ECG signal labeling method, software tool and system based on key points and quality label assistance

Abstract

The application provides an ECG signal labeling method, a software tool and a system based on key points and quality label assistance. The method comprises: standardizing an ECG signal to form a data set S0 to be labeled; performing data filtering and noise reduction preprocessing on S0; calculating key points (R-Peak, P-Peak and T-Peak) of the ECG signal; training a quality evaluation model and calculating a quality label of the ECG signal based on the model; building a visual labeling platform, loading the key points and the quality label calculated in the previous step to the labeling platform as auxiliary prompts for manual labeling; manually labeling R-Peak, P-Peak and T-Peak, noise levels and arrhythmia characteristics; and saving the labeling results in a standardized format. The scheme of the application can be applied to quality evaluation of ECG signals collected in a complex noise environment, and has important significance for research in the direction of ECG signal processing, medical signal intelligent analysis algorithm and the like by using a visual labeling tool and auxiliary labeling information to assist in building an ECG related data set in a man-machine interactive strategy.

Description

Technical Field

[0001] This invention relates to the fields of medical signal processing and software engineering technology, specifically to ECG signal annotation methods, software tools, and systems based on key points and quality labels. Background Technology

[0002] ECG signals are an important medical tool for detecting cardiac electrical activity, commonly used to diagnose heart disease, monitor patient health, and study cardiac electrophysiology. However, for certain tasks, ECG signals may contain a large amount of redundant information, making it necessary to locate key points such as the R-Peak. Furthermore, the quality of ECG signals can be affected by various interference factors, such as muscle movement, electrode detachment, and noise. In recent years, deep learning methods have been widely applied to ECG signal processing, especially convolutional neural networks, which can automatically learn features from raw ECG signals and be used to determine signal quality.

[0003] However, previously available ECG datasets with keypoint and quality annotations are scarce. This is largely due to the excessive complexity of manually annotating keypoints and quality, making it difficult to construct a large ECG dataset with these annotations. This invention proposes an ECG signal annotation method, software tools, and system based on keypoint and quality labels. This method can be applied to the quality assessment of ECG signals acquired in complex noisy environments. Furthermore, the visual annotation tool with auxiliary references greatly assists in the construction of ECG-related datasets, ECG signal analysis, and research on intelligent medical signal analysis algorithms. Summary of the Invention

[0004] This invention provides a method, software tool, and system for ECG signal annotation based on key points and quality labels, characterized in that the method includes:

[0005] Step 1: Standardize the ECG signal to form the dataset S0 to be labeled;

[0006] Step 2: Perform data filtering and noise reduction preprocessing on S0;

[0007] Step 3: Calculate the key points of the ECG signal (R-Peak, P-Peak, and T-Peak);

[0008] Step 4: Train the quality assessment model and calculate the quality label of the ECG signal based on the model;

[0009] Step 5: Build a visual annotation platform and load the key points and quality labels calculated in Steps 3 and 4 into the annotation platform as auxiliary prompts for manual annotation.

[0010] Step 6: Human-computer interaction to label R-Peak, P-Peak, and T-Peak, noise level, and arrhythmia characteristics;

[0011] Step 7: Save the annotation results in a standardized format.

[0012] Specifically, in step 1, the signal sequence s with an arbitrary sampling rate is first... i Resampled into a signal sequence s′ with sampling rate f0 i Then, it is divided into unit sequences s' in units of t0 seconds. i =[s′ i1 ,s′ i2 ,…,s′ in ],s′ ij Let j = 1, 2, ..., n, and length f0*t0 be given. Then s′ ij Normalization yields Where m ij v ij s′ ij The mean and variance;

[0013] Specifically, in step 1, the standardization of the ECG signal to form the dataset S0 to be labeled includes: standardizing the original ECG signal sequence to obtain an ECG signal segment s with a mean of 0, a variance of 1, a sampling rate of f0, and a length of f0*t0. 11 ,s 12 ,…,s 1n ,s 21 ,s 22 ,…,s 2n ,……,s m1 ,s m2 ,…,s mn ;

[0014] Specifically, in step 2, the data filtering and noise reduction preprocessing of S0 includes: first passing it through a baseline drift filter and then through a high-frequency suppression filter to filter out baseline drift and high-frequency noise in the ECG signal, thereby obtaining the filtered and noise-reduced ECG signal. in σ(*) represents the high-frequency suppression filter and the baseline drift filter, respectively;

[0015] Specifically, in step 2, when setting the baseline drift suppression filter, the suppression frequency is set to 0.56Hz; when setting the high-frequency suppression filter, the suppression frequency is set to 200Hz.

