Skip to main content
Download PDF

Detection of atrial fibrillation using a nonlinear Lorenz Scattergram and deep learning in primary care

BMC Primary Care volume 25, Article number: 267 (2024) Cite this article

Abstract

Background

Atrial fibrillation (AF) is highly correlated with heart failure, stroke and death. Screening increases AF detection and facilitates the early adoption of comprehensive intervention. Long-term wearable devices have become increasingly popular for AF screening in primary care. However, interpreting data obtained by long-term wearable ECG devices is a problem in primary care. To diagnose the disease quickly and accurately, we aimed to build AF episode detection model based on a nonlinear Lorenz scattergram (LS) and deep learning.

Methods

The MIT-BIH Normal Sinus Rhythm Database, MIT-BIH Arrhythmia Database and the Long-Term AF Database were extracted to construct the MIT-BIH Ambulatory Electrocardiograph (MIT-BIH AE) dataset. We converted the long-term ECG into a two-dimensional LSs. The LSs from MIT-BIH AE dataset was randomly divided into training and internal validation sets in a 9:1 ratio, which was used to develop and internally validated model. We built a MOBILE-SCREEN-AF (MS-AF) dataset from a single-lead wearable ECG device in primary care for external validation. Performance was quantified using a confusion matrix and standard classification metrics.

Results

During the evaluation of model performance based on the LS, the sensitivity, specificity and accuracy of the model in diagnosing AF were 0.992, 0.973, and 0.983 in the internal validation set respectively. In the external validation set, these metrics were 0.989, 0.956, and 0.967, respectively. Furthermore, when evaluating the model’s performance based on ECG records in the MS-AF dataset, the sensitivity, specificity and accuracy of model diagnosis paroxysmal AF were 1.000, 0.870 and 0.876 respectively, and 0.927, 1.000 and 0.973 for the persistent AF.

Conclusions

The model based on the nonlinear LS and deep learning has high accuracy, making it promising for AF screening in primary care. It has potential for generalization and practical application.

Peer Review reports

Background

Atrial fibrillation (AF) is one of the world’s most challenging cardiovascular diseases of the 21st century and is strongly associated with stroke, heart failure and death [1, 2]. However, approximately 30% of AF patients can be asymptomatic or subclinical [3]. Although the relationship between AF screening and clinical outcomes is controversial [4]. Various devices can improve the detection of atrial fibrillation and facilitate the early adoption of comprehensive intervention, to reduce complications [5]. In recent years, wearable technologies for AF screening have been rapidly developing, with advantages such as home use, portability and long-term monitoring [1]. There are two types of wearable ECG devices, one is the device that monitors the electrical activity of the heart, and the other is the device based on photoplethysmography (PPG) [6]. However, the diagnosis of arrhythmias using PPG devices needs to be confirmed by ECG, which limits their practical application in clinical practice. Therefore, the devices that obtain ECG signals has become the focus of this study. A randomized clinical trial showed that long-term wearable ECG devices increased AF detection in the community by 4.1% [7].

Interpreting long-term ECG data obtained by wearable devices is time-consuming, which restrict the rapid diagnosis of AF by general practitioners in primary care [8, 9]. In recent years, there has been a growing number of machine learning models that have been developed to enable rapid diagnosis of long-term ECG [10]. These models are constructed from the ECG characteristics of atrial and ventricular electrical activity [11,12,13]. Santala et al. showed that an automated artificial intelligence arrhythmia detection achieved a sensitivity of 100% and a specificity of 95.4% for detection of AF from 24-hour single-lead heart belt recordings in a clinical setting [14]. However, the study did not provide specific information about the intelligent model used, making it difficult to replicate and validate. Hartikainen et al. reported that COSEn and AFEvidence algorithms was achieved with sensitivity of 95.3%/96.3% and specificity of 95.5/98.2% for detection of AF from chest strap [15]. However, These models use traditional shallow machine learning algorithms, which require manual feature extraction and are easily overfit [16]. Deep learning, including convolutional neural networks (CNNs) and recurrent neural networks (RNNs), is an emerging branch of machine learning. It uses an end-to-end learning mechanism to achieve direct output prediction without human intervention [17]. Several studies have found that deep learning can generally perform better than traditional ECG modeling algorithms [18,19,20]. A study published in The Lancet found that CNNs can discover ECG features that are difficult to discern by the human eye and demonstrated the advantages of CNNs in ECG image analysis [21].

However, the heart is essentially a nonlinear chaotic system, and the current research did not consider the nonlinear characteristics. The Lorenz scattergram (LS) is an ECG analysis technique based on chaos theory, which converts traditional linear ECG into a nonlinear two-dimensional image. In LS, homogenous heartbeats gather together to form an “attractor“ [22]. The attractor shows a characteristic pattern and is used for arrhythmia diagnosis [22]. In recent years, various diagnostic models based on an LS have emerged that are only constructed by shallow machine learning algorithms [13, 23, 24]. Kisohara et al. [25] used CNNs and a LS to establish an AF diagnostic model and found that the optimal segment window length for detecting AF with 32 × 32-pixel LS is 85 beats. However, this study only included sinus rhythm and persistent AF in the model training and did not include paroxysmal AF or other arrhythmias. The generalization ability of this model in the clinic weakened. Therefore, it is necessary to establish a deep learning-based model that can diagnose persistent and paroxysmal AF on long-term ECG and validate it in primary care.

