Semi-Supervised Domain Adaptation for Individual Identification from Electrocardiogram Signals

Abstract

When acquiring electrocardiogram (ECG) signals, the placement of electrode patches is crucial for acquiring electrocardiographic signals. Constant displacement positions are essential for ensuring the consistency of the ECG signal when used for individual identification. However, achieving constant placement of ECG electrode patches in every trial for data acquisition is challenging. This is because different individuals may attach patches, and even when the same person attaches them, it may be difficult to specify the exact position. Therefore, gathering ECG data from various locations is necessary. However, this process requires a substantial amount of labor and time, owing to the requirement for multiple attempts. Nonetheless, persisting with these efforts enables the endurance of some ECG differences. To reduce labor and time, we propose a semi-supervised domain adaptation for individual identification using ECG signals. The method operates with a full set of original ECG signals and a small set of ECG signals from different placements to account for the differences between the signals in the generative adversarial network (CycleGAN). Specifically, to train the CycleGAN, the ECG signals were transformed into time–frequency representations, and the trained generator was used to generate ECG signals to expand the small set of ECG signals from different placements. Subsequently, both the original and generated signals were used to train the classifier for individual identification. This scenario can also be applied to the classification of ECG signals from different sensors. The PTB-ECG dataset was used for this experiment. We found that the proposed method showed higher accuracy than when only the original ECG signals were used for the training classifier.

1. Introduction

An electrocardiogram (ECG) is a small electrical biosignal produced by the heartbeat and captured from the body’s surface via attaching electrode patches [1]. ECG can be used for biometrics because individuals have varying body states, including differences in weight, height, and muscle ratios. These variations render an ECG signal unique for each person. There are several benefits to using ECG signals in biometrics. The use of ECG for biometrics eliminates the need to remember passwords, carry keys or cards, and reduces environmental constraints. Moreover, using biometrics, especially ECG, ensures the real-time verification of a user’s vital signs and the awareness of biometric trials. Unlike facial images or voices, which can be recorded remotely, ECG data cannot be captured from a distance and remains concealed externally, providing a strong security advantage. Biometrics using an ECG can be applied to various applications, including online banking systems, smartphone or door unlocking, and login processes [2,3,4,5,6,7,8].

ECG signals are easily affected by noises such as movement, aspiration, muscle activation state, the power for sensors, and the placement of patches. To capture good ECG signals, a subject should remain calm and relaxed. Theccurate analysis of ECG data demands expertise, and automated interpretation algorithms may encounter difficulties in effectively addressing a variety of patterns and abnormalities. ECG signals contain sensitive medical information, requiring the protection of patient privacy and compliance with data protection regulations. The ECG system has a tendency to overfit to small personal datasets, making it challenging to effectively generalize to newly-collected, unlabeled data [9,10].

The placement of the electrode patches is crucial for measuring ECG signals. The consistent placement of electrode patches is essential to ensuring consistency in ECG signals for individual identification [11,12,13,14]. However, achieving consistent placement in each data-acquisition attempt is challenging. Different individuals may attach patches, and even the same person may struggle to specify the exact position consistently. Therefore, it is necessary to collect ECG data from various electrode placements. However, this process requires multiple attempts and consumes a considerable amount of labor and time [15,16,17,18,19]. Despite the effort required, persisting with these endeavors helps address the differences in ECG signals [20,21,22].

