Identifying Atrial Fibrillation using Integrated Methods of 12-lead and Single-lead ECG during Normal Sinus Rhythm based on Artificial Intelligence

Identifying Atrial Fibrillation using Integrated Methods of 12-lead and Single-lead ECG during Normal Sinus Rhythm based on Artificial Intelligence | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Identifying Atrial Fibrillation using Integrated Methods of 12-lead and Single-lead ECG during Normal Sinus Rhythm based on Artificial Intelligence Myungeun Lee, Hyeonwoo Choi, Young Ho Lee, Ki Hong Lee This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7462212/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract The use of artificial intelligence (AI) with electrocardiogram (ECG) data has shown promise in detecting atrial fibrillation (AF). Although single-lead ECGs allow convenient and simple rhythm monitoring, arrhythmias prediction using AI is limited due to single-channel utilization. We aimed to improve the capability of AI algorithms for AF identification in integrated models with 12-lead and single-lead ECG during normal sinus rhythm (NSR). A total of 7,199 single-lead mobile ECGs were acquired from 6,806 patients. Four deep learning models, i.e., EfficientNet-B4, residual neural networks (Restnet-50), Attention Restnet-50, and long short-term memory (LSTM), were employed to analyze the dataset. To develop an integrated model, an LSTM-based generative adversarial network was used to generate 12-lead ECGs from single-lead ECGs. The generated ECGs were then applied to the identification algorithm to predict AF. The integrated ECG-based model achieved an accuracy of 0.974, precision of 0.975, recall of 0.973, and F1-score of 0.974 for the training dataset with EfficientNet-B4. The area under the receiver operating characteristic curve (AUC) value for identifying AF was 0.98 with the integrated model, 0.91 with a 12-lead ECG, and 0.88 with a single-lead ECG. The integrated ECG-based model has the advantage of analyzing both standard 12-lead ECG and single-lead ECG signals. The findings underscore the potential of the integrated model in identifying AF using NSR ECGs without the limitations of relying solely on 12-lead or single-lead data. A GUI format focusing on user convenience may be used to apply the integrated ECG-based model to clinical settings. Health sciences/Cardiology Biological sciences/Computational biology and bioinformatics artificial intelligence atrial fibrillation electrocardiography mobile applications single-lead 12-lead Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Introduction Atrial fibrillation (AF) is a significant risk factor for thromboembolism-related ischemic stroke, which can be effectively prevented by anticoagulants. However, screening for AF is difficult due to a significant number of patients with paroxysmal and asymptomatic characteristics. To address existing issues, electrocardiogram (ECG) interpretation has predominantly relied on the prediction of arrhythmias using 12-lead ECGs. Consequently, most of the research on AF has focused on using 12-lead ECGs for the identification or prediction of AF. 1,2 However, nowadays, studies employing single-lead ECGs have been increasing, which could identify/predict potential AF during normal sinus rhythm (NSR) using a mobile ECG device. In particular, there has been an increased interest in the measurement of mobile ECG signals using wearable devices such as smartwatches. 3,4 Although these wearable devices typically offer limited information, their advantage lies in the recording of data in real time. In our previous study 5 , the use of a single-lead mobile ECG device was convenient for the attachment of electrodes to the body, demonstrating high versatility for AF prediction. However, considering the paroxysmal nature of AF, long-term ECG monitoring is often essential for its detection. Various artificial intelligence (AI) methods, ranging from convolutional neural networks (CNNs) to the more recent transformer encoder models, have been implemented in ECG identification/prediction, which showed promising results. 5,6,7,8,9 To reduce measurement and analysis time, and achieve cost-effectiveness, we present an integrated ECG-based model with the advantage of analyzing both standard 12-lead ECG and single-lead ECG signals, rather than just single-lead signals. In this study, the generation of 12-lead ECG signals from single-lead counterparts was explored to enhance the level of detail available for AF prediction. To achieve this, we employed a bidirectional long short-term memory (LSTM)-based generative adversarial network. We also developed a deep learning model integrating standard 12-lead ECG signals generated from single-lead mobile ECG signals for AF identification, which significantly improved the performance of the model compared with that of a model based solely on single-lead mobile ECG signals. With enhanced AF predictability, the implementation of this deep learning model with integrated ECG signals could serve as a convenient, pre-emptive assistive tool to provide probabilistic predictions for paroxysmal AF. Moreover, the integrated ECG-based deep learning model could identify potential AF in patients using a single-lead mobile ECG device, which is user-friendly, affordable, and easily accessible. Methods Describe The methodology consists of five fundamental components: mobile ECG dataset, preprocessing procedures, ECG generation, deep learning models, and outputs (Figure 1) . The reliability of our analysis was ensured through the meticulous diagnosis and labeling of all mobile ECGs by clinical experts. The study protocol received approval from the Chonnam National University Hospital Institutional Review Board (IRB) in South Korea (IRB No. CNUH-2021-176). Written informed consent was obtained from all patients. All experiments were carried out in accordance with relevant guidelines and regulations. Data collection. Mobile ECGs were prospectively and consecutively collected from consenting patients who visited the Departments of Cardiology and Neurology at a tertiary teaching institution. From September 2021 to December 2022, 7,199 single-lead ECGs from 6,806 patients were acquired using the mobiCARE MC100 device (Seers Technology, Seongnam, Korea). The ECG device (mobiCARE MC100) features a single lead and is a patch-type device that adheres to the chest rather than being handheld. The ECGs were recorded at a sampling rate of 256 Hz and consisted of single-lead recordings, each lasting approximately 60 s. The baseline characteristics of the patients are presented in Table 1 . Based on demographic data, 60.3% of the patients were male, with a mean age of 67.1 years. Patient medical histories included hypertension (79.0%), diabetes mellitus (26.2%), dyslipidemia (59.4%), myocardial infarction (24.7%), angina pectoris (43.5%), aortic disease (2.8%), peripheral artery disease (0.4%), arrhythmia (42.8%), sudden cardiac arrest (1.7%), heart failure (12.7%), hypertrophic cardiomyopathy (1.0%), dilated cardiomyopathy (1.9%), and cerebrovascular disease (6.9%). We applied specific exclusion criteria to ensure the suitability of the data for analysis as follows: ECGs showing AF or non-sinus rhythm and those without a confirmed patient identification number (PID) or containing incomplete information. Consequently, a total of 1,425 ECGs from 1,032 patients were excluded from the study, comprising 493 ECGs from 100 patients with missing digital information or unverified PID and 932 ECGs from 932 patients with confirmed AF or non-sinus rhythm (Figure 2) . For analysis, we utilized 5,774 NSR ECGs obtained from 5,774 patients. Among these NSR ECGs, 1,192 NSR ECGs from 1,192 patients with a history of AF were categorized as "AF", whereas 4,582 ECGs from 4,582 patients were categorized as "Healthy". The presence of AF history was determined based on at least one documented 12-lead ECG showing AF. Notably, a history of AF without a documented 12-lead ECG was not considered to indicate a history of AF. Data preprocessing. To reduce noise in the ECG data, a 0.5 Hz high-pass Butterworth filter was applied to eliminate low-frequency interference. The implementation of this filter was executed utilizing the neurokit2 package. 10 After the identification of the R-peak within the processed signal, the data underwent segmentation, with each segment containing five consecutive R-peaks, thereby ensuring uniformity across the segments. During the preprocessing phase, segments with fewer than five R-peaks, abnormally extended R-peaks, or delayed waveforms due to suboptimal measurement were excluded from the data analysis. Resampling was performed to standardize the length of the segmented data, thereby maintaining consistency across the dataset. Through this process, a total of 32,066 data segments were generated (Figure 3) . To address the data imbalance issue and enhance the generalizability of the developed model while mitigating bias, we employed a random undersampling technique. 11,12 This technique removed 18,462 pieces of data, ultimately resulting in a balanced dataset of 13,604 pieces, consisting of 6,802 NSR ECGs and 6,802 AF cases. To facilitate the training, validation, and evaluation of the deep learning model, the dataset was partitioned into the training, internal validation, and testing datasets at a ratio of 6:2:2. Data generation. Before generating 12-lead ECGs from single-lead ECGs, the process of PID matching was conducted between the two datasets. Data from a 12-lead ECG during NSR were used, which were obtained 6 months before and after the date of single-lead ECG measurement. Consistency between the two datasets was ensured by subjecting them to the same preprocessing process as used for the single-lead ECG. The model used for generative learning was the bidirectional LSTM model 11 , which allows information to flow bidirectionally in time, enhancing the learning of the entire sequence data. Deep learning model. EfficientNet-B4 represents a deep learning architecture designed to optimize both performance and efficiency through structural scalability. 