Methods

ECG dataset

This is a cross-sectional study. We extracted the MIT-BIH Normal Sinus Rhythm Database [26], the MIT-BIH Arrhythmia Database [27] and the Long-Term AF Database [28] from PhysioNet and constructed the MIT-BIH Ambulatory Electrocardiograph (MIT-BIH AE) dataset for training and internal validating the diagnosis model. The AF records were labeled as AF. Sinus rhythm and other arrhythmias were labeled as non-AF.

We built a MOBILE-SCREEN-AF (MS-AF) dataset, which came from a single-lead wearable ECG device in a community hospital in Chengdu, Sichuan and was used for external validation. This hospital is located in the suburbs of Chengdu. It has jurisdiction over 16 communities and serves a population of 138,700. We prospectively enrolled participants. The inclusion criteria for this study were as follows: (i) patients with one or more of the following 7 high-risk factors for atrial fibrillation: age ≥ 65 years, hypertension, diabetes, obesity (Chinese standard, BMI ≥ 28 kg/m2 [29]), smoking, drinking, and myocardial infarction; and (ii) Voluntary participation and provision of health examination information. The exclusion criteria were as follows: (i) patients with a pacemaker or ICD; (ii) patients with a history of severe skin allergy (Severe skin allergy is defined as the presence of one or more symptoms, including localized or systemic skin breakdown, loss of consciousness, decreased blood pressure, shortness of breath, fast and weak pulse, nausea or vomiting.); and (iii) patients who were expected to wear the device for less than 24 h. For participants who provided consent, we equipped them with wearable ECG devices. The device is a patch that is used with wet electrodes. The device, provided by Chengdu Synwing Technology Co., Ltd. (Medical device registration number: 20,212,070,096), is capable of continuous monitoring for up to 72 h. The device is showed in Appendix Fig. 1. The positions and diagnostic labels of R waves were first performed by a computer software. ECG technician cleaned up single-lead ECG data. The cleanup process consists of three parts. Firstly, he assessed whether the actual monitoring time is equal to or greater than 24 h. Secondly, he evaluated the quality of the ECG waveforms, which included two steps. The first step was noise reduction processing of the ECG, based on the high-pass filtering, low-pass filtering and band-stop filtering techniques. The second step was to assess whether there are still unidentifiable ECG segments in the denoised ECG. An unidentifiable ECG segment is defined as a segment in which the precise location of the R wave cannot be accurately determined. If the proportion of unidentifiable ECG segments exceeds 10% of the actual monitoring time, the record was marked as low quality and was excluded. Third, he evaluated whether the computer software correctly labeled the R-wave location and diagnostic label (AF or non-AF). And corrected any erroneous labels. Subsequently, a cardiologist and electrocardiologist from a tertiary hospital independently reviewed the diagnostic marks of R waves, and inconsistencies were determined after discussion. The AF records were labeled as AF. Sinus rhythm and other arrhythmias were labeled as non-AF. This study was approved by the Ethics Committee of West China Hospital of Sichuan University (2018 − 454).

Lorenz Scattergram

The LS was constructed from ECG segments by determining a point (RRn, RRn + 1) in the two-dimensional coordinate system (RRn, RRn + 1 is the adjacent RR interval on the ECG). The RR interval was iterated successively to form an LS [22]. The LS preparation process was as follows: (1) The ECG recording R-wave position and diagnostic markers were converted into a text file with the suffix.rri3. (2) The text files containing R-wave positions and diagnostic markers were converted into TFRecord data streams. TFRecord is a binary file provided by TensorFlow to store data, which is convenient for copying, moving or cutting data without requiring another marker file. (3) The TFRecord data stream was continuously cut into units of 85 R waves to obtain data segments. (4) The TFRecord data segments were converted into an LS with a resolution of 32 × 32. (5) The LS was considered to be AF if the R waves of 30 consecutive seconds or more were marked as AF in the 85 R waves; otherwise, it was non-AF. Figure 1 shows an LS of AF and non AF (sinus rhythm) with 85 R waves.

We converted the ECG into a two-dimensional LS according to the aforementioned procedure. Data augmentation was used to expand the sample size for the MIT-BIH AE dataset. We used 85 R waves as the window width and 9 R waves as the step length and continuously cut the TFRecord data stream. LSs from the MIT-BIH AE dataset were randomly divided into training and validation sets in a 9:1 ratio. Data augmentation was not used in the MS-AF dataset, and LSs from the MS-AF dataset constituted the external validation set.

Fig. 1
figure 1

Lorenz scattergram with 85 R waves. A LS of AF. B LS of non AF (sinus rhythm)

Full size image

Deep learning model

The CNN in this study comprised an input layer, convolutional layer, pooling layer, fully connected layer and output layer. The detailed process of model development was shown in the supplementary information. And the structure of the CNN in this study was described in Appendix Table 1. During the training and validation phases, we noticed no significant improvement in accuracy and loss value after 20 epochs. We used 20 epochs to train the model. The accuracy and loss function value of the diagnostic model of each epoch in CNN training and validation were shown in Appendix Fig. 2 and Appendix Fig. 3. We further visually analyzed the convolution kernel of the three convolution layers. The image features of the convolution kernel of the convolutional layers of the model were shown in Appendix Fig. 4, Appendix Fig. 5 and Appendix Fig. 6.

Evaluation metrics

We selected sensitivity (Sen), specificity (Spe), accuracy (Acc), positive predictive value (PPV), negative predictive value (NPV), positive likelihood ratio (+ LR), negative likelihood ratio (-LR), receiver operating characteristic curve (ROC) and area under the curve (AUC) to evaluate model diagnostic performance [30, 31].