[0016] Specifically, in step 3, calculating the key points (R-Peak, P-Peak, and T-Peak) of the ECG signal includes: using any physiological signal analysis library, such as the Neurokit2 library, to calculate the positions of the three key points R-Peak, P-Peak, and T-Peak of the ECG signal and storing them in different lists, and ECG signal segments. R-Peak is [R ij1 ,R ij2 ,…R ijk P-Peak is saved in the list [P ij1 ,P ij2 ,…P ijl In the list [T], T-Peak is stored. ij1 ,T ij2 ,…T ijq In the context of ], k, l, q are respectively The total number of R-Peaks, P-Peaks, and T-Peaks within the base;

[0017] Specifically, in step 4, training the quality assessment model and calculating the quality label of the ECG signal on this model includes: dividing the ECG signal segments... Inputting into the quality assessment model M yields the quality label Q. ij ∈{0,1}, 0 represents high-quality ECG signal segments, and 1 represents low-quality ECG signal segments. The construction of the quality assessment model uses key point detection task as pre-training, that is, firstly, the model is pre-trained using different public datasets with key point annotations, and then the model is fine-tuned on the dataset with quality annotations with quality assessment as the task. That is, firstly, the key point detection task of two different datasets (e.g., QT[1] and MIT-BIH[2]) is used to pre-train the feature extraction network of the quality assessment task. This can enable the model to obtain prior knowledge of key point recognition in the subsequent quality assessment task, and better assist the model in completing the quality assessment task. Then, the pre-trained feature extraction network is connected to a new quality assessment classifier, and the network is retrained on the dataset with noisy annotations with the quality assessment task to achieve better quality assessment performance.

[0018] Specifically, in step 5, a visual annotation platform is built, and the key points and quality labels calculated in steps 3 and 4 are loaded into the annotation platform as auxiliary prompts for manual annotation. This includes: the visual interface simultaneously loads ECG signal segments that allow human-computer interaction for annotation of key points and ECG signal segments with key point auxiliary prompts. The former can provide mouse click annotation of key point features, while the latter is the same ECG signal as the former. The key point positions obtained in step 3 are used to highlight the key points to assist in the accurate annotation of manual key points.

[0019] Specifically, in step 6, the human-computer interaction labels R-Peak, P-Peak, and T-Peak, noise level, and arrhythmia characteristics, including: the mouse can switch between three modes: R, P, and T. In each mode, clicking on a point in the ECG signal segment will save the number of that point in the array of the corresponding mode (R-Peak, P-Peak, and T-Peak). There are also corresponding human-computer interaction buttons and keyboard shortcuts for noise level and arrhythmia characteristics, which can quickly label the feature tags corresponding to the ECG signal segment and save them in a standardized format.

[0020] Specifically, in step 7, the annotation results are saved in a standardized format, including: unless otherwise specified, the annotation tool will automatically save it in a dictionary format as a JSON file with the same name as the ECG file. The format can be, but is not limited to, the following: {"R":[],"P":[],"T":[],"Noise_Level":0,"A_Exist":false}, where "R", "P", and "T" store the position numbers of the three key points, "Noise_Level" stores the noise level of the ECG signal, and "A_Exist" represents whether there is a heart rate arrhythmia.

[0021] This invention addresses the severe lack of keypoint and quality labels in current ECG datasets. Considering the susceptibility of ECG signals to various noise interferences, it simplifies and improves data annotation efficiency from a signal processing and software engineering perspective. By utilizing convolutional neural networks and pre-trained models of existing public ECG datasets as prior knowledge, it enhances annotation accuracy. This tool not only simplifies the construction of larger-scale ECG datasets but also obtains more accurate feature labels than without auxiliary annotation. This is of great significance for the analysis and research of ECG signals. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0023] Figure 1 A flowchart illustrating the working steps of the ECG signal annotation method, software tools, and system based on key points and quality labels;

[0024] Figure 2 Flowchart of the training process for a quality assessment classification model;

