Automatic electrocardiogram diagnosis method

Abstract

The present invention discloses an automatic electrocardiogram diagnosis method, which includes the following steps: first, reading a plurality of heart disease data, and obtaining a specific electrocardiogram measurement record of a corresponding patient based on the heart disease data, and excluding a non-specific electrocardiogram measurement record of the corresponding patient. Using machine learning, according to the heart disease data, corresponding feature data is identified in the specific electrocardiogram measurement record of the corresponding patient. Then, the electrocardiogram measurement record to be identified is read, and the feature data is used to determine whether the electrocardiogram measurement record to be identified has heart disease, thereby effectively improving the accuracy of heart disease judgment.

Description

Technical Field

[0001] The present invention relates to an electrocardiogram diagnosis method, and in particular to an automatic electrocardiogram diagnosis method. Background Art

[0002] With the advancement of science and technology and medical technology, human life expectancy is gradually increasing. At present, my country has gradually entered an aging society, the proportion of the elderly population is rising year by year, and heart disease has become one of the top invisible killers of the elderly.

[0003] Arrhythmia is a common symptom of heart-related diseases, so monitoring heart-related diseases is an important issue for the elderly. Currently, the detection of heart rate / ECG is mainly monitored by ECG measuring equipment, which is then interpreted by doctors.

[0004] However, since everyone is born with different physical constitutions, relying on traditional ECG comparison methods to interpret ECGs and use them as a basis for disease diagnosis often leads to inappropriate judgments.

[0005] In addition, the traditional way of using big data analysis to interpret the user's electrocardiogram not only requires a huge amount of data, but also often causes errors in interpretation due to data differences, thereby reducing the accuracy of electrocardiogram interpretation. The electrocardiogram judgment results obtained by the computer are also presented in a different way from the doctor's judgment results, which also causes trouble for doctors and patients. Summary of the invention

[0006] In view of this, the present invention discloses an automatic electrocardiogram diagnosis method, which can reduce the time and software and hardware resources required for learning, and thus quickly and correctly perform electrocardiogram disease marking and diagnosis, thereby improving medical quality.

[0007] According to an embodiment disclosed in the present invention, there is an automatic electrocardiogram diagnosis method, which includes the following steps: first, a plurality of heart disease data are read, and based on the heart disease data, specific electrocardiogram measurement records of the corresponding patient are obtained, and non-specific electrocardiogram measurement records of the corresponding patient are excluded. By using machine learning, corresponding feature data is identified in the specific electrocardiogram measurement records of the corresponding patient based on the heart disease data. Then, the electrocardiogram measurement record to be identified is read, and the feature data is used to determine whether the electrocardiogram measurement record to be identified has heart disease.

[0008] In some embodiments, the aforementioned exclusion of non-specific ECG measurement records corresponding to the patient is based on big data analysis to exclude ECG measurement records of leads with low correlation with heart disease data.

[0009] In some embodiments, the aforementioned exclusion of ECG measurement records of leads with low correlation with heart disease data based on big data analysis means recording ECG measurement records of leads with a recognition accuracy rate lower than 50% as non-specific ECG measurement records.

[0010] In some embodiments, the aforementioned exclusion of ECG measurement records of leads with low relevance to heart disease data based on big data analysis further includes recording ECG measurement records of leads with an identification accuracy rate higher than 90% as specific ECG measurement records.

[0011] In some embodiments, the feature data is used to determine whether the electrocardiogram measurement record to be identified has a heart disease, and the determination of whether the electrocardiogram measurement record to be identified has a heart disease is made based only on a specific electrocardiogram measurement record of the electrocardiogram measurement record to be identified.

[0012] In some embodiments, using machine learning further includes marking corresponding feature data, wherein marking the corresponding feature data is to cut the specific electrocardiogram measurement record according to the time coordinate of the maximum absolute value of each heartbeat voltage.

[0013] In some embodiments, the characteristic data corresponding to the mark also includes taking a specific electrocardiogram measurement record, with the cutting position where the absolute value of the heartbeat voltage corresponding to the heart disease data is the maximum as the origin, taking two-thirds of the previous cutting position, and taking two-thirds of the next cutting position as the marking interval of the corresponding characteristic data.

