System and method

Abstract

To provide means for detecting with high accuracy that heart sound is different from heart sound at normal time.SOLUTION: A system compares with target heart sound data, a comparison heart sound model in which machine learning is performed on heart sound data as a heart rate within a prescribed range relative to a heart rate of target heart sound data as heart sound data of an inspection target. When the comparison heart sound model does coincide with the target heart sound data, the system determines that the target heart sound data are different from those at normal time, and when determining that the target heart sound data are different from those at normal time, the system displays a waveform based on the target heart sound data and a waveform based on the comparison heart sound model.SELECTED DRAWING: Figure 3

Description

TECHNICAL FIELD

[0001] The present invention relates to a system and a method for detecting that a heart sound is different from a normal heart sound.

BACKGROUND ART

[0002] As a conventional technique for detecting that a heart sound is different from a normal heart sound (hereinafter, the fact that a heart sound is different from a normal heart sound is also referred to as "abnormality"), there is the following technique. That is, a heart sound waveform of a subject is collected, the collected heart sound waveform is overwritten to reduce noise, and a basic heart sound waveform is extracted and recorded. Then, the phases of the basic heart sound waveform and the heart sound waveform at the time of measurement are aligned, and it is determined whether the heart sound is abnormal based on the difference between the basic heart sound waveform and the heart sound waveform at the time of measurement.

[0003] The abstract of Patent Document 1 discloses the following content. (1) In the waveform extraction unit 42 (SA3), a heart sound waveform in a predetermined section including the II sound generated for each beat is extracted as an extracted waveform WE II , and in the phase alignment unit 58 (SA9), the phases of the extracted waveform WE II are aligned to match each other, and in the adding unit 60 (SA9), the phase-aligned extracted waveforms WE II are added, whereby normal heart sounds and heart murmurs generated for each beat are superimposed, and the external noise is averaged. (2) Then, in the representative extracted waveform determination unit 62 (SA12), the representative extracted waveform WR is determined based on the waveform obtained by addition in the adding unit 60 (SA9), so that heart disease can be accurately determined using the representative extracted waveform WR.

PRIOR ART DOCUMENTS

PATENT DOCUMENTS

[0004]

PATENT DOCUMENT 1

SUMMARY OF THE INVENTION

[0005] However, in order to prepare training data for detecting abnormalities in heart sounds using machine learning and deep learning, and to classify heart sound data using this training data, it is more accurate to not only match the phases but also create training data using features (such as FFT (fast Fourier transform) data) that have a similar period to the waveform of normal heart sounds, and then compare the training data with the measured heart sound data to detect whether or not the heart sounds are abnormal.

[0006] The object of the present invention is to provide a means for detecting with high accuracy when heart sounds differ from normal heart sounds. [Means for solving the problem]

[0007] The system of the first invention is characterized by comparing the target heart sound data, which is heart sound data of the subject to examination, with a comparative heart sound model that has been machine-learned on heart sound data with a heart rate within a predetermined range, and determining that the target heart sound data is different from normal if the comparative heart sound model and the target heart sound data do not match.

[0008] In this invention, the target heart sound data is compared with a comparative heart sound model, which has been machine-learned on heart sound data with heart rates within a predetermined range. If the comparative heart sound model and the target heart sound data do not match, it is determined that the target heart sound data is different from normal heart sounds. In this way, since the determination of whether or not a heart sound is different from normal is made based on a comparison with heart sounds with similar heart rates (waveform periods), it is possible to detect that a heart sound is different from normal heart sounds with high accuracy.

[0009] The system of the second invention is characterized in that, in the system of the first invention, it stores normal heart sound data, accepts input of the heart rate of target heart sound data, selects data from the stored heart sound data that is within a predetermined range of the heart rate of the target heart sound data that was accepted as input, and performs machine learning on the set of selected data, which is training data, to form a comparative heart sound model.

[0010] The system of the third invention is characterized in that, in the system of the first invention, it stores multiple heart sound data under normal conditions, accepts input of the heart rate of a target heart sound data, selects heart sound data from the multiple stored heart sound data based on the heart rate of the target heart sound data that was accepted as input, selects data from the selected heart sound data that has a heart rate within a predetermined range from the heart rate of the target heart sound data that was accepted as input, and performs machine learning on the set of selected data, which is training data, to form a comparative heart sound model.

[0011] The system of the fourth invention is characterized in that, in the system of the second or third invention, features are extracted from training data, and machine learning is performed on the extracted features to form a comparative heart sound model.