[0025] Figure 3A flowchart illustrating the key point annotation process using a visual annotation tool;

[0026] Figure 4 Examples of human-computer interaction interfaces. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Currently, there are very few publicly available ECG datasets with keypoint and quality annotations, and ECG signals are susceptible to various types of noise degradation. Manually annotating keypoints and quality levels is also quite complex, making the construction of ECG datasets difficult. This invention addresses these problems by leveraging the application of deep learning in signal processing to propose a keypoint and quality label-assisted ECG signal annotation method, software tools, and system. The specific steps are as follows:

[0029] S101: Standardize the ECG signal to form the dataset S0 to be labeled. Specifically, first, the signal sequence s at any sampling rate is... i Resampled into a signal sequence s′ with a sampling rate of f0Hz i Then, it is divided into unit sequences s' in units of t0 seconds. i =[s′ i1 ,s′ i2 ,…,s′ in ],s′ ij Let j = 1, 2, ..., n, and length f0*t0 be given. Then s′ ij Normalization yields Where m ij v ij s′ ij The mean and variance are used to obtain, for example, an ECG signal segment s' with a mean of 0, a variance of 1, a sampling rate of 360Hz, and a length of 10 seconds. 11 ,s' 12 ,…,s' 1n ,s' 21 ,s' 22 ,…,s' 2n ,……,s' m1 ,s' m2 ,…,s' mn Each sample segment is 3600 units long.

[0030] S102: Perform data filtering and noise reduction preprocessing on S0. Specifically, it first passes through a baseline drift filter and then a high-frequency suppression filter to remove baseline drift and high-frequency noise from the ECG signal, resulting in the filtered and noise-reduced ECG signal. in σ(*) represents the high-frequency suppression filter and the baseline drift filter, respectively. When setting the baseline drift suppression filter, the suppression frequency is set to 0.56Hz; when setting the high-frequency suppression filter, the suppression frequency is set to 200Hz.

[0031] S103: Calculate the key points (R-Peak, P-Peak, and T-Peak) of the ECG signal. Specifically, any physiological signal analysis library can be used, such as the Neurokit2 library, to calculate the locations of the three key points (R-Peak, P-Peak, and T-Peak) of the ECG signal and store them in separate lists for each ECG signal segment. R-Peak is [R ij1 ,R ij2 ,…R ijk P-Peak is saved in the list [P ij1 ,P ij2 ,…P ijl In the list [T], T-Peak is stored. ij1 ,T ij2 ,…T ijq In the context of ], k, l, q are respectively The total number of R-Peak, P-Peak and T-Peak.

[0032] S104: Train the quality assessment model and calculate the quality label of the ECG signal based on this model. Specifically, the ECG signal segments... Inputting into the quality assessment model M yields the quality label Q. ij ∈{0,1}, 0 represents high-quality ECG signal segments, and 1 represents low-quality ECG signal segments. The construction of the quality assessment model uses the key point detection task as pre-training. That is, the model is first pre-trained using different public datasets with key point annotations, and then fine-tuned on the quality assessment task on the dataset with quality annotations. That is, the key point detection task of two different datasets (e.g., QT[1] and MIT-BIH[2]) is used to pre-train the feature extraction network of the quality assessment task. This allows the model to obtain prior knowledge of key point recognition in the subsequent quality assessment task, and better assist the model in completing the quality assessment task. Then, the pre-trained feature extraction network is connected to a new quality assessment classifier, and the network is retrained on the dataset with noisy annotations using the quality assessment task to achieve better quality assessment performance.

[0033] S105: Build a visual annotation platform and load the keypoints and quality labels calculated in S103 and S104 into the annotation platform as auxiliary prompts for manual annotation. Specifically, the visual interface simultaneously loads ECG signal segments that allow for human-computer interaction in annotating keypoints and ECG signal segments with keypoint auxiliary prompts. The former allows for mouse clicks to annotate keypoint features, while the latter is the same ECG signal segment as the former, using the keypoint positions obtained in S103 for keypoint highlighting to assist in the accurate annotation of manual keypoints.

[0034] S106: Human-computer interaction annotation of R-Peak, P-Peak, and T-Peak, noise level, and arrhythmia characteristics. Specifically, the mouse can switch between three modes: R, P, and T. Clicking on a point in the ECG signal segment in each mode will save the point's index in the corresponding mode's (R-Peak, P-Peak, and T-Peak) array. There are also corresponding human-computer interaction buttons and keyboard shortcuts for noise level and arrhythmia characteristics, which can quickly label the feature tags corresponding to the ECG signal segment and save them in a standardized format.