The performance evaluation of the model was divided into two levels. In the first level, based on the LS with 85 R waves, the predicted classification of the model was compared with the real classification of the LS to evaluate the model performance. The second level was based on long-range ECG records. Each ECG record generated several LSs with 85 R waves through sequential data cutting. The model predicted the diagnosis of ECG records according to the LS prediction results. The prediction results of the ECG record of the model were compared with the real diagnosis results of the ECG records to evaluate the model performance.

Statistical analysis

Data were analyzed by SPSS 26.0. If the measurement data met the normal distribution, they were described by the mean±standard deviation. We used the median and quartile if the normal distribution was not satisfied. Enumeration data were described by frequency and percentage. The computer configuration was as follows: MacBook Air (2017), 1.8 GHz dual-core Intel Core i5 processor, 8 GB 1600 MHz DDR3 memory and an Intel HD Graphics 6000 1536 MB graphics card. GPU: NVIDIA TESLA P100. We used Python (version 3.8.8) for programming. The Keras software package was used for constructing and optimizing the CNN, sklearn.metrics was used to create confusion matrices, the roc_curve function was used to create ROC and calculate the AUC. The optimal threshold of the model was determined by the maximum Youdan index. The np.argmax function was used to find the maximum Youdan index and to determine the ROC cutoff value.

Results

Baseline characteristics of datasets

The MIT-BIH Normal Sinus Rhythm Database includes 18 ECG records. The MIT-BIH Arrhythmia Database includes 48 ECG records, and the Long-Term AF Database includes 84 ECG records. In the MS-AF dataset, a total of 138 patients used single-lead wearable ECG devices between September 8 and December 31, 2021. Twenty-five patients were excluded, 8 patients did not have complete health information, 3 patients had an actual measurement time of less than 24 h, and 14 patients (10.77%) had poor-quality ECG data. Finally, 113 patients were included in the analysis. The characteristics of the ECG datasets are summarized in Table 1.

Table 1 Baseline characteristics of the ECG datasets
Full size table

Characteristics of the lorenz scattergram

A total of 1,345,849 LSs were obtained from the MIT-BIH AE dataset. The training set comprised 1,211,264 LSs, and the internal validation set comprised 134,585 LSs. In the training set, there were 575,820 cases (47.54%) of AF and 635,444 cases (52.46%) of non-AF. In the internal validation set, 66,949 cases (49.74%) of LSs were AF, and 67,636 cases (50.26%) were non-AF. A total of 207,540 LSs were transformed from the MS-AF dataset to form the external validation set. There were 67,263 cases (32.41%) of AF and 140,277 cases (67.59%) of non-AF.

The cutoff value of the model

The first cutoff value of the model was to distinguish whether the LS was AF or not. The ROC curves of the model for diagnosing AF in all LSs from the training and internal validation sets were shown in Fig. 2(A). The cutoff value of the model for diagnosing whether LS was AF or non-AF was 0.610. A probability value ≥ 0.610 predicted AF, and < 0.610 predicted non-AF. The second cutoff value was to distinguish whether a single-lead ECG record was AF or not. The ROC curves of the model for diagnosing AF in the MIT-BIH AE dataset were shown in Fig. 2(B). The cutoff value of the model for diagnosing whether the ECG record was AF was 0.007. A probability value ≥ 0.007 predicted AF, and < 0.007 predicted non-AF. The third cutoff value was to distinguish whether a single-lead ECG recording is persistent AF or paroxysmal AF. The ROC of the model for diagnosing paroxysmal AF in the MIT-BIH AE dataset of AF records is shown in Fig. 2(C). The cutoff value of the model for diagnosing whether the ECG record was paroxysmal AF or persistent AF was 0.948, 0.007 ≤ probability value < 0.948 predicts paroxysmal AF, and ≥ 0.948 predicts persistent AF.

Fig. 2
figure 2

Receiver operating characteristic curves. A ROC of the diagnostic model in all LSs from the training and internal validation set. B ROC of the diagnostic model in the MIT-BIH AE dataset of ECG records. C ROC curves of the diagnostic model in the MIT-BIH AE dataset of AF records

Full size image

Performance evaluation in the lorenz scattergram

The confusion matrix for diagnostic model in the internal and external validation sets was shown in Table 2. In the internal validation set, the sensitivity of the model for diagnosing AF was 0.992, the specificity was 0.973, and the accuracy was 0.983. In the external validation set, the sensitivity of the model for diagnosing AF was 0.989, the specificity was 0.956, and the accuracy was 0.967. Table 3 showed the performance of the model in the internal and external validation sets. The ROC curves of the model for diagnosing AF in the internal and external validation sets were shown in Fig. 3.

Table 2 The confusion matrix for the model in the internal and external validation sets
Full size table
Table 3 Performance of the model in the internal and external validation sets
Full size table
Fig. 3
figure 3

Receiver operating characteristic curves. A ROC of the model for diagnosing AF in the internal validation set. B ROC of the model for diagnosing AF in the external validation set

Full size image

Performance evaluation in ECG records

In the 113 ECG records from the MS-AF dataset. The sensitivity of the model diagnosis of paroxysmal AF was 1.000, the specificity was 0.870, and the accuracy was 0.876. The sensitivity of the model diagnosis of persistent AF was 0.927, the specificity was 1.000, and the accuracy was 0.973. Table 4 shows the performance of the model in ECG records of the MS-AF dataset.