The differences in signals caused by variations in electrode placement make it challenging for neural networks to classify signals correctly. This is referred to as domain adaptation. Numerous studies have been conducted on domain adaptation. In particular, several studies on domain adaptation in ECG applications are described. Bazi [23] preprocessed signals using both a notch filter and a high-pass filter. The preprocessed signal was then extracted into ECG morphology features and ECG wavelet features, including the QRS width. In addition, the features were normalized to achieve a uniform length across the ECG cycles. The domain transfer support vector machine (SVM) and importance-weighted kernel logistic regression were introduced for domain adaptation. Liu [9] presented a deep learning approach for ECG classification based on adversarial domain adaptation. This approach involved multiscale feature extraction, domain discrimination, and classification. The feature extraction module was constructed using three separate convolutional blocks to broaden the width of the features. The domain-discrimination module consisted of three convolutional blocks and a fully connected layer. The classification module enhanced feature diversity via merging temporal features with those extracted using deep learning. Wang [24] introduced a domain-adaptive ECG arrhythmia classification model constructed on a convolutional network that employed unsupervised domain adaptation. Two distinct objective functions were proposed using the observed clustering characteristics of the data. Cluster-aligning loss was designed to align the distributions of training and test data, whereas cluster-maintaining loss aimed to enhance the structural information and discriminative aspects of the features. Yin [25] introduced a self-adjustable domain adaptation strategy aimed at mitigating overfitting and harnessing unlabeled data. The dataset was expanded using data augmentation via incorporating the actual ECG and radar data. To use unlabeled data, self-adjustable domain adaptation combined transfer learning and self-organizing maps for label prediction. To mitigate overfitting, the domain adaptation algorithm was integrated with one-class classification. Carrera [26] proposed a domain-adaptation solution that modeled heartbeats using dictionaries obtained using sparse representations and further transforms them into user-specific dictionaries based on the heart rate. This method involved learning appropriate linear transformations from a vast dataset that includes ECG traces. Chen [27] presented an adaptive region aggregation network that employed adversarial training to address the domain shift issue in ECG delineation. Utilizing a regional aggregation network to learn domain-invariant features, its performance was enhanced in both the source and target domains. A neural network structure composed of 1D convolutional layers and three pooling layers was employed to extract features from the ECG signals. A 1D discriminator was used to differentiate the domains of the feature maps, and ECG delineation was performed using an auxiliary branch. The generator learned to minimize the difference between the source and target domains, thereby narrowing their distribution in the feature space. A predictor was used to predict the positions of nine characteristic points. Shang [28] introduced an ECG classification model designed to learn domain-agnostic features. Moreover, this model is adaptable to ECG data with a few leads. The model employed a modified ResNet with squeeze-and-excitation attention blocks for deep feature extraction. A multisource adversarial network was trained via combining handcrafted features. Efficient domain-invariant features were used to diagnose cardiac abnormalities. Rafi [29] introduced a source-free ECG adaptation framework for patient-specific ECG classification. This framework addresses privacy concerns during adaptation because it does not require source data. To address imbalanced classes due to limited data, it employed a generative model to synthesize patient-specific ECG data, thereby generating additional source data. Subsequently, it employed a local structure clustering method to align the target ECG features with neighboring features that exhibit similarities. After effectively capturing the clustered target features, it employed both the generated source samples and a classifier trained on the source data to generalize the performance of the model for unseen data classification. Deng [30] introduced a multi-source unsupervised domain-adaptation neural network to efficiently utilize diverse source data in ECG classification and enhance model generalization. The model is distinguished by a two-branch domain adaptation and a sample imbalance-aware mixing strategy, facilitating the integration of information across domains. One branch was designed to capture domain-invariant representations, whereas the other was customized to extract domain-specific features. These two branches align the ECG data from the target domain with the individual source domains, resulting in improved discriminative features. Natarajan [31] presented a domain-adaptation approach to assess and mitigate the potential causes of performance degradation in the process of generalization from laboratory to real-world settings. This method was applied to detect cocaine use using wearable ECG sensor data. He [32] proposed a multimodal image fusion and continual test-time adaptation-based method for heartbeat classification. The ECG data were transformed into three-channel color images using a multimodal image fusion framework. The challenge of class imbalance was addressed via employing batch-weighted loss functions. A pretrained source model was adapted to the target domain using a continual test-time adaptation approach. Lee [10] conducted a study on biometric identification utilizing photoplethysmogram and ECG signals. The research proposed both unsupervised and semi-supervised adversarial learning techniques for cross-domain adaptation. CycleGAN and MobileNetV2 were employed for domain adaptation and biometric identification.

