.7.1 library for model development.

3. Results and Discussion

In this study, the results are from two approaches: ECG denoising and delineation. For ECG denoising, we have experimented two channels of the NSTDB (Noise1 and Noise2 for training and testing set, respectively) and one channel (.pu0) of the QTDB (14 records for testing set and the remaining records for 70% training and 30% validation sets). For the ECG denoising process, this study was only concerned with BW noise. For ECG delineation, among 200 records, there were 23 records for the testing set and the remaining records for 90% training and 10% validation sets. The record distributions of the NSTDB, QTDB, and LUDB are listed in

Table 2. The record distributions of NSTDB, QTDB, and LUDB.

Database	Aim	ECG Records
		Training Set	Validation Set	Testing Set (Unseen)
NSTDB	Denoising	Noise1 (BW)	-	Noise 2 (BW)
QTDB	Denoising and delineation (unseen)	Remaining records	Remaining records	MIT-BIH Arrhythmia (sel123 and sel233), MIT-BIH ST Change (sel302 and sel307), MIT-BIH Supraventricular Arrhythmia (sel820 and sel853), MIT-BIH Normal Sinus Rhythm (sel16420 and sel1679), European ST-T database (sele0106 and sele0121), Sudden death patients (sel32 and sel49), MIT-BIH Long-Term ECG (sel14046 and sel15814).
LUDB	Delineation	Remaining records	Remaining records	Sinus rhythm (2), Sinus bradycardia (1), Sinus arrhythmia (22), Sinus tachycardia (70), Irregular sinus rhythm (108), Atrial Fibrillation (8, 38, 44, 51, 83, 88, 93, 95, 96, 101, 110, 112, 129, 173), and Atrial Flutter (35, 52, 103).

.

For the results analysis of iCOR, we divided them by two approaches that are described below:

3.1. ECG Denoising

To evaluate the effectiveness of the proposed model in ECG denoising, this study conducted a series of experiments and analyzed the results with two scenarios of data distribution, namely shuffled and non-shuffled. In the shuffled scenario, the training and validation datasets are separated randomly using a fixed seed to ensure the experimental results are reproducible. Conversely, in the non-shuffled scenario, the training and validation datasets are separated sequentially. This categorization is performed to verify the model’s robustness to varying data distributions. Three feature engineering (FE) methods are employed during the padding process at the ECG signal noise removal stage: raw padding, amplitude, and amplitude–isoelectric. Padding is a technique used to ensure that all input sequences have the same length, which is essential for accurate feature extraction and ECG signal delineation. This helps in preserving the morphology of the ECG signal, which is critical for diagnosing heart conditions. The performance of various model noise removal processes is depicted in

Table 3. The ECG signal denoising results with two scenarios of data distribution for training, validation, and testing sets.

Data	FE	Metrics Evaluation	Mean ± Standard Deviation
			Training	Validation	Testing
Non-shuffled	Raw	SSD	1665.42 ± 5994.15	1399.01 ± 2697.22	163.23 ± 370.85
		SNR	14.17 ± 10.91	15.83 ± 10.84	17 ± 9.61
		MAD	1.77 ± 2.48	1.75 ± 2.15	0.84 ± 0.79
		PRD	209.08 ± 208.99	130.09 ± 167.42	103.75 ± 120.34
		COS_SIM	0.58 ± 0.33	0.62 ± 0.33	0.68 ± 0.29
	Amplitude	SSD	4.79 ± 9.83	24.32 ± 7.96	6.22 ± 7.3
		SNR	23.89 ± 7.94	37.01 ± 7.94	22.77 ± 7.3
		MAD	0.31 ± 0.22	0.35 ± 0.25	0.41 ± 0.29
		PRD	45.83 ± 33.9	43.86 ± 27.32	53.1 ± 30.21
		COS_SIM	0.91 ± 0.1	0.91 ± 0.1	0.89 ± 0.1
	Amplitude–isoelectric	SSD	3.86 ± 8.89	5.1 ± 9.28	5.03 ± 7.81
		SNR	25.15 ± 7.77	25.39 ± 7.75	23.39 ± 7.06
		MAD	0.31 ± 0.22	0.35 ± 0.26	0.41 ± 0.3
		PRD	39.67 ± 28.3	38.98 ± 27.59	50.39 ± 28.85
		COS_SIM	0.93 ± 0.09	0.93 ± 0.1	0.91 ± 0.1
Shuffled	Raw	SSD	878 ± 4158.58	843.78 ± 4547.76	83.47 ± 117.33
		SNR	11.66 ± 8.02	11.64 ± 8.01	12.16 ± 7.02
		MAD	1.76 ± 2.08	1.74 ± 2.02	1.09 ± 0.6
		PRD	114.03 ± 64.23	114.46 ± 65.08	118.86 ± 57.36
		COS_SIM	0.65 ± 0.28	0.65 ± 0.28	0.61 ± 0.25
	Amplitude	SSD	4.48 ± 8.26	4.32 ± 7.88	5.49 ± 7.32
		SNR	24.34 ± 7.84	37.02 ± 7.88	22.77 ± 7.32
		MAD	0.32 ± 0.22	0.35 ± 0.25	0.43 ± 0.29
		PRD	44.78 ± 34.2	45.21 ± 35.38	56.98 ± 37.66
		COS_SIM	0.92 ± 0.09	0.91 ± 0.1	0.9 ± 0.1
	Amplitude–isoelectric	SSD	3.6 ± 7.15	3.6 ± 7.11	4.81 ± 7.29
		SNR	25.55 ± 7.75	25.46 ± 7.81	23.46 ± 7.22
		MAD	0.31 ± 0.22	0.35 ± 0.26	0.43 ± 0.3
		PRD	37.72 ± 27.3	37.77 ± 27.2	53.05 ± 38.42
		COS_SIM	0.94 ± 0.08	0.94 ± 0.09	0.91 ± 0.1

