KR102704349B1 - Method and apparatus for QRS detection of ECG signal for cardiac diagnosis - Google Patents

Abstract

A method for detecting QRS of an electrocardiogram signal for cardiac diagnosis according to one embodiment of the present invention comprises: (A) a step in which a normalization unit normalizes an input original electrocardiogram (ECG) signal; (B) a step in which a U-Net model generates a positive output indicating a predicted location of a QRS complex from the normalized ECG signal; and (C) a step in which a density-based spatial clustering unit (DBSCAN) applying noise clusters the positive output of the U-Net model to detect a QRS complex.

Description

{Method and apparatus for QRS detection of ECG signal for cardiac diagnosis}

The present invention relates to QRS detection of electrocardiogram signals for cardiac diagnosis.

An electrocardiogram (ECG) is an integrated representation of cardiac activity transmitted to the body surface by bioelectricity, which is an important noninvasive tool for diagnosing cardiac diseases. It directly reflects the rhythm of cardiac activity and contains a variety of information in terms of pathology and physiology.

In clinical practice, doctors usually diagnose heart disease by observing and analyzing ECG signals. When screening a large number of diagrams, doctors may ignore subtle changes in the signals due to excessive workload. And the level of expertise of doctors is different. The accuracy of diagnosis is greatly influenced by the expertise of doctors. In this case, an intelligent and automatic diagnostic method is required to analyze ECG to reduce the workload of doctors and improve the accuracy and reliability of diagnosis.

In the past few decades, many inventions have focused on ECG diagnosis. In many diagnostic and analysis methods for ECG signals, QRS complex detection is a preliminary and important step that directly affects further measurement, recognition, and diagnosis. The QRS complex is the most prominent and important part of the ECG cycle. It reflects potential changes during depolarization of the left and right ventricles. Among the QRS complexes, the R-peak is the peak of the QRS complex and is often considered as an indication of an independent beat. By finding the R-peak of the ECG signal, the location of the QRS complex, P wave, and T wave can be easily obtained.

In recent years, inventors have invented many algorithms for QRS detection in ECG signals, including wavelet transform, Hilbert transform, regular grammar, template matching and convolutional neural network (CNN). Among them, wavelet analysis method can extract the frequency domain characteristics of ECG signals, but it requires artificial selection of morphology, and improper selection of morphology may lead to poor results.

In addition, existing QRS detection methods usually require complex noise removal steps and rely heavily on manual settings or selections, including feature types and feature selection, parameter settings, etc., which may result in poor adaptability to various types of signals and interferences.

KR 10-2021-0109879 A

The problem to be solved by the present invention is to provide a method and device for detecting the QRS of an electrocardiogram signal for cardiac diagnosis, which has good adaptability to various types of signals and interferences and can accurately detect the QRS complex, without requiring a complex noise removal step and not greatly relying on manual settings or selections, including feature types and feature selection, parameter settings, etc.

A method for detecting QRS of an electrocardiogram signal for cardiac diagnosis according to one embodiment of the present invention for solving the above problem is as follows.

(A) A normalization step for normalizing an input original electrocardiogram (ECG) signal;

(B) a step of the U-Net model generating a positive output indicating the predicted location of the QRS complex from the normalized electrocardiogram signal; and

(C) A density-based spatial clustering unit (DBSCAN) with noise applied may include a step of clustering positive outputs of the U-Net model to detect QRS complexes.

In a method for detecting QRS of an electrocardiogram signal for cardiac diagnosis according to one embodiment of the present invention,

In the above step (C), the density-based spatial clustering unit applying the noise can detect the QRS complex by clustering the positive output of the U-Net model into multiple clusters and then determining the midpoint of the multiple clusters as the location of the R-peak.

In addition, in the QRS detection method of the electrocardiogram signal for cardiac diagnosis according to one embodiment of the present invention, the U-Net model,

A coding path section that repeatedly performs convolution and Max-pooling on the normalized ECG signal to obtain a feature map of the normalized ECG signal; and

The method may include a decoding path section that repeatedly performs upsampling and convolution on a feature map output from the above coding path section to generate a positive output indicating a predicted location of the QRS complex.

Additionally, in a method for detecting QRS of an electrocardiogram signal for cardiac diagnosis according to one embodiment of the present invention, the convolution may include a 13×1 convolution.

