Specification device and computer program

Abstract

To provide a technology capable of accurately specifying a generation timing of specific waves on an electrocardiogram.SOLUTION: A specification device specifies information related to a generation timing of specific waves on an electrocardiogram. The specification device includes an acquisition part acquiring specific electrocardiogram of a subject and an operation part. The operation part acquires temporal phase information expressing cyclic generation of specific waves of the specific electrocardiogram on the basis of machine learning that uses learning data including learning electrocardiogram and a generation timing of the specific waves of the learning electrocardiogram, and specifies timing information expressing generation timing of the specific waves included in the specific electrocardiogram from the acquired phase information.SELECTED DRAWING: Figure 8

Description

[Technical field]

[0001] The technology disclosed herein relates to technology for identifying information related to the timing of occurrence of specific waves on an electrocardiogram. [Background technology]

[0002] When diagnosing cardiac disease using an electrocardiogram, it is necessary to interpret the activity of the heart's stimulation conduction system from the electrocardiogram. Even specialized medical personnel such as doctors and medical technicians may take time to interpret an electrocardiogram. For this reason, technology for automatically analyzing electrocardiograms has been developed (for example, Patent Document 1). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2020-130772 A Summary of the Invention [Problem to be solved by the invention]

[0004] When analyzing an electrocardiogram, characteristic waves in the cardiac conduction system are read, such as P waves, which indicate the excitation process from the sinus node to the atrioventricular node, QRS waves, which indicate the excitation process from the bundle of His to the Purkinje fibers, and T waves, which indicate the process of ventricular excitation disappearing. However, some of the waves may not appear clearly, and there is a problem in that the waves that do not appear clearly are difficult to detect.

[0005] This specification discloses a technique that can accurately identify the occurrence timing of a specific wave on an electrocardiogram. [Means for solving the problem]

[0006] In a first aspect of the technology disclosed in this specification, a determination device determines information related to the occurrence timing of a specific wave on an electrocardiogram. The determination device includes an acquisition unit that acquires a specific electrocardiogram of a subject, and a calculation unit. The calculation unit acquires time-dependent phase information representing the periodic occurrence of a specific wave on the specific electrocardiogram based on machine learning using learning data including a learning electrocardiogram and the occurrence timing of the specific wave on the learning electrocardiogram, and determines timing information representing the occurrence timing of the specific wave included in the specific electrocardiogram from the acquired phase information.

[0007] It is known that certain waves on an electrocardiogram (e.g., P waves, QRS waves, T waves, etc.) occur periodically. By using machine learning, it is possible to obtain time-dependent phase information that indicates the periodic occurrence of a certain wave on a specific electrocardiogram, and to identify the occurrence timing based on the periodicity of the certain wave. Therefore, it is possible to accurately identify the occurrence timing of a certain wave on a specific electrocardiogram.

[0008] This specification also discloses a computer program for identifying information related to the occurrence timing of a specific wave on an electrocardiogram. The computer program causes a computer to function as an acquisition unit that acquires, based on machine learning using learning data including a learning electrocardiogram and the occurrence timing of the specific wave on the learning electrocardiogram, time-varying phase information that represents the periodic occurrence of a specific wave on the specific electrocardiogram, and an identification unit that identifies, from the acquired phase information, timing information that represents the occurrence timing of the specific wave included in the specific electrocardiogram. [Brief description of the drawings]

[0009] [Figure 1] FIG. 2 is a block diagram showing the functions of a specific device according to an embodiment. [Diagram 2] Diagram showing the architecture of a machine learning model. [Diagram 3] A diagram showing periodically occurring P waves in terms of time t and angle θ. [Figure 4] An electrocardiogram showing overlapping P waves and T waves, and a diagram showing the timing of P wave occurrence. [Diagram 5]1 is a flowchart showing an example of a machine learning process. [Figure 6] 11 is a flowchart showing an example of a process for identifying the timing of occurrence of a P wave in a subject's electrocardiogram using a DNN trained by machine learning processing. [Figure 7] 1A and 1B are diagrams for explaining the correction process, in which (a) shows an electrocardiogram including a portion where P waves overlap with T waves, (b) shows the phase trajectory and phase orbital of phase θ before the correction process, and (c) shows the phase trajectory and phase orbital of phase θ after the correction process. [Figure 8] FIG. 4 is a diagram showing a schematic diagram of a difference between the present embodiment and a comparative example. [Figure 9] 1 is a table showing the recall, precision, and F-score for the comparative example, example 1, and example 2. [Figure 10] FIG. 1 shows the verification results of the comparative example, Example 1, and Example 2, where (a) shows the electrocardiogram of the subject, (b) shows the phase trajectory of the P-wave occurrence cycle of the correct label, (c) shows the phase trajectory of the P-wave occurrence cycle of Example 1, (d) shows the phase trajectory of the P-wave occurrence cycle of Example 2, (e) shows the occurrence timing of the P-wave of the correct label, (f) shows the occurrence timing of the P-wave of the comparative example, (g) shows the occurrence timing of the P-wave of Example 1, and (h) shows the occurrence timing of the P-wave of Example 2. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0010] The main features of the embodiments described below are listed below. Note that the technical elements described below are independent technical elements that exhibit technical usefulness alone or in various combinations, and are not limited to the combinations described in the claims at the time of filing.

