Optimal Length of R-R Interval Segment Window for Lorenz Plot Detection of Paroxysmal Atrial Fibrillation by Machine Learning

doi:10.21203/rs.3.rs-19433/v1

Background

Machine learning of R-R interval Lorenz plot (LP) images is a promising method for the detection of atrial fibrillation (AF) in long-term ECG monitoring, but the optimal length of R-R interval segment window for the LP images is unknown. We examined the performance of LP AF detection by differing the segment length using convolutional neural network (CNN). LP images with a 32 x 32-pixel resolution of non-overlapping R-R interval segments with lengths of 10, 20, 50, 100, 200, and 500 beats were created from 24-h ECG data in 52 patients with chronic AF and 58 non-AF controls as training data and in 53 patients with paroxysmal AF and 52 non-AF controls as test data. For each segment length, classification models were made by 5-fold cross-validation subsets of the training data and its classification performance was examined with the test data.

Results

In machine learning with the training data, the averages of cross-validation scores were 0.995 and 0.999 for 10 and 20-beat LP images, respectively, and >0.999 for 50 to 500-beat images. The classification of test data showed good performance for all segment lengths with an accuracy from 0.970 to 0.988. Positive likelihood ratio for detecting AF segments, however, showed a convex parabolic curve linear relationship to log segment length with a peak ratio of 111 at 100 beats, while negative likelihood ratio showed monotonous increase with increasing segment length.

Conclusions

This study suggests that the optimal R-R interval segment window length that maximizes the positive likelihood ratio for detecting paroxysmal AF with 32 x 32-pixel LP image is about 100 beats.

Biomedical Engineering

Artificial intelligence

atrial fibrillation

convolutional neural network

Holter electrocardiogram

Lorenz plot

machine learning

paroxysmal atrial fibrillation

Atrial fibrillation (AF) is known as a major cause of acute cerebral infarction [1]. Because AF often causes embolization in the main trunk of the cerebral arteries, it leads to wide areas of infarct and its prognosis is generally poor [2]. Patients with AF can prevent serious cerebral infarction by taking oral anticoagulants suppressing intra-atrial thrombus formation [3]. Thus, it is critically important to detect the presence of AF. AF, especially paroxysmal AF (PAF) is generally diagnosed by long-term ECG monitoring [4], where it is detected by algorithmically analyzing the digital ECG or R-R interval data [5–8]. To characterize the R-R interval variations with AF, many studies have proposed various indices, which include Markov process model [5], coefficient of variance [6], symbolic dynamics [7], Shannon entropy [8], and turning point ratio of R-R intervals and their combinations [8]. All of these studies used the MIT-BIH AF and other Physionet ECG databases [http://physionet.org/physiobank/database/] and reported sensitivity and specificity of 84-97% for AF segments with a length from 17 to 128 beats. There is little information, however, about the performance when they are applied to 24-h Holter ECG and thus, their practical utility for the detection of PAF is unknown. Furthermore, the performance of these algorithms may be inferior to classification by the human eye. They could fail to classify the episodes of PAF from other arrhythmias with irregular R-R intervals [9], but most of them are easily identified by humans from visual images of ECG waveforms or R-R interval trajectories such as Lorenz plot (LP) [10, 11]. Recently, machine-learning-based approaches have become popular in medical fields [12]. These approaches are expected to improve AF detection because they are advantageous over traditional algorithms in image identifications.