14 This model adapts the intricate layers of existing deep learning frameworks, offering scalability across various sizes. By simultaneously adjusting the scaling of depth, width, and resolution, superior performance can be achieved while maintaining a lower computational load using EfficientNet-B4 compared with previous models. In this study, EfficientNet-B4 was used to detect potential AF during NSR. A dimensionality augmentation method was employed to effectively process the ECG signal, allowing it to be treated as an image. To compare different models, a comparative experiment was conducted using Resnet-50, Attention Resnet-50, and LSTM. Supplementary 1 presents detailed information on the Resnet-50, Attention Resnet-50, and LSTM models. Model optimization and implementation. The deep learning model was trained using the AdamW 15,16 optimizer. Specifically, the learning rate was set to 0.0001, and the weights were adjusted to 0.001 to facilitate the proper management of weight updates throughout the learning iteration. The batch size was set to 128, and most experiments lasted between 200 and 300 epochs. To mitigate the risk of overfitting, cross-entropy loss was employed as the designated loss function. All training procedures were executed using an Nvidia H100 graphics processing unit (GPU) with Python 3.10 and the PyTorch 2.0.1 framework. External validation dataset. The external validation dataset comprised 72 h single-lead Holter ECGs from 160 distinct patients. From May 2022 to June 2023, a total of 160 ECGs were obtained from 160 patients with palpitations. The same device was used for all data acquisition. The ECGs were sampled at a rate of 256 Hz. The data were segmented based on R-peak intervals to maintain consistency with the training data. The ECGs of 54 patients with missing data or unmatched PID, 54 patients with abnormal heart rhythm, or those judged to be in AF were excluded. After applying the exclusion criteria, the ECG data of 101 patients were used for external validation. This dataset was segmented based on R-R intervals, resulting in a total of 2,354 data segments. Through undersampling, a total of 1,606 data segments were retrained. Results To evaluate and compare the performance of the AI-enhanced ECG-based models in identifying AF during NSR, a comprehensive statistical analysis was conducted. Four deep learning models (EfficientNet-B4, Restnet-50, Attention Restnet-50, and LSTM) were employed to analyze data refined by specific exclusion criteria. All four AI methods were evaluated based on their precision, recall, F1-score, area under the receiver operating characteristic curve (AUC), and accuracy. Data generation results. Figure 4 presents the generated 12-lead ECG signals alongside the corresponding real 12-lead ECG signals. The red line represents the original ECG, and the blue line represents the generated ECG. The quality of the reconstructed 12-lead ECG was assessed through both quantitative and qualitative evaluations. For quantitative evaluation, the coefficient of determination (R² score) was used, where the score ranges from 0 to 1, with values closer to 1 indicating a stronger correlation with the data. The average R² score for ECG signals generated from single-lead ECG signals was approximately 0.6, which can be interpreted as above average. Qualitative evaluation involved assessing the morphological similarity between the original and reconstructed 12-lead ECGs. Four expert clinicians visually scored each ECG over more than 1 month, evaluating parameters such as the T-wave inversion, S-T segment, QRS voltage, R-R interval, and axis for each reconstructed ECG. Performance comparison of different deep learning models. The performance of the four deep learning models in utilizing the generated data to identify potential AF was compared, and the results are presented in Table 2 . For the training dataset, the EfficientNet-B4 model demonstrated precision, recall, F1-score, AUC, and accuracy of 97.5%, 97.3%, 97.4%, 99.7%, and 97.4%, respectively, in AF identification. The Resnet-50 model achieved a precision of 98.2%, a recall of 98.1%, an F1-score of 98.2%, an AUC of 99.8%, and an accuracy of 98.2%. On the other hand, the Attention Resnet-50 model demonstrated a precision of 78.3%, a recall of 80.8%, an F1-score of 79.5%, an AUC of 87.8%, and an accuracy of 73.9%. The LSTM model showed a precision of 64.5%, a recall of 60.3%, an F1-score of 62.3%, an AUC of 68.6%, and an accuracy of 63.7%. Overall, the integrated ECG-based models showed improved performance. Specifically, the performance of the Resnet-50 model was better than that of the single-lead ECG-based model for the training dataset, improving the F1-score from 79.3% to 98.2%. The accuracy was increased from 80.6% to 99.8%. The EfficientNet-B4 model demonstrated precision, recall, F1-score, AUC, and accuracy of 72.7%, 73.5%, 73.1%, 80.1%, and 72.4%, respectively, in AF identification in the testing dataset. The Resnet-50 model achieved a precision of 73.1%, a recall of 74.24%, an F1-score of 73.7%, an AUC of 79.4%, and an accuracy of 72.9%, whereas the Attention Resnet-50 model demonstrated a precision of 65.4%, a recall of 68.7%, an F1-score of 67%, an AUC of 71.4%, and an accuracy of 65.5%. Similarly, the LSTM model showed a precision of 63.1%, a recall of 60.7%, an F1-score of 61.9%, an AUC of 66.1%, and an accuracy of 61.9%. Overall, improved results were obtained using integrated ECG compared with previous results obtained using only single-lead ECG 5 for both the testing dataset and the training dataset. For the internal validation dataset, EfficientNet-B4 showed precision, recall, F1-score, AUC, and accuracy of 71.7%, 73.9%, 72.8%, 79.5%, and 72.6%, respectively. The Resnet-50 model demonstrated a precision of 70.2%, a recall of 72.0%, an F1-score of 71.1%, an AUC of 77.1%, and an accuracy of 71%. The Attention Resnet-50 model exhibited a precision of 63.5%, a recall of 67.9%, an F1-score of 65.6%, an AUC of 70.3%, and an accuracy of 64.8%. The LSTM model demonstrated a precision of 60.1%, a recall of 56.5%, an F1-score of 58.3%, an AUC of 63.8%, and an accuracy of 59.9%. In the case of internal validation, compared with an F1-score value of 69.1% in our previous study with Resnet-50, we achieved F1-scores of 72.8% and 71.1% with EfficientNet-B4 and Resnet-50, respectively, in the current study. Confusion matrices were used to visually assess and compare the classification performance of each model. The confusion matrix results for EfficientNet-B4, Resnet-50, Attention Resnet-50, and LSTM using the testing dataset are shown in Figure 5 . Receiver operating characteristic (ROC) curves and AUC values for the identification of potential AF in the training dataset, internal validation dataset, and testing dataset are presented in Figure 6 . In the determination of AF during NSR, EfficientNet-B4 demonstrated the highest AUC at 0.801, followed by Resnet-50 at 0.794, Attention Resnet-50 at 0.714, and LSTM at 0.661 for the testing dataset. For the training dataset, the AUCs of EfficientNet-B4, Resnet-50, Attention Resnet-50, and LSTM were 0.997, 0.998, 0.878, and 0.686, respectively. For the internal validation dataset, the AUCs of EfficientNet-B4, Resnet-50, Attention Resnet-50, and LSTM were 0.795, 0.771, 0.703, and 0.638, respectively. Performance on external validation set. For the external validation dataset, EfficientNet-B4 showed precision, recall, F1-score, AUC, and accuracy of 76.7%, 63.1%, 69.2%, 76.2%, and 71.9%, respectively. The Resnet-50 model demonstrated a precision of 58.2%, a recall of 87.1%, an F1-score of 69.8%, an AUC of 67%, and an accuracy of 62.3%. The Attention Resnet-50 model exhibited a precision of 58.2%, a recall of 43.3%, an F1-score of 49.7%, an AUC of 58.8%, and an accuracy of 56.1%. The LSTM model demonstrated a precision of 51.4%, a recall of 95%, an F1-score of 66.7%, an AUC of 45%, and an accuracy of 52.6%. Discussion In this study, we developed and evaluated a deep learning model integrating 12-lead ECG signals generated from single-lead ECG recordings for identifying potential AF during NSR. The proposed deep learning model achieved an AUC of 0.99, with an internal validation of 0.79. An AUC value of 0.7 or higher for the testing/validation dataset generally indicates good performance. Our results were also compared with those of a previous study 5 , in which AF was detected using only 12-lead or single-lead ECG data. The comparison revealed that the integrated ECG-based model could produce reliable results. The findings demonstrated that potential AF could be identified using integrated ECG data rather than relying only on 12-lead or single-lead ECG data. Until recently, only a limited number of studies utilized single-lead ECG data obtained from single-lead mobile ECG devices for the classification of possible AF during NSR. Most studies have predominantly concentrated on utilizing 12-lead ECG data to identify AF. 1,2,17,18 The benefits of a single-lead mobile ECG device include convenient body attachment and high versatility, allowing extended ECG monitoring and real-time data collection. 3 However, considering the paroxysmal nature of AF, long-term ECG monitoring is frequently necessary for its identification. 19 The results of this study demonstrated the reliability of single-lead ECG data collected with a single-lead mobile ECG device, suggesting its potential as a convenient and highly effective tool for ECG data collection. In addition, this study focused on data augmentation and 12-lead signals generated from single-lead ECG signals, which is consistent with the increasing trend of research on generative models in recent studies. 