Table 4 Performance of the model in ECG records of the MS-AF dataset
Full size table

Discussion

In this study, we developed a deep learning model for detecting AF episodes based on nonlinear LS. The proposed model was external validated in a clinical ECG dataset. We found that the sensitivity, specificity and accuracy of the model in diagnosing paroxysmal AF were 100.0%, 87.0% and 87.6%, respectively, in long-term ECG records. The sensitivity, specificity and accuracy were 92.7%, 100.0% and 97.3%, respectively, for diagnosing persistent AF in long-term ECG records.

In recent years, various machine learning algorithms have emerged for diagnosing AF, including shallow machine learning algorithms and deep learning algorithms [32]. Shallow machine learning algorithms need to manually extract features, which are prone to overfitting. Deep learning algorithms automatically extract features based on raw data, which is conducive to recognizing of high-dimensional features [32]. Deep learning algorithms include CNN, long short-term memory(LSTM) and hybrid deep models(CNN-LSTM) [16]. Xia et al. [33] first proposed a CNN model for diagnosing AF, converting the 5-second ECG segment into a two-dimensional matrix by short-term Fourier transform and stationary wavelet transform and inputting a two-dimensional matrix into CNN to build a model. The model’s sensitivity and specificity for diagnosing AF were 98.34/98.79% and 97.87/98.24%, respectively, in the MIT-BIH AF database. However, this model was not validated in a clinic, and nonlinear ECG characteristics were not considered [34, 35]. Mittal et al. used a 2-part(a QRS complex detector and an AF detector) deep neural network filter to identify AF episodes in ECG data from implantable loop recorder [36]. This study demonstrated deep neural network can filter up to 66% of all false-positive AF events, and reduce the burden on clinicians. However, the model was fed with PDF reports (low resolution) instead of ECG signals, which affected the performance of the model. Another study used LSTM [37] to diagnose AF and achieved a sensitivity and specificity of 99.8% and 99.6% in MIT-BIH Atrial Fibrillation Database. Andersen et al. [19] used a CNN-LSTM for AF diagnosis, which achieved a sensitivity and specificity of 98.98% and 96.95%, respectively in Physionet dataset (MIT-BIH AF Database, the MIT-BIH Arrhythmia Database and the MIT-BIH NSR Database). The diagnostic performance of the LSTM and CNN-LSTM models is similar to that of the CNN model, but model training has a heavy calculation burden and high time consumption. In our study, we used a CNN to build the model, which has a simple structure, short training time, simple GPU configuration, good performance in primary care.

The common AF diagnosis model was mainly constructed according to the characteristics of atrial fibrillation f waves and irregular RR intervals [11, 38,39,40]. Nuryani et al. [11] extracted the number of P/f-wave peaks and the P/f-wave width and used the fuzzy inference system to construct a diagnostic model. The sensitivity, specificity, and accuracy of diagnosing AF were 77.89%, 60.40%, and 75.90%, respectively in MIT-BIH arrhythmia dataset. The diagnostic performance was low because the low peak value of f wave is difficult to extract, while the clinical validation process was lacking. Logan et al. [41] extracted the RR interval variance in the 10-second ECG segment to build an AF detection algorithm, with a sensitivity of 96.0% and a specificity of 89.0% in the MIT-BIH AF database. This model’s performance has been improved, but the model was also not validated in the clinic. Du et al. [38] combined the three features of “average number of f waves in TQ interval”, “maximum difference of RR interval” and “standard deviation of RR interval” of 6-second ECG segments for AF diagnosis. The accuracy of classifying AF or non-AF was 93.67%, and the specificity was 94.13% in the multiple MIT-BIH databases. However, the model was only validated in the open source database, and there might be overfitting and a lack of clinical verification. Importantly, none of the above models consider the nonlinear characteristics of ECG data.

Heart rhythm is essentially a nonlinear chaotic system. Advances in nonlinear analysis techniques of electrical signal processing could lead to a better arrhythmia diagnosis [42]. The LS is a nonlinear analytical technique based on chaos theory that can fully reflect the nonlinear heart characteristics and transform the traditional linear electrocardiogram into a nonlinear two-dimensional image [22]. Based on the LS, other scattergram analysis techniques have appeared in recent years. Examples include the RR difference scattergram [43], RdR scattergram [44] and three-dimensional scattergram [45]. Lian et al. [46] established a diagnostic model based on the RdR scattergram of different RR intervals (32, 64, and 128 RR intervals), which yielded excellent sensitivity and specificity for window sizes of 32 (94.4% and 92.6%, respectively), 64 (95.8% and 94.3%), and 128 (95.9% and 95.4%) in 4 PhysioNet databases(MIT-BIH AF database, MIT-BIH arrhythmia database, MIT-BIH normal sinus rhythm database, and normal sinus rhythm RR interval database). However, it does not use real-world ECG data for verification, nor does it consider the diagnostic value of the model in long-range ECG recordings. Mark Lown et al. [13] established a AF diagnostic model based on the decorrelated Lorenz plot of 60 consecutive RR intervals and support vector machine. The sensitivity and specificity of the model were 100.0% and 97.6% in the validation data from heart rate monitor device (Polar-H7). However, the model only considered its diagnostic performance in short-range ECG (60 RR intervals). Its diagnostic performance in long range ECG was not further verified. On the other hand, the validation data were collected in a relaxed state without an interference signal. The performance of model was limited in long-term ECG data because of long-term ECGs with rich noise and baseline drift [47]. Single-lead wearable ECG devices lack supplementary information from other leads, and are more susceptible to noise and baseline drift, resulting in poor quality. A study indicated that wearable devices based on PPG signals had poor quality, resulting in an inability to determine diagnoses in as high as 32.2% of the data [48]. Another study found that approximately 20% of data from the heart belt ECG could not be interpreted because of poor quality [14]. In our study, the proportion of low-quality ECG data was 10.77%. This percentage is lower than other studies. This may be attributed to differences in the evaluation methods for data quality. Further research is warranted.