[0012] The system of the fifth invention is characterized in that, in the system of the second or third invention, the heart rate is labeled for each beat of resting heart sound data, and the resting heart sound data with the heart rate labeled is stored.

[0013] The system of the sixth invention is characterized in that, in the system of the second invention, the average heart rate of each heart sound data is labeled for multiple heart sound data under normal conditions, and the normal heart sound data with the average heart rate labeled is stored.

[0014] The system of the seventh invention is characterized in that, in the system of the sixth invention, the deviation of each heart sound data from the average heart rate is labeled for multiple heart sound data from normal conditions, and the system stores the heart sound data from normal conditions with the average heart rate and the deviation labeled.

[0015] The system of the eighth invention is characterized in that, in the system of the second invention, it accepts input of a predetermined range of heart rate for data to be selected from stored heart sound data.

[0016] The system of the ninth invention is characterized in that, in the system of the first invention, a comparison heart sound model is compared with target heart sound data, the degree of agreement between the comparison heart sound model and the target heart sound data is calculated, and if the degree of agreement between the comparison heart sound model and the target heart sound data does not exceed a predetermined value, it is determined that the target heart sound data is different from that of normal.

[0017] The system of the tenth invention is characterized in that, in the system of the first invention, if it is determined that the target heart sound data is different from that of normal, it notifies that the heart sound is different from that of normal.

[0018] The system of the eleventh invention is characterized in that, in the system of the first invention, when it is determined that the target heart sound data is different from normal, it displays a waveform based on the target heart sound data and a waveform based on a comparison heart sound model.

[0019] The system of the twelfth invention is characterized in that, in the system of the first invention, the target heart sound data is data from which features have been extracted from the heart sound data.

[0020] The method of the 13th invention is characterized by comparing the target heart sound data, which is heart sound data of the subject to examination, with a comparative heart sound model that has been machine-learned on heart sound data with a heart rate within a predetermined range, and determining that the target heart sound data is different from normal if the comparative heart sound model and the target heart sound data do not match. [Effects of the Invention]

[0021] According to the present invention, it is possible to detect with high accuracy that a heart sound is different from a normal heart sound. [Brief explanation of the drawing]

[0022] [Figure 1] This is a diagram showing heart sound data and the like labeled with the average heart rate and deviation. [Figure 2] This is a diagram showing selection of data from normal-time heart sound data and the like. [Figure 3] This is a diagram showing comparison between target heart sound data and a comparative heart sound model, and the like. [Figure 4] This is a flowchart showing the processing operation of a PC when comparing target heart sound data and a comparative heart sound model. [Modes for Carrying Out the Invention]

[0023] Hereinafter, embodiments of the present invention will be described. The system according to the embodiments of the present invention detects that a heart sound is different from a normal-time heart sound. Hereinafter, the fact that a heart sound is different from a normal-time heart sound is also referred to as "abnormality of heart sound". The system is configured by, for example, a personal computer (PC). For example, an application program for detecting an abnormality of a heart sound is installed in the PC, and by executing the application program, the PC functions as a system for detecting an abnormality of a heart sound. A digital stethoscope for recording the heart sound of a subject is connected to the PC. The digital stethoscope records the heart sound of the subject for whom it is determined whether the heart sound is abnormal.

[0024] Heart sound data, which is data of heart sounds recorded by a digital stethoscope or the like, is stored in a storage unit such as a hard disk drive of the PC constituting the system. The heart sound data is stored for each subject to be examined for whether the heart sound is abnormal. A plurality of pieces of continuous normal-time heart sound waveform data are stored as the heart sound data. Here, "heart sound data" refers to continuous heart sound data.

[0025] Figure 1 shows files 1 to 10 as heart sound data (resting subject heart sound database). After data collection, the heart sound data is labeled by a PC with the average heart rate and deviation. For example, in Figure 1, file 1 is labeled with "60" as the average heart rate and "±1.0" as the deviation of the heart sounds contained in file 1.

[0026] The PC accepts input of the heart rate of the target heart sound data, which is the heart sound data of the subject being examined, in order to detect whether the heart sounds are abnormal or not. For example, suppose the heart rates of the heart sound data recorded by a digital stethoscope are "62", "80", and "100" (three heart sound data). If the user wants to detect whether the heart sound data with a heart rate of "62" is abnormal or not, the user inputs "62" as the heart rate into the PC (see Figure 1). The PC accepts the input "62" as the heart rate of the heart sound data to be examined.