[0011] In a second aspect of the technology disclosed in the present specification, in the first aspect, the phase information may include at least one of a sine wave and a cosine wave whose phase is 0 at the occurrence timing of the specific wave. With this configuration, the occurrence timing of the periodic specific wave can be accurately represented using at least one of a sine wave and a cosine wave. This makes it possible to accurately identify the occurrence timing of the specific wave of the specific electrocardiogram.

[0012] In a third aspect of the technology disclosed in the present specification, in the second aspect, the calculation unit may use phase information to identify a trajectory of changes in sine and cosine values ​​due to a transition of a phase indicating 0 at the occurrence timing of a specific wave in a specific electrocardiogram, and may correct the identified trajectory to identify timing information representing the occurrence timing of the specific wave. When a specific wave overlaps with another wave, it may be difficult to identify the occurrence timing. In this case, the above trajectory does not become circular. According to the above configuration, by correcting the above trajectory, the occurrence timing of a specific wave can be identified even for timings where a trajectory is not generated accurately.

[0013] In a fourth aspect of the technology disclosed in the present specification, in any one of the first to third aspects, the learning electrocardiogram may include a pseudo first learning electrocardiogram including a specific wave and a second learning electrocardiogram measured in advance from a person. In the machine learning, the calculation unit may execute a first learning step using the first learning electrocardiogram and the occurrence timing of the specific wave in the first learning electrocardiogram, and a second learning step using the second learning electrocardiogram and the occurrence timing of the specific wave in the second learning electrocardiogram. An electrocardiogram actually measured from a person (patient) includes a portion where the specific wave is difficult to distinguish. For this reason, if machine learning is executed using only electrocardiograms measured from a person, there is a possibility that the machine learning will not converge. By performing machine learning in two stages: a first learning step using a pseudo first learning electrocardiogram that includes specific waves, and a second learning step using a second learning electrocardiogram that has been measured in advance from a person, it is possible to perform machine learning (pre-learning) using an accurately periodically changing waveform, and then perform machine learning using an electrocardiogram measured from a person.

[0014] In a fifth aspect of the technique disclosed in the present specification, in any one of the first to fourth aspects, the specific wave may be a P wave. P waves have smaller potential changes on an electrocardiogram than QRS waves and T waves, and are therefore more difficult to distinguish from an electrocardiogram than other waves. By focusing on the periodicity of P waves to identify the occurrence timing of P waves, it is possible to accurately identify the occurrence timing of P waves that are difficult to distinguish from an electrocardiogram waveform. EXAMPLES

[0015] The identification device 10 according to the embodiment will be described with reference to the drawings. The identification device 10 identifies the occurrence timing of a specific wave (e.g., P wave, QRS wave, T wave, etc.) on an electrocardiogram. An electrocardiogram is a measurement of an electrical signal of the stimulation conduction system of the subject's heart. The waveform of an electrocardiogram includes characteristic waveforms such as a P wave indicating an excitation process from the sinus node to the atrioventricular node, a QRS wave indicating an excitation process from the bundle of His to the Purkinje fiber, and a T wave indicating a process in which the excitation of the ventricle disappears. When analyzing an electrocardiogram, it is necessary to detect waveforms such as a P wave, a QRS wave, and a T wave to grasp the stimulation conduction system of the heart. Thus, analysis of an electrocardiogram requires specialized knowledge to interpret the characteristic waveforms, and the electrocardiogram is analyzed by a medical professional having specialized knowledge such as a doctor or a medical technician (hereinafter, simply referred to as a "specialist"). In this specification, the term "electrocardiogram" may refer to the waveform itself indicating the electrical signal of the stimulation conduction system of the subject's heart, or may refer to data indicating the waveform.