We propose a semi-supervised domain adaptation method for individual identification using ECG signals. This method utilizes a complete set of original ECG signals along with a limited set of signals from different placements. This approach addresses certain signal variations within a generative adversarial network (CycleGAN). Specifically, for the CycleGAN training, the ECG signals were transformed into time-frequency representations, and the trained generator generated additional ECG signals to augment a small set of signals from various placements. Both the original and generated signals were employed to train the classifier for individual identification. This approach is also adaptable to classifying ECG signals from different sensors. The PTB-ECG dataset was used for the experimental setup. The PTB-ECG dataset was acquired using a total of 15 leads, including 12 standard leads and Frank XYZ leads, and included data from 290 healthy individuals and individuals with various heart conditions. The results showed an improvement in accuracy with the proposed method compared with using only the original ECG signals for classifier training. Certain pairs of source and target data were employed to train the CycleGAN, enabling the acquisition of features necessary for converting source-domain data into target-domain data. The trained CycleGAN generates distinct yet consistent time-frequency representations, which are domain-specific ECG variations, from a full set of source-domain data. These generated ECG variations are advantageous for improving generalization in neural network training.

This study describes a semi-supervised domain adaptation method for individual identification using ECG signals. Section 2 describes the study methods, and Section 3 presents the proposed method for individual identification with domain adaptation. Section 4 describes the experiments and results of this study, and Section 5 presents our conclusions.

2. Methods

2.1. Signal to Image via Continuous Wavelet Transform

The continuous wavelet transform offers a dual analysis of signals, providing insights within both the time and frequency domains. This method rectifies the inherent limitations of the Fourier transform, which adheres to a fixed scale. Unlike the Fourier transform, in which low-resolution and high-resolution regions are examined at the same scale, the wavelet transform conducts a multiscale analysis. This enables us to uncover intricate details, even when examining high-resolution segments. Via applying a mother wavelet to the signal, the wavelet transform decomposes the signal into wavelets of various scales. Upon the reintegration of these dissected components, the original signals are restored. Equation (1) shows the wavelet transform formula, with $f\left(t\right)$ representing the subject signal for decomposition, and ${\mathsf{\psi }}_{a,b}\left(t\right)$ denoting the mother wavelet. The ‘a’ factor corresponds to the scale, while ‘b’ accounts for translation. Among the well-recognized mother wavelets, the Morse wavelet, expressed in Equation (2) [33,34,35,36], stands out as an example.

$$\begin{gathered} \mathrm{T}\left(\mathrm{a},\mathrm{b}\right)=\frac{1}{\sqrt{a}}{{\displaystyle \int }}_{-\infty }^{\infty }f\left(t\right) \psi \left(\frac{t-b}{a}\right)dt \\ \mathrm{a}ϵR^{+}-\left\{0\right\}, \mathrm{b}ϵ\mathrm{R} \end{gathered}$$

(1)

$${\mathsf{\Psi }}_{P,\gamma }\left(\omega \right)=\mathrm{U}\left(\mathsf{\omega }\right)a_{P,\gamma }{\omega }^{\frac{P^2}{\gamma }}{e}^{-{\omega }^{\gamma }}$$

(2)

2.2. CycleGAN (Cycle Generative Adversarial Network)

A generative adversarial network (GAN) is a deep learning network that can generate data with features similar to those of the input data used in training. It typically consists of two networks: a generator and a discriminator. The generator is trained to generate data similar to the input data used for training. The discriminator is trained to distinguish between the real data used in generator training and the data generated by the generator.

The CycleGAN is a prominent model for image-to-image translation. This addresses the drawbacks of traditional image-to-image translation, which requires paired image data for training purposes. Training a GAN with unpaired images can lead to mode collapse. This problem occurs when the model strongly converges to a single mode, producing identical outputs instead of even learning across a given data distribution. To address this issue, a cyclic structure using two GANs is proposed. One generator transforms data from domain A to domain B, whereas the other generator transforms data from domain B to domain A. Meanwhile, one discriminator classifies the data transformed from domain A to domain B and real data from domain B, and another discriminator classifies the data transformed from domain B to domain A and real data from domain A. Generators are trained to generate data that deceive the discriminators, whereas discriminators are trained to distinguish between the data generated by the generators and real data. Each generator aims to make the discriminator classify the generated data as real, and each discriminator strives not to be deceived by the generator.