AF detections methods by machine learning are divided into two categories depending on the type of image used, i.e., ECG waveform or inter-beat interval LP (Figure 1) [13]. The methods using ECG images have the advantage that p-wave information can be used to distinguish AF from other arrhythmias. Xia et al. [14] used convolutional neural network (CNN) for two-dimensional matrix images derived from 5-s ECG waveform. Their two CNN models corresponding to short-term Fourier and stationary wavelet transforms achieved an accuracy of 98.29% and 98.63%, respectively. Fan et al [15] proposed multiscale CNNs that had two-stream convolutional networks with different filter sizes. They reported 96.99% and 98.13% of classification accuracy for ECG recordings of 5 and 20 s, respectively. These studies with CNNs demonstrated high performance of AF detection from short ECG waveforms. The performance of the ECG-waveform-based methods, however, depends on the signal quality of ECG and on the cardiac electric axis that affects the waveform and amplitude of P waves and f waves [16,17]. The recording of adequate ECG signals seems hard to be guaranteed even in clinical practice [18]. On the other hand, there are a few studies to apply machine learning on inter-beat interval LP images [13]. In contrast to ECG waveform that could appear as different shapes among subjects even for the same type of cardiac rhythm/arrhythmia, LP shows closer shapes for the same type of rhythm/arrhythmia when a sufficient data length is used for the plot. Actually, LP is often useful to detect PAF episodes that are overlooked by the visual inspection of ECG waveforms [10, 11]. The further advantages of the LP-image-based method are robustness to noises and potential applicability to other signals such as pulse waves. Given the rapid spreading of wearable pulse wave sensors, this method could become the mainstream of AF detection. This method, however, has at least two disadvantages; i.e., the difficulty in distinguish AF from other arrhythmias and low temporal resolution. Since detailed diagnosis of arrhythmia types requires analysis of the ECG waveforms, the former may not be a critical disadvantage. Whereas, temporal resolution is important because it can affect the performance of detecting short PAF episodes. To obtain a useful LP images for classification, a longer data segment (R-R interval sequence) is required (Figure 2), but at the expense of lower temporal resolution. At present, however, there is no study to report the R-R interval segment window length to optimize the performance of LP-based AF detection.

We therefore conducted this study to examine the effect of segment length of LP image on AF detection performance. CNN was used for generating classification models and 24-h ECG data of AF and non-AF cases were obtained from the Holter ECG database of the Allostatic State Mapping by Ambulatory ECG Repository (ALLSTAR) project [19]. As training data to develop classification models, we used R-R interval data from chronic AF (CAF) and non-AF cases, while as test data to evaluate the classification performance of the models, we used R-R interval data from PAF and non-AF cases. The reasons why we used asymmetric data for the training and test were as follows. We performed the classification on an LP image base, not on a case base in this study. We needed to balance the number of LP images between AF and non-AF in the training data for a wide range of segment length. To generate an enough number of LP images during confirmed AF periods, we used CAF data on the assumption that most of the LP features of CAF are also in PAF. On the other hand, because we aimed at determining the optical segment length for detecting PAF, we generated LP images form PAF and non-AF cases for the test data. We examined segment lengths between 10 and 500 beats. On each segment length, we performed the machine learning and examined the classification performance.

Characteristics of data

The characteristics of the subjects contributed to the training and test data are shown in Table 1. The median (IQR) duration of PAF in subjects with PAF in test data was 139.9 (10.3-579.3) min/24 h. Table 2 shows the characteristics of LP images with different segment lengths generated from the training and test data.

Machine learning with training data

The machine learning of the training data took the maximum epoch of 204 on the training of the 500-beat LP images and the minimum epoch of 96 on the 50-beat LP images. Training time length tended to increase with decreasing segment length and took 13342.05 s (3.71 h) for the training of the 10-beat LP images. The cross-validation scores were 0.995 and 0.999 for 10 and 20-beat LP images, respectively, and >0.999 for 50 to 500-beat images. The confusion matrix in the training data for each segment length is shown as Table 3.

Classification performance of test data

Table 3 also showed the confusion matrix of the test data for each segment length. Figure 5 showed receiver-operating characteristic (ROC) curves for each segment length. The area under the curve (AUC) was high for all segment lengths and the ROC curves were mostly overlapped. Figures 6A and 6B show the relationships between classification performance measures and segment length. A maximum accuracy of 0.988 and a minimum accuracy of 0.970 were obtained at 100 and 10 beats, respectively. Sensitivity, negative predictive value, and AUC of ROC curve increased monotonously with decreasing segment length. Whereas, specificity, positive predictive value, and accuracy showed convex parabolic curve linear relationships with logarithmic segment length with peaks at a segment length of 100 beats (common log value of 2.0). At this point, sensitivity and specificity reached 0.968 and 0.991, respectively.

Positive likelihood ratio also showed a convex curve linear relationship with log segment length with a peak of 111 at 100 beats (Figure 6C). Whereas, negative likelihood ratio decreased monotonously with decreasing segment length (Figure 6D). It showed 0.032 at 100 beats and reached 0.006 at 10 beats.