20,21,22 A platform for predicting ventricular arrhythmias has recently been introduced. 23 If our proposed AF prediction model could be integrated into this platform, it is expected that a wide range of arrhythmias could be detected. Our deep learning model integrating 12-lead ECGs generated from single-lead ECGs was uniquely designed to predict AF occurrence even during NSR. The proposed model could be better for the prediction of arrhythmias, which improves on the disadvantages of using only 12-lead or single-lead ECG signals. Although the generated ECG signals generally reflected waveform changes in the original ECG signals, in the case of lead III, the correlation was low. The generated signals did not accurately capture subtle waveform changes in areas with significant fluctuations. The data collection process involved 6,806 patients, which led to the acquisition and analysis of 7,199 mobile ECGs. An issue frequently encountered with various medical datasets is the problem of data imbalance, characterized by the underrepresentation of specific health conditions in classification algorithms. This issue was apparent in our study, especially due to the limited availability of ECGs labeled as AF. The data were primarily categorized as “Healthy,” potentially impeding the identification of AF. To address the data imbalance issue and improve the generalization ability of our model while reducing bias, we utilized a random under-sampling technique. Moreover, ECG signals that were excessively noisy or contained a high number of artifacts were not included in the training dataset. For the external validation dataset, the same methodology was applied as used for our training, testing, and internal validation datasets. The 72 h Holter ECGs from 160 distinct palpitation patients were divided into R-R peak segments, and random sampling without replacement was conducted to obtain a representative subset for validation. This method of random sampling without replacement ensures a more unbiased analysis by avoiding potential bias due to the researchers’ preferences or assumptions. Moreover, four deep learning models (EfficientNet-B4, Resnet-50, Attention Resnet-50, and LSTM) were utilized to identify potential AF by integrating 12-lead ECGs generated from single-lead ECGs during NSR, and their performance was compared. As shown in Table 2, our proposed EfficientNet-B4 model was excellent with precision, recall, F1-score, AUC, and accuracy of 97.5%, 97.3%, 97.4%, 99.7%, and 97.4%, respectively. The Resnet-50 model with the training dataset also demonstrated superior performance, with precision, recall, F1-score, AUC, and accuracy of 98.2%, 98.1%, 98.2%, 99.8%, and 98.2%, respectively. In particular, the Resnet-50 model achieved better results compared with those obtained with our previous single-lead ECG-based model 3 for the training dataset, improving the F1-score from 79.3% to 98.2%. The results suggest that the overall performance of the integrated ECG-based models was better than that of the single-lead ECG-based model. There are some limitations in this study. First, when converting single-lead ECG data to 12-lead ECG data, it would be ideal to obtain simultaneously measured data; however, this was not realistically feasible. Therefore, to ensure the accuracy of the 12-lead ECG data, 12-lead ECG data with NSR status within the past 6 months were used as the reference point. To overcome this limitation, we plan to obtain single-lead ECG data and 12-lead ECG data simultaneously. During the 72 h measurement period, there were no waveforms or significant outliers observed, which could be attributed to incorrect patient movement or improper machine attachment during external verification. In addition, due to the large amount of data, it was not possible to segment all 72 h of data based on R-R intervals. Even for data classified as NSR, there were many outliers; thus, relevant parts had to be excluded. If more data could be obtained in the future, we plan to conduct the experiment again, excluding all the above cases. Further research is necessary to improve the performance and generalizability of AI models in clinical settings. Although our proposed deep learning model, which utilizes integrated ECG signals, demonstrated promising outcomes in the identification of AF with an AUC of 80.1% for the testing dataset, its efficacy was comparatively lower for the external validation dataset. This could have implications with the small amount of external testing data available. Consequently, we intend to acquire additional data and carry out further testing. Moreover, previous studies have indicated that the R-R interval, P-wave morphology, and QT interval are significant factors affecting the risk of AF. 24,25 The study utilized ECGs to train the deep learning model. However, future studies could enhance model accuracy by incorporating additional relevant features as indicators of AF. Conclusion In the case of the integrated ECG-based model, deep learning was applied to predict potential AF, showing high accuracy and F1-score. In comparison with existing results obtained for single-lead ECG signals with the Resenet-50 model, the accuracy was improved by around 14%. In addition, the results of the EfficientNet-B4 model proposed in this study were better than those of the Restnet-50 model. However, the LSTM model showed somewhat worse results. Nevertheless, the integrated ECG-based model has the advantage of analyzing both standard 12-lead ECG and single-lead ECG signals rather than analyzing only single-lead ECG signals. The use of the proposed model in clinical applications may reduce time and cost. Finally, a GUI format focusing on user convenience can be found in Supplementary Figure 1, which may be used to apply the proposed integrated ECG-based model to actual clinical settings. Declarations ACKNOWLEDGMENTS Special thanks to Dr. Seong Won Jeon, Dr. Changhyun Kim, and Dr. Dong Kyun Kim for the diagnosis and labeling mobile electrocardiograms. Disclosure The authors have no potential conflicts of interest to disclose. Author Contributions Lee M and Choi H performed data curation, formal analysis, validation and visualization; Lee M and Lee KH conducted the investigation; Lee M, Choi H, Lee YH and Lee KH designed the methodology; Choi H developed the software; Lee KH acquired funding; Lee M wrote the original draft; Lee M and Lee KH reviewed and edited the manuscript. All authors reviewed the manuscript. Data availability The mobile and 12-lead ECG electrocardiogram dataset generated and analyzed during the current study is not publicly available due to privacy and ethical restrictions stipulated by the Institutional Review Board of Chonnam National University Hospital. However, it is available from the corresponding author upon reasonable request. Funding: This work was supported by a grant (BCRI25041) of Chonnam National University Hospital Biomedical Research Institute, in part a grant of Establishment of K-Health National Medical Care Service and Industrial Ecosystem funded by the Ministry of Science and ICT (MSIT, Korea) Balanced National Development Account. [Project Name: Establishment of K-Health National Medical Care Service and Industrial Ecosystem/Project Number: ITAH0603230110010001000100100], and in part by Institute of Information & communications Technology Planning & Evaluation (IITP) under the Artificial Intelligence Convergence Innovation Human Resources Development (IITP-2023-RS-2023-00256629) grant funded by the Korea government (MSIT). References Attia, Z. I. et al. An artificial intelligence-enabled ECG algorithm for the identification of patients with atrial fibrillation during sinus rhythm: a retrospective analysis of outcome prediction. Lancet 394 , 861–867. DOI:10.1016/S0140-6736(19)31721-0 (2019). Baek, Y. et al. A new deep learning algorithm of 12-lead electrocardiogram for identifying atrial fibrillation during sinus rhythm. Sci. Rep . 11 , 12818. https://doi.org/10.1038/s41598-021-92172-5 (2021). Sana, F. et al. Wearable devices for ambulatory cardiac monitoring: JACC state-of-the-art review. J. Am. Coll. Cardiol 75 , 1582–1592. DOI:10.1016/j.jacc.2020.01.046 (2020). Perez, M. V. et al. Large-scale assessment of a smartwatch to identify atrial fibrillation. N. Engl. J. Med. 381 , 1909–1917. DOI: 10.1056/NEJMoa1901183 (2019). Kim, J. et al. Identification of atrial fibrillation with single-lead mobile ECG during normal sinus rhythm using deep learning. J. Korean Med. Sci. 39 , https://doi.org/10.3346/jkms.2024.39.e56 (2024) Yıldırım, Ö., Pławiak, P., Tan, R.-S. & Acharya, U. R. Arrhythmia detection using deep convolutional neural network with long duration ECG signals. Comput. Biol. Med. 102 , 411–420. https://doi.org/10.1016/j.compbiomed.2018.09.009 (2018). Sherin, M., Chandra, K. & Kenneth, B. A novel application of deep learning for single-lead ECG classification. Comput. Biol. Med . 99 , 53–62. DOI:10.1016/j.compbiomed.2018.05.013 (2018). Petmezas, G., Leandros, S., Vassilis, K. et al. State-of-the-art deep learning methods on electrocardiogram data: systematic review. JMIR Med. Inform. 10 , e38454. doi:10.2196/38454 (2022). Baek, Y. S., Kwon, S., Seng, C. Y. et al. Artificial intelligence-enhanced 12-lead electrocardiography for identifying atrial fibrillation during sinus rhythm (AIAFib) trial: protocol for a multicenter retrospective study. Front. Cardiovasc. Med. 10 , 1258167. doi:10.3389/fcvm.2023.1258167 (2023). Makowski, D., Pham, T., Lau, Z. J. et al. NeuroKit2: a Python toolbox for neurophysiological signal processing. Behav. Res. Methods 53 , 1689–1696. https://doi.org/10.3758/s13428-020-01516-y (2021). Lee, W. & Seo, K. Downsampling for binary classification with a highly imbalanced dataset using active learning. Big Data Res . 