Our model was established based on CNN and LS adequately considered the nonlinear ECG data characteristics and achieved good performance in long-range ECG in the real world. The sensitivity of the model is high, which makes it unlikely that AF was missed. It can be used for screening in community settings. In addition, the specificity of the model is high and false positive is not easy to occur. The high performance of the model allows for an earlier and more accurate diagnosis of AF, enabling timely delivery of intervention and appropriate treatment strategies. However, whether this model can actually bring benefits to the subjects needs to be further studied. Moreover, we can establish similar LSs based on the peak of the PPG signal in theory. Subsequently, the generated graph can be fed into existing models to diagnose AF. This would make the algorithm much wider applicable.

Limitations

There were several limitations to this study. First, our study used the CNN to establish a diagnosis model. This model is a black box that has poor interpretability. Explainable models are more popular in general, which can prevent some of the poor decisions in medicine that are caused by black box models [49]. Second, the research subject of the MS-AF dataset was collected from a primary care setting. However, we did not continuously include subjects due to interruptions caused by the COVID-19 pandemic, which caused selection bias for subjects. This is also the reason why the rate of AF in the MS-AF dataset was as high as 40.65%. Third, the number of patients with paroxysmal AF was only five patients in the MS-AF dataset, which is a low number and a low proportion of AF patients in the current study. This may affect the test results of the model. Fourth, limited by the types and number of arrhythmias in the database, the current model can only distinguish between non-AF, paroxysmal AF and persistent AF. In the classification of non-AF, it also includes other arrhythmias, such as atrial flutter. Compared to sinus rhythm, these conditions are associated with an increased risk of stroke or other complications. Hence, the presence of non-AF does not indicate that the subjects are completely free from underlying pathological conditions. With the accumulation of ECG data, we can develop machine learning diagnostic models for other arrhythmias to help general practitioners to identify more high-risk groups. Fifth, the database in this study may not fully capture the diversity of primary care population. More diverse and representative datasets from primary care settings are needed to validate and extend our findings. And we lacked a direct comparison with current standard-of-care methods and other machine learning algorithms within the same dataset.

Conclusion

We developed a novel model for detecting AF episode based on a nonlinear LS and deep learning. The model had high accuracy and had application value in AF screening in primary care.

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Abbreviations

AF:

Atrial fibrillation

LS:

Lorenz scattergram

CNN:

Convolutional neural network

MIT-BIH AE:

MIT-BIH Ambulatory Electrocardiograph

MS-AF:

MOBILE-SCREEN-AF

Sen:

Sensitivity

Spe:

Specificity

Acc:

Accuracy

PPV:

Positive predictive value

NPV:

Negative predictive value

+LR:

Positive likelihood ratio

-LR:

negative likelihood ratio

ROC:

Receiver operating characteristic curve

AUC:

Area under the curve

LSTM:

Long short-term memory

References

  1. Hindricks G, Potpara T, Dagres N, Arbelo E, Bax JJ, Blomstrom-Lundqvist C, et al. 2020 ESC guidelines for the diagnosis and management of atrial fibrillation developed in collaboration with the European Association for Cardio-Thoracic Surgery (EACTS): the Task Force for the diagnosis and management of atrial fibrillation of the European Society of Cardiology (ESC) developed with the special contribution of the European Heart Rhythm Association (EHRA) of the ESC. Eur Heart J. 2021;42(5):373–498. https://doi.org/10.1093/eurheartj/ehaa612.

    Article  PubMed  Google Scholar 

  2. Ziff OJ, Carter PR, McGowan J, Uppal H, Chandran S, Russell S, et al. The interplay between atrial fibrillation and heart failure on long-term mortality and length of stay: insights from the, United Kingdom ACALM registry. Int J Cardiol. 2018;252:117–21. https://doi.org/10.1016/j.ijcard.2017.06.033.

    Article  PubMed  Google Scholar 

  3. Friberg L, Rosenqvist M, Lindgren A, Terént A, Norrving B, Asplund K. High prevalence of atrial fibrillation among patients with ischemic stroke. Stroke. 2014;45(9):2599–605. https://doi.org/10.1161/strokeaha.114.006070.

    Article  CAS  PubMed  Google Scholar 

  4. Svennberg E, Tjong F, Goette A, Akoum N, Di Biase L, Bordachar P, et al. How to use digital devices to detect and manage arrhythmias: an EHRA practical guide. EP Europace. 2022;24(6):979–1005. https://doi.org/10.1093/europace/euac038.

    Article  Google Scholar 

  5. Brandes A, Stavrakis S, Freedman B, Antoniou S, Boriani G, Camm AJ, et al. Consumer-led screening for Atrial Fibrillation: Frontier Review of the AF-SCREEN International collaboration. Circulation. 2022;146(19):1461–74. https://doi.org/10.1161/circulationaha.121.058911.

    Article  PubMed  PubMed Central  Google Scholar 

  6. Neha, Sardana HK, Kanwade R, Tewary S. Arrhythmia detection and classification using ECG and PPG techniques: a review. Phys Eng Sci Med. 2021;44(4):1027–48. https://doi.org/10.1007/s13246-021-01072-5.