[0016] An electrocardiogram is often measured over a relatively long period of time. Therefore, analysis takes time. When analyzing an electrocardiogram, it is necessary to identify characteristic waveforms such as P waves, QRS waves, and T waves of the electrocardiogram. The identifying device 10 of this embodiment accurately identifies the occurrence timing of a specific wave of the electrocardiogram. In particular, identifying P waves originating from the atrium is important in diagnosing arrhythmias such as atrial fibrillation, supraventricular extrasystoles, and atrioventricular block. However, P waves have smaller potential changes (smaller waveforms) on the electrocardiogram compared to QRS waves and T waves. Therefore, it is difficult to identify P waves from the waveform of the electrocardiogram. An example in which the identifying device 10 identifies the occurrence timing of P waves will be described below.

[0017] 1, the identifying device 10 includes a calculation unit 12 and a communication unit 40. The calculation unit 12 and the communication unit 40 are connected to each other by wiring such as a bus bar so as to be able to communicate with each other.

[0018] The calculation unit 12 is configured using a computer including a memory 14 and a CPU 30. A calculation program is stored in the memory 14, and the CPU 30 executes the calculation program, causing the CPU 30 to function as a learning processing unit 32, a specification unit 34, and the like shown in FIG. 1. The processing of the learning processing unit 32 and the specification unit 34 will be described in detail later. The memory 14 includes a learning data storage unit 16, a trained DNN storage unit 18, and a subject information storage unit 20.

[0019] The learning data storage unit 16 stores data used in machine learning processing. The data used in machine learning includes first learning data and second learning data. The first learning data is used in a first learning process performed before executing machine learning. The second learning data is used when performing machine learning again after the first learning process. The data used in machine learning is acquired from an external device via the communication unit 40 and stored in the learning data storage unit 16.

[0020] The trained DNN storage unit 18 stores information learned by machine learning. In this embodiment, the P wave occurrence period is learned by machine learning using a deep neural network (DNN). The trained DNN storage unit 18 stores the DNN learned by machine learning.

[0021] The subject information storage unit 20 stores information about the subject. The information about the subject includes the subject's electrocardiogram (hereinafter also referred to as a specific electrocardiogram) and information about the occurrence timing of the P wave of the specific electrocardiogram identified by the identification unit 34. The subject's electrocardiogram (specific electrocardiogram) is measured using an electrocardiogram measuring device such as a Holter electrocardiograph. The calculation unit 12 acquires the specific electrocardiogram from an external device such as an electrocardiogram measuring device via the communication unit 40 and stores it in the subject information storage unit 20. The calculation unit 12 stores information about the occurrence timing of the P wave of the specific electrocardiogram identified by the identification unit 34 in the subject information storage unit 20.

[0022] The communication unit 40 includes a communication interface for the specific device 10 to communicate with an external device in a wired or wireless manner. The calculation unit 12 can acquire various data from the external device via the communication unit 40. Specifically, the calculation unit 12 acquires data to be used for machine learning and an electrocardiogram (specific electrocardiogram) of the subject, which are transmitted from the external device, via the communication unit 40. The calculation unit 12 stores the acquired data to be used for machine learning in the learning data storage unit 16, and stores the acquired specific electrocardiogram in the subject information storage unit 20.

[0023] Next, a process of identifying the occurrence timing of P waves in the electrocardiogram of the subject by the identification device 10 will be described. The process of identifying the occurrence timing of P waves in the electrocardiogram of the subject by the identification device 10 is executed by a machine learning process of learning the occurrence period of P waves by machine learning, and an identification process of identifying the occurrence timing of P waves in the electrocardiogram of the subject by using the DNN learned by the machine learning process.

[0024] Here, the machine learning model of this embodiment will be described. FIG. 2 shows the architecture of the machine learning model of this embodiment. As shown in FIG. 2, the machine learning model of this embodiment is composed of BLSTM (short for Bidirectional Long Short-Term Memory), FPN (short for Feature Pyramid Networks), and a linearly connected layer. BLSTM is a bidirectional long-short-term memory that is effective for feature extraction of time series data. FPN is a feature pyramid network used for image recognition. The input layer is a one-dimensional time series signal of an electrocardiogram having a length of 1248 (about 10 seconds) sampled at 125 Hz. The output layer is a two-dimensional time series of sine and cosine values ​​with the same phase as the input layer. In addition, each intermediate layer has the same time length as the input layer and the output layer. The machine learning model of this embodiment uses the first-order time series signal of an electrocardiogram to output two-dimensional time series data of sine and cosine values ​​from a deep neural network (DNN) composed of BLSTM, FPN, a linearly connected layer, and BLSTM. In addition, since publicly known DNN technology can be applied to the DNN composed of BLSTM, FPN, and linearly coupled layers, detailed explanations are omitted.