[0014] In some embodiments, the characteristic data corresponding to the mark further includes, when the mark interval is less than 2.5 seconds, adding an average time forward and backward to make the mark interval equal to 2.5 seconds.

[0015] In some embodiments, the characteristic data corresponding to the mark also includes randomly multiplying the time coordinate of the mark interval by any value from 0.75 to 1.25 to generate auxiliary characteristic data, and when the mark interval of the auxiliary characteristic data is less than 2.5 seconds, the subsequent data of the auxiliary characteristic data is padded with zeros to 2.5 seconds.

[0016] In some embodiments, the corresponding feature data is identified by using a shallow neural network model to identify the corresponding feature data.

[0017] In some embodiments, the automatic electrocardiogram diagnosis method further comprises outputting the heart disease to be identified by the electrocardiogram measurement records according to the heartbeat, rhythm and pattern classification.

[0018] In some embodiments, the automatic electrocardiogram diagnosis method further includes utilizing multi-label learning to simultaneously learn and judge multiple heart diseases.

[0019] In some embodiments, the automatic electrocardiogram diagnosis method further includes outputting a plurality of heart diseases and corresponding probabilities.

[0020] Therefore, the automatic ECG diagnosis method disclosed in the present invention can improve the accuracy of learning and judgment, reduce the number of ECG measurement records required, and reduce the software and hardware requirements for computing data, saving a lot of resources such as electricity, making it easier for doctors and patients to quickly and correctly understand heart disease and obtain relevant advice, thereby effectively improving the quality of medical care. In addition, through multi-label learning, multiple heart diseases and their corresponding probabilities can be output for reference by doctors and patients, effectively improving the accuracy of doctors' disease judgments and increasing doctors' willingness to use the method. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to make the above and other purposes, features, advantages and embodiments of the present invention more clearly understood, the following description is made in conjunction with the accompanying drawings:

[0022] Figure 1 It is a schematic diagram of an automatic electrocardiogram diagnosis method according to an embodiment of the present invention.

[0023] Figure 2 is in accordance with Figure 1 Schematic diagram of the steps of identifying feature data in the automatic electrocardiogram diagnosis method.

[0024] Figure 3 is a schematic diagram of a specific electrocardiogram measurement record.

[0025] Figure 4 It is a schematic diagram of marking feature data.

[0026] Description of main reference numerals:

[0027] 100-Automatic ECG Diagnosis Method, 110-160-Steps, 210-270-Steps, 300-Specific ECG Measurement Record, 310-Lead V1 ECG Waveform, 320-Positive S Valley Waveform, 330-Lead II ECG Waveform, 401-Block, 410-Cutting Line, 420-Origin, 430-X Point, 440-Z Point, 450-Marking Interval, 451-Front Interval, 452-Back Interval. DETAILED DESCRIPTION

[0028] The following is a detailed description of the embodiments with the accompanying drawings, but the embodiments provided are not intended to limit the scope of the present invention, and the description of the structural operation is not intended to limit the order of its execution. Any device with equal functions produced by the re-combination of components is within the scope of the present invention. In addition, the drawings are for illustration purposes only and are not drawn according to the original size. For ease of understanding, the same components or similar components in the following description will be described with the same symbols.

[0029] In addition, the terms used throughout the specification and the scope of the patent application generally have the ordinary meaning of each term used in this field, in the content disclosed herein and in the specific content, unless otherwise specified. Certain terms used to describe the present invention will be discussed below or elsewhere in this specification to provide additional guidance to those skilled in the art in the description of the present invention.

[0030] The terms “first”, “second”, etc. used in this document do not specifically refer to an order or sequence, nor are they used to limit the present invention. They are merely used to distinguish components or operations described with the same technical terms.

[0031] Secondly, the words "include", "including", "have", "contain" and the like used in this article are all open terms, which mean including but not limited to.

[0032] Figure 1 is a schematic diagram of an automatic electrocardiogram diagnosis method according to an embodiment of the present invention, and Figure 2 is in accordance with Figure 1 Schematic diagram of the steps of identifying feature data of the automatic electrocardiogram diagnosis method, Figure 3 is a schematic diagram of a specific electrocardiogram measurement record. Figure 4 It is a schematic diagram of marking feature data.