In addition, in the QRS detection method of an electrocardiogram signal for cardiac diagnosis according to one embodiment of the present invention, the U-Net model may include a dropout layer to avoid overfitting.

According to one embodiment of the present invention for solving the above problem, a QRS detection device of an electrocardiogram signal for cardiac diagnosis is provided.

A normalization unit that normalizes the input original electrocardiogram (ECG) signal;

A U-Net model that generates a positive output indicating the predicted location of the QRS complex from the normalized electrocardiogram signal; and

It may include a density-based spatial clustering unit (DBSCAN) that applies noise to cluster the positive output of the above U-Net model to detect QRS complexes.

In a QRS detection device for an electrocardiogram signal for cardiac diagnosis according to one embodiment of the present invention, the density-based spatial clustering unit applying the noise can detect the QRS complex by clustering the positive output of the U-Net model into a plurality of clusters and then determining the midpoint of the plurality of clusters as the location of the R-peak.

In addition, in the QRS detection device of the electrocardiogram signal for cardiac diagnosis according to one embodiment of the present invention, the U-Net model,

A coding path section for repeatedly performing convolution and Max-pooling on the normalized ECG signal to obtain a feature map of the normalized ECG signal; and

The method may include a decoding path section that repeatedly performs upsampling and convolution on a feature map output from the above coding path section to generate a positive output indicating a predicted location of the QRS complex.

Additionally, in the QRS detection device of an electrocardiogram signal for cardiac diagnosis according to one embodiment of the present invention, the convolution may include a 13×1 convolution.

In addition, in the QRS detection device of an electrocardiogram signal for cardiac diagnosis according to one embodiment of the present invention, the U-Net model may include a dropout layer to avoid overfitting.

According to a method and device for detecting QRS of an electrocardiogram signal for cardiac diagnosis according to one embodiment of the present invention, a complex noise removal step is not required and there is no significant dependence on manual settings or selections, including feature types and feature selection, parameter settings, etc., so that the adaptability to various types of signals and interferences is good and the QRS complex can be accurately detected.

FIG. 1 is a flowchart of a method for detecting QRS of an electrocardiogram signal for cardiac diagnosis according to one embodiment of the present invention.
Figure 2 is a diagram illustrating an ECG signal in the preprocessing stage, where (a) is an original ECG signal and (b) is a diagram illustrating a normalized ECG signal.
Figure 3 is a diagram illustrating the structure of the U-Net model.
Figure 4 illustrates clusters of U-Net outputs for QRS detection in the present invention, where the blue solid line represents the original ECG signal, the asterisks represent points in the clusters, the color of the asterisks represents the classes after clustering, the black dashed lines represent manually identified R-peak locations, and the black upward pointing triangles represent the R-peak location results.
FIG. 5 is a diagram illustrating QRS detection in an ECG signal in the present invention, wherein (a) is a raw ECG signal, (b) is a normalized ECG signal, (c) is a prediction of a U-Net model, (d) is a U-Net positive output using a threshold of 0.9, (e) is a diagram illustrating QRS detection, and the blue solid lines in (a), (e), and (f) represent the original ECG signal, the orange solid line in (c) represents the output of the U-Net model, the black solid line in (d) represents binarized prediction values using a threshold of 0.9, the asterisks in (e) represent points within clusters, and the color of the asterisks represents classes after clustering, the black dashed lines in (f) represent R-peak locations manually identified in MITDB, and the black upward-pointing triangles represent the results of detecting the locations of R-peaks using the present invention.
FIG. 6 is a diagram showing examples of QRS detection in ECG signals of NSR, LBBB, RBBB, PVC, and APB, where (a) is an ECG signal of NSR from record 205, (b) is an ECG signal of LBBB from record 109, (c) is an ECG signal of RBBB from record 118, (d) is an ECG signal of PVC from record 223, and (e) is an ECG signal of PVC from record 223, and for each part, the black solid line is the original ECG signal, the orange solid line represents the output of the U-Net model, the black dashed line represents an R-peak manually identified in MITDB, and the black upward-pointing triangle represents the position result of the R-peak detected using the present invention, and the R-peak detected within the gray marked area is regarded as a TP.
FIG. 7 is a diagram illustrating QRS detection in an ECG signal with a large peak T-wave and severe noise, (a) an ECG signal with a large peak T-wave from Record 113, (b) an ECG signal with severe noise from Record 209, and for each part, the blue solid line is the original ECG signal, the orange solid line represents the output of the U-Net model, the black dashed line represents the R-peak location manually identified from MITDB, the black upward-pointing triangle represents the result of the R-peak location using the present invention, and the detected R-peaks within the gray marked area are regarded as TP.
FIG. 8 is a drawing illustrating a QRS detection device of an electrocardiogram signal for cardiac diagnosis according to one embodiment of the present invention.