[0025] In addition, the sine and cosine values ​​created from the occurrence time of the P wave identified on the electrocardiogram are used as the correct answer label for the DNN. Specifically, a specialist visually identifies the waveform of the P wave on the electrocardiogram, and identifies the occurrence time of the P wave on the electrocardiogram from the identified waveform of the P wave. Next, as shown in Figure 3, a graph is generated that shows the periodically occurring P waves in terms of time t and angle θ. Next, the occurrence time t of the kth P wave is calculated. k At multiple points on the graph (t k The phase θ(t) for continuous time progression is determined by applying cubic spline interpolation to the vectors θ(t) ( , 2kπ). The sine and cosine values ​​of the determined phase θ(t) are calculated and used to obtain the ground truth label.

[0026] By executing machine learning using the above machine learning model, it is possible to learn the occurrence period of P waves in an electrocardiogram. P waves have the property of occurring periodically and regularly. On the other hand, since P waves have a small potential change, if they overlap with other large waveforms such as QRS waves and T waves, or overlap with noise such as body movement, the P waves may not appear clearly in the electrocardiogram. FIG. 4 shows an example in which a P wave overlaps with a T wave. In the electrocardiogram shown in FIG. 4, the P wave occurring for the fourth time overlaps with a T wave, making it difficult to identify the P wave from the electrocardiogram. A specialist can empirically identify P waves that do not appear clearly in an electrocardiogram based on electrophysiological knowledge that slight changes in waveform and P waves have strong periodicity. Therefore, in this embodiment, focusing on the periodicity of P waves, sine and cosine values ​​representing the occurrence period of P waves in an electrocardiogram are learned by machine learning using correct answer labels by experts. Then, the trained DNN is used to calculate sine and cosine values ​​(phase information over time) that represent the occurrence period of P waves in the electrocardiogram of the subject, and the calculated sine and cosine values ​​are used to identify the occurrence timing of the P waves. Here, "over time" refers to continuous or intermittent (which may be regular or irregular).

[0027] The machine learning process will be described with reference to Fig. 5. As shown in Fig. 5, first, the learning processing unit 32 acquires first learning data (S12). The first learning data is stored in the learning data storage unit 16. The learning processing unit 32 acquires the first learning data from the learning data storage unit 16.

[0028] The first learning data is data used in the first learning process described later, and includes a pseudo input signal and its correct answer label. In this embodiment, machine learning is performed using the above-mentioned machine learning model. However, since the above-mentioned machine learning model has a high degree of freedom, if machine learning is performed using only a part (only data of a limited size) of an electrocardiogram (i.e., raw data) measured from a person, learning may not converge. Therefore, in order to learn in advance the occurrence period of P waves to be learned by the machine learning process, a first learning process, which is a pre-learning process, is performed using an artificially generated pseudo input signal instead of raw data.

[0029] The pseudo input signal and the correct label included in the first learning data are created as follows. First, the probability distribution of the P-wave interval is defined as a normal distribution having an average P-wave interval randomly selected from 0.6 to 1.0 seconds corresponding to a normal heart rate and a standard deviation of 3.75% of the average P-wave interval, each time a pseudo input signal with a time length of 1248 is created. Next, a series of P-wave intervals randomly sampled from the defined probability distribution is used to generate a graph (i.e., a graph as shown in FIG. 3) showing the P-wave occurrence period in terms of time t and angle θ, to determine the phase θ(t), and the sine and cosine values ​​are calculated to obtain the correct label.

[0030] Next, a pseudo input signal is created using the calculated sine value. Specifically, the sine value is multiplied by an amplitude randomly selected from the interval [0,1], and a signal is generated by adding white noise with an amplitude equal to or less than twice the amplitude of the signal. A plurality of pseudo input signals are created. For example, in this embodiment, 900 pseudo input signals are created.

[0031] Next, the learning processing unit 32 executes a first learning process using the first learning data acquired in step S12 (S14). In the first learning process, most of the multiple pseudo input signals are used as learning data, and some are used as verification data. For example, in this embodiment, of the 900 pseudo input signals, 800 are used as learning data, and 100 are used as verification data. The verification data is randomly selected from the 900 pseudo input signals.