28 , 100314. doi:10.1016/j.bdr.2022.100314 (2022). Qazi, N. & Raza, K. Effect of feature selection, SMOTE and under sampling on class imbalance classification. In 2012 UKSim 14th International Conference on Computer Modelling and Simulation 145–150. DOI : 10.1109/UKSim.2012.116 (2012). Hochreiter, S. & Schmidhuber, J. Long short-term memory. Neural Comput . 9 , 1735–1780. DOI : 10.1162/neco.1997.9.8.1735 (1997). Tan, M. & Le, Q. EfficientNet: rethinking model scaling for convolutional neural networks. In Proceedings of the 36th International Conference on Machine Learning (eds. Chaudhuri, K. & Salakhutdinov, R.) Proc. Mach. Learn. Res . 97 , 6105–6114 (2019). Reyad, M., Sarhan, A. M. & Arafa, M. A modified Adam algorithm for deep neural network optimization. Neural Comput. Appl . 35 , 17095–17112. https://doi.org/10.1007/s00521-023-08568-z (2023). Jia, X., Feng, X., Yong, H. & Meng, D. Weight decay with tailored Adam on scale-invariant weights for better generalization. IEEE Trans. Neural Netw. Learn. Syst. 35 . DOI:10.1109/TNNLS.2022.3213536 (2022). Somani, S. et al. Deep learning and the electrocardiogram: review of the current state-of-the-art. Europace 23 , 1179–1191. DOI:10.1093/europace/euaa377 (2021). Ribeiro, A. H. et al. Automatic diagnosis of the 12-lead ECG using a deep neural network. Nat. Commun. 11 , 1760. https://doi.org/10.1038/s41467-020-15432-4 (2020). Nigusse, A. B., Mengistie, D. A., Malengier, B., Tseghai, G. B. & Van Langenhove, L. Wearable smart textiles for long-term electrocardiography monitoring—a review. Sensors 21 , 4174. https://doi.org/10.3390/s21124174 (2021). Seo, H. C., Yoon, G. W., Joo, S. & Na, G. B. Multiple electrocardiogram generator with single-lead electrocardiogram. Comput. Methods Programs Biomed. 221 , 106858. https://doi.org/10.1016/j.cmpb.2022.106858 (2022). Shin, H., Sun, S., Lee, J. & Kim, H. Complementary photoplethysmogram synthesis from electrocardiogram using generative adversarial network. IEEE Access 9 , 70639–70649. DOI:10.1109/ACCESS.2021.3078534 (2021). Yoon, G. W. & Joo, S. Classification feasibility test on multi-lead electrocardiography signals generated from single-lead electrocardiography signals. Sci . Rep . 14 , 1888. https://doi.org/10.1038/s41598-024-52216-y (2024). Compagnucci, P. et al. Implantable defibrillator-detected heart failure status predicts ventricular tachyarrhythmias. J. Cardiovasc. Electrophysiol. 34 , 1257–1267. DOI: 10.1111/jce.15898 (2023). Coppola, E. E., Gyawali, P. K., Vanjara, N., Giaime, D. & Wang, L. Atrial fibrillation classification from a short single lead ECG recording using hierarchical classifier. IEEE Computing in Cardiology (CinC) 354–425. DOI: 10.22489/CinC.2017.354-425 (2017). Yazdani, S., Laub, P., Luca, A. & Vesin, J. M. Heart rhythm classification using short-term ECG atrial and ventricular activity analysis. IEEE Computing in Cardiology (CinC) 67–120. DOI : 10.22489/CinC.2017.067-120 (2017). Additional Declarations No competing interests reported. Supplementary Files 20250818IntegrationECGSupplementaryFig.1.docx 20250818IntegrationECGSupplementaryMethods.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7462212","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":525061123,"identity":"61720f72-de6c-4c18-a8ae-03c46d37e975","order_by":0,"name":"Myungeun Lee","email":"","orcid":"","institution":"Chonnam National University","correspondingAuthor":false,"prefix":"","firstName":"Myungeun","middleName":"","lastName":"Lee","suffix":""},{"id":525061127,"identity":"3f1d78f5-ab9e-4497-b7c4-93b2b553df4b","order_by":1,"name":"Hyeonwoo Choi","email":"","orcid":"","institution":"Chonnam National University Hospital","correspondingAuthor":false,"prefix":"","firstName":"Hyeonwoo","middleName":"","lastName":"Choi","suffix":""},{"id":525061129,"identity":"c78c1dd2-ddb2-4516-8b97-d578c5237fef","order_by":2,"name":"Young Ho Lee","email":"","orcid":"","institution":"Chonnam National University Hospital","correspondingAuthor":false,"prefix":"","firstName":"Young","middleName":"Ho","lastName":"Lee","suffix":""},{"id":525061130,"identity":"e31a00ba-9b04-4ded-b37a-2c1f831bb064","order_by":3,"name":"Ki Hong Lee","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAuklEQVRIiWNgGAWjYDCCAyCCzUaOZC1pxiRrOZzYQLQOvtvHLz74UMacvp39jOHjCgY7OV1CmiXP5RQbzjjHlruzJ8fY8AxDsrHZAQJaDM7wpEnztvHkbjiQlibZwHAgcRuRWiTSDc4/S/9JpBb2Y0AtBgkGN5KPMRKlRfIMDzPQLwmGG248PizZYECEX/jOsD8Ehth/eYPziY0fGyrs5AhqYWDgMUB2J0HlIMD+gChlo2AUjIJRMIIBAGHSQ/OBxFnlAAAAAElFTkSuQmCC","orcid":"","institution":"Chonnam National University Hospital","correspondingAuthor":true,"prefix":"","firstName":"Ki","middleName":"Hong","lastName":"Lee","suffix":""}],"badges":[],"createdAt":"2025-08-26 11:08:16","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7462212/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7462212/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":93062985,"identity":"b43f723f-5b6a-492e-a253-bf3aaf069377","added_by":"auto","created_at":"2025-10-08 16:38:06","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":4539499,"visible":true,"origin":"","legend":"\u003cp\u003eProposed method for AF identification using integrated with 12-lead and mobile ECG in normal sinus rhythm. \u003cstrong\u003e(A)\u003c/strong\u003e Input stage with raw ECG signal. \u003cstrong\u003e(B)\u003c/strong\u003ePreprocessing stage. \u003cstrong\u003e(C)\u003c/strong\u003e Generation of 12-lead ECG signals from single-lead ECG signals. \u003cstrong\u003e(D)\u003c/strong\u003e Application of EfficientNet-B4. \u003cstrong\u003e(E)\u003c/strong\u003eFinal output stage. ECG = electrocardiogram, AF = atrial fibrillation, Conv = convolutional layer, ReLU = rectified linear unit.\u003c/p\u003e","description":"","filename":"Figure.1dataflow.png","url":"https://assets-eu.researchsquare.com/files/rs-7462212/v1/322f7d73924a221d52b99f6c.png"},{"id":93063381,"identity":"f098c9a6-e93b-42ff-9ef7-edfa1fa3e0b6","added_by":"auto","created_at":"2025-10-08 16:38:15","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":95036,"visible":true,"origin":"","legend":"\u003cp\u003eFlow diagram of patient data selection. PID = patient identification number, AF = atrial fibrillation, ECG = electrocardiogram, NSR = normal sinus rhythm.\u003c/p\u003e","description":"","filename":"Figure.2Maindataset.png","url":"https://assets-eu.researchsquare.com/files/rs-7462212/v1/5eb251acd61da7c949040081.png"},{"id":93063251,"identity":"b84fbd83-bebd-44dc-badc-c88c0789c2de","added_by":"auto","created_at":"2025-10-08 16:38:11","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":40827606,"visible":true,"origin":"","legend":"\u003cp\u003eProcess of preprocessing of patient data. \u003cstrong\u003e(A)\u003c/strong\u003e Raw ECG signal. \u003cstrong\u003e(B)\u003c/strong\u003e Denoising of ECG signal. \u003cstrong\u003e(C)\u003c/strong\u003e R-R interval of ECG segments. ECG = electrocardiogram.\u003c/p\u003e","description":"","filename":"Figure.3dataprocessing.png","url":"https://assets-eu.researchsquare.com/files/rs-7462212/v1/2b7ded428a617a37abb1750f.png"},{"id":93063143,"identity":"aa9649ea-34b9-4fae-832d-ef8d28d896f3","added_by":"auto","created_at":"2025-10-08 16:38:08","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":21182742,"visible":true,"origin":"","legend":"\u003cp\u003eGenerated 12-lead ECG signals from single-lead ECG signals.\u003c/p\u003e","description":"","filename":"Figure.4Generation12lead.png","url":"https://assets-eu.researchsquare.com/files/rs-7462212/v1/962e52f45de762585ca2b853.png"},{"id":93062841,"identity":"29546bb4-0328-468c-8e3f-14a6436273c0","added_by":"auto","created_at":"2025-10-08 16:38:04","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":8887209,"visible":true,"origin":"","legend":"\u003cp\u003ePerformance evaluation of deep learning models with confusion matrices using the testing dataset. \u003cstrong\u003e(A)\u003c/strong\u003e Confusion matrix of EfficientNet-B4. \u003cstrong\u003e(B)\u003c/strong\u003e Confusion matrix of Resnet-50\u003cstrong\u003e. (C)\u003c/strong\u003e Confusion matrix of Attention Resnet-50\u003cstrong\u003e. (D)\u003c/strong\u003eConfusion matrix of LSTM.\u003c/p\u003e\n\u003cp\u003eResnet = residual neural network, LSTM = long short-term memory.\u003c/p\u003e","description":"","filename":"Figure.5originalconfusionmatrix.png","url":"https://assets-eu.researchsquare.com/files/rs-7462212/v1/0242f4ac055cdb3d499a48d0.png"},{"id":93062810,"identity":"e77d6682-c6f1-4519-a798-d0872ea86372","added_by":"auto","created_at":"2025-10-08 16:38:03","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":29524739,"visible":true,"origin":"","legend":"\u003cp\u003eROC curves demonstrating the performance of deep learning models using integrated ECG for AF identification during NSR. \u003cstrong\u003e(A)\u003c/strong\u003e Training dataset\u003cstrong\u003e. (B)\u003c/strong\u003e Testing dataset. \u003cstrong\u003e(C)\u003c/strong\u003e Internal validation dataset.\u003c/p\u003e\n\u003cp\u003eROC = receiver operating characteristic, AUC = area under the receiver operating characteristic curve, Resnet = residual neural network, LSTM = long short-term memory.\u003c/p\u003e","description":"","filename":"Figure.6generationroccurves.png","url":"https://assets-eu.researchsquare.com/files/rs-7462212/v1/64978b202d915dbb75c3c6b6.png"},{"id":93061651,"identity":"97706d8e-3362-4c3f-b5d9-5d51c2b60ab6","added_by":"auto","created_at":"2025-10-08 16:20:16","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":514989,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7462212/v1/045247bb-4cd7-4acc-a862-aea427419aca.pdf"},{"id":93062937,"identity":"7cbce1a7-4224-4bea-8f76-051aebce64b6","added_by":"auto","created_at":"2025-10-08 16:38:05","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":356166,"visible":true,"origin":"","legend":"","description":"","filename":"20250818IntegrationECGSupplementaryFig.1.docx","url":"https://assets-eu.researchsquare.com/files/rs-7462212/v1/2cb831750d636a4918afc12e.docx"},{"id":93063212,"identity":"129d272c-3626-4510-89b0-a42e9550650f","added_by":"auto","created_at":"2025-10-08 16:38:09","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":16872,"visible":true,"origin":"","legend":"","description":"","filename":"20250818IntegrationECGSupplementaryMethods.docx","url":"https://assets-eu.researchsquare.com/files/rs-7462212/v1/9d3bb75b8d73276f58cf2c12.