    Article  Google Scholar 

  7. Steinhubl SR, Waalen J, Edwards AM, Ariniello LM, Mehta RR, Ebner GS, et al. Effect of a home-based wearable continuous ECG Monitoring Patch on detection of undiagnosed Atrial Fibrillation: the mSToPS Randomized Clinical Trial. JAMA. 2018;320(2):146–55. https://doi.org/10.1001/jama.2018.8102.

    Article  PubMed  PubMed Central  Google Scholar 

  8. Ball J, Carrington MJ, McMurray JJ, Stewart S. Atrial fibrillation: profile and burden of an evolving epidemic in the 21st century. Int J Cardiol. 2013;167(5):1807–24. https://doi.org/10.1016/j.ijcard.2012.12.093.

    Article  PubMed  Google Scholar 

  9. Salinas-Martínez R, de Bie J, Marzocchi N, Sandberg F. Detection of brief episodes of Atrial Fibrillation based on Electrocardiomatrix and convolutional neural network. Front Physiol. 2021;12:673819. https://doi.org/10.3389/fphys.2021.673819.

    Article  PubMed  PubMed Central  Google Scholar 

  10. Isaksen JL, Baumert M, Hermans ANL, Maleckar M, Linz D. Artificial intelligence for the detection, prediction, and management of atrial fibrillation. Herzschrittmacherther Elektrophysiol. 2022;33(1):34–41. https://doi.org/10.1007/s00399-022-00839-x.

    Article  PubMed  PubMed Central  Google Scholar 

  11. Nuryani N, Harjito B, Yahya I, Lestari A, Anzihory E, Farosi K. Atrial fibrillation detection using swarm fuzzy inference system and Electrocardiographic P-Wave features. Procedia Comput Sci. 2015;72:154–61. https://doi.org/10.1016/j.procs.2015.12.116.

    Article  Google Scholar 

  12. Asgari S, Mehrnia A, Moussavi M. Automatic detection of atrial fibrillation using stationary wavelet transform and support vector machine. Comput Biol Med. 2015;60:132–42. https://doi.org/10.1016/j.compbiomed.2015.03.005.

    Article  PubMed  Google Scholar 

  13. Lown M, Brown M, Brown C, Yue AM, Shah BN, Corbett SJ, et al. Machine learning detection of Atrial Fibrillation using wearable technology. PLoS ONE. 2020;15(1):e0227401. https://doi.org/10.1371/journal.pone.0227401.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  14. Santala OE, Halonen J, Martikainen S, Jäntti H, Rissanen TT, Tarvainen MP, et al. Automatic Mobile Health Arrhythmia Monitoring for the detection of Atrial Fibrillation: prospective feasibility, Accuracy, and user experience study. JMIR Mhealth Uhealth. 2021;9(10):e29933. https://doi.org/10.2196/29933.

    Article  PubMed  PubMed Central  Google Scholar 

  15. Hartikainen S, Lipponen JA, Hiltunen P, Rissanen TT, Kolk I, Tarvainen MP, et al. Effectiveness of the Chest Strap Electrocardiogram to detect Atrial Fibrillation. Am J Cardiol. 2019;123(10):1643–8. https://doi.org/10.1016/j.amjcard.2019.02.028.

    Article  PubMed  Google Scholar 

  16. Olier I, Ortega-Martorell S, Pieroni M, Lip GYH. How machine learning is impacting research in atrial fibrillation: implications for risk prediction and future management. Cardiovascular Res. 2021;117(7):1700–17. https://doi.org/10.1093/cvr/cvab169.

    Article  CAS  Google Scholar 

  17. Hosny A, Parmar C, Quackenbush J, Schwartz LH, Aerts H. Artificial intelligence in radiology. Nat Rev Cancer. 2018;18(8):500–10. https://doi.org/10.1038/s41568-018-0016-5.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  18. Hong S, Zhou Y, Shang J, Xiao C, Sun J. Opportunities and challenges of deep learning methods for electrocardiogram data: a systematic review. Comput Biol Med. 2020;122:103801.

    Article  PubMed  Google Scholar 

  19. Andersen RS, Peimankar A, Puthusserypady S. A deep learning approach for real-time detection of atrial fibrillation. Expert Syst Appl. 2019;115:465–73. https://doi.org/10.1016/j.eswa.2018.08.011.

  20. Murat F, Sadak F, Yildirim O, Talo M, Murat E, Karabatak M, et al. Review of deep learning-based atrial fibrillation detection studies. Int J Environ Res Public Health. 2021;18(21):11302.

    Article  PubMed  PubMed Central  Google Scholar 

  21. Attia ZI, Noseworthy PA, Lopez-Jimenez F, Asirvatham SJ, Deshmukh AJ, Gersh BJ, 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 (London England). 2019;394(10201):861–7. https://doi.org/10.1016/s0140-6736(19)31721-0.

    Article  PubMed  Google Scholar 

  22. Esperer HD, Esperer C, Cohen RJ, Cardiac Arrhythmias Imprint Specific Signatures on Lorenz Plots. Ann Noninvasive Electrocardiol. 2008;13(1):44–60. https://doi.org/10.1111/j.1542-474X.2007.00200.x.

    Article  PubMed  PubMed Central  Google Scholar 

  23. Zhang L, Guo T, Xi B, Fan Y, Wang K, Bi J, et al. Automatic recognition of cardiac arrhythmias based on the geometric patterns of Poincaré plots. Physiol Meas. 2015;36(2):283–301. https://doi.org/10.1088/0967-3334/36/2/283.