.

Table 3 indicates the highest SNR improvement is achieved with the amplitude–isoelectric line centered padding method with the shuffle scenario. It can be observed that the model of ECG signal denoising demonstrated outstanding performance with the following average evaluation metrics and standard deviations (Stdev): SSD (4.81 ± 7.29), SNR (23.46 ± 7.22), MAD (0.43 ± 0.30), PRD (53.05 ± 38.42), and COS_SIM (0.90 ± 0.10). The shuffled scenario approaches help in understanding the model’s generalization capabilities when exposed to diverse data. Shuffling ensures that the model encounters a wide range of examples during each epoch, enhancing its learning process. This diversity in training examples typically leads to better overall performance metrics. The sample plots of ECG signals in shuffled and non-shuffled scenarios with raw padding, amplitude, and amplitude–isoelectric methods are presented in Figure 4, Figure 5 and Figure 6, respectively.

The isoelectric line is the baseline of the ECG signal, representing periods of electrical inactivity in the heart. Combining amplitude and isoelectric line padding involves adjusting the signal’s amplitude and aligning it with the isoelectric line before padding. This method helps preserve both the relative magnitude and baseline of the ECG signals, leading to more accurate denoising. The ECG signal denoising with multi-layer filter visualization is shown in Figure 4, Figure 5 and Figure 6, where the red line is $y$ prediction, the blue line is $y$ test, and the red region is an error prediction (error region). The error region displays the size of the difference between the ground truth and the predicted results. The ECG signal denoising result with the amplitude–isoelectric line centered method produces good metrics and has the ability to reduce the noise in the signal, however, during the shuffled scenario, the performance results for training and validation can vary significantly (refer to Table 3). Using distributed data in the shuffled scenario has produced good testing performance.

3.2. ECG Delineation

Based on the ECG denoising process using the best multi-layer filter model, a noise-free ECG signal can be produced. Following this, the ECG delineation process is carried out to separate the P-QRS-T-wave morphology through a classification scheme. The proposed parameters for the ECG delineation model are batch size of 16, learning rate of 0.0001, 100 epochs, an adaptive moment estimation (ADAM) optimizer, and a sparse categorical cross-entropy loss function. From the best model of ECG denoising in the FE amplitude–isoelectric centered method, transfer learning is performed where part of the first hidden layer of the model is frozen, while part of the last hidden layer is set to trainable. In the output layer, BiLSTM followed by dense layers with a Softmax activation function forms a delineation model. The performance results of ECG delineation with MLF-BiLSTM in record-based and beat-based approaches are visualized in Figure 7. Figure 7 shows the classification performance for unseen LUDB datasets yields above 90.4% and 98% accuracy for record-based and beat-based approaches, respectively. Beat-based approaches outperformed the record-based approach in overall performance metric results. Similar results were shown in an unseen set of the QTDB, while beat-based approaches performed well. As seen in Figure 7, the F1 score helps to understand the trade-off between precision and recall, making it a valuable metric. The delineation model with iCOR demonstrates effective signal delineation on both LUDB and QTDB datasets. It achieves an F1 score of above 85.3% for the P-wave class, 87.7% for the QRS-complex class, and 89% for the T-wave class with the beat-based approach. These results indicate that the model is robust and performs well across various sources of ECG signal data. Despite the differences in ECG signal recording characteristics between the QTDB and LUDB in terms of frequency, recording length, equipment used, and recorded leads, the proposed method effectively classifies the morphology of ECG signals.