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Identifying Atrial Fibrillation using Integrated Methods of 12-lead and Single-lead ECG during Normal Sinus Rhythm based on Artificial Intelligence","fulltext":[{"header":"Introduction","content":"\u003cp\u003eAtrial fibrillation (AF) is a significant risk factor for thromboembolism-related ischemic stroke, which can be effectively prevented by anticoagulants. However, screening for AF is difficult due to a significant number of patients with paroxysmal and asymptomatic characteristics.\u003c/p\u003e\n\u003cp\u003eTo address existing issues, electrocardiogram (ECG) interpretation has predominantly relied on the prediction of arrhythmias using 12-lead ECGs. Consequently, most of the research on AF has focused on using 12-lead ECGs for the identification or prediction of AF.\u003csup\u003e1,2\u003c/sup\u003e However, nowadays, studies employing single-lead ECGs have been increasing, which could identify/predict potential AF during normal sinus rhythm (NSR) using a mobile ECG device. In particular, there has been an increased interest in the measurement of mobile ECG signals using wearable devices such as smartwatches.\u003csup\u003e3,4\u0026nbsp;\u003c/sup\u003eAlthough these wearable devices typically offer limited information, their advantage lies in the recording of data in real time. In our previous study\u003csup\u003e5\u003c/sup\u003e, the use of a single-lead mobile ECG device was convenient for the attachment of electrodes to the body, demonstrating high versatility for AF prediction. However, considering the paroxysmal nature of AF, long-term ECG monitoring is often essential for its detection. Various artificial intelligence (AI) methods, ranging from convolutional neural networks (CNNs) to the more recent transformer encoder models, have been implemented in ECG identification/prediction, which showed promising results.\u003csup\u003e5,6,7,8,9\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003eTo reduce measurement and analysis time, and achieve cost-effectiveness, we present an integrated ECG-based model with the advantage of analyzing both standard 12-lead ECG and single-lead ECG signals, rather than just single-lead signals. In this study, the generation of 12-lead ECG signals from single-lead counterparts was explored to enhance the level of detail available for AF prediction. To achieve this, we employed a bidirectional long short-term memory (LSTM)-based generative adversarial network. We also developed a deep learning model integrating standard 12-lead ECG signals generated from single-lead mobile ECG signals for AF identification, which significantly improved the performance of the model compared with that of a model based solely on single-lead mobile ECG signals. With enhanced AF predictability, the implementation of this deep learning model with integrated ECG signals could serve as a convenient, pre-emptive assistive tool to provide probabilistic predictions for paroxysmal AF. Moreover, the integrated ECG-based deep learning model could identify potential AF in patients using a single-lead mobile ECG device, which is user-friendly, affordable, and easily accessible.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003eDescribe The methodology consists of five fundamental components: mobile ECG dataset, preprocessing procedures, ECG generation, deep learning models, and outputs \u003cstrong\u003e(Figure 1)\u003c/strong\u003e. The reliability of our analysis was ensured through the meticulous diagnosis and labeling of all mobile ECGs by clinical experts.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe study protocol received approval from the Chonnam National University Hospital Institutional Review Board (IRB) in South Korea (IRB No. CNUH-2021-176). Written informed consent was obtained from all patients. All experiments were carried out in accordance with relevant guidelines and regulations.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData collection.\u0026nbsp;\u003c/strong\u003eMobile ECGs were prospectively and consecutively collected from consenting patients who visited the Departments of Cardiology and Neurology at a tertiary teaching institution. From September 2021 to December 2022, 7,199 single-lead ECGs from 6,806 patients were acquired using the mobiCARE MC100 device (Seers Technology, Seongnam, Korea). The ECG device (mobiCARE MC100) features a single lead and is a patch-type device that adheres to the chest rather than being handheld. The ECGs were recorded at a sampling rate of 256 Hz and consisted of single-lead recordings, each lasting approximately 60 s.\u003c/p\u003e\n\u003cp\u003eThe baseline characteristics of the patients are presented in\u0026nbsp;\u003cstrong\u003eTable 1\u003c/strong\u003e. Based on demographic data, 60.3% of the patients were male, with a mean age of 67.1 years. Patient medical histories included hypertension (79.0%), diabetes mellitus (26.2%), dyslipidemia (59.4%), myocardial infarction (24.7%), angina pectoris (43.5%), aortic disease (2.8%), peripheral artery disease (0.4%), arrhythmia (42.8%), sudden cardiac arrest (1.7%), heart failure (12.7%), hypertrophic cardiomyopathy (1.0%), dilated cardiomyopathy (1.9%), and cerebrovascular disease (6.9%).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eWe applied specific exclusion criteria to ensure the suitability of the data for analysis as follows: ECGs showing AF or non-sinus rhythm and those without a confirmed patient identification number (PID) or containing incomplete information. Consequently, a total of 1,425 ECGs from 1,032 patients were excluded from the study, comprising 493 ECGs from 100 patients with missing digital information or unverified PID and 932 ECGs from 932 patients with confirmed AF or non-sinus rhythm \u003cstrong\u003e(Figure 2)\u003c/strong\u003e. For analysis, we utilized 5,774 NSR ECGs obtained from 5,774 patients. Among these NSR ECGs, 1,192 NSR ECGs from 1,192 patients with a history of AF were categorized as \"AF\", whereas 4,582 ECGs from 4,582 patients were categorized as \"Healthy\". The presence of AF history was determined based on at least one documented 12-lead ECG showing AF. Notably, a history of AF without a documented 12-lead ECG was not considered to indicate a history of AF.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData preprocessing.\u0026nbsp;\u003c/strong\u003eTo reduce noise in the ECG data, a 0.5 Hz high-pass Butterworth filter was applied to eliminate low-frequency interference. The implementation of this filter was executed utilizing the neurokit2 package.\u003csup\u003e10\u003c/sup\u003e After the identification of the R-peak within the processed signal, the data underwent segmentation, with each segment containing five consecutive R-peaks, thereby ensuring uniformity across the segments. During the preprocessing phase, segments with fewer than five R-peaks, abnormally extended R-peaks, or delayed waveforms due to suboptimal measurement were excluded from the data analysis. Resampling was performed to standardize the length of the segmented data, thereby maintaining consistency across the dataset. Through this process, a total of 32,066 data segments were generated \u003cstrong\u003e(Figure 3)\u003c/strong\u003e.\u003c/p\u003e\n\u003cp\u003eTo address the data imbalance issue and enhance the generalizability of the developed model while mitigating bias, we employed a random undersampling technique.\u003csup\u003e11,12\u0026nbsp;\u003c/sup\u003eThis technique removed 18,462 pieces of data, ultimately resulting in a balanced dataset of 13,604 pieces, consisting of 6,802 NSR ECGs and 6,802 AF cases. To facilitate the training, validation, and evaluation of the deep learning model, the dataset was partitioned into the training, internal validation, and testing datasets at a ratio of 6:2:2.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData generation.\u0026nbsp;\u003c/strong\u003eBefore generating 12-lead ECGs from single-lead ECGs, the process of PID matching was conducted between the two datasets. Data from a 12-lead ECG during NSR were used, which were obtained 6 months before and after the date of single-lead ECG measurement. Consistency between the two datasets was ensured by subjecting them to the same preprocessing process as used for the single-lead ECG. The model used for generative learning was the bidirectional LSTM model\u003csup\u003e11\u003c/sup\u003e, which allows information to flow bidirectionally in time, enhancing the learning of the entire sequence data.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDeep learning model.\u0026nbsp;\u003c/strong\u003eEfficientNet-B4 represents a deep learning architecture designed to optimize both performance and efficiency through structural scalability.\u003csup\u003e14\u003c/sup\u003e This model adapts the intricate layers of existing deep learning frameworks, offering scalability across various sizes. By simultaneously adjusting the scaling of depth, width, and resolution, superior performance can be achieved while maintaining a lower computational load using EfficientNet-B4 compared with previous models. In this study, EfficientNet-B4 was used to detect potential AF during NSR. A dimensionality augmentation method was employed to effectively process the ECG signal, allowing it to be treated as an image. To compare different models, a comparative experiment was conducted using Resnet-50, Attention Resnet-50, and LSTM. Supplementary 1 presents detailed information on the Resnet-50, Attention Resnet-50, and LSTM models.