In this study, we examined effects of the length of R-R interval segment window on the performance of LP-image-based PAF detection. We used machine learning with CNN for generating classification models. In general, the differences in LP images between non-AF and AF become clearer as the R-R interval segment becomes longer (Figure 2). Therefore, we originally intended to search for the minimum segment length required for properly detecting AF. In the training data consisting of CAF and non-AF, as expected, the longer the segment length, the higher the precision of AF classification by the CNN model. In the test data consisting of PAF and non-AF, however, the positive likelihood ratio of AF detection showed a convex curve with a peak at 100 beats. This indicates that, in PAF detection using LP images, if the segment length of the RR interval is too long, the classification performance is rather reduced, and there is an optimal segment length for that. As the result of analysis, we found that the positive likelihood ratio for detecting PAF with 32 × 32-pixel LP image reached a peak value of 111 at a segment length of 100 beats.

Although the LP of R-R intervals has long been recognized as a useful tool to detect AF [10, 11], the present study is the first to report the effects of segment length on PAF detection by LP. Anan et al. [10] used LP with R-R intervals of one-h segments to estimate the functional refractory period of atrioventricular conduction. In a study of the circadian variation of atrioventricular conduction properties during AF, Hayano et al. [11] used LP with 512 beats. In a recent study, we reported the performance of AF detection with LP image using CNN [20]. We used a segment length of 600 beats for LP images with a 32 × 32-pixel resolution and a 5-bit brightness level. We observed that the CNN model achieved 100% accuracy in the five-fold cross-validation of the training data consisted of CAF and non-AF. In the test data consisted of 24-h data including PAF and those without AF, the model detected the 24-h data including PAF with 100% accuracy on a case base, although not on an LP-image base. The previous finding of 100% classification accuracy for the training data was reproduced in the present study in the classification of the training data with 200-beat and 500-beat LP images. None of these earlier studies, however, have reported the rationale for choosing those segment lengths for LP images.

In this study, we employed the likelihood ratio for evaluating the performance of classification models and used regression analysis for determining the relationships between the performance measures and LP segment length. The reason for using the likelihood ratio was that while other measures of performance, such as sensitivity, specificity, predictive values, and accuracy, are influenced by the prior probability of events, the likelihood ratio is known to be robust to the probability [21]. In the present study, the LP images were classified as AF if the segments included PAF even partly. Therefore, the probability that a segment is classified as AF increased with the length of segment. Actually, the ratio of LP images classified as AF in the test data was 16.1% and 14.5% for 500-beat and 10-beat LP images (Table 2). On the other hand, the reason for using regression analysis for the relationships between performance measures and segment length was that the observed performance measures were the products from the data randomly sampled through the probabilistic segmentation of R-R intervals. Thus, the values were considered to vary around the true values that need to be estimated statistically by regressions.

In the present study, we found that the optimal segment length of LP images for detecting PAF is around 100 beats, when LP images with a 32 × 32-pixel resolution and a 3-bit brightness level are used. Several mechanisms may be considered for this finding. The mechanism by which the lower limit of segment length exists may be straightforward. Assuming that AF is a random sequence of R-R intervals, a certain number of points are required stochastically for the LP to form a characteristic shape to AF. Actually, the confusion matrices for the training data showed that longer segment lengths improved classification performance (Table 3). In contrast, the mechanisms by which the upper limit of segment length exists may be more complicated. First, if all LP images for PAF consisted only of AF beats, there may be no clear difference in the appearance of the LP image between PAF and CAF, but they actually included segments consist of AF, sinus rhythm, and premature beats. The number of premature beats in an LP image increases as the segment length increases (Table 2), which may form complex LP images that are difficult to distinguish from those including PAF episode(s) shorter than the segment length. Second, the observed decreases in the sensitivity and negative predictive value (Figure 6) may be the consequences of increasing ratio of LP images with AF (prior probability) with increasing segment length (Table 2). These above suggest that the existence of the optical point may be explained as a result of a trade-off between two different segment length effects that change in the opposite direction each other for the classification performance.