[0033] like Figure 1 As shown, an automatic electrocardiogram diagnosis method 100 includes the following steps: first, in step 110, heart disease data is read, and then, in step 120, specific electrocardiogram measurement records of the corresponding patient are obtained according to the heart disease data, and non-specific electrocardiogram measurement records of the corresponding patient are excluded. The specific electrocardiogram measurement records refer to electrocardiogram data of multiple leads that are relevant to the heart disease patient.

[0034] See also Figure 3, which is a schematic diagram of a specific electrocardiogram measurement record 300. When the heart disease is right bundle branch block (RBBB), the electrocardiogram data of the 12 leads measured simultaneously will only produce obvious waveform features in the electrocardiogram waveforms of leads V1 and V6, such as a waveform 320 of a positive S Valley in the electrocardiogram waveform 310 of lead V1. However, the electrocardiogram data of other leads, such as the electrocardiogram waveform 330 of lead II, are not obvious or have no corresponding changes. In this case, when the subsequent mechanical learning process and the judgment of whether there is such a heart disease are performed in a traditional manner, they are often affected by the electrocardiogram data of other leads, thereby reducing the accuracy of learning and judgment.

[0035] Therefore, the automatic ECG diagnosis method 100 can exclude such ECG data first, so that it is recorded as a non-specific ECG measurement record. Not only is it excluded before learning, but also when the ECG measurement record to be identified is subsequently judged, such ECG data is also excluded. In this way, not only the accuracy of learning can be effectively improved, but also the accuracy of judgment can be improved. What is more noteworthy is that, since traditional machine learning needs to read ECG data of 12 leads, this automatic ECG diagnosis method 100 only needs to read the patient's specific ECG measurement record, thus greatly reducing the amount of ECG data required for the learning process.

[0036] Taking right bundle branch block (RBBB) as an example, it is only necessary to judge the ECG data of two leads, namely the ECG data of lead V1 and lead V6, to effectively learn the feature data. In subsequent judgments, when judging whether a heart disease such as right bundle branch block (RBBB) is present, it is only necessary to judge the ECG data of two leads and exclude interference from the ECG data of other leads. This not only speeds up the judgment efficiency, but also improves the accuracy of the judgment, and can also improve the accuracy of automatic ECG diagnosis.

[0037] In some embodiments, specific electrocardiogram measurement records of corresponding patients are obtained based on heart disease data, and non-specific electrocardiogram measurement records of corresponding patients are excluded, and screening can be performed through settings of a doctor.

[0038] In some embodiments, specific ECG measurement records of the corresponding patient are obtained based on heart disease data, and non-specific ECG measurement records of the corresponding patient are excluded. Big data analysis can be used to set ECG data of certain heart diseases as specific ECG measurement records, and other ECG data can be set as non-specific ECG measurement records and excluded.

[0039] In some embodiments, based on big data analysis, ECG measurement records of leads with low relevance to heart disease data are excluded, that is, ECG measurement records of leads with an identification accuracy lower than 50% are recorded as non-specific ECG measurement records, and ECG measurement records of leads with an identification accuracy higher than 90% can also be recorded as specific ECG measurement records.

[0040] Next, in step 130, machine learning is used to identify corresponding characteristic data in the specific electrocardiogram measurement records of the corresponding patient based on the heart disease data. Steps 110 to 130 are steps for learning characteristic data, which can be performed continuously and repeatedly to improve the accuracy of the judgment.

[0041] See also Figure 2 and Figure 4 As shown in the figure, the recognition feature data of step 130 using machine learning also includes the feature data corresponding to the label. First, step 210 cuts the specific electrocardiogram record, which is to cut the specific electrocardiogram measurement record according to the time coordinate of the maximum absolute value of each heartbeat voltage.

[0042] See also Figure 4 , the characteristic data is marked with the premature ventricular contraction (PVC) recorded by the specific electrocardiogram measurement. The electrocardiogram is divided into a plurality of blocks 401 with the cutting line 410 as the boundary. Step 220, the cutting line 410 of the heartbeat voltage corresponding to the heart disease data is used as the origin 420. Next, step 230, two-thirds of the distance from the previous cutting line 410, such as the X point 430 in the figure, to the origin 420 is used as the starting point of the marking interval 450, and another step 240, two-thirds of the distance from the next cutting line 410, such as the Z point 440 in the figure, to the origin 420 is used as the end point of the marking interval 450, so as to form the marking interval 450 surrounded by the dotted line in the middle of the figure.