The overall loss for CycleGAN is defined as a combination of the adversarial and cycle consistency losses, as shown in Equation (3). Parameter lambda $\left(\mathsf{\lambda }\right)$ controls the relative importance of these two losses. CycleGAN uses the least-squares loss function for both the generator and discriminator. The discriminator loss, as depicted in Equation (4), minimizes the square sum of the differences between the predicted and expected values of real and fake data. Here, $\mathrm{X}$ represents the real input data, and ${\mathrm{X}}^{\prime }$ represents the generated data. $\mathrm{D}\left(\mathrm{X}\right)$ represents discriminator prediction when $\mathrm{X}$ is an input. The generator loss, as shown in Equation (5), encourages the discriminator to recognize the generated data as real data: $\mathrm{D}\left({\mathrm{X}}^{\prime }\right)$ represents the discriminator’s prediction when ${\mathrm{X}}^{\prime }$ is input. The cycle consistency loss in Equation (6) quantifies the difference between the original data and the data reconstructed back to the original domain. Here, $\mathrm{X}$ and $\mathrm{Y}$ represent data from two different domains, and ${\mathrm{G}}_{\mathrm{X}}\left(\mathrm{Y}\right)$ represents the output of the generator when $\mathrm{Y}$ is the input to produce data in the $\mathrm{X}$ domain [37].

$$\mathrm{Total} \mathrm{Loss}=\mathrm{Adversarial} \mathrm{Loss}+ \mathsf{\lambda }\times \mathrm{Cycle} \mathrm{Consistency} \mathrm{Loss}$$

(3)

$$\mathrm{Discriminator} \mathrm{Loss}={\left(\mathrm{D}\left(\mathrm{X}\right)-1\right)}^2+\mathrm{D}{\left({\mathrm{X}}^{\prime }\right)}^2$$

(4)

$$\mathrm{Generator} \mathrm{Loss}={\left(\mathrm{D}\left({\mathrm{X}}^{\prime }\right)-1\right)}^2$$

(5)

$$\mathrm{Cycle} \mathrm{Consistency} \mathrm{Loss}=\left|\mathrm{Y}-{\mathrm{G}}_Y\left({\mathrm{G}}_X\left(\mathrm{Y}\right)\right)\right|+\left|\mathrm{X}-{\mathrm{G}}_X\left({\mathrm{G}}_Y\left(\mathrm{X}\right)\right)\right|$$

(6)

2.3. Convolutional Neural Network for Image Classification

Convolutional neural networks (CNNs) are primarily employed to recognize patterns within images. They employ convolutional operations to extract features from images and utilizes subsampling to decrease data dimensionality. In addition, they typically integrate fully connected layers at the end of the architecture to generate classification results as the output. CNNs are distinguished from other neural networks because of their ability to learn the filters utilized in the convolution layers, enabling them to acquire feature extraction and classification skills simultaneously within a single network. CNNs can be designed in various configurations via arranging and composing different layers. It is common to utilize pre-trained models, given the potential performance variations among different CNN structures. These pre-trained models, which have already demonstrated impressive experimental results on large-scale datasets, have finely tuned initial parameters. This translates into competitive performance, even when the training iterations and data volumes are limited. Notable pre-trained models include VGGNet, GoogLeNet, ResNet, and DenseNet [38,39,40].

3. Proposed Domain Adaptation Method for Individual Identification Using ECG Signals

This section describes the details of the proposed domain adaptation method for individual identification using ECG signals. The model consists of three modules: the transformation of the ECG signal to an image, semi-supervised CycleGAN-based domain adaptation, and a ResNet-based classifier.

3.1. First Module for Individual Identification Using ECG

The first module converts ECG signals into images. To convert the ECG signals into images, we employed a continuous wavelet transformation (CWT). This method transformed the 1D electrocardiogram signal into a time-frequency representation, extending various features for ECG analysis. In particular, it offered a more efficient feature representation, owing to the diverse analysis scales based on the signal resolution. Utilizing the CWT enabled the conversion of 1D signals into 2D images, opening up various extensions for 2D-based deep learning neural network applications. Figure 1 illustrates an example of the time-frequency representation of ECG signals using CWT. In Figure 1, the first and second columns display the signals from channels 1 and 2, respectively, whereas the rows represent the signals from different individuals.