The objects, specific advantages and novel features of the present invention will become more apparent from the following detailed description and preferred embodiments taken in conjunction with the accompanying drawings.

Prior to this, the terms or words used in this specification and claims should not be interpreted in their usual or dictionary meanings, but should be interpreted in a meaning and concept that conforms to the technical idea of the present invention based on the principle that the inventor can appropriately define the concept of the term in order to explain his or her own invention in the best way.

When adding reference numbers to components of each drawing in this specification, it should be noted that identical components are given the same numbers as much as possible even if they are shown in different drawings.

Additionally, terms such as “first,” “second,” “one side,” “the other side,” etc. are used to distinguish one component from another, and the components are not limited by these terms.

Hereinafter, in describing the present invention, detailed descriptions of related known technologies that may unnecessarily obscure the gist of the present invention will be omitted.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the attached drawings.

QRS complex detection or QRS detection is an important task in electrocardiogram signal analysis, and it is a preliminary and essential step for further recognition and diagnosis. In a method for QRS detection of an electrocardiogram signal for cardiac diagnosis according to an embodiment of the present invention, a U-Net based method for QRS detection is proposed. The QRS detection method of an electrocardiogram signal for cardiac diagnosis according to an embodiment of the present invention consists of three steps: normalization as a preprocessing step, a U-Net model, and density-based spatial clustering of applications with noise (DBSCAN).

Normalization is performed using the Z-score method in preprocessing. In the QRS detection method of the electrocardiogram signal for cardiac diagnosis according to one embodiment of the present invention, location prediction is performed using the U-Net model. The U-Net output is then processed and clustered based on the threshold by DBSCAN. Finally, the middle point of the cluster is considered as the R peak of the QRS complex.

Meanwhile, the QRS detection method of the electrocardiogram signal for cardiac diagnosis according to one embodiment of the present invention demonstrates high accuracy for the ECG signal in the MIT-BIH Arrhythmia Database (MITDB). The experimental results show an average sensitivity of 99.98%, a positive predictive value of 99.95%, an accuracy of 99.93%, and an F1 score of 99.97%. The QRS detection method of the electrocardiogram signal for cardiac diagnosis according to one embodiment of the present invention is similar to other existing methods in overall performance and is much superior in accuracy and F1 score.

The method and device for QRS detection of electrocardiogram signals for cardiac diagnosis according to one embodiment of the present invention aim to design QRS detection to achieve state-of-the-art performance. The proposed method is based on U-Net and density-based spatial clustering (DBSCAN) with noise.

QRS detection algorithm using CNN achieves high accuracy with few preprocessing steps and manual settings. Therefore, we first adopt U-Net, a widely used CNN, to predict ECG signals. The prediction results are classified into clusters to find R-peaks. The number of QRS complexes is generally unknown and their lengths are different. Therefore, DBSCAN is adopted for classification, which does not require the number of clusters to be set in advance and can find clusters of all sizes and shapes.

Referring to FIG. 8, a QRS detection device of an electrocardiogram signal for cardiac diagnosis according to one embodiment of the present invention may include a normalization unit (800) that normalizes an input raw electrocardiogram (ECG) signal, a U-Net model (802) that generates a positive output indicating a predicted location of a QRS complex from the normalized electrocardiogram signal, and a density-based spatial clustering unit (DBSCAN) (804) that applies noise and clusters the positive output of the U-Net model (802) to detect the QRS complex.

In addition, referring to FIG. 1 and FIG. 8, a method for detecting QRS of an electrocardiogram signal for cardiac diagnosis according to an embodiment of the present invention may include a step (S100) in which a normalization unit (800) normalizes an input original electrocardiogram (ECG) signal, a step (S102) in which a U-Net model (802) generates a positive output indicating a predicted location of a QRS complex from the normalized ECG signal, and a step (S104) in which a density-based spatial clustering unit (DBSCAN) (804) applying noise clusters the positive output of the U-Net model (802) to detect a QRS complex.