This study has the following limitations. First, although the CAF segments in the training data and the PAF segments included transient atrial flutter or atrial flutter-fibrillation, we included patients with pure atrial flutter in neither training nor test data. Consequently, the results of the present study cannot be extended to the classification between sinus rhythm (SR) and atrial flutter. Second, we used asymmetric datasets for the training and test data, i.e., CAF for the training data and PAF for the test data. Thus, swapping the training and test data may give different results. We did not examine it because the study purpose was the detection of PAF but not CAF. It may be important, however, to study using mixed data of CAF and PAF for both training and test, if an enough number of segment data with confirmed rhythm was available, which was not possible in the present study due to the necessity to balance the number of LP images between AF and non-AF in the training data. Third, among the segment lengths selected in this study, a segment length of 100 beats was found to be optimal for the classification of LP images in test data, but the precise optimal segment length was unknown because we selected six segment lengths arbitrarily. However, the relationship between classification performance and segment length was examined by regression analysis. Given the stochastic feature of performance measures depend on data segmentation patterns, the estimated optimal segment length of 100 beats seems a reasonable approximation. Fourth, although we used 5-fold cross-validation, there is no guarantee that it is the best approach to finding the optimal segment length. We calculated five predictive values for each LP image in test data using five classification models and averaged the five predictive values. We used this as the final predictive value for each LP image, and examined effects from different segment lengths. It was possible that each classifier doesn't show the same trend for each segment length, but we would eventually have to average that trend for each segment length in order to obtain the effect of each segment length. However, there is no prior report about the choice of an appropriate classification method to this issue. Finally, we used LP images with a 32 × 32-pixel resolution and a 3-bit brightness level. These parameters were determined considering the computational power of a currently commercialized Holter ECG analyzers, but using LP images with different resolution and density levels may give different results.

We examined the effects of segment window length for generating R-R interval LP images on the performance of PAF detection by CNN. We found that the segment length is an important determinant of the performance and that the optimal segment length that maximizes the positive likelihood ratio for detecting PAF with 32 × 32-pixel LP image is about 100 beats.

Data selections

Study data were obtained from the ALLSTAR database [19] that was consisting of more than 400,000 data of 24-h ambulatory ECG consecutively referred to the Suzuken Holter ECG analysis centers at Tokyo, Nagoya, and Sapporo in Japan (Suzuken Co., Ltd., Nagoya, Japan). All Holter ECG data in the database were sampled at 125 Hz and 10-bit resolution (0.02 mV/digit). The database was used for extracting both training and test data. The use of this database for this study has been approved by the Institutional Review Board of Nagoya City University Graduate School of Medical Sciences and Nagoya City University Hospital (No. 709).

Training data

As the sources of the training data, we extracted 24-h Holter ECG data in 58 subjects with persistent SR >99% of 24-h total beats without PAF and 52 subjects with CAF. The cardiac rhythms were diagnosed and confirmed independently by multiple laboratory technicians and cardiologists. The training data were used for the machine learning to develop classification models.

Test data

As the sources of the test data, we extracted 24-h Holter ECG data of other 52 subjects with SR without PAF and of 53 subjects with PAF, independently of the training data. The cardiac rhythms were diagnosed and confirmed independently by multiple ECG technicians and cardiologists. The onset and offset points of each episode of PAF were determined by agreement between the ECG technicians and confirmed by clinical laboratory technicians. The authors in this research were not involved in these ECG assessments. The test data were used for evaluating the performance of classification model obtained from the training data.

Lorenz plot image

We used only R-R interval data from each 24-h ECG. The detection of QRS waves and the measurement of R-R intervals were performed with Holter ECG scanners (Cardy Analyzer 05, Suzuken Co., Ltd., Nagoya, Japan). For both training and test data, 24-h R-R interval data were split into consecutive non-overlapping segments of 10, 20, 50, 100, 200, and 500 beats for generating LP images. These segment lengths were chosen so that they were approximately equally spaced on log scales.