3.2. Second Module for Individual Identification Using ECG

The second module involved the transformation of the ECG signals into 2D images using CycleGAN to change the domains. CycleGAN was originally designed for image-to-image translation; however, it can also be applied to domain transformations. In traditional image-to-image translation, the model requires symmetrical image pairs that differ only in their domains, making it challenging, and sometimes impossible, to obtain such data. However, CycleGAN allows for a random composition of image pairs, minimizing the need for additional human labor in domain adaptation, thus making it an efficient model for domain transfer. Once the CycleGAN training was completed, data with the characteristics of the target domain were generated using the generator. If the CycleGAN performed well, it produced results that are similar to those of the target domain; however, if its performance was subpar, the results may be less similar to those of the target domain. Using less similar data in the training of an individual identification model can negatively impact recognition performance. Therefore, a weighted average of the generated and original data was applied to reduce the differences between them [37]. Figure 2 illustrates the CycleGAN generator for domain transfer between channels.

3.3. Third Module for Individual Identification Using ECG

The third module is a classifier that uses CNN for individual identification. To perform individual identification using ECG, the ECG signals were first transformed into a time-frequency representation, and these images were utilized to train the CycleGAN to enable domain transformation. Once the CycleGAN training was complete, domain transformation was performed using the generator, and the data transformed from the original domain, along with a small amount of original data, were used to train a CNN classifier. ResNet was employed for classification using a CNN. ResNet is a prominent pretrained model in the field of CNNs and is designed as a deep neural network with skip connections that add the upper-level output to the lower-level output. It was designed in this manner because deep layers are required to facilitate neural network learning, even for complex problems. However, as the depth of the layers increases, the weight values tend to approach zero during the learning process. This design was intended to address this issue. Additionally, ResNet pre-training on a large-scale image dataset ensured that the initial weight values were optimized for efficient feature extraction [39]. Figure 3 illustrates the ResNet-based individual identification model.

4. Experiments and Results

4.1. ECG Datasets for Individual Identification

To assess the performance of the proposed method, we used the PTB (Physikalisch-Technische Bundesanstalt) electrocardiogram (ECG) database. The PTB-ECG database comprises 549 high-resolution ECG recordings obtained from 290 subjects, including healthy individuals and patients with various heart conditions. The recordings were obtained from 15 leads, including 12 standard and Frank XYZ leads. The participants’ ages ranged from 17 to 87 years (209 males and 81 females among them). Each subject’s data were acquired in the form of 1–5 record files. The ECG signals were sampled at 1000 samples/s with a 16-bit resolution. Each record included medical summaries encompassing age, sex, and diagnosis. There are nine diagnostic categories for PTB-ECG, including various heart conditions, as listed in

Table 1. Diagnostic categories of PTB-ECG.

Diagnostic Categories
Myocardial infarction
Cardiomyopathy/Heart failure
Bundle branch block
Dysrhythmia
Myocardial hypertrophy
Valvular heart disease
Myocarditis
Miscellaneous
Healthy controls

[41,42].

4.2. Experiments and Results

To perform the experiments, this study used a computer equipped with an Intel^® Xeon(R) CPU E5-1650 v3 3.5 GHz, Windows 10 64 bit, 32 GB RAM (Random Access Memory), an NVIDIA GeForce GTX Titan X, and Matlab 2022b.

We analyzed the 1-channel signals of the PTB-ECG data to find the R-peaks using the Pan–Tompkins algorithm. The number of R-peaks found for each subject was counted, and subjects with too few R-peaks were excluded from the dataset. Among the 290 participants, 51 were excluded, resulting in 239 participants for the dataset. Each data sample was centered around the R-peak and truncated to a length of approximately 800, with approximately 400 data points on each side. To standardize the number of data samples per participant, 100 samples were collected from each participant. Among these, 70 were used for training and the remaining 30 were used for validation. No preprocessing was performed on the data other than R-peak detection and segmentation. The training dataset consisted of 16,730 samples (239 subjects × 70 samples), and the validation dataset consisted of 7170 samples (239 subjects × 30 samples). The prepared data were used for domain adaptation with 1-channel and 2-channel ECG signals. It was assumed that the 1-channel ECG signal represented the source domain with a large amount of data for neural network training, and the 2-channel ECG signal represented the target domain with limited data. Therefore, all 70 samples of 1-channel data were used for each subject in the training set, whereas only 12 of the 70 two-channel data samples were used for 2-channel data. To prepare the data for input to the CycleGAN, a time-frequency representation was applied to convert the ECG signals into images. The time-frequency representation used was a scalogram based on the continuous wavelet transform (CWT). The input size for the CycleGAN was set to 256 × 256, and the model was trained for epochs ranging from 100 to 600 in increments of 100. Owing to the challenges of training all the classes in a single CycleGAN, a separate model was trained for each class. Figure 4 illustrates the training courses of CycleGAN on class 1, while Figure 5 shows the training courses of CycleGAN on class 2.