In the above step S104, the density-based spatial clustering unit (804) to which the noise is applied can detect the QRS by clustering the positive output of the U-Net model into multiple clusters and then determining the midpoint of the multiple clusters as the location of the R-peak.

Referring to FIG. 3, the U-Net model (802) may include a coding path unit (300) for repeatedly performing convolution and Max-pooling on the normalized electrocardiogram signal to obtain a feature map of the normalized electrocardiogram signal, and a decoding path unit (302) for repeatedly performing upsampling and convolution on the feature map output from the coding path unit (300) to generate a positive output indicating a predicted location of the QRS complex.

As illustrated in FIG. 1, a QRS detection method of an electrocardiogram signal for cardiac diagnosis according to an embodiment of the present invention consists of three steps: normalization (S100), U-Net (S102), and DBSCAN (S104). FIG. 1 shows a flowchart of a QRS detection method including all three steps and their interactions. In the preprocessing step, only the normalization step is required. Then, the normalized ECG signal is processed using the trained U-Net. In the decision step, the DBSCAN method is adopted to automatically localize the QRS complex and detect the R-peak.

Normalization

The raw ECG signal usually has individual differences. In this case, the present invention first normalizes the original ECG signal by using the Z-score as a linear scaling method. Normalization is easy to implement and can avoid the influence of non-standardized data on deep neural networks. The signal is normalized according to the following mathematical formula.

Here, ECG is the raw ECG signal, ECG _mean is the mean of the original ECG signal, and ECG _std is the standard deviation of the original ECG signal. In addition, the raw ECG signal is divided into standard-sized segments containing 5000 sample points. No additional noise removal step is included in the preprocessing. Figure 2 shows the normalized ECG signal.

U-Net

The U-Net model includes an encoding path (300) and a decoding path (302) as shown in Fig. 3. The network has three max-pooling layers, three upsampling layers, and 15 convolutions. In order to extract the characteristics of the temporal structure from the ECG signal, 13×1 convolutions were adopted for all convolutions of the U-Net model instead of the 3×3 convolutions of the existing U-Net model.

The main reason is that even a 3x3 area can exhibit a large change in a small area if the resolution of the input image is relatively low. However, since the ECG signal generally has a low frequency and a low sampling rate, it is difficult for the three-point area to form a specific waveform and is susceptible to interference from noise. In the present invention, a relatively large-sized convolution is used to effectively extract the features of the QRS complex.

U-Net has a symmetric structure. The coding path (300) is a repetitive structure. Each repetitive structure (303) includes two convolutions (304) and a max-pooling layer (308). The size of the convolutions (304) is 13×1 and the activation function is ReLu. After two convolutions, there is a 2×1 max-pooling layer (308) with a step size of 2.

After each downsampling, the number of channels is doubled. In addition, unlike the existing U-Net model, padding is adopted to obtain feature maps of the same size to avoid clipping operation when concatenating feature maps.

In the decoding path (302), deconvolution (upsampling) (310) is used at each stage. Using deconvolution reduces the number of feature channels by half and doubles the size of the feature map. After deconvolution, the result is connected to the feature map of the corresponding stage in the coding path (300), and two 13×1 convolutions (304) are performed. In the last convolution layer (312), a 1×1 convolution is used, and a sigmoid function is used as an activation function.

In CNN, as the number of convolutional layers increases, the model representation and feature extraction capabilities become stronger. However, if the model is too complex, a large amount of computation and overfitting may occur. Therefore, to avoid the overfitting problem, dropout layers (306) with ratios of 0.2, 0.5, and 0.2 are added after the third, seventh, and eleventh convolutions, respectively.

Dropout randomly discards nodes of a neural network with a certain probability value to prevent the neural network from overfitting. When a node is randomly selected as a dropout point in an iteration, the output of the node is set to 0 in the forward process of the neural network in this iteration, and the weights and bias terms are not updated in the backward process. Therefore, the node is randomly disabled and does not participate in the training process, so that the joint adaptability of the neural node is weakened and the generalization ability is improved. In the algorithm of the present invention, the output of the U-Net model is thresholded using a preset threshold interval to achieve a positive output.

Density-based spatial clustering of applications with noise (DBSCAN)

To detect QRS complexes, the positive output of the U-Net model must be classified. A simple counting method for classification can lead to localization errors because the positive output of the U-Net can form two rectangular waves for one QRS region when the signal is noisy.