For each trial, LP was generated for each data segment by plotting all R-R intervals but the first one as y values against preceding R-R intervals as x values [10, 11]. The obtained LPs were converted into the monochrome images of a 32 × 32-pixel resolution with 3-bit scale for brightness level, resulting in a temporary resolution of 80 ms and a dynamic range between 0 and 2,560 ms for both the x and y axes. When (x, y) values scaled out of the range, the data were plotted at the edge of the image. The number of plots in each pixel was counted and was used as the brightness level of the pixel. When the number of plots was >7, the level was set at 7. Figure 3 shows the example of LP images.

Machine learning with training data

The machine learning was performed by a CNN using keras (version 2.2.4) open-source python machine learning library, using the Microsoft Windows 10 operating system on a computer equipped with an 8-core Intel Xeon E3-1275 V3 processor with 32-GB memory. The computer was also equipped with an NVIDIA RTX 2080 Ti graphic board with 27-GB memory.

The CNN consisted of an input layer, 1st convolution layer, 2nd convolution layer, 1st max pooling layer, 3rd convolution layer, 2nd max pooling layer, 1st dropout layer, a flatten layer, 1st dense layer, 2nd dropout layer, and 2nd dense layer (Figure 4A). In all convolution layers, the kernel size was set at 3 × 3 and the numbers of output filter were set at 32, 64, and 128 in the first, second, and third convolution layers, respectively. In all pooling layers, the height and width were set to be half each. The dropout rate was set at 0.25. In the 1st and 2nd dense layers, the numbers of unit were set at 128 and 2, respectively. Rectified linear unit was used for the activation function of all convolution layers and 1st dense layer and softmax for the activation function of the 2nd dense layer in order to classify LP images. Binary cross entropy was used for loss function and stochastic gradient descent (SGD) for optimizer. The hyperparameters of SGD were as follows: the learning rate = 0.001, the learning rate decay = 0.000001, and the momentum accelerating SGD = 0.9.

We enrolled 5-fold cross-validation for the CNN. Briefly, the LP images of training data were divided randomly into five subsets. For each cross-validation, one subset out of five was selected in order and used as a validation subset, and the remaining 4 subsets were used as training subsets (Figure 4B). Consequently, we obtained five datasets with different training subsets and validation subsets. We set the batch size to 32. Accuracy and loss were updated for each batch. The number of the maximum epochs was set at 50 times per training. If the validation loss didn’t improve during 10 epochs, training stopped early. As a result, we obtained five models for each segment length. The validation accuracy and the validation loss of each dataset were updated for each epoch. We obtained the final validation accuracy and the final validation loss of each dataset in the last epoch.

Classification accuracy in the training data was evaluated by a cross-validation score and the confusion matrices of the entire training data. For the cross-validation score, predictive values were calculated by applying each of the five classification models to each LP image in the corresponding data subset from which the model was derived. LP images were classified as AF when the predictive value was ≥0.5 and as non-AF when the value <0.5. The classification accuracy of each model was calculated and the cross-validation score was obtained as the average accuracy of the five models. On the other hand, the confusion matrices for the entire training data were obtained as follows. For each segment length, five models were applied to each LP image in the training data. LP images were classified as AF when the average of five values was ≥0.5 and as non-AF when it was <0.5. The classification results were summarized as a confusion matrix for each segment length.

Classification of test data

We calculated five predictive values for each LP image in test data using five classification models. We averaged the five predictive values as the final predictive value for each LP image (Figure 4B). A LP image was classified as AF when the final predictive value was ≥0.5 and as non-AF when the value <0.5. A confusion matrix for each segment length was calculated against the grand truth.

Statistical analysis

The receiver operating characteristics curve and the area under the curve were calculated for the final predictive value. From the classification results of the test data, the sensitivity, specificity, positive predictive value, negative predictive value, accuracy, and positive and negative likelihood ratios of classifications were calculated for each segment length. The relationships between these performance measures and segment length were analyzed by regression curve fitting. The logarithmic transformation of axis was used as necessary to the better fitting of regression curves. Among these indices of classification performance, the optimal segment length was determined based on the likelihood ratios that are known to provide a fair evaluation for classification performance independent of prior probability [21]. These analyses were performed with numpy (version 1.15.4) and scikit-learn (version 0.20.2) open-source python machine learning library.