    Article  CAS  PubMed  Google Scholar 

  24. Cui X, Chang E, Yang W-H, Jiang BC, Yang AC, Peng C-K. Automated detection of Paroxysmal Atrial Fibrillation using an information-based Similarity Approach. Entropy. 2017;19(12):677.

    Article  Google Scholar 

  25. Kisohara M, Masuda Y, Yuda E, Ueda N, Hayano J. Optimal length of R-R interval segment window for Lorenz plot detection of paroxysmal atrial fibrillation by machine learning. Biomed Eng Online. 2020;19(1):49. https://doi.org/10.1186/s12938-020-00795-y.

    Article  PubMed  PubMed Central  Google Scholar 

  26. Goldberger AL, Amaral LA, Glass L, Hausdorff JM, Ivanov PC, Mark RG, et al. PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals. Circulation. 2000;101(23):E215–20. https://doi.org/10.1161/01.cir.101.23.e215.

    Article  CAS  PubMed  Google Scholar 

  27. Moody GB, Mark RG. The impact of the MIT-BIH arrhythmia database. IEEE Eng Med Biol Mag. 2001;20(3):45–50. https://doi.org/10.1109/51.932724.

    Article  CAS  PubMed  Google Scholar 

  28. Petrutiu S, Sahakian AV, Swiryn S. Abrupt changes in fibrillatory wave characteristics at the termination of paroxysmal atrial fibrillation in humans. Europace. 2007;9(7):466–70. https://doi.org/10.1093/europace/eum096.

    Article  PubMed  Google Scholar 

  29. China WGoOi. Guidelines for prevention and control of overweight and obesity in Chinese adults. Acta Nutrimenta Sinica. 2004;26(1):1–4.

    Google Scholar 

  30. Alzubaidi L, Zhang J, Humaidi AJ, Al-Dujaili A, Duan Y, Al-Shamma O, et al. Review of deep learning: concepts, CNN architectures, challenges, applications, future directions. J Big Data. 2021;8(1):53. https://doi.org/10.1186/s40537-021-00444-8.

    Article  PubMed  PubMed Central  Google Scholar 

  31. M H, M.N S. A Review on Evaluation Metrics for Data Classification Evaluations. Int J Data Min Knowl Manage Process. 2015;5(2).

  32. Liaqat S, Dashtipour K, Zahid A, Assaleh K, Arshad K, Ramzan N. Detection of atrial fibrillation using a machine learning approach. Information. 2020;11(12):549.

    Article  Google Scholar 

  33. Xia Y, Wulan N, Wang K, Zhang H. Detecting atrial fibrillation by deep convolutional neural networks. Comput Biol Med. 2018;93:84–92. https://doi.org/10.1016/j.compbiomed.2017.12.007.

    Article  PubMed  Google Scholar 

  34. Jo Y-Y, Cho Y, Lee SY, Kwon J-m, Kim K-H, Jeon K-H, et al. Explainable artificial intelligence to detect atrial fibrillation using electrocardiogram. Int J Cardiol. 2021;328:104–10. https://doi.org/10.1016/j.ijcard.2020.11.053.

    Article  PubMed  Google Scholar 

  35. Lai D, Bu Y, Su Y, Zhang X, Ma CS. Non-standardized Patch-based ECG lead together with Deep Learning Based Algorithm for Automatic Screening of Atrial Fibrillation. IEEE J Biomedical Health Inf. 2020;24(6):1569–78. https://doi.org/10.1109/JBHI.2020.2980454.

    Article  Google Scholar 

  36. Mittal S, Oliveros S, Li J, Barroyer T, Henry C, Gardella C. AI Filter improves positive predictive value of Atrial Fibrillation detection by an Implantable Loop recorder. JACC: Clin Electrophysiol. 2021;7(8):965–75. https://doi.org/10.1016/j.jacep.2020.12.006.

    Article  PubMed  Google Scholar 

  37. Faust O, Shenfield A, Kareem M, San TR, Fujita H, Acharya UR. Automated detection of atrial fibrillation using long short-term memory network with RR interval signals. Comput Biol Med. 2018;102:327–35. https://doi.org/10.1016/j.compbiomed.2018.07.001.

    Article  PubMed  Google Scholar 

  38. Du X, Rao N, Qian M, Liu D, Li J, Feng W, et al. A novel method for real-time atrial fibrillation detection in electrocardiograms using multiple parameters. Annals Noninvasive Electrocardiology: Official J Int Soc Holter Noninvasive Electrocardiol Inc. 2014;19(3):217–25. https://doi.org/10.1111/anec.12111.

    Article  Google Scholar 

  39. Jiang K, Huang C, Ye SM, Chen H. High accuracy in automatic detection of atrial fibrillation for Holter monitoring. J Zhejiang Univ Sci B. 2012;13(9):751–6. https://doi.org/10.1631/jzus.B1200107.

    Article  PubMed  PubMed Central  Google Scholar 

  40. Huang C, Ye S, Chen H, Li D, He F, Tu Y. A novel method for detection of the transition between atrial fibrillation and sinus rhythm. IEEE Trans Biomed Eng. 2011;58(4):1113–9. https://doi.org/10.1109/tbme.2010.2096506.

    Article  PubMed  Google Scholar 

  41. Logan B, Healey J. Robust detection of atrial fibrillation for a long term telemonitoring system. In: Computers in Cardiology, 2005. 2005: 619 – 22.

  42. Zhang L, Zhao D. The clinical significance of Lorenz scatter plots in detecting ventricular parasystoles. Int J Clin Exp Med. 2020;13(6):4243–50.