For a more comprehensive analysis, this study detected atrial fibrillation (AF) and atrial flutter (AFL) with the proposed iCOR model. Figure 8 presents the sample results of ECG delineation between ground truth and iCOR prediction in AF and AFL records, respectively. The results showed there is no significant gap between ground truth and iCOR prediction for P-QRS-T-wave classification. The characteristics of both are no visible P-waves, while in AF, the atria beat irregularly, and in AFL, the atria beat regularly. We have tested 14 AF records (8, 38, 44, 51, 83, 88, 93, 95, 96, 101, 110, 112, 129, and 173) and 3 AFL records (35, 52, and 103). The experimental results of both records are presented in

Table 4. The experimental results of AF detection with iCOR.

AF Records	Class	Performance Results (%)
		Accuracy	Sensitivity	Specificity	Precision	F1
8	P-wave	99.7	0	99.7	0	0
	QRS-complex	96.5	77.8	99.6	96.9	86.3
	T-wave	96.8	91.3	97.9	90.3	90.8
38	P-wave	94.4	0	94.4	0	0
	QRS-complex	95.5	76	99.5	96.6	85
	T-wave	84.7	16.8	99.7	91.4	28.3
44	P-wave	97.8	0	97.8	0	0
	QRS-complex	96.9	80.2	99.6	97.2	87.9
	T-wave	94	73.7	98	87.7	80.1
51	P-wave	97.2	0	97.2	0	0
	QRS-complex	96.5	77.2	99.9	99.1	86.8
	T-wave	95.6	76.4	98.8	91.6	83.3
83	P-wave	96.3	0	96.3	0	0
	QRS-complex	89.1	40.3	100	99.8	57.5
	T-wave	87.1	77	89.9	67.7	72
88	P-wave	96.6	0	96.6	0	0
	QRS-complex	97.9	89.7	99.1	93.5	91.6
	T-wave	81.4	24	95.6	57.7	33.9
93	P-wave	97.1	0	97.1	0	0
	QRS-complex	94.3	62.5	99.7	97.7	76.3
	T-wave	93.7	85.5	95.4	79.2	82.2
95	P-wave	99.6	0	99.6	0	0
	QRS-complex	93.7	53.7	99.2	90.2	67.3
	T-wave	89.4	73.8	93.3	73.8	73.8
96	P-wave	96.1	0	96.1	0	0
	QRS-complex	98.2	92.8	99	92.5	92.6
	T-wave	87.4	40.3	98.9	89.7	55.6
101	P-wave	93.8	0	93.8	0	0
	QRS-complex	96.7	80	99.6	97.2	87.7
	T-wave	89.4	63.8	96.6	83.8	72.4
110	P-wave	98.9	0	98.9	0	0
	QRS-complex	95.2	69.6	99.6	96.4	80.8
	T-wave	88.3	56.6	97.8	88.3	69
112	P-wave	92.1	0	92.1	0	0
	QRS-complex	98.1	91.9	99.2	95	93.4
	T-wave	87	37	97.2	73.5	49.2
129	P-wave	97.7	0	97.7	0	0
	QRS-complex	97.8	77.3	99.7	95.9	85.6
	T-wave	94.7	83.7	96.9	84.2	84
173	P-wave	96.3	0	96.3	0	0
	QRS-complex	97.9	88.1	99.4	95.6	91.7
	T-wave	93.4	73.4	98.7	93.9	82.4

and

Table 5. The experimental results of AFL detection with iCOR.

AFL Records	Class	Accuracy	Sensitivity	Specificity	Precision	F1
35	P-wave	92.3	0	92.3	0	0
	QRS-complex	95.1	83	98.9	96.1	89.1
	T-wave	80.7	36.2	99.8	98.8	53
52	P-wave	96.9	0	96.9	0	0
	QRS-complex	97.3	79.4	99.7	97.3	87.4
	T-wave	90.5	72.8	94.2	71.8	72.3
103	P-wave	86.2	0	86.2	0	0
	QRS-complex	96.9	91	97.9	87.8	89.4
	T-wave	85.6	49.6	95.1	72.6	58.9

for AF and AFL, respectively. The experiment of 14 AF records with iCOR showed that the sensitivity, precision, and F1 score of P-waves are zero (0). P-waves will not be seen on the ECG in patients with AF.