[0035] S107: Save the annotation results in a standardized format. Specifically, unless otherwise specified, the annotation tool will automatically save the data in a dictionary format as a JSON file with the same name as the ECG file. The format can be, but is not limited to, the following: {"R":[],"P":[],"T":[],"Noise_Level":0,"A_Exist":false}, where "R", "P", and "T" store the position numbers of the three key points, "Noise_Level" stores the noise level of the ECG signal, and "A_Exist" indicates whether there is a heart rate arrhythmia.

[0036] like Figure 2 The diagram shows the training process for the quality assessment classification model, which specifically includes:

[0037] S201: The feature extraction network of the quality assessment model is pre-trained using a public dataset with keypoint annotations for a keypoint detection task. Specifically, the feature extraction network employs a backbone network with strong keypoint localization capabilities, such as the hourglass network used for pose estimation in computer vision, connected to a four-classifier to predict the category of each point in the signal: R-Peak, P-Peak, T-Peak, and Others. The purpose of pre-training the network using a keypoint detection task is to enable the network to recognize keypoints in ECG signals, which is strongly correlated with the quality assessment task and is a highly effective auxiliary task.

[0038] S202: Use a keypoint detection task on another publicly available dataset with keypoint annotations to pre-train the feature extraction network of the quality evaluation model. Specifically, the feature extraction network is further trained using a keypoint detection task on another dataset. The aim is to enable the network to learn features from datasets with more diverse distributions, making the network's keypoint recognition capabilities more robust.

[0039] S203: Construct a quality assessment dataset S1 by using specific noise data to corrupt the high-quality signal of a portion of the clean ECG dataset. Specifically, four types of noise—baseline drift (bw), electrode movement (em), muscle artifacts (ma), and Gaussian white noise (gs)—are used to corrupt the high-quality signal at a certain signal-to-noise ratio. The noise components are as follows: noise = 0.1 × bw + 0.2143 × em + 0.6143 × ma + 0.714 × gs. When noise is superimposed on the clean ECG signal, the low-quality signal is superimposed at a signal-to-noise ratio of -11 dB. To simulate actual acquisition conditions, the high-quality signal also needs a certain degree of noise corruption, which is superimposed at a signal-to-noise ratio of -2 dB.

[0040] S204: A classifier for training a quality assessment model using the quality assessment task on the S1 dataset. Specifically, during training, the feature extraction network trained on the keypoint detection task is completely frozen and does not participate in gradient backpropagation; only the binary classifier for quality assessment is trained.

[0041] S205: Fine-tuning the classifier of the quality assessment model using a self-collected dataset for a quality assessment task. Specifically, the classifier is fine-tuned using binary cross-entropy loss, while the feature extraction network remains frozen to preserve its ability to identify key points. Training the network sequentially on two different datasets exposes the model to data from more diverse distributions, improving the robustness and accuracy of the quality assessment classification performance.

[0042] like Figure 3 The diagram shown is a flowchart of the key point annotation process using a visual annotation tool, which includes:

[0043] S301: Determine the keypoint annotation type (RPT). Specifically, the visual annotation tool provides three mouse annotation modes: R-Peak, P-Peak, and T-Peak. You can select and switch modes using the three buttons at the top. Taking R-Peak mode as an example, in this mode, the ECG sequence points clicked by the left mouse button will be stored in the R array, which means that R-Peak points have been annotated. P-Peak and T-Peak are the same.

[0044] S302: Get cursor coordinates event, get cursor click event. Specifically, getting cursor coordinates means obtaining the cursor's exact position within the annotation tool interface, and getting cursor click event means determining whether the user has clicked the left mouse button at the current moment.

[0045] S303: Locate all points within a certain range of the cursor coordinates in the ECG signal and highlight them in red. If a cursor click event occurs, save the sequence number of the red highlighted points in the list corresponding to the current mode. Specifically, for the current cursor position coordinates, the tool first checks if there are any ECG signal sequence points within an area with a radius of r centered on the cursor coordinates. If so, it changes the color attribute of these points from blue to red and redraws them, highlighting the ECG signal sequence points near the cursor. For cursor click events, if a click event occurs, the tool first checks if there are any already highlighted sequence points. If so, it stores these points in the list of the current RPT mode, indicating that the user has performed a key point annotation operation. If there are no highlighted points nearby, no operation is performed.