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eModel optimization and implementation.\u0026nbsp;\u003c/strong\u003eThe deep learning model was trained using the AdamW\u003csup\u003e15,16\u003c/sup\u003e optimizer. Specifically, the learning rate was set to 0.0001, and the weights were adjusted to 0.001 to facilitate the proper management of weight updates throughout the learning iteration. The batch size was set to 128, and most experiments lasted between 200 and 300 epochs. To mitigate the risk of overfitting, cross-entropy loss was employed as the designated loss function. All training procedures were executed using an Nvidia H100 graphics processing unit (GPU) with Python 3.10 and the PyTorch 2.0.1 framework.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eExternal validation dataset.\u0026nbsp;\u003c/strong\u003eThe external validation dataset comprised 72 h single-lead Holter ECGs from 160 distinct patients. From May 2022 to June 2023, a total of 160 ECGs were obtained from 160 patients with palpitations. The same device was used for all data acquisition. The ECGs were sampled at a rate of 256 Hz. The data were segmented based on R-peak intervals to maintain consistency with the training data. The ECGs of 54 patients with missing data or unmatched PID, 54 patients with abnormal heart rhythm, or those judged to be in AF were excluded. After applying the exclusion criteria, the ECG data of 101 patients were used for external validation. This dataset was segmented based on R-R intervals, resulting in a total of 2,354 data segments. Through undersampling, a total of 1,606 data segments were retrained.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eTo evaluate and compare the performance of the AI-enhanced ECG-based models in identifying AF during NSR, a comprehensive statistical analysis was conducted. Four deep learning models (EfficientNet-B4, Restnet-50, Attention Restnet-50, and LSTM) were employed to analyze data refined by specific exclusion criteria. All four AI methods were evaluated based on their precision, recall, F1-score, area under the receiver operating characteristic curve (AUC), and accuracy.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData generation results. Figure 4\u003c/strong\u003e presents the generated 12-lead ECG signals alongside the corresponding real 12-lead ECG signals. The red line represents the original ECG, and the blue line represents the generated ECG.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe quality of the reconstructed 12-lead ECG was assessed through both quantitative and qualitative evaluations. For quantitative evaluation, the coefficient of determination (R² score) was used, where the score ranges from 0 to 1, with values closer to 1 indicating a stronger correlation with the data.\u003c/p\u003e\n\u003cp\u003eThe average R² score for ECG signals generated from single-lead ECG signals was approximately 0.6, which can be interpreted as above average. Qualitative evaluation involved assessing the morphological similarity between the original and reconstructed 12-lead ECGs. Four expert clinicians visually scored each ECG over more than 1 month, evaluating parameters such as the T-wave inversion, S-T segment, QRS voltage, R-R interval, and axis for each reconstructed ECG.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePerformance comparison of different deep learning models.\u0026nbsp;\u003c/strong\u003eThe performance of the four deep learning models in utilizing the generated data to identify potential AF was compared, and the results are presented in \u003cstrong\u003eTable 2\u003c/strong\u003e. For the training dataset, the EfficientNet-B4 model demonstrated precision, recall, F1-score, AUC, and accuracy of 97.5%, 97.3%, 97.4%, 99.7%, and 97.4%, respectively, in AF identification. The Resnet-50 model achieved a precision of 98.2%, a recall of 98.1%, an F1-score of 98.2%, an AUC of 99.8%, and an accuracy of 98.2%. On the other hand, the Attention Resnet-50 model demonstrated a precision of 78.3%, a recall of 80.8%, an F1-score of 79.5%, an AUC of 87.8%, and an accuracy of 73.9%. The LSTM model showed a precision of 64.5%, a recall of 60.3%, an F1-score of 62.3%, an AUC of 68.6%, and an accuracy of 63.7%. Overall, the integrated ECG-based models showed improved performance. Specifically, the performance of the Resnet-50 model was better than that of the single-lead ECG-based model for the training dataset, improving the F1-score from 79.3% to 98.2%. The accuracy was increased from 80.6% to 99.8%.\u003c/p\u003e\n\u003cp\u003eThe EfficientNet-B4 model demonstrated precision, recall, F1-score, AUC, and accuracy of 72.7%, 73.5%, 73.1%, 80.1%, and 72.4%, respectively, in AF identification in the testing dataset. The Resnet-50 model achieved a precision of 73.1%, a recall of 74.24%, an F1-score of 73.7%, an AUC of 79.4%, and an accuracy of 72.9%, whereas the Attention Resnet-50 model demonstrated a precision of 65.4%, a recall of 68.7%, an F1-score of 67%, an AUC of 71.4%, and an accuracy of 65.5%. Similarly, the LSTM model showed a precision of 63.1%, a recall of 60.7%, an F1-score of 61.9%, an AUC of 66.1%, and an accuracy of 61.9%. Overall, improved results were obtained using integrated ECG compared with previous results obtained using only single-lead ECG \u003csup\u003e5\u003c/sup\u003e for both the testing dataset and the training dataset.\u003c/p\u003e\n\u003cp\u003eFor the internal validation dataset, EfficientNet-B4 showed precision, recall, F1-score, AUC, and accuracy of 71.7%, 73.9%, 72.8%, 79.5%, and 72.6%, respectively. The Resnet-50 model demonstrated a precision of 70.2%, a recall of 72.0%, an F1-score of 71.1%, an AUC of 77.1%, and an accuracy of 71%. The Attention Resnet-50 model exhibited a precision of 63.5%, a recall of 67.9%, an F1-score of 65.6%, an AUC of 70.3%, and an accuracy of 64.8%. The LSTM model demonstrated a precision of 60.1%, a recall of 56.5%, an F1-score of 58.3%, an AUC of 63.8%, and an accuracy of 59.9%. In the case of internal validation, compared with an F1-score value of 69.1% in our previous study with Resnet-50, we achieved F1-scores of 72.8% and 71.1% with EfficientNet-B4 and Resnet-50, respectively, in the current study.\u003c/p\u003e\n\u003cp\u003eConfusion matrices were used to visually assess and compare the classification performance of each model. The confusion matrix results for EfficientNet-B4, Resnet-50, Attention Resnet-50, and LSTM using the testing dataset are shown in \u003cstrong\u003eFigure 5\u003c/strong\u003e.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eReceiver operating characteristic (ROC) curves and AUC values for the identification of potential AF in the training dataset, internal validation dataset, and testing dataset are presented in \u003cstrong\u003eFigure 6\u003c/strong\u003e. In the determination of AF during NSR, EfficientNet-B4 demonstrated the highest AUC at 0.801, followed by Resnet-50 at 0.794, Attention Resnet-50 at 0.714, and LSTM at 0.661 for the testing dataset.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFor the training dataset, the AUCs of EfficientNet-B4, Resnet-50, Attention Resnet-50, and LSTM were 0.997, 0.998, 0.878, and 0.686, respectively. For the internal validation dataset, the AUCs of EfficientNet-B4, Resnet-50, Attention Resnet-50, and LSTM were 0.795, 0.771, 0.703, and 0.638, respectively.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePerformance on external validation set.\u0026nbsp;\u003c/strong\u003eFor the external validation dataset, EfficientNet-B4 showed precision, recall, F1-score, AUC, and accuracy of 76.7%, 63.1%, 69.2%, 76.2%, and 71.9%, respectively. The Resnet-50 model demonstrated a precision of 58.2%, a recall of 87.1%, an F1-score of 69.8%, an AUC of 67%, and an accuracy of 62.3%. The Attention Resnet-50 model exhibited a precision of 58.2%, a recall of 43.3%, an F1-score of 49.7%, an AUC of 58.8%, and an accuracy of 56.1%. The LSTM model demonstrated a precision of 51.4%, a recall of 95%, an F1-score of 66.7%, an AUC of 45%, and an accuracy of 52.6%.\u0026nbsp;\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, we developed and evaluated a deep learning model integrating 12-lead ECG signals generated from single-lead ECG recordings for identifying potential AF during NSR. The proposed deep learning model achieved an AUC of 0.99, with an internal validation of 0.79. An AUC value of 0.7 or higher for the testing/validation dataset generally indicates good performance. Our results were also compared with those of a previous study \u003csup\u003e5\u003c/sup\u003e, in which AF was detected using only 12-lead or single-lead ECG data. The comparison revealed that the integrated ECG-based model could produce reliable results. The findings demonstrated that potential AF could be identified using integrated ECG data rather than relying only on 12-lead or single-lead ECG data.\u003c/p\u003e\n\u003cp\u003eUntil recently, only a limited number of studies utilized single-lead ECG data obtained from single-lead mobile ECG devices for the classification of possible AF during NSR. Most studies have predominantly concentrated on utilizing 12-lead ECG data to identify AF.\u003csup\u003e1,2,17,18\u003c/sup\u003e The benefits of a single-lead mobile ECG device include convenient body attachment and high versatility, allowing extended ECG monitoring and real-time data collection.\u003csup\u003e3\u003c/sup\u003e However, considering the paroxysmal nature of AF, long-term ECG monitoring is frequently necessary for its identification.