From a clinical perspective, AF and AFL have many similarities. Both are fast heart rhythm disturbances affecting the upper chambers and can lead to symptoms such as palpitations, fatigue, and reduced exercise capacity. AF is characterized by irregular atrial activity in both atria, while AFL is identified as a macro-reentrant tachyarrhythmia with consistent, regular atrial activity in both atria. AF and AFL showed absent P-waves in ECG signals. Table 5 presents absent P-waves in AFL records (35, 52, and 103). From the results of Table 4 and Table 5, our proposed iCOR has successfully detected the characteristic of AF and AFL with absent P-wave morphology in ECG signals.

Besides the absent P-wave, this study visualized the irregular/regular RR-interval to differentiate AF and AFL. The regular RR-interval is 0.6–1.2 s. The sample visualization of AF and AFL detection with iCOR from an RR-interval perspective is presented in Figure 9. Figure 9 presents samples of irregular rhythm (abnormal beat-to-beat, red color) and regular rhythm (normal beat-to-beat, blue color) in AF and AFL records, respectively. As seen in Figure 9, the irregular rhythms are mostly visualized in AF records (records 38 and 110), and regular rhythms are mostly visualized in AFL records (records 52 and 103).

In recent years, there have been previous studies that provide end-to-end learning for ECG morphology classification (refer to

Table 6. The benchmarking studies of ECG denoising and delineation process with DL approach.

Authors	Method		Database	Accuracy (%)
	ECG Denoising	ECG Delineation		P-Wave	QRS-Complex	T-Wave
Tutuko et al. [21]	DAE	BiLSTM	QTDB (normal sinus records only)	99.81	99.86	99.35
Chen et al. [40]	Regularized latent space (encoder–decoder)	ECGVEDNET	ICDIRS	-	86.28	89.94
Present study	iCOR		LUDB	98	98.8	98
			QTDB (normal and abnormal records)	98.7	98.1	97

). Tutuko et al. [21] proposed the denoising autoencoder (DAE) for denoising algorithms and convolutional-bidirectional long short-term memory (ConvBiLSTM) for ECG delineation to classify ECG waveforms in terms of the P-QRS-T-wave. The denoising reconstruction using unsupervised learning based on the encoder–decoder process can be proposed to improve the drawbacks. First, the ECG signals are reduced to a low-dimensional vector in the encoder. Second, the decoder reconstructs the signals. Finally, the reconstructed signals of ECG can be processed by ConvBiLSTM. The results achieved an accuracy of 99.81%, 99.86%, and 99.35% for P-wave, QRS-complex, and T-wave, respectively. Though the performance results are outstanding, they only concern the normal sinus records for generating the ConvBiLSTM model. Chen et al. [40] conducted experiments on a large-scale 12-lead ECG dataset, ICDIRS, using a variational encoder–decoder network, ECGVEDNET, specifically designed to address morphological variations in ECG signals. In ECGVEDNET, a well-regularized latent space is created, where the latent ECG features adhere to a regular distribution, exhibiting reduced morphology variations compared to the raw data. Additionally, a transfer learning framework is introduced to leverage the knowledge gained from ICDIRS and apply it to smaller datasets. On ICDIRS, ECGVEDNET achieves accuracies of 86.28% for QRS onset and 89.94% for T peak. From the previous studies, it can be claimed that iCOR performs excellently in classifying three primary ECG waveforms, i.e., P-wave, QRS-complex and T-wave. iCOR outperformed the accuracy of P-wave, QRS-complex and T-wave in the previous studies [21,40].

The performance results of this experiment study look promising, however, there are limitations of this study:

• This study was only concerned with BW noise for the ECG denoising process using an MLF;
• The testing of iCOR was only explored in noisy ECG signals.

4. Conclusions

In this study, we proposed a novel approach for ECG signal delineation using an MLF combined with a BiLSTM model, namely iCOR. Our method addresses the limitations of previous models, which often performed sub-optimally on diverse datasets. Our approach achieves a simpler yet effective delineation model by utilizing the MLF for simultaneous noise removal and feature extraction and integrating these features into the BiLSTM for classification. The evaluation of combined LUDB and QTDB datasets demonstrates the robustness of our model, achieving high F1 scores. It can be observed that the classification performance for unseen sets of LUDB datasets yields above 90.4% and 98% accuracy for record-based and beat-based approaches, respectively. Beat-based approaches outperformed the record-based approach in overall performance metric results. Similar results were shown in an unseen set of the QTDB, in which beat-based approaches performed with accuracy above 97%. These results indicate that the proposed model can effectively delineate ECG signals from various sources, making it a versatile tool for clinical applications. The performance results of iCOR across different datasets underscore its potential for reliable and precise cardiac health monitoring. Future work will focus on further refining the model and exploring its application to other biomedical signal-processing tasks.