[0046] S304: After the cursor moves away from the highlighted point by a certain distance, the highlighted point automatically reverts to a normal color sequence point. Selecting the deletion mode with the cursor and clicking on a saved key point location will remove that point from the corresponding list. Specifically, if a previously highlighted ECG signal sequence point is no longer near the cursor, its color attribute will be changed from red to blue, and the highlighting will be canceled. Additionally, if some incorrect key points are marked, the deletion mode can be selected above the annotation tool. In this mode, the program will sequentially search the three RPT lists for the clicked point. If it exists, all points will be removed; otherwise, no operation will be performed.

[0047] like Figure 4 The example shown is a human-computer interaction interface, which specifically includes:

[0048] S401: The human-computer interaction interface is divided into three horizontal columns. Specifically, the top column contains the annotation mode selection buttons. Clicking one of the three buttons (RPT, RPT, etc.) will switch the current mouse annotation mode to the corresponding R-Peak, P-Peak, or T-Peak mode. The Del button represents the mouse mode for deleting selected annotation points, allowing deletion of incorrectly annotated key points. Clicking the Quality button in the middle of the top column allows switching the quality characteristic label of the ECG signal segment between High (high quality) and Low (low quality). The text box to the right of the button displays the label of this characteristic in real time, enabling signal quality annotation. Similarly, the Arrhythmia button on the right side of the top column... The Exist button controls the arrhythmia labeling of the signal. Yes indicates the presence of an arrhythmia, and No indicates the absence of an arrhythmia. The middle column visualizes ECG signal segments with key points highlighted; these key points are enlarged and highlighted. As shown in the example, all three types of key points are highlighted. The bottom column represents ECG segments that allow for interactive annotation of key points. The data itself is exactly the same as the middle column. Users can use the highlighted parts in the middle column as supplementary prompts, along with their own judgment, to more accurately label the various features of the ECG signal. After annotating a signal segment, pressing the S key on the keyboard will standardize the annotation results and save them to a specified file.

[0049] The ECG signal annotation method, software tool, and system based on key points and quality labels of this invention can be applied to the quality assessment of ECG signals acquired in complex noisy environments. Furthermore, the visual annotation tool with auxiliary references simplifies the construction of larger-scale ECG-related datasets and yields more accurate feature labels than without auxiliary annotation. This is of great significance for research in ECG signal processing, intelligent medical signal analysis algorithms, and other fields.

Claims

1. An ECG signal labeling method based on key point and quality label assistance, characterized in that, The method comprises the following steps: Step 1, standardize the ECG signal to form a data set S0 to be labeled, specifically: first, the signal sequence s of any sampling rate is converted into a signal sequence s' of a sampling rate f0 i Resample the signal sequence s' into a signal sequence s' of a sampling rate f0 i Then cut into unit sequences s' with t0 seconds as a unit i s' i1 s' i2 s' in ], S' ij j = 1, 2, …, n length f0*t0, normalize S' ij Where m ij v ij are the mean and variance of s' ij , respectively;​ Step 2, data filtering and noise reduction preprocessing is performed on S0, specifically, first passing through a baseline drift filter and then passing through a high-frequency suppression filter, filtering out the baseline drift and high-frequency noise of the ECG signal to obtain the filtered and noise-reduced ECG signal wherein , σ(*) are the high-frequency suppression filter and the baseline drift filter, respectively, and when setting the baseline drift filter, the suppression frequency is recommended but not limited to 0.56 Hz; when setting the high-frequency suppression filter, the suppression frequency is recommended but not limited to 200 Hz; Step 3, calculating the key points R-Peak, P-Peak and T-Peak of the ECG signal; Step 4, training a quality evaluation model and calculating the quality label of the ECG signal based on the model, specifically: inputting the ECG signal segment into the quality evaluation model M to obtain the quality label Q ij ∈{0,1}, 0 represents a high-quality ECG signal segment, and 1 represents a low-quality ECG signal segment; Step 5, building a visual labeling platform, loading the key points and quality labels calculated in steps 3 and 4 into the labeling platform as auxiliary prompts for manual labeling; Step 6, human-computer interaction labeling R-Peak, P-Peak and T-Peak, noise level and arrhythmia features; Step 7, saving the labeling results in a standardized format.