\u003csup\u003e19\u003c/sup\u003e The results of this study demonstrated the reliability of single-lead ECG data collected with a single-lead mobile ECG device, suggesting its potential as a convenient and highly effective tool for ECG data collection. In addition, this study focused on data augmentation and 12-lead signals generated from single-lead ECG signals, which is consistent with the increasing trend of research on generative models in recent studies.\u003csup\u003e20,21,22\u003c/sup\u003e A platform for predicting ventricular arrhythmias has recently been introduced.\u003csup\u003e23\u003c/sup\u003e If our proposed AF prediction model could be integrated into this platform, it is expected that a wide range of arrhythmias could be detected.\u003c/p\u003e\n\u003cp\u003eOur deep learning model integrating 12-lead ECGs generated from single-lead ECGs was uniquely designed to predict AF occurrence even during NSR. The proposed model could be better for the prediction of arrhythmias, which improves on the disadvantages of using only 12-lead or single-lead ECG signals. Although the generated ECG signals generally reflected waveform changes in the original ECG signals, in the case of lead III, the correlation was low. The generated signals did not accurately capture subtle waveform changes in areas with significant fluctuations.\u003c/p\u003e\n\u003cp\u003eThe data collection process involved 6,806 patients, which led to the acquisition and analysis of 7,199 mobile ECGs. An issue frequently encountered with various medical datasets is the problem of data imbalance, characterized by the underrepresentation of specific health conditions in classification algorithms. This issue was apparent in our study, especially due to the limited availability of ECGs labeled as AF. The data were primarily categorized as “Healthy,” potentially impeding the identification of AF. To address the data imbalance issue and improve the generalization ability of our model while reducing bias, we utilized a random under-sampling technique. Moreover, ECG signals that were excessively noisy or contained a high number of artifacts were not included in the training dataset.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFor the external validation dataset, the same methodology was applied as used for our training, testing, and internal validation datasets. The 72 h Holter ECGs from 160 distinct palpitation patients were divided into R-R peak segments, and random sampling without replacement was conducted to obtain a representative subset for validation. This method of random sampling without replacement ensures a more unbiased analysis by avoiding potential bias due to the researchers’ preferences or assumptions.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eMoreover, four deep learning models (EfficientNet-B4, Resnet-50, Attention Resnet-50, and LSTM) were utilized to identify potential AF by integrating 12-lead ECGs generated from single-lead ECGs during NSR, and their performance was compared. As shown in Table 2, our proposed EfficientNet-B4 model was excellent with precision, recall, F1-score, AUC, and accuracy of 97.5%, 97.3%, 97.4%, 99.7%, and 97.4%, respectively. The Resnet-50 model with the training dataset also demonstrated superior performance, with precision, recall, F1-score, AUC, and accuracy of 98.2%, 98.1%, 98.2%, 99.8%, and 98.2%, respectively. In particular, the Resnet-50 model achieved better results compared with those obtained with our previous single-lead ECG-based model \u003csup\u003e3\u003c/sup\u003e for the training dataset, improving the F1-score from 79.3% to 98.2%. The results suggest that the overall performance of the integrated ECG-based models was better than that of the single-lead ECG-based model.\u003c/p\u003e\n\u003cp\u003eThere are some limitations in this study. First, when converting single-lead ECG data to 12-lead ECG data, it would be ideal to obtain simultaneously measured data; however, this was not realistically feasible. Therefore, to ensure the accuracy of the 12-lead ECG data, 12-lead ECG data with NSR status within the past 6 months were used as the reference point. To overcome this limitation, we plan to obtain single-lead ECG data and 12-lead ECG data simultaneously. During the 72 h measurement period, there were no waveforms or significant outliers observed, which could be attributed to incorrect patient movement or improper machine attachment during external verification. In addition, due to the large amount of data, it was not possible to segment all 72 h of data based on R-R intervals. Even for data classified as NSR, there were many outliers; thus, relevant parts had to be excluded. If more data could be obtained in the future, we plan to conduct the experiment again, excluding all the above cases.\u003c/p\u003e\n\u003cp\u003eFurther research is necessary to improve the performance and generalizability of AI models in clinical settings. Although our proposed deep learning model, which utilizes integrated ECG signals, demonstrated promising outcomes in the identification of AF with an AUC of 80.1% for the testing dataset, its efficacy was comparatively lower for the external validation dataset. This could have implications with the small amount of external testing data available. Consequently, we intend to acquire additional data and carry out further testing. Moreover, previous studies have indicated that the R-R interval, P-wave morphology, and QT interval are significant factors affecting the risk of AF.\u003csup\u003e24,25\u0026nbsp;\u003c/sup\u003eThe study utilized ECGs to train the deep learning model. However, future studies could enhance model accuracy by incorporating additional relevant features as indicators of AF.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn the case of the integrated ECG-based model, deep learning was applied to predict potential AF, showing high accuracy and F1-score. In comparison with existing results obtained for single-lead ECG signals with the Resenet-50 model, the accuracy was improved by around 14%. In addition, the results of the EfficientNet-B4 model proposed in this study were better than those of the Restnet-50 model. However, the LSTM model showed somewhat worse results. Nevertheless, the integrated ECG-based model has the advantage of analyzing both standard 12-lead ECG and single-lead ECG signals rather than analyzing only single-lead ECG signals. The use of the proposed model in clinical applications may reduce time and cost. Finally, a GUI format focusing on user convenience can be found in Supplementary Figure 1, which may be used to apply the proposed integrated ECG-based model to actual clinical settings.\u003c/p\u003e\n\u003cp\u003e\u003cbr\u003e\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eACKNOWLEDGMENTS\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eSpecial thanks to\u0026nbsp;Dr. Seong Won Jeon, Dr. Changhyun Kim, and Dr. Dong Kyun Kim for the diagnosis and labeling mobile electrocardiograms.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDisclosure\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors\u0026nbsp;have no potential conflicts of interest to disclose.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor Contributions \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eLee M and Choi H performed data curation, formal analysis, validation and visualization; Lee M and Lee KH conducted the investigation; Lee M, Choi H, Lee YH and Lee KH designed the methodology; Choi H developed the software; Lee KH acquired funding; Lee M wrote the original draft; Lee M and Lee KH reviewed and edited the manuscript. All authors reviewed the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe mobile and 12-lead ECG electrocardiogram dataset generated and analyzed during the current study is not publicly available due to privacy and ethical restrictions stipulated by the Institutional Review Board of Chonnam National University Hospital. However, it is available from the corresponding author upon reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding:\u0026nbsp;\u003c/strong\u003eThis work was supported by a grant (BCRI25041) of Chonnam National University Hospital Biomedical Research Institute, in part a grant of Establishment of K-Health National Medical Care Service and Industrial Ecosystem funded by the Ministry of Science and ICT (MSIT, Korea) Balanced National Development Account. [Project Name: Establishment of K-Health National Medical Care Service and Industrial Ecosystem/Project Number: ITAH0603230110010001000100100], and in part by Institute of Information \u0026amp; communications Technology Planning \u0026amp; Evaluation (IITP) under the Artificial Intelligence Convergence Innovation Human Resources Development (IITP-2023-RS-2023-00256629) grant funded by the Korea government (MSIT).\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAttia, Z. I. et al. An artificial intelligence-enabled ECG algorithm for the identification of patients with atrial fibrillation during sinus rhythm: a retrospective analysis of outcome prediction. \u003cem\u003eLancet\u003c/em\u003e \u003cstrong\u003e394\u003c/strong\u003e, 861\u0026ndash;867. DOI:10.1016/S0140-6736(19)31721-0 (2019).\u003c/li\u003e\n\u003cli\u003eBaek, Y. et al. A new deep learning algorithm of 12-lead electrocardiogram for identifying atrial fibrillation during sinus rhythm. \u003cem\u003eSci. Rep\u003c/em\u003e. \u003cstrong\u003e11\u003c/strong\u003e, 12818. https://doi.org/10.1038/s41598-021-92172-5 (2021).\u003c/li\u003e\n\u003cli\u003eSana, F. et al. Wearable devices for ambulatory cardiac monitoring: JACC state-of-the-art review. \u003cem\u003eJ. Am. Coll. Cardiol\u003c/em\u003e \u003cstrong\u003e75\u003c/strong\u003e, 1582\u0026ndash;1592. DOI:10.1016/j.jacc.2020.01.046 (2020).