In the algorithm, DBSCAN is selected as the clustering method. DBSCAN is a density-based spatial clustering algorithm. It does not require a preset number of clusters and can find clusters of any size and shape. This feature is suitable for QRS detection because the number of QRS complexes is generally unknown and their lengths are different. Compared with partition-based clustering methods and hierarchical clustering methods, DBSCAN requires less knowledge to determine input parameters and can form clusters when the density of the corresponding area is sufficiently high.

The main process of DBSCAN starts from a random data point that has not been accessed and extracts neighbors of this point with a distance ε. If there are enough points in the neighborhood, the clustering process starts and the current data point becomes the first point of a new cluster. Otherwise, the point is marked as noise.

For the first point in a new cluster, the points in the neighborhood with a distance ε belong to this cluster. This process is repeated until all new points are added to the cluster. After all points in the cluster are determined, new points are extracted and processed. This process is repeated until all points are marked as visited. Finally, all points are accessed and each point is marked as belonging to a cluster or noise.

Figure 4 shows an example of clustering of U-Net positive outputs for QRS detection. After clustering, the midpoint of all clusters is determined as the location of the R-peak.

Experimental Results

The proposed QRS detection algorithm is implemented in Python, and the U-Net model of the algorithm is trained and tested in Keras2.2.4. The experimental computer included CPU i5-6500 3.20GHz, 8GB RAM, GTX 1080Ti GPU, and 16GB graphics memory.

Database

In our experiments, the MIT-BIH Arrhythmia Database (MITDB) was used as a training and test database. This database is one of the internationally recognized ECG databases and is widely used in scientific inventions of ECG signal processing.

The MITDB contains 48 ECG records from 47 subjects divided into two groups. One group contains 23 ECG records randomly selected from the ECG database, and the other group contains ECG records with rare changes that are of great importance in clinical practice. Each record is approximately 30 minutes long. It consists of two leads with a sampling rate of 360 Hz. One lead is MLII and the other is lead V1, V2, V4 or V5. In this experiment, only ECG data with lead MLII are used.

evaluation

To properly evaluate the experimental results, MITDB calculates four evaluation metrics: sensitivity, positive accuracy, accuracy, and F1 score. These metrics are defined as follows:

Here, TP is true positive, FP is false positive, and FN is false negative. Se, PPV, Acc, and F1 represent sensitivity, positive predictive value, accuracy, and F1 score, respectively.

Results and Discussion

For MITDB, we use records 100, 103, 105, 106, 108, 111, 113, 115, 116, 117, 119, 122, 124, 200, 201, 203, 207, 208, 209, 210, 212, 214, 217, 220, 222, 228, 230, 231, and 233 for training, and records 101, 107, 109, 112, 114, 118, 121, 123, 202, 205, 213, 215, 219, 223, 232, and 234 for testing. The U-Net model is first trained by the training set.

Figure 5 shows the steps of the testing process. First, the original ECG signal is normalized, and then the trained U-Net model is used to classify the normalized signal. The U-Net output is binarized with a threshold of 0.9 taken from the manual setting. Then the binary signal is *-* classified by DBSCAN. Finally, the midpoint of the cluster is calculated to achieve R-peak.

The detected R-peaks are considered as TPs when they are located at a distance of ±75 ms from the actual location of the R-peaks (27 points). The QRS detection results for all training and test sets are shown in Tables 1 and 2.

The average sensitivity, positive predictive value, accuracy and F1 score for the test set are up to 99.98%, 99.95%, 99.93% and 99.97%, respectively. The accuracy values of three records in the test set are less than 99.90%. The main reasons are as follows.

(1) The records are segmented into a standard size of 5000 samples in the preprocessing step, but the location of the R-peak is not considered in the segmentation, which leads to poor detection when the R-peak is on the boundary of the segment.

(2) Some features of arrhythmia are not detected by the proposed method's R-peak due to the lack of training sets.

(3) QRS detection may be incorrect due to noise and waveform distortion in the recording.

Table 3 compares the proposed method with the existing MITDB-based method. The bold numbers in Table 3 represent the best results for the evaluation metrics. The proposed algorithm of the present invention achieves similar detection results compared to the other 11 methods. In terms of overall performance, the proposed method is competitive and outperforms the existing method in terms of accuracy and F1 score. Some of the methods in Table 3 are tested on all records in MITDB, and the results of the proposed method are on the test set selected from MITDB.