AF: atrial fibrillation

ALLSTAR: Allostatic State Mapping by Ambulatory ECG Repository

AUC: area under the curce

CAF: chronic atrial fibrillation

CNN: convolutional neural network

LP: Lorenz plot

PAF: paroxysmal atrial fibrillation

ROC: receiver-operating characteristic

SGD: stochastic gradient descent

SR: sinus rhythm

Ethics approval and consent to participate

The study protocol and the use of the ALLSTAR database in this study have been approved by the Institutional Review Board of Nagoya City University Graduate School of Medical Sciences and Nagoya City University Hospital (No. 709).

Consent for publication

Not applicable.

Availability of data and materials

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

Competing interests

The authors declare that they have no competing interests.

Funding

This study was performed as a part of the collaborative study between Suzuken Co., Ltd. and Nagoya City University Graduate School of Medical Sciences.

Authors’ contributions

MK: concept/design, data analysis/interpretation, drafting article, statistics, and approval of article. YM: concept/design, data collection, data analysis/interpretation, critical revision of article, and approval of article. NU: concept/design, data collection, data interpretation, critical revision of article, and approval of article. EY: concept/design, data interpretation, critical revision of article, and approval of article. JH: concept/design, data collection, data interpretation, statistics, drafting article, critical revision of article, and approval of article.

Acknowledgements

Not applicable.

Affiliations

Department of Medical Education, Nagoya City University Graduate School of Medical Sciences, Kawasumi 1 Mizuho-cho, Mizuho-ku, Nagoya 4678602, Japan

Masaya Kisohara, Norihiro Ueda, Junichiro Hayano

Kenz Division, Suzuken Co. Ltd., Suzuken Tomei building, Himewaka-cho 6, Meito-ku, Nagoya 4650045, Japan

Yuto Masuda

Department of Electrical Engineering, Tohoku University Graduate School of Engineering, Aoba 6-6-05 Aramaki,,Aoba-ku, Sendai, 980-8759, Japan

Emi Yuda

Corresponding author

Correspondence to Junichiro Hayano.

[1] P. A. Wolf, R. D. Abbott, and W. B. Kannel, “Atrial fibrillation as an independent risk factor for stroke: The Framingham study,” Stroke, vol. 22, no. 8, pp. 983–988, 1991, doi: 10.1161/01.STR.22.8.983.

[2] K. Kimura, S. Kazui, K. Minematsu, and T. Yamaguchi, “Hospital-based prospective registration of acute ischemic stroke and transient ischemic attack in Japan,” J. Stroke Cerebrovasc. Dis., vol. 13, no. 1, pp. 1–11, Jan. 2004, doi: 10.1016/j.jstrokecerebrovasdis.2003.11.025.

[3] “Risk Factors for Stroke and Efficacy of Antithrombotic Therapy in Atrial Fibrillation: Analysis of Pooled Data From Five Randomized Controlled Trials,” Arch. Intern. Med., vol. 154, no. 13, pp. 1449–1457, Jul. 1994, doi: 10.1001/archinte.1994.00420130036007.

[4] C. T. January et al., 2014 AHA/ACC/HRS guideline for the management of patients with atrial fibrillation: Executive summary: A report of the American College of cardiology/American heart association task force on practice guidelines and the heart rhythm society, vol. 130, no. 23. 2014.

[5] G. B. Moody and R. G. Mark, “a New Method for Detecting Atrial Fibrillation Using R-R Intervals,” a New Method for Detecting Atrial Fibrillation Using R-R Intervals. pp. 227–230, 1983.

[6] K. Tateno and L. Glass, “Automatic detection of atrial fibrillation using the coefficient of variation and density histograms of RR and ΔRR intervals,” Med. Biol. Eng. Comput., vol. 39, no. 6, pp. 664–671, 2001, doi: 10.1007/BF02345439.

[7] X. Zhou, H. Ding, B. Ung, E. Pickwell-MacPherson, and Y. Zhang, “Automatic online detection of atrial fibrillation based on symbolic dynamics and Shannon entropy,” Biomed. Eng. Online, vol. 13, no. 1, 2014, doi: 10.1186/1475-925X-13-18.