[0032] In the first learning process, machine learning is performed using the machine learning model shown in FIG. 2 above, using a pseudo input signal as the input layer. An example of the conditions for the first learning process is as follows. In the first learning process, the parameters of ResNet-34 pre-trained using ImageNet are used as the initial parameters of the FPN. In addition, absolute mean error is used as the loss function, and Adam is used as the optimization method, with a learning rate of 0.0001 and a batch size of 10. Early Stopping is used as the stopping condition for learning, and learning is stopped when the loss on the validation data does not decrease for five epochs. Calculations during learning are performed on a Linux (registered trademark) server using Pytorch.

[0033] Next, the learning processing unit 32 acquires second learning data (S16). The second learning data is stored in the learning data storage unit 16. The learning processing unit 32 acquires the second learning data from the learning data storage unit 16.

[0034] The second learning data is data used in the second learning process described later, and includes learning electrocardiogram data and its correct answer label. The learning electrocardiogram data is clinical electrocardiogram data of a patient with second degree atrioventricular block, partially including a section of normal sinus rhythm. A time series signal of length 1248 sampled at 125 Hz of the above patient is used as the learning electrocardiogram data. A plurality of learning electrocardiogram data are created. For example, in this embodiment, 170 examples of learning electrocardiogram data are created from 170 examples of clinical electrocardiogram data of 17 patients. The correct answer label is the occurrence timing of P waves visually identified from the electrocardiogram by a specialized medical professional.

[0035] Next, the learning processing unit 32 executes the second learning process using the second learning data acquired in step S16 (S18). In the second learning process, the DNN of the first learning process (pre-learning process) executed in step S14 is re-learned with the machine learning model shown in FIG. 2 above using human clinical electrocardiogram data (i.e., electrocardiogram data included in the second learning data). Among the parameters of the DNN of the pre-learning process, the linear combination layer and the BLSTM after the linear combination layer are fixed and re-learning is executed. In the second learning process, most of the multiple electrocardiogram data are used as learning data, and a part of them are used as verification data. For example, in this embodiment, of the 170 electrocardiogram data, 150 are used as learning data and 20 are used as verification data. The verification data is randomly selected from the 170 electrocardiogram data. The second learning process is executed using the same conditions as the first learning process in step S14. By executing the first learning process and the second learning process, it is possible to learn outputs of sine and cosine values ​​representing the occurrence period of P waves from an electrocardiogram.

[0036] Next, the learning processing unit 32 stores the DNN learned in step S18 in the learned DNN memory unit 18 (S20).

[0037] Next, a description will be given of a specification process for specifying the occurrence timing of the P wave of the electrocardiogram of the subject using the DNN learned by the machine learning process. As shown in FIG. 6, first, the specification unit 34 acquires a specific electrocardiogram stored in the subject information storage unit 20 (S100). As described above, the specific electrocardiogram is, for example, an electrocardiogram of the subject measured using an electrocardiogram measuring device such as a Holter electrocardiograph. The specific electrocardiogram is stored in the subject information storage unit 20. The specification unit 34 acquires the specific electrocardiogram from the subject information storage unit 20.

[0038] Next, the identification unit 34 acquires the trained DNN stored in the trained DNN storage unit 18 (S110).

[0039] Next, the identification unit 34 uses the trained DNN acquired in step S110 to calculate the sine and cosine values ​​representing the occurrence period of the P wave of the specific electrocardiogram acquired in step S100 (S120).

[0040] Next, the identification unit 34 corrects the sine and cosine values ​​representing the occurrence period of the P wave calculated in step S120 (S130). As described later, in this embodiment, the phase θ is calculated by arctan from two-dimensional time series data of the sine and cosine values ​​related to the occurrence period of the P wave in the electrocardiogram of the subject, and the timing at which the phase θ becomes 0 is identified as the occurrence timing of the P wave. The phase trajectory representing the trajectory of the temporal change of the phase θ can be expressed as a circular trajectory of (cosθ, sinθ) in two-dimensional space (see FIG. 7). The phase trajectory is constrained on a unit circle in the correct answer label. On the other hand, the phase trajectory of the phase θ calculated from the sine and cosine values ​​output using a DNN learned by machine learning may not be constrained on a unit circle. For example, when one P wave is not identified as shown in FIG. 4, the period of the occurrence timing of the P wave is shifted. As a result, the center point of the phase trajectory may not be fixed and may deviate from the unit circle. In such a case, it may not be possible to accurately identify the occurrence timing of the P wave. Therefore, the identification unit 34 corrects the phase trajectory of the phase θ calculated from the sine and cosine values ​​calculated in step S120. Specifically, the identification unit 34 applies baseline correction using pEMD (short for pseudo empirical mode decomposition) to the phase trajectory of the phase θ calculated from the output sine and cosine values. That is, the identification unit 34 detects maximum and minimum points for each of the output sine and cosine values. Then, the identification unit 34 applies spline interpolation to each of the maximum and minimum points of the sine and cosine values ​​to obtain a baseline that varies depending on the average of the calculated maximum and minimum lines, and subtracts it from the original signal. As a result, the position of the center point of the phase trajectory of the phase θ calculated from the output sine and cosine values ​​is corrected.