Using the trained CycleGAN generator, all 70 data samples from channel 1 were input, resulting in 70 transformed channel 2 data samples. These generated channel 2 data samples were used to train the neural network for individual identification. To fully exploit the available data, a small amount of the original channel 2 data was also included in the training set. The model for individual identification used ResNet-101. The model was trained using the Adam optimization algorithm with a batch size of 20. Cross-entropy was employed as the loss function, and the initial weights were pre-trained using the ImageNet dataset. The training began with an initial learning rate of 0.0001 and a momentum value of 0.9. Additionally, a learning rate decay of 0.2 was applied every five epochs, L2 regularization with a coefficient of 0.0001 was applied every five epochs, and the training was limited to a maximum of three epochs. Data-augmentation techniques were not used during the training. All the weights of the pre-trained model were set to be trainable.

Table 2. Accuracies on the test data for individual identification with original and generated data.

Epochs of CycleGAN	Test Accuracy
100	90.92
200	88.42
300	88.87
400	85.77
500	90.67
600	91.61

shows the accuracy of the test data for individual identification using the original and generated data.

When the input size of the CycleGAN was set to 256 and the number of epochs was 600, the highest accuracy achieved was 91.61%. It was observed that the accuracy generally fluctuated according to the training epochs of the CycleGAN. When the ResNet-101 model was trained with only a small amount of original 2-channel data, the accuracy was 94.52%. Although there were instances in which training with both the generated 2-channel data and the original data resulted in similar or slightly higher accuracy, most of the time, the accuracy was lower compared with training with only the original data.

Table 3. Comparison of test accuracies when training the classification model with generated data versus without.

Method	Test Accuracy
ResNet-101 only with original data	94.52
ResNet-101 with original and generated data	91.61

shows a comparison of the test accuracies when training the classification model with and without the generated data.

Owing to the fluctuating accuracy when training the individual identification model with both the generated and original data, we analyzed the directly generated data. Figure 6 illustrates a comparison of the original and generated data using the CycleGAN to transform channel 1 into channel 2 ECG data based on 600 epochs of the CycleGAN training. The transformation from 1-channel to 2-channel data using the CycleGAN exhibited varying restoration capabilities across different classes, often resulting in instances in which the similarity between the transformed and original data was reduced. If the generated data exhibited considerable disparity from the original data, their integration into the training process could introduce disturbances, leading to a general reduction in accuracy.

To reduce the difference between the generated and original data, a weighted average of the original and generated data was calculated. Starting with an equal ratio of the generated data to the original data, the weights for the original data were gradually increased as training data for the individual identification model. For each class, 12 original and 70 generated data samples were used. A weighted average was computed over the entire set of generated data for each original dataset, effectively expanding the data to 840 $\left(12\times 70\right)$ samples per class. All 12 original and 840 expanded data samples were used for training in each class. As there were 239 classes, the total number of data samples used for training was 203,628 ($\left(12+840\right)\times 239$).

Table 4. Comparison of test accuracies according to the ratio of weighted average.

Epochs of CycleGAN	Ratio Of Weighted Average (Original: Generated)	Test Accuracy
100	1:1	77.15
	2:1	88.30
	3:1	94.06
	4:1	89.83
200	1:1	71.24
	2:1	91.20
	3:1	91.94
	4:1	94.97
300	1:1	80.28
	2:1	92.78
	3:1	93.22
	4:1	92.37
400	1:1	79.76
	2:1	88.87
	3:1	92.40
	4:1	93.08
500	1:1	91.39
	2:1	93.79
	3:1	95.40
	4:1	93.47
600	1:1	91.92
	2:1	95.27
	3:1	97.06
	4:1	94.76

presents a comparison of the test accuracies according to the ratio of the weighted average. When the ratio of the original data was too high, the significance of applying the generated data using domain transfer diminished; when the ratio of the original data was too low, the difference between the original and generated data becames substantial, leading to a decrease in recognition accuracy.