[8] S. Dash, K. H. Chon, S. Lu, and E. A. Raeder, “Automatic real time detection of atrial fibrillation,” Ann. Biomed. Eng., vol. 37, no. 9, pp. 1701–1709, 2009, doi: 10.1007/s10439-009-9740-z.

[9] A. García-Alberola et al., “RR interval variability in irregular monomorphic ventricular tachycardia and atrial fibrillation,” Circulation, vol. 93, no. 2, pp. 295–300, 1996, doi: 10.1161/01.CIR.93.2.295.

[10] T. Anan, K. Sunagawa, H. Araki, and M. Nakamura, “Arrhythmia analysis by successive RR plotting,” J. Electrocardiol., vol. 23, no. 3, pp. 243–248, 1990, doi: https://doi.org/10.1016/0022-0736(90)90163-V.

[11] J. Hayano, S. Sakata, A. Okada, S. Mukai, and T. Fujinami, “Circadian rhythms of atrioventricular conduction properties in chronic atrial fibrillation with and without heart failure,” J. Am. Coll. Cardiol., vol. 31, no. 1, pp. 158–166, 1998, doi: 10.1016/S0735-1097(97)00429-4.

[12] S. Sharifi, A. Pakdel, M. Ebrahimi, J. M. Reecy, S. F. Farsani, and E. Ebrahimie, “Integration of machine learning and metaanalysis identifies the transcriptomic bio-signature of mastitis disease in cattle,” PLoS One, vol. 13, no. 2, pp. 1–18, 2018, doi: 10.1371/journal.pone.0191227.

[13] M. Kisohara, Y. Masuda, E. Yuda, and J. Hayano, “Neural Network Detection of Atrial Fibrillation by Lorenz Plot Images of Interbeat Interval Variation,” IEEE 7th Global Conference on Consumer Electronics, GCCE 2018

[14] Y. Xia, N. Wulan, K. Wang, and H. Zhang, “Detecting atrial fibrillation by deep convolutional neural networks,” Comput. Biol. Med., vol. 93, pp. 84–92, 2018, doi: https://doi.org/10.1016/j.compbiomed.2017.12.007.

[15] X. Fan, Q. Yao, Y. Cai, F. Miao, F. Sun, and Y. Li, “Multiscaled Fusion of Deep Convolutional Neural Networks for Screening Atrial Fibrillation from Single Lead Short ECG Recordings,” IEEE J. Biomed. Heal. Informatics, vol. 22, no. 6, pp. 1744–1753, 2018, doi: 10.1109/JBHI.2018.2858789.

[16] J. Slocum, A. Sahakian, and S. Swiryn, “Diagnosis of atrial fibrillation from surface electrocardiograms based on computer-detected atrial activity,” J. Electrocardiol., vol. 25, no. 1, pp. 1–8, 1992, doi: https://doi.org/10.1016/0022-0736(92)90123-H.

[17] D. E. Lake and J. R. Moorman, “Accurate estimation of entropy in very short physiological time series: The problem of atrial fibrillation detection in implanted ventricular devices,” Am. J. Physiol. - Hear. Circ. Physiol., vol. 300, no. 1, pp. 319–325, 2011, doi: 10.1152/ajpheart.00561.2010.

[18] R. He et al., “Automatic detection of atrial fibrillation based on continuous wavelet transform and 2D convolutional neural networks,” Front. Physiol., vol. 9, no. AUG, pp. 1–11, 2018, doi: 10.3389/fphys.2018.01206.

[19] J. Hayano, M. Kisohara, Y. Yoshida, H. Sakano, and E. Yuda, “Association of heart rate variability with regional difference in senility death ratio: ALLSTAR big data analysis,” SAGE Open Med., vol. 7, p. 205031211985225, 2019, doi: 10.1177/2050312119852259.

[20] J. Hayano, M. Kisohara, Y. Masuda, and E. Yuda. (2019, March). In “Detection of paroxysmal atrial fibrillation by Lorenz plot imaging of ECG RR intervals,” International Forum on Medical Imaging in Asia 2019, SPIE Proceedings Vol. 11050: p. 110501O.

[21] W. W. Flemons and M. R. Littner, “Measuring Agreement Between Diagnostic Devices,” Chest, vol. 124, no. 4, pp. 1535–1542, Oct. 2003, doi: 10.1378/chest.124.4.1535.