[0043] In some embodiments, when the length of the marking interval 450 is less than 2.5 seconds, a time is added on average to expand the time length of the marking interval 450 to 2.5 seconds. Referring to step 250, a front interval 451 and a rear interval 452 are added to increase the time of the marking interval 450 to 2.5 seconds. Therefore, the marking interval 450 can effectively cover the data and time of the entire feature data, effectively avoiding the distortion or fragmentation of the feature data caused by the traditional method of cutting with standard length.

[0044] In some embodiments, identifying feature data may further include step 260, generating auxiliary feature data, which may use random numbers to generate auxiliary feature data, thereby increasing effective data of the machine learning process.

[0045] For example, it can compress or expand the time coordinate of the original data to any value between 0.75 and 1.25 of the original value, such as the original time coordinate*(1+(random value (0-1)-0.5) / 2) to compress or expand the time coordinate of the original data, thereby generating auxiliary feature data, increasing the effective samples of machine learning, and improving the accuracy of recognition.

[0046] In some embodiments, when the marking interval of the auxiliary feature data is less than 2.5 seconds, the subsequent data of the auxiliary feature data are padded with zeros to 2.5 seconds.

[0047] In some embodiments, when the marking interval of the auxiliary feature data is greater than 2.5 seconds, the data of the first 2.5 seconds is taken as the auxiliary feature data.

[0048] In addition, step 270 uses a shallow neural network model to analyze the aforementioned specific electrocardiogram record to improve the accuracy of recognition. Since the automatic electrocardiogram diagnosis method 100 uses specific electrocardiogram records for learning and recognition, the complexity of learning can be simplified, and the accuracy of recognition can be effectively improved without using deep learning. In addition, since the automatic electrocardiogram diagnosis method 100 uses specific electrocardiogram records for learning and recognition, it can achieve high-precision recognition capabilities with a small amount of calculations, effectively reduce the requirements for software and hardware required for calculations, and save a lot of resources such as electricity.

[0049] Next, see Figure 1 , step 140, reading the electrocardiogram measurement record to be identified, and, step 150, using the characteristic data, determining whether the electrocardiogram measurement record to be identified has a heart disease. When making a determination, the automatic electrocardiogram diagnosis method 100 further uses specific electrocardiogram measurement records to make a determination when a possible heart disease is identified, and excludes non-specific electrocardiogram measurement records, so as to further improve the accuracy of the heart disease determination, without being affected by the non-specific electrocardiogram measurement records and reducing the accuracy of the determination.

[0050] Then, step 160 outputs the diagnosis result, which includes outputting the heart disease data of the electrocardiogram measurement record to be identified in a hierarchical classification manner in terms of heartbeat, rhythm and pattern. Among them, the heartbeat classification can include results such as normal heartbeat, tachycardia or bradycardia. The classification of rhythm includes results such as normal sinus rhythm, sinus arrhythmia, atrial rhythm, atrial fibrillation, sinus bradycardia, junctional rhythm, ventricular rhythm, ventricular tachycardia or pacemaker pacing rhythm. The morphological classification includes heart diseases such as ST change, T-wave inversion, right bundle branch block (RBBB), left bundle branch block (LBBB), myocardial infarction, left ventricular hypertrophy, etc.

[0051] In some embodiments, the automatic ECG diagnosis method 100 can also output the probability of occurrence of each heart disease through multi-label learning, and can adjust the probability threshold as needed to control the specificity and sensitivity of the heart disease, and finally output the possible rhythm and type category of the ECG based on the probability distribution.

[0052] For example, through multi-label learning, multiple heart diseases can be learned and judged at the same time, and by adjusting the probability threshold, multiple possible heart diseases and their corresponding probabilities can be output for reference by doctors and patients. For example, when the probability threshold is set at 75%, when the electrocardiogram data is judged to have a heart rhythm disease, such as sinus arrhythmia and sinus bradycardia, the probability of which is greater than 75%, the output diagnosis result further outputs the probability of the patient having sinus arrhythmia and the corresponding probability of sinus bradycardia. When the electrocardiogram data is judged to have a heart morphological disease, such as T-wave inversion and myocardial infarction, the output diagnosis result further outputs the probability of the patient having T-wave inversion and the corresponding probability, and the corresponding probability of myocardial infarction. Therefore, doctors can more clearly and accurately judge the patient's heart disease and physical condition.