MITDB records contain five typical ECG signals: normal sinus rhythm (NSR), left bundle branch block beat (LBBB), right bundle branch block beat (RBBB), premature ventricular contractions (PVC), and atrial premature beats (APB).

Fig. 6 shows the QRS detection results for these five ECG signals. The method according to the present invention accurately detects QRS complexes in these five types of ECG signals. The R-peak detected within the gray area in Fig. 6 is regarded as TP. And the gray width area is 150 ms.

To test the robustness of the algorithm of the present invention, three records with serious noise or changes are selected from MITDB as shown in Fig. 7. The selected signals are taken from records 113 and 209, respectively. Record 113 has a high peak of T-wave, and record 209 has severe noise.

Note that the signal in Fig. 6b has a baseline drift problem. From Fig. 6b and Fig. 7, it can be seen that the method according to the present invention is stable for QRS detection of ECG signals with baseline drift, high peak T wave and severe noise. In addition, time-consuming tests are performed. The average processing time for the records of the test set is 16.42 minutes, and the average processing time for a segment of 5000 samples is 1.23 seconds.

The main idea of the method by the present invention is to automatically extract ECG signal features using the U-Net model, and then detect the location of the QRS complex by DBSCAN. Existing QRS detection algorithms generally require additional preprocessing steps for noise reduction and the characteristics and parameters of the QRS complex need to be preset, so they have poor adaptability to various types of signals and noises. The U-Net-based method proposed in the present invention has fewer preprocessing steps and higher detection accuracy, making it suitable for QRS detection in various ECG signals.

In this invention, we propose a method of merging U-Net model and DBSCAN and demonstrate the feasibility of this method for QRS detection in ECG signals. Unlike the existing methods, the U-Net model can automatically extract features with several preprocessing steps and manual settings. After U-Net prediction, DBSCAN is suitable for clustering to find R-peaks without knowing the number and length of QRS complexes. The experiments show that this method can accurately and automatically detect QRS complexes. The average sensitivity, positive predictive value, accuracy and F1 score of the proposed method in MITDB are up to 99.98%, 99.95%, 99.93% and 99.97%, respectively. The proposed approach yields similar results compared to other existing methods. In terms of overall performance, the method by the present invention obtains much better results in terms of accuracy and F1 score. The experimental results show that the method by the present invention is feasible for QRS detection in ECG signals.

Although the present invention has been described in detail through specific examples, this is intended to specifically explain the present invention, and the present invention is not limited thereto, and it will be apparent that modifications or improvements can be made by those skilled in the art within the technical spirit of the present invention.

All simple modifications or changes of the present invention fall within the scope of the present invention, and the specific protection scope of the present invention will be made clear by the appended claims.

[References]

[1] Arzeno NM, Deng Z, Poon C (2008) Analysis of first-derivative based QRS detection algorithms. IEEE Trans Biomed Eng 55(2):478-484. https://doi.org/10.1109/TBME.2007.912658

[2] Benitez DS, Gaydecki PA, Zaidi A, Fitzpatrick AP (2000) A new QRS detection algorithm based on the Hilbert transform. In: Computers in Cardiology. vol 27 (Cat. 00CH37163), 24-27 Sept (2000 2000), pp379-382. https://doi.org/10.1109/CIC.2000.898536

[3] Cai W, Hu D (2020) QRS complex detection using novel deep learning neural networks. IEEE Access 8: 97082-97089. https://doi.org/10.1109/access.2020.2997473

[4] Ester M, Kriegel H-P, Sander J, Xu XA (1996) Density-based algorithm for discovering clusters in large spatial databases with noise. In: The 2nd International Conference on Knowledge Discovery and Data Mining, Portland, WA, pp 226-231

[5] Farashi S (2016) A multiresolution time-dependent entropy method for QRS complex detection. Biomed Signal Process Control 24:63-71. https://doi.org/10.1016/j.bspc.2015.09.008

[6] Goldberger AL, Amaral LAN, Glass L, Hausdorff JM, Ivanov PC, Mark RG, Mietus JE, Moody GB, Peng CK, Stanley HE (2000) PhysioBank, PhysioToolkit, and PhysioNet - Components of a new research resource for complexity. physiologic signals. Circulation 101(23):E215-E220. https://doi.org/10.1161/01.CIR.101.23.e215

[7] Hamdi S, Ben Abdallah A, Bedoui MH (2017) Real time QRS complex detection using DFA and regular grammar. Biomed Eng Online 16:1-20. https://doi.org/10.1186/s12938-017-0322-2