\u003c/li\u003e\n\u003cli\u003ePerez, M. V. et al. Large-scale assessment of a smartwatch to identify atrial fibrillation. \u003cem\u003eN. Engl. J. Med.\u003c/em\u003e \u003cstrong\u003e381\u003c/strong\u003e, 1909\u0026ndash;1917. DOI: 10.1056/NEJMoa1901183 (2019).\u003c/li\u003e\n\u003cli\u003eKim, J. et al. Identification of atrial fibrillation with single-lead mobile ECG during normal sinus rhythm using deep learning. \u003cem\u003eJ. Korean Med. Sci.\u003c/em\u003e \u003cstrong\u003e39\u003c/strong\u003e, https://doi.org/10.3346/jkms.2024.39.e56 (2024)\u003c/li\u003e\n\u003cli\u003eYıldırım, \u0026Ouml;., Pławiak, P., Tan, R.-S. \u0026amp; Acharya, U. R. Arrhythmia detection using deep convolutional neural network with long duration ECG signals. \u003cem\u003eComput. Biol. Med.\u003c/em\u003e \u003cstrong\u003e102\u003c/strong\u003e, 411\u0026ndash;420. https://doi.org/10.1016/j.compbiomed.2018.09.009 (2018).\u003c/li\u003e\n\u003cli\u003eSherin, M., Chandra, K. \u0026amp; Kenneth, B. A novel application of deep learning for single-lead ECG classification. \u003cem\u003eComput. Biol. Med\u003c/em\u003e. \u003cstrong\u003e99\u003c/strong\u003e, 53\u0026ndash;62. DOI:10.1016/j.compbiomed.2018.05.013 (2018).\u003c/li\u003e\n\u003cli\u003ePetmezas, G., Leandros, S., Vassilis, K. et al. State-of-the-art deep learning methods on electrocardiogram data: systematic review. \u003cem\u003eJMIR Med. Inform.\u003c/em\u003e \u003cstrong\u003e10\u003c/strong\u003e, e38454. doi:10.2196/38454 (2022).\u003c/li\u003e\n\u003cli\u003eBaek, Y. S., Kwon, S., Seng, C. Y. et al. Artificial intelligence-enhanced 12-lead electrocardiography for identifying atrial fibrillation during sinus rhythm (AIAFib) trial: protocol for a multicenter retrospective study. \u003cem\u003eFront. Cardiovasc. Med.\u003c/em\u003e \u003cstrong\u003e10\u003c/strong\u003e, 1258167. doi:10.3389/fcvm.2023.1258167 (2023).\u003c/li\u003e\n\u003cli\u003eMakowski, D., Pham, T., Lau, Z. J. et al. NeuroKit2: a Python toolbox for neurophysiological signal processing. \u003cem\u003eBehav. Res. Methods\u003c/em\u003e \u003cstrong\u003e53\u003c/strong\u003e, 1689\u0026ndash;1696. https://doi.org/10.3758/s13428-020-01516-y (2021).\u003c/li\u003e\n\u003cli\u003eLee, W. \u0026amp; Seo, K. Downsampling for binary classification with a highly imbalanced dataset using active learning. \u003cem\u003eBig Data Res\u003c/em\u003e. \u003cstrong\u003e28\u003c/strong\u003e, 100314. doi:10.1016/j.bdr.2022.100314 (2022).\u003c/li\u003e\n\u003cli\u003eQazi, N. \u0026amp; Raza, K. Effect of feature selection, SMOTE and under sampling on class imbalance classification. In 2012 UKSim 14th International Conference on Computer Modelling and Simulation 145\u0026ndash;150. DOI\u003cstrong\u003e:\u003c/strong\u003e10.1109/UKSim.2012.116 (2012).\u003c/li\u003e\n\u003cli\u003eHochreiter, S. \u0026amp; Schmidhuber, J. Long short-term memory. \u003cem\u003eNeural Comput\u003c/em\u003e. \u003cstrong\u003e9\u003c/strong\u003e, 1735\u0026ndash;1780. DOI\u003cstrong\u003e:\u003c/strong\u003e10.1162/neco.1997.9.8.1735 (1997).\u003c/li\u003e\n\u003cli\u003eTan, M. \u0026amp; Le, Q. EfficientNet: rethinking model scaling for convolutional neural networks. In Proceedings of the 36th International Conference on Machine Learning (eds. Chaudhuri, K. \u0026amp; Salakhutdinov, R.) \u003cem\u003eProc. Mach. Learn. Res\u003c/em\u003e. \u003cstrong\u003e97\u003c/strong\u003e, 6105\u0026ndash;6114 (2019).\u003c/li\u003e\n\u003cli\u003eReyad, M., Sarhan, A. M. \u0026amp; Arafa, M. A modified Adam algorithm for deep neural network optimization. \u003cem\u003eNeural Comput. Appl\u003c/em\u003e. \u003cstrong\u003e35\u003c/strong\u003e, 17095\u0026ndash;17112. https://doi.org/10.1007/s00521-023-08568-z (2023).\u003c/li\u003e\n\u003cli\u003eJia, X., Feng, X., Yong, H. \u0026amp; Meng, D. Weight decay with tailored Adam on scale-invariant weights for better generalization. \u003cem\u003eIEEE Trans. Neural Netw. Learn. Syst.\u003c/em\u003e \u003cstrong\u003e35\u003c/strong\u003e. DOI:10.1109/TNNLS.2022.3213536 (2022).\u003c/li\u003e\n\u003cli\u003eSomani, S. et al. Deep learning and the electrocardiogram: review of the current state-of-the-art.\u003cem\u003e Europace\u003c/em\u003e \u003cstrong\u003e23\u003c/strong\u003e, 1179\u0026ndash;1191. DOI:10.1093/europace/euaa377 (2021).\u003c/li\u003e\n\u003cli\u003eRibeiro, A. H. et al. Automatic diagnosis of the 12-lead ECG using a deep neural network. \u003cem\u003eNat. Commun.\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e, 1760. https://doi.org/10.1038/s41467-020-15432-4 (2020).\u003c/li\u003e\n\u003cli\u003eNigusse, A. B., Mengistie, D. A., Malengier, B., Tseghai, G. B. \u0026amp; Van Langenhove, L. Wearable smart textiles for long-term electrocardiography monitoring\u0026mdash;a review. \u003cem\u003eSensors\u003c/em\u003e \u003cstrong\u003e21\u003c/strong\u003e, 4174. https://doi.org/10.3390/s21124174 (2021).\u003c/li\u003e\n\u003cli\u003eSeo, H. C., Yoon, G. W., Joo, S. \u0026amp; Na, G. B. Multiple electrocardiogram generator with single-lead electrocardiogram. \u003cem\u003eComput. Methods Programs Biomed.\u003c/em\u003e \u003cstrong\u003e221\u003c/strong\u003e, 106858. https://doi.org/10.1016/j.cmpb.2022.106858 (2022).\u003c/li\u003e\n\u003cli\u003eShin, H., Sun, S., Lee, J. \u0026amp; Kim, H. Complementary photoplethysmogram synthesis from electrocardiogram using generative adversarial network. \u003cem\u003eIEEE Access\u003c/em\u003e \u003cstrong\u003e9\u003c/strong\u003e, 70639\u0026ndash;70649. DOI:10.1109/ACCESS.2021.3078534 (2021).\u003c/li\u003e\n\u003cli\u003eYoon, G. W. \u0026amp; Joo, S. Classification feasibility test on multi-lead electrocardiography signals generated from single-lead electrocardiography signals. \u003cem\u003eSci\u003c/em\u003e. \u003cem\u003eRep\u003c/em\u003e. \u003cstrong\u003e14\u003c/strong\u003e, 1888. https://doi.org/10.1038/s41598-024-52216-y (2024).\u003c/li\u003e\n\u003cli\u003eCompagnucci, P. et al. Implantable defibrillator-detected heart failure status predicts ventricular tachyarrhythmias. \u003cem\u003eJ. Cardiovasc. Electrophysiol.\u003c/em\u003e \u003cstrong\u003e34\u003c/strong\u003e, 1257\u0026ndash;1267. DOI: 10.1111/jce.15898 (2023).\u003c/li\u003e\n\u003cli\u003eCoppola, E. E., Gyawali, P. K., Vanjara, N., Giaime, D. \u0026amp; Wang, L. Atrial fibrillation classification from a short single lead ECG recording using hierarchical classifier. \u003cem\u003eIEEE\u003c/em\u003e\u003cem\u003e Computing in Cardiology (CinC)\u003c/em\u003e 354\u0026ndash;425. DOI: 10.22489/CinC.2017.354-425 (2017).\u003c/li\u003e\n\u003cli\u003eYazdani, S., Laub, P., Luca, A. \u0026amp; Vesin, J. M. Heart rhythm classification using short-term ECG atrial and ventricular activity analysis. \u003cem\u003eIEEE\u003c/em\u003e\u003cem\u003e Computing in Cardiology\u003c/em\u003e (CinC) 67\u0026ndash;120. DOI\u003cstrong\u003e: \u003c/strong\u003e10.22489/CinC.2017.067-120 (2017).\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"artificial intelligence, atrial fibrillation, electrocardiography, mobile applications, single-lead, 12-lead","lastPublishedDoi":"10.21203/rs.3.rs-7462212/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7462212/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"The use of artificial intelligence (AI) with electrocardiogram (ECG) data has shown promise in detecting atrial fibrillation (AF). Although single-lead ECGs allow convenient and simple rhythm monitoring, arrhythmias prediction using AI is limited due to single-channel utilization. We aimed to improve the capability of AI algorithms for AF identification in integrated models with 12-lead and single-lead ECG during normal sinus rhythm (NSR). A total of 7,199 single-lead mobile ECGs were acquired from 6,806 patients. Four deep learning models, i.e., EfficientNet-B4, residual neural networks (Restnet-50), Attention Restnet-50, and long short-term memory (LSTM), were employed to analyze the dataset. To develop an integrated model, an LSTM-based generative adversarial network was used to generate 12-lead ECGs from single-lead ECGs. The generated ECGs were then applied to the identification algorithm to predict AF. The integrated ECG-based model achieved an accuracy of 0.974, precision of 0.975, recall of 0.973, and F1-score of 0.974 for the training dataset with EfficientNet-B4. The area under the receiver operating characteristic curve (AUC) value for identifying AF was 0.98 with the integrated model, 0.91 with a 12-lead ECG, and 0.88 with a single-lead ECG. The integrated ECG-based model has the advantage of analyzing both standard 12-lead ECG and single-lead ECG signals. The findings underscore the potential of the integrated model in identifying AF using NSR ECGs without the limitations of relying solely on 12-lead or single-lead data. A GUI format focusing on user convenience may be used to apply the integrated ECG-based model to clinical settings.","manuscriptTitle":"Identifying Atrial Fibrillation using Integrated Methods of 12-lead and Single-lead ECG during Normal Sinus Rhythm based on Artificial Intelligence","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-10-08 16:02:04","doi":"10.21203/rs.3.rs-7462212/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"a3867444-0663-4930-b8fd-aecaf7ff43d9","owner":[],"postedDate":"October 8th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":55798959,"name":"Health sciences/Cardiology"},{"id":55798960,"name":"Biological sciences/Computational biology and bioinformatics"}],"tags":[],"updatedAt":"2025-12-17T09:53:11+00:00","versionOfRecord":[],"versionCreatedAt":"2025-10-08 16:02:04","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7462212","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7462212","identity":"rs-7462212","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}