Table 1 Characteristics of subjects selected for the training and test data

	Training data		Test data
	Non-AF	CAF	Non-AF	PAF
n	58	52	52	53
Male, %	36.2	61.5	42.3	62.3
Age, y	61.4 ± 12.5	78.0 ± 8.4	65.7 ± 11.8	68.2 ± 9.5
Total PAF duration, min/24 h	-	-	0	139.9 (10.3-579.3)
Total PAF beat, /24 h	-	-	0	12,666 (1,263-51,983)
APC, /24 h	31 (10-61)	-	54 (11-150)	316 (88-1849)
VPC, /24 h	5 (1-39)	48 (9-196)	10 (1-24)	6 (1-28)
24-h Mean RRI, ms	844 ± 105	789 ± 141	873 ± 135	801 ± 119
24-h SD of RRI, ms	138 ± 36	217 ± 57	134 ± 43	185 ± 57

Data are frequency (%), median (IQR), or mean ± SD.

AF = atrial fibrillation; CAF = chronic AF; PAF = paroxysmal AF; RRI = R-R interval; SR = sinus rhythm, APC = atrial premature contraction; VPC = ventricular premature contraction.

Table 2 Characteristics of LP images of training and test data

Segment window length, beat	LP images of training data				LP images of test data
	Non-AF		AF		Non-AF		AF
	LP n (%)	Premature beat* n (%)	LP n (%)	Premature beat* n (%)	LP n (%)	Premature beat* n (%)	LP n (%)	Premature beat* n (%)
10	590,602 (50.7)	1.2 (12.1)	574,014 (49.3)	1.1 (11.4)	934,803 (85.5)	1.9 (18.7)	159,003 (14.5)	1.2 (12.1)
20	295,284 (50.7)	1.4 (6.75)	286,996 (49.3)	1.3 (6.48)	467,009 (85.4)	2.7 (13.5)	79,865 (14.6)	1.5 (7.73)
50	118,099 (50.7)	1.6 (3.13)	114,781 (49.3)	1.7 (3.37)	186,432 (85.2)	4.5 (8.95)	32,291 (14.8)	2.7 (5.31)
100	59,035 (50.7)	1.8 (1.75)	57,373 (49.3)	2.2 (2.22)	92,967 (85.0)	6.5 (6.53)	16,365 (15.0)	4.1 (4.06)
200	29,506 (50.7)	2.0 (1.00)	28,670 (49.3)	3.1 (1.57)	46,268 (84.7)	9.5 (4.75)	8,367 (15.3)	6.7 (3.33)
500	11,783 (50.7)	2.6 (0.51)	11,453 (49.3)	5.3 (1.05)	18,314 (83.9)	15.9 (3.17)	3,511 (16.1)	15.2 (3.03)

AF = atrial fibrillation; LP = Lorenz plot.

* The mean number and ratio of premature beats in the LP images including at least one premature beat.

Table 3 Confusion matrix for each segment window length

			Ground Truth
Segment window length, beats			Training data		Test data
Segment window length, beats			Non-AF	AF *¹	Non-AF	AF *²
Classification	10	Non-AF	588036	3341	903219	930
	10	AF	2566	570673	31584	158073
	20	Non-AF	294986	329	458259	692
	20	AF	298	286667	8750	79173
	50	Non-AF	118089	11	184063	671
	50	AF	10	114770	2369	31620
	100	Non-AF	59032	2	92160	527
	100	AF	3	57371	807	15838
	200	Non-AF	29505	0	45759	423
	200	AF	1	28670	509	7944
	500	Non-AF	11783	0	18006	281
	500	AF	0	11453	308	3230

AF = atrial fibrillation; LP = Lorenz plot.

*¹ AF LP images obtained from CAF cases.

*² AF LP images obtained from PAF cases.

Optimal Length of R-R Interval Segment Window for Lorenz Plot Detection of Paroxysmal Atrial Fibrillation by Machine Learning

Status:

Journal Publication

Version 1

Abstract

Figures

Background

Results

Discussions

Limitations

Conclusions

METHODS

Abbreviations

Declarations

References

Tables

Status:

Journal Publication

Version 1