[8] Hannun AY, Rajpurkar P, Haghpanahi M, Tison GH, Bourn C, Turakhia MP, Ng AY (2019) Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network. Nat Med 25(1):65-65+. https://doi.org/10.1038/s41591-018-0268-3

[9] Jia M, Li F, Wu J, Chen Z, Pu Y (2020) Robust QRS detection using high-resolution wavelet packet decomposition and time-attention convolutional neural network. IEEE Access 8:16979-16988. https://doi.org/10.1109/access.2020.2967775

[10] Kiranyaz S, Ince T, Gabbouj M (2016) Real-time patient-specific ECG classification by 1-D convolutional neural networks. IEEE Trans Biomed Eng 63(3):664-675. https://doi.org/10.1109/tbme.2015.2468589

[11] Labati RD, Munoz E, Piuri V, Sassi R, Scotti F (2019) Deep-ECG: Convolutional Neural Networks for ECG biometric recognition. Pattern Recogn Lett 126:78-85. https://doi.org/10.1016/j.patrec.2018.03.028

[12] Lee JS, Lee SJ, ChoiM, Seo M, Kim SW (2019) QRS detection method based on fully convolutional networks for capacitive electrocardiogram. Expert Syst Appl 134:66-78. https://doi.org/10.1016/j.eswa.2019.05.033

[13] Manikandan MS, Soman KP (2012) A novel method for detecting R-peaks in electrocardiogram (ECG) signal. Biomed Signal Process Control 7(2):118-128. https://doi.org/10.1016/j.bspc.2011.03.004

[14] Merah M, Abdelmalik TA, Larbi BH (2015) R-peaks detection based on stationary wavelet transform. Comput Methods Prog Biomed 121(3):149-160. https://doi.org/10.1016/j.cmpb.2015.06.003

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

[16] Nakai Y, Izumi S, Nakano M, Yamashita K, Fujii T, Kawaguchi H, Yoshimoto M (2014) Noise tolerant QRS detection using template matching with short-term autocorrelation. In: 36th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, 26-30 Aug. 2014 2014. pp 34-37. https://doi.org/10.1109/EMBC.2014.6943522

[17] Nayak C, Saha SK, Kar R, Mandal D (2019) An optimally designed digital differentiator based preprocessor for R-peak detection in electrocardiogram signal. Biomed Signal Process Control 49:440-464. https://doi.org/10.1016/j.bspc.2018.09.005

[18] Oh SL, Ng EYK, Tan RS, Acharya UR (2018) Automated diagnosis of arrhythmia using combination of CNN and LSTM techniques with variable length heart beats. Comput Biol Med 102:278-287. https://doi.org/10.1016/j.compbiomed.2018.06.002

[19] Pan J, Tompkins WJ (1985) A real-time QRS detection algorithm. IEEE Trans Biomed Eng BME 32(3): 230-236. https://doi.org/10.1109/TBME.1985.325532

[20] Peimankar A, Puthusserypady S (2021) DENS-ECG: A deep learning approach for ECG signal delineation. Expert Syst Appl 165. https://doi.org/10.1016/j.eswa.2020.113911

[21] Phukpattaranont P (2015) QRS detection algorithm based on the quadratic filter. Expert Syst Appl 42(11): 4867-4877. https://doi.org/10.1016/j.eswa.2015.02.012

[22] Qin Q, Li J, Yue Y, Liu C (2017) An adaptive and time-efficient ECG R-peak detection algorithm. J Healthc Eng 2017:1-14. https://doi.org/10.1155/2017/5980541

[23] Rakshit M, Das S (2017) An efficient wavelet-based automated R-peaks detection method using Hilbert transform. Biocybern Biomed Eng 37(3):566-577. https://doi.org/10.1016/j.bbe.2017.02.002

[24] Ronneberger O, Fischer P (2015) Brox TU-Net: Convolutional Networks for Biomedical Image Segmentation. In: Navab N, Hornegger J, Wells WM, Frangi AF (eds) Medical image computing and computer-assisted intervention - MICCAI 2015. Springer International Publishing, Cham, pp 234-241.