[0053] Therefore, the output of the relevant diagnostic results is closer to the doctor's diagnostic method, making the diagnostic results of the automatic electrocardiogram diagnostic method 100 closer to the doctor's diagnostic results, rather than just a combination of many possible heart diseases, but unable to improve the accuracy of the diagnosis.

[0054] In summary, the automatic ECG diagnosis method disclosed in the present invention can improve the accuracy of learning and judgment, reduce the number of ECG measurement records required, and reduce the software and hardware requirements for computing data, saving a lot of resources such as electricity, making it easier for doctors and patients to quickly and correctly understand heart disease and obtain relevant advice, thereby effectively improving the quality of medical care. In addition, through multi-label learning, multiple heart diseases and their corresponding probabilities can be output for reference by doctors and patients, effectively improving the accuracy of doctors' disease judgments and increasing doctors' willingness to use the method.

[0055] Although the present invention has been disclosed in the above embodiments, it is not intended to limit the present invention. Any technician in this field can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be based on the definition of the claims.

Claims

1. An automatic electrocardiogram diagnosis method, characterized in that: Include: Read multiple heart disease data; Acquire specific electrocardiogram measurement records of corresponding patients according to the plurality of heart disease data, and exclude non-specific electrocardiogram measurement records of the corresponding patients; Using machine learning, based on the plurality of heart disease data, identifying corresponding feature data in the specific electrocardiogram measurement records of the corresponding patient; Read the electrocardiogram measurement records to be identified; as well as Using the multiple feature data, determine whether the electrocardiogram measurement record to be identified has heart disease, wherein the use of machine learning also includes marking the corresponding feature data, and the marking of the corresponding feature data is to cut the multiple specific electrocardiogram measurement records at the time coordinate of the maximum value of the absolute value of each heartbeat voltage, and the marking of the corresponding feature data also includes: taking the multiple specific electrocardiogram measurement records, with the cutting position where the maximum value of the absolute value of the heartbeat voltage corresponding to the heart disease data is located as the origin, taking two-thirds of the previous cutting position, and taking the next cutting position Set two-thirds as the marking interval of the corresponding characteristic data, and when the marking interval is less than 2.5 seconds, add an average time forward and backward to make the marking interval equal to 2.5 seconds, wherein the exclusion of the non-specific ECG measurement records of the corresponding patient is based on big data analysis, and the ECG measurement records of the leads with low correlation with the multiple heart disease data are excluded, wherein the ECG measurement records of the leads with an identification accuracy rate lower than 50% are recorded as non-specific ECG measurement records, and the ECG measurement records of the leads with an identification accuracy rate higher than 90% are recorded as specific ECG measurement records.

2. The automatic electrocardiogram diagnosis method according to claim 1, characterized in that: The method of using the plurality of characteristic data to determine whether the electrocardiogram measurement record to be identified has a heart disease is to determine whether the electrocardiogram measurement record to be identified has a heart disease only based on a specific electrocardiogram measurement record of the electrocardiogram measurement record to be identified.

3. The automatic electrocardiogram diagnosis method according to claim 1, characterized in that: The marking of the corresponding characteristic data also includes: randomly multiplying the marking interval by any value from 0.75 to 1.25 to generate auxiliary characteristic data, and when the marking interval of the auxiliary characteristic data is less than 2.5 seconds, padding the subsequent data of the auxiliary characteristic data with zeros to 2.5 seconds.

4. The automatic electrocardiogram diagnosis method according to claim 1, characterized in that: The corresponding feature data is identified by using a shallow neural network model to identify the multiple corresponding feature data.

5. The automatic electrocardiogram diagnosis method according to claim 1, characterized in that: It also includes: classifying the heart disease recorded in the electrocardiogram to be identified by heartbeat, rhythm and type level.

6. The automatic electrocardiogram diagnosis method according to claim 5, characterized in that: It also includes: using multi-label learning to learn and judge multiple heart diseases at the same time.

7. The automatic electrocardiogram diagnosis method according to claim 6, characterized in that: Also includes: output of multiple heart diseases and corresponding probabilities.

Images