2. The keypoint and quality label aided based ECG signal labeling method of claim 1, wherein, In step 1, the ECG signal is standardized to form a data set S0 to be labeled, specifically: after the original ECG signal sequence is standardized, an ECG signal segment s' of f0*t0 seconds in length is obtained, with a mean of 0, a variance of 1, a sampling rate of f0, and a unit sequence cut into t0 seconds 11 ,S′ 12 ,…,s′ 1n ,s′ 21 ,s′ 22 ,…,s′ 2n ,……,S′ m1 ,S′ m2 ,…,S′ mn .

3. The keypoint and quality label aided ECG signal labeling method of claim 1, wherein, In step 3, the key points R-Peak, P-Peak and T-Peak of the ECG signal are calculated, specifically: using the physiological signal analysis library Neurokit2, the positions of the three types of key points R-Peak, P-Peak and T-Peak of the ECG signal are calculated and saved in different lists respectively, and the R-Peak of the ECG signal segment is [R ij1 ,R ij2 ,…R ijk ], the P-Peak is saved in the list [P ij1 ,P ij2 ,…P ijl ], and the T-Peak is saved in the list T ij1 ,T ij2 ,…T ijq ] wherein k, l, q are respectively the total number of R-Peak, P-Peak and T-Peak within.

4. The keypoint and quality label aided ECG signal labeling method of claim 1, wherein, In step 4, the training quality evaluation model, and based on the model, the quality label of the ECG signal is calculated, specifically: the ECG signal segment is input into the quality evaluation model M to obtain the quality label Q ij ∈{0,1}, 0 represents a high-quality ECG signal segment, and 1 represents a low-quality ECG signal segment, and the construction of the quality evaluation model uses a key point detection task as pre-training, that is, the model is first pre-trained using different public data sets with key point annotations, and then the model is fine-tuned on the self-collected data set with quality evaluation as the task.

5. The keypoint and quality label aided ECG signal labeling method of claim 4, wherein, First, the feature extraction network for the quality evaluation task is pre-trained using the key point detection task of different data sets, which enables the model to obtain prior knowledge of key point recognition in the subsequent quality evaluation task, better assisting in completing the quality evaluation task. Then, a new quality evaluation classifier is connected to the pre-trained feature extraction network, and the network is retrained using the quality evaluation task on the noisy labeled data set to achieve better quality evaluation performance.

6. The keypoint and quality label aided based ECG signal labeling method of claim 1, wherein, In step 5, a visual labeling platform is built, and the key points and quality labels calculated in steps 3 and 4 are loaded into the labeling platform as auxiliary prompts for manual labeling. Specifically, the ECG signal segment for human-computer interaction labeling key points and the ECG signal segment with key point auxiliary prompts are loaded into the visual interface at the same time. The former provides a mouse click labeling key point feature operation, and the latter uses the key point positions obtained in step 3 for key point highlighting to assist accurate manual key point labeling.

7. The keypoint and quality label aided based ECG signal labeling method of claim 1, wherein, In step 6, R-Peak, P-Peak and T-Peak, noise level and arrhythmia features are labeled by human-computer interaction. Specifically, the mouse switches between R, P and T modes. In each mode, clicking a point in the ECG signal segment saves the point number in the corresponding R-Peak, P-Peak and T-Peak arrays. Noise level and arrhythmia features also have corresponding human-computer interaction buttons and keyboard shortcuts for quickly labeling the feature labels of the ECG signal segment, and are saved in a standardized format.

8. The keypoint and quality label aided based ECG signal labeling method of claim 1, wherein, In step 7, if not specified, the labeling tool will automatically save in dictionary format as a json file with the same name as the ECG file. The format is as follows: {"R":[],"P":[],"T":[],"Noise_Level":0,"A_Exist":false}, where "R", "P", "T" store the position numbers of the three key points, "Noise_Level" stores the noise level of the ECG signal, and "A_Exist" represents whether there is arrhythmia.

9. A software tool for implementing the method of any one of claims 1-8.

10. A system for implementing the method of any one of claims 1-8.

Images