[25] Sarlija M, Jurisic F, Popovic S (2017) A convolutional neural network based approach to QRS detection. In: Kovacic S, Loncaric S, Kristan M, Struc V, Vucic M (eds) Proceedings of the 10th International Symposium on Image and Signal Processing and Analysis. International Symposium on Image and Signal Processing and Analysis, pp 121-125

[26] Sharma T, Sharma KK (2017) QRS complex detection in ECG signals using locally adaptive weighted total variation denoising. Comput Biol Med 87:187-199. https://doi.org/10.1016/j.compbiomed.2017.05.027

[27] Wang X, Zou QQRS (2019) Detection in ECG signal based on residual network. In: 2019 IEEE 11th International Conference on Communication Software and Networks (ICCSN), 12-15 June 2019, pp 73-77. https://doi.org/10.1109/ICCSN.2019.8905308

[28] Xiang Y, Lin Z, Meng J (2018) Automatic QRS complex detection using two-level convolutional neural network. Biomed Eng Online 17:1-17. https://doi.org/10.1186/s12938-018-0441-4

[29] Xu X, Liu Y (2004) ECG QRS complex detection using slope vector waveform (SVW) algorithm. Conference proceedings: Annual International Conference of the IEEE Engineering in Medicine and Biology Society IEEE Engineering in Medicine and Biology Society Annual Conference 2004, pp. 3597-3600

[30] Yang H, Huang M, Cai Z, Yao Y, Liu CA, Faster R (2019) CNN-based real-time QRS detector. In: 2019 Computing in cardiology (CinC), 8-11 Sept. 2019. pp 4. https://doi.org/10.23919/CinC49843.2019.9005798

[31] Zhou Y, Hu X, Tang Z, Ahn AC (2016) Sparse representation-based ECG signal enhancement and QRS detection. Physiol Meas 37(12):2093-2110. https://doi.org/10.1088/0967-3334/37/12/2093

300: Coding path section 302: Decoding path section
800: Normalization part 802: U-Net model
804: Density-based spatial clustering with noise

Claims (10)

(A) A normalization step in which the input original electrocardiogram (ECG) signal is normalized using the Z-score method;
(B) a step of the U-Net model processing and generating a positive output indicating the predicted location of the QRS complex from the normalized electrocardiogram signal based on a threshold; and
(C) a density-based spatial clustering unit (DBSCAN) applying noise includes a step of clustering positive outputs of the U-Net model to detect QRS complexes,
In the above step (C), the density-based spatial clustering unit applying the noise detects the QRS complex by clustering the positive output of the U-Net model into multiple clusters and then determining the midpoint of the multiple clusters as the location of the R-peak.
The above U-Net model is,
A coding path section that repeatedly performs convolution and Max-pooling on the normalized ECG signal to obtain a feature map of the normalized ECG signal; and
A method for detecting QRS of an electrocardiogram signal for cardiac diagnosis, comprising a decoding path section that repeatedly performs upsampling and convolution on a feature map output from the above coding path section to generate a positive output indicating a predicted location of the QRS complex.
delete delete In claim 1,
A method for detecting QRS of an electrocardiogram signal for cardiac diagnosis, wherein the above convolution includes a 13×1 convolution. In claim 1,
The above U-Net model is a method for detecting QRS of an electrocardiogram signal for cardiac diagnosis, including a dropout layer to avoid overfitting. A normalization unit that normalizes the input original electrocardiogram (ECG) signal using the Z-score method;
A U-Net model that generates a positive output indicating the predicted location of the QRS complex from the normalized electrocardiogram signal; and
It includes a density-based spatial clustering unit (DBSCAN) that applies noise to detect QRS complexes by clustering the positive output of the above U-Net model,
The density-based spatial clustering unit applying the above noise processes the positive output of the U-Net model based on a threshold value and clusters it into multiple clusters, and then detects the QRS complex by determining the midpoint of the multiple clusters as the location of the R-peak.
The above U-Net model is,
A coding path section for repeatedly performing convolution and Max-pooling on the normalized ECG signal to obtain a feature map of the normalized ECG signal; and
A QRS detection device of an electrocardiogram signal for cardiac diagnosis, comprising a decoding path section that repeatedly performs upsampling and convolution on a feature map output from the above coding path section to generate a positive output indicating a predicted location of the QRS complex. delete delete In claim 6,
A QRS detection device of an electrocardiogram signal for cardiac diagnosis, wherein the above convolution includes a 13×1 convolution. In claim 6,
The above U-Net model is a QRS detection device of an electrocardiogram signal for cardiac diagnosis, which includes a dropout layer to avoid overfitting.

Images