    Google Scholar 

  43. Hargittai S. Is it possible to detect atrial fibrillation by simply using RR intervals? In: Computing in Cardiology 2014. IEEE; 2014: 897–900.

  44. Lu H, Lu X, Wang C, Hua Y, Tian J, Liu S. Heart rate variability study based on a novel RdR RR intervals scatter plot. Sheng Wu Yi Xue Gong Cheng Xue Za Zhi = Journal of Biomedical Engineering = Shengwu. Yixue Gongchengxue Zazhi. 2014;31(4):747–50.

    Google Scholar 

  45. Hu M, Han W, Liu J, Jang C, Wang S. Extraction and quantification clusters of three-dimensional Lorenz plots. MEDINFO 2017: Precision Healthcare through Informatics. IOS; 2017. pp. 118–21.

  46. Lian J, Wang L, Muessig D. A simple method to detect Atrial Fibrillation using RR intervals. Am J Cardiol. 2011;107(10):1494–7. https://doi.org/10.1016/j.amjcard.2011.01.028.

    Article  PubMed  Google Scholar 

  47. Jerard JM, Thilagaraj V, Pandiaraj M, Easwaran K, Govindan M, Elamaran P. Reconfigurable Architectures with high-frequency noise suppression for Wearable ECG devices. J Healthc Eng. 2021;2021:1552641. https://doi.org/10.1155/2021/1552641.

    Article  Google Scholar 

  48. Avram R, Ramsis M, Cristal AD, Nathan V, Zhu L, Kim J, et al. Validation of an algorithm for continuous monitoring of atrial fibrillation using a consumer smartwatch. Heart Rhythm. 2021;18(9):1482–90. https://doi.org/10.1016/j.hrthm.2021.03.044.

    Article  PubMed  Google Scholar 

  49. Rudin C. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nat Mach Intell. 2019;1(5):206–15. https://doi.org/10.1038/s42256-019-0048-x.

    Article  PubMed  PubMed Central  Google Scholar 

Download references

Acknowledgements

We thank American Journal Experts for providing language editing.

Funding

This work was supported financially by grants from Sichuan Science and Technology Program (No. 2023YFS0027, 2023YFS0240, 2023YFS0074, 2023NSFSC1652, 2022YFS0279, 2021YFQ0062, 2022JDRC0148), Sichuan Provincial Health Commission (No. ZH2022-101), Sichuan University West China Nursing Discipline Development Special Fund Project (No. HXHL21016).

Author information

Authors and Affiliations

  1. General Practice Ward/International Medical Center Ward, General Practice Medical Center, West China Hospital, Sichuan University, Chengdu, China

    Yi Yao, Yu Jia, Miaomiao Wu, Songzhu Wang, Haiqi Song, Xiang Fang, Xiaoyang Liao & Qian Zhao

  2. Teaching&Research Section, General Practice Medical Center, West China Hospital, Sichuan University, Chengdu, China

    Yi Yao, Yu Jia, Miaomiao Wu, Songzhu Wang, Haiqi Song, Xiang Fang, Xiaoyang Liao & Qian Zhao

  3. General Practice Medical Center and General Practice Research Institute, West China Hospital, Sichuan University, Chengdu, China

    Yi Yao, Yu Jia, Miaomiao Wu, Songzhu Wang, Haiqi Song, Xiang Fang, Xiaoyang Liao & Qian Zhao

  4. Department of Emergency Medicine and Laboratory of Emergency Medicine, West China Hospital, Sichuan University, Chengdu, China

    Dongze Li

Authors
  1. Yi Yao
    View author publications

    You can also search for this author in PubMed Google Scholar

  2. Yu Jia
    View author publications

    You can also search for this author in PubMed Google Scholar

  3. Miaomiao Wu
    View author publications

    You can also search for this author in PubMed Google Scholar

  4. Songzhu Wang
    View author publications

    You can also search for this author in PubMed Google Scholar

  5. Haiqi Song
    View author publications

    You can also search for this author in PubMed Google Scholar

  6. Xiang Fang
    View author publications

    You can also search for this author in PubMed Google Scholar

  7. Xiaoyang Liao
    View author publications

    You can also search for this author in PubMed Google Scholar

  8. Dongze Li
    View author publications

    You can also search for this author in PubMed Google Scholar

  9. Qian Zhao
    View author publications

    You can also search for this author in PubMed Google Scholar

Contributions

XL, DL, QZ and YY initiated the research. XL, DL, QZ and YY contributed to design of the study. MW, HS, XF and SW organized the database. YY performed the analysis. YY wrote the first draft of the manuscript. YJ and MW wrote sections of the manuscript. DL and QZ supervised the entire process. All authors contributed to manuscript revision, read, and approved the submitted version.

Corresponding authors

Correspondence to Dongze Li or Qian Zhao.

Ethics declarations

Ethical approval

This research was approved by the Ethics Committee of West China Hospital of Sichuan University (2018 − 454) and has been performed in accordance with the Declaration of Helsinki. The data used in this study were anonymized before its use. According to China legislation, no informed consent by participants were needed for the open source databases (MIT-BIH AE dataset). And written informed consent of MS-AF dataset was obtained from individual or guardian participants.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.

Additional information

Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Supplementary Material 1

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Yao, Y., Jia, Y., Wu, M. et al. Detection of atrial fibrillation using a nonlinear Lorenz Scattergram and deep learning in primary care. BMC Prim. Care 25, 267 (2024). https://doi.org/10.1186/s12875-024-02407-3

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://doi.org/10.1186/s12875-024-02407-3

Keywords

BMC Primary Care

ISSN: 2731-4553

Contact us