[0027] The PC also accepts input for a predetermined range (acceptable range) of heart rate for the data to be selected from the stored heart sound data. For example, if the heart rate of the target heart sound data to be detected as abnormal is "62", and the user wants to select data within the range of "62" to "±0", they would input "±0" as the predetermined range (acceptable range) (see Figure 1). Note that "inputting the heart rate" and "inputting the predetermined range (acceptable range)" also include selecting a desired value from multiple values.

[0028] In this way, the user can set the heart rate of the target heart sound data, which is the heart sound data of the subject being examined (heart sound data to be examined), and a predetermined range (acceptable range) of heart rates for data selected from the stored heart sound data, in order to detect whether or not the heart sounds are abnormal. The PC then accepts the settings for the heart rate and the acceptable range (hereinafter, these together will also be referred to as the "matching heart rate").

[0029] The PC selects a heart rate from among its stored heart rate data that falls within a predetermined range from the heart rate of the target heart rate data for which it received input. In other words, the PC refers to a heart rate data from among its stored heart rate data that is highly likely to contain the matching heart rate. For example, if the PC receives input of "62" as the heart rate and "±0" as the predetermined range (acceptable range), as shown in Figure 1, it selects file 2, which has an average heart rate of "60" and a deviation of "±3.0", file 4, which has an average heart rate of "60" and a deviation of "±2.0", and file 7, which has an average heart rate of "60" and a deviation of "±5.0", from among its stored heart rate data (files 1 to 10).

[0030] Figure 2 shows the contents of the selected file 7. The file contains waveform data of continuous heart sounds during rest, with the heart rate labeled for each beat. In Figure 2, the heart rate is labeled as "60", "62", "62", etc. for each beat. The PC selects data from the selected heart sound data that falls within a predetermined range from the heart rate of the target heart sound data that received input. Specifically, the PC collects multiple beats of data that are within an acceptable range from the resting heart rate data. For example, as shown in Figure 2, the PC selects data with a heart rate of "62" from the heart rate data.

[0031] As shown in Figure 3, the PC extracts features from the training data, which is a set of selected data. Features are extracted using methods such as the Fast Fourier Transform (FFT). The PC then performs machine learning on the extracted features to create a comparison heart sound model for comparison with the target heart sound data. Meanwhile, the PC extracts features from the subject's heart sound data, recorded using a digital stethoscope, etc., using methods such as FFT, and uses these extracted features as the target heart sound data (extracted data), which is the subject's heart sound data (the heart sound data being examined).

[0032] The PC compares the target heart sound data (extracted data) with a comparison heart sound model. If the target heart sound data does not match the comparison heart sound model, the PC determines that the target heart sound data is abnormal (different from normal). Specifically, the PC compares the comparison heart sound model with the target heart sound data and calculates the degree of agreement between the comparison heart sound data and the target heart sound data. If the degree of agreement between the comparison heart sound model and the target heart sound data does not exceed a predetermined value, the PC determines that the target heart sound data is abnormal.

[0033] Here, an anomaly detection algorithm is used to create a comparative heart sound model and to compare the comparative heart sound model with the target heart sound data. An example of an anomaly detection algorithm is an anomaly detection method using an autoencoder.

[0034] If the PC determines that the target heart sound data is abnormal, it will notify the user that the heart sound is abnormal. For example, the PC will display an alert (warning) on ​​the screen indicating that the heart sound is abnormal. In addition, if the PC determines that the target heart sound data is abnormal, it will display a comparison of the waveform based on the target heart sound data with the waveform based on comparison heart sound data.

[0035] Figure 4 is a flowchart showing the PC's processing operation when comparing target heart sound data with a comparison heart sound model. The PC determines whether the degree of agreement between the target heart sound data and the comparison heart sound model exceeds 0.9 (a predetermined value) (S1). If the PC determines that the degree of agreement between the target heart sound data and the comparison heart sound model exceeds 0.9 (S1: Yes), it displays an alert (warning) indicating that the heart sounds are abnormal (S2). Next, the PC displays a comparison of the waveform based on the target heart sound data (examination waveform) and the waveform based on the comparison heart sound model (learning waveform) (S3). If the PC determines that the degree of agreement between the target heart sound data and the comparison heart sound model does not exceed 0.9 (S1: No), it terminates the process. The numerical value of 0.9 (a predetermined value) used to determine whether the target heart sound data is abnormal can be changed as appropriate. The final definitive diagnosis is made by a physician.