With the CycleGAN trained for 600 epochs and a weighted average ratio of 3:1 between the original and generated data, a maximum recognition accuracy of 97.06% was achieved.

Table 5. Comparison of test accuracy with other methods.

Method	Epochs of CycleGAN	Accuracy	Precision	Recall	F1-Score
CycleGAN-based MobileNet-V2 [10]	100	94.11	95.63	94.11	93.78
	200	92.62	94.35	92.62	92.18
	300	90.06	93.06	90.06	89.30
	400	91.70	94.24	91.70	91.56
	500	93.25	94.96	93.25	92.69
	600	93.43	95.14	93.43	93.03
Ours	600	97.06	97.32	97.06	96.82

compares the test accuracy with that of the other methods. Based on this, superior performance was observed compared to previous studies. MobileNetV2 was trained using the Adam optimization algorithm with a batch size of 20. Cross-entropy was employed as the loss function, and the initial weights were pre-trained using the ImageNet dataset. The training began with an initial learning rate of 0.0001 and a momentum value of 0.9. Additionally, a learning rate decay of 0.2 was applied every five epochs, L2 regularization with a coefficient of 0.0001 was applied every five epochs, and the training was limited to a maximum of three epochs. Data-augmentation techniques were not used during the training. All the weights of the pre-trained model were set to be trainable. When the denominator was zero while calculating certain metrics using true positive, true negative, false positive, and false negative, the resulting metric was set to 0.5.

To separately assess performance between men and women, data from only 20 men and 20 women were constructed separately. Only the configuration of the data was changed, and the rest of the experimental process remained the same.

Table 6. Comparison of test accuracy between men and women.

Method	Gender	Epochs of CycleGAN	Accuracy
Without generated data	Men	-	99.33
	Women		97.83
With generated data	Men	600	99.17
	Women		98.67

displays the comparison of test accuracy between men and women. It is noted that this experiment was conducted with a small size of ECG data. There were no particularly noteworthy changes observed between men and women.

This method may have certain potential limitations. The input data should adhere to normalization as closely as possible. In some cases, strict normalization may be necessary. The ECG signal is a special case that can benefit from normalization. This is because the ECG signal can be organized based on fiducial points such as R-peaks. This process requires high computation, depending on the number of classes, as the current CycleGAN is restricted to learning a one-to-one transformation between two classes. Therefore, if the application necessitates many classes, the CycleGAN model is also required as it is.

The ECG signal constitutes personal medical data. Utilizing personal ECG data for research purposes requires careful consideration and adherence to research ethics regulations. It is essential to safeguard personal data, obtain consent from providers, and clearly communicate the purpose of data usage. The data should be presented anonymously, preventing the identification of individuals associated with the data. The generated data, produced using this domain adaptation method, will replace the actual personal data and decrease the reliance on real personal ECG data.

5. Conclusions

We propose a semi-supervised domain adaptation approach for individual identification using ECG signals. When acquiring ECG signals, the placement of electrode patches plays a critical role. Consistent electrode placement is essential to ensure the reliability of ECG signals for individual identification. However, maintaining a consistent placement of ECG electrode patches during data acquisition is challenging. Therefore, it is necessary to collect ECG data from various electrode placements. However, this process is labor-intensive and time-consuming because it often requires multiple attempts. To streamline this process and reduce the required labor and time, a semi-supervised domain adaptation approach for individual identification using ECG signals can be effectively used. This method works with a comprehensive dataset of original ECG signals and a smaller set of ECG signals from various placements to accommodate signal variations in the CycleGAN. In this approach, the ECG signals were transformed into time-frequency representations to train the CycleGAN. The trained generator was then used to produce ECG signals via expanding the dataset to include signals from different placements. To make the signals of the original and generated data similar, a weighted average was applied. Both original and generated signals were used to train the classifier for individual identification. This approach can also be applied to the classification of ECG signals from different sensors. For this experiment, we used the PTB-ECG dataset. Our findings indicate that the proposed method yields a higher accuracy than using only the original ECG signals for training the classifier and other methods. In a future study, we will investigate methods to enhance the domain transfer between the generated and original data with the aim of improving the classification accuracy. This involves analyzing the generated data and modifying the structure of the generator.