[0036] As explained above, in this embodiment, the target heart sound data is compared with a comparative heart sound model, which has been machine-learned on heart sound data with heart rates within a predetermined range, against the target heart sound data, which is the heart sound data to be examined. If the comparative heart sound model and the target heart sound data do not match, the target heart sound data is judged to be abnormal (different from normal). In this way, since the determination of whether or not the heart sound is different from normal is made based on a comparison with heart sounds with similar heart rates (waveform periods), it is possible to detect that the heart sound is different from the normal heart sound with high accuracy.

[0037] Although embodiments of the present invention have been described above, the applicable forms of the present invention are not limited to those described above, and modifications can be made as appropriate without departing from the spirit of the invention. [Industrial applicability]

[0038] The present invention can be suitably employed in a system and method for detecting when heart sounds differ from normal heart sounds.

Claims

1. From the normal heart sound data, select the normal heart sound data whose heart rate is within a predetermined range relative to the heart rate of the target heart sound data which is the heart sound data to be examined. We perform machine learning on the selected data set, which is the training data, to create a comparative heart sound model. The target heart sound data is input into the created comparative heart sound model, and the degree of agreement is calculated. A system characterized by determining that the target heart sound data is different from normal heart sound data if the calculated degree of agreement does not exceed a predetermined value.

2. From the resting heart sound data, select resting heart sound data whose heart rate is within a predetermined range relative to the heart rate of the target heart sound data which is the heart sound data to be examined. According to the anomaly detection algorithm, Machine learning is performed on the selected data set, which is the training data, to create a comparative heart sound model. Input the target heart sound data into the created comparative heart sound model and calculate the degree of agreement. A system characterized by determining that the target heart sound data is different from normal heart sound data if the calculated degree of agreement does not exceed a predetermined value.

3. It stores normal heart sound data, From the stored heart sound data, select data with a heart rate within a predetermined range. The system according to claim 1, characterized in that machine learning is performed on training data, which is a set of selected data, to create a comparative heart sound model.

4. It stores multiple heart sound data from normal conditions. Select a heart sound data from among several stored heart sound data. From the selected heart sound data, select data with a heart rate within a specified range. The system according to claim 1, characterized in that machine learning is performed on training data, which is a set of selected data, to form a comparative heart sound model.

5. It stores normal heart sound data, The system accepts input of the heart rate for the target heart sound data. From the stored heart sound data, select data whose heart rate falls within a predetermined range from the heart rate of the target heart sound data for which input was received. The system according to claim 1, characterized in that machine learning is performed on training data, which is a set of selected data, to form a comparative heart sound model.

6. It stores multiple heart sound data from normal conditions. The system accepts input of the heart rate for the target heart sound data. From among multiple stored heart sound data, the system selects heart sound data based on the heart rate of the target heart sound data for which input was received. From the selected heart sound data, select the data whose heart rate falls within a predetermined range from the heart rate of the target heart sound data for which input was received. The system according to claim 1, characterized in that machine learning is performed on training data, which is a set of selected data, to form a comparative heart sound model.

7. Extract features from the training data, The system according to claim 5 or 6, characterized in that machine learning is performed on the extracted features to form a comparative heart sound model.

8. The heart rate is labeled for each individual heartbeat in the normal heart sound data. The system according to claim 5 or 6, characterized in that it stores resting heart sound data with labeled heart rate.

9. For multiple heart sound data points under normal conditions, the average heart rate of each heart sound data point is labeled. The system according to claim 5, characterized in that it stores resting heart sound data labeled with the average heart rate.

10. For multiple heart sound data points under normal conditions, the deviation of each heart sound data point from the average heart rate is labeled. The system according to claim 9, characterized in that it stores resting heart sound data labeled with the average heart rate and deviation.

11. The system according to claim 5, characterized in that it accepts input of a predetermined range of heart rate for data to be selected from stored heart sound data.

12. The system according to claim 1, characterized in that the target heart sound data is data from which features have been extracted from heart sound data.

13. By computer, From the resting heart sound data, select resting heart sound data whose heart rate falls within a predetermined range relative to the target heart sound data, which is the heart sound data to be examined. We perform machine learning on the selected data set, which is the training data, to create a comparative heart sound model. The target heart sound data is input into the created comparative heart sound model, and the degree of agreement is calculated. A method characterized by determining that the target heart sound data is different from normal heart sound data if the calculated degree of agreement does not exceed a predetermined value.

Images