[0041] The correction process will be further described with reference to FIG. 7. FIG. 7(a) shows the electrocardiogram of the subject (specifically, the electrocardiogram from 6.0 to 9.4 seconds in FIG. 10(a)), and at the time indicated by the arrow, the P wave overlaps with the T wave, and the P wave does not appear clearly. FIG. 7(b) shows, on the left side, the phase orbit of the phase θ calculated from the sine value and the cosine value representing the occurrence period of the P wave before the above correction process is performed, and on the right side, the phase orbit of the phase θ is shown as a circular orbit of (cosθ, sinθ) for the region A sandwiched between the dashed lines in the left side. Since the P wave does not appear clearly in the region A (see FIG. 7(a)), the phase orbit does not appear as a unit circle, as shown in the left side of FIG. 7(b). Although the circular orbit in the right side of FIG. 7(b) draws a closed trajectory after one revolution, the origin is not included inside the circle that has been rotated once. Therefore, if the above correction process is not performed, the occurrence of a P wave is not detected in region A. On the other hand, as shown in the right diagram of Fig. 7(c), the orbit of phase θ in region A when the above correction process is performed includes the origin inside a circle that makes one rotation. In this way, by performing the correction process of step S130, it is possible to identify the occurrence timing of a P wave that does not appear clearly on the electrocardiogram.

[0042] Next, the determination unit 34 determines the occurrence timing of the P wave using the sine and cosine values ​​corrected in step S130 (S140). Specifically, the determination unit 34 calculates the phase θ from the sine and cosine values ​​corrected in step S130 by arctan, and determines the timing at which the phase θ becomes 0 as the occurrence timing of the P wave. The determination unit 34 stores the determined occurrence timing of the P wave in the subject information storage unit 20.

[0043] In this embodiment, the occurrence period of the P wave of the electrocardiogram is learned by machine learning, and the learned DNN is used to identify the occurrence timing of the P wave of the electrocardiogram (specific electrocardiogram) of the subject. For example, if the occurrence timing of the P wave of the electrocardiogram of the subject is identified by learning the characteristics of the waveform of the P wave by machine learning without focusing on the periodicity of the P wave, the occurrence of the P wave may not be detected if the P wave does not appear clearly because it overlaps with other large waveforms such as QRS waves and T waves, or overlaps with noise such as body movement. In this embodiment, the occurrence timing of the P wave is identified by focusing on the periodicity of the P wave. Therefore, it is possible to identify the occurrence timing of the P wave that cannot be clearly determined from the waveform of the electrocardiogram, and the accuracy of identifying the occurrence timing of the P wave from the electrocardiogram of the subject can be improved.

[0044] In this embodiment, in step S130, a correction process is performed to correct the phase trajectory of the phase θ calculated from the output sine value and cosine value, but the present invention is not limited to this configuration. The correction process in step S130 may be omitted. As described above, in this embodiment, machine learning is performed to learn the occurrence period of P waves in an electrocardiogram, focusing on the periodicity of P waves. Therefore, even if the correction process in step S130 is omitted in the identification process, the occurrence timing of P waves can be accurately identified from the electrocardiogram of the subject, compared to the case where the occurrence timing of P waves is identified using information that has learned the characteristics of the waveform of P waves without focusing on the periodicity of P waves.

[0045] In addition, in a verification conducted by the present inventors, it has been confirmed that the timing of occurrence of a P wave can be accurately identified from the electrocardiogram (specific electrocardiogram) of the subject. In the verification, the above machine learning process was executed. As described above, in the identification process, even if the correction process of step S130 is omitted, the timing of occurrence of a P wave can be more accurately identified. In the following, an example in which the above machine learning process is executed and the above identification process is executed while omitting the correction process of step S130 is referred to as Example 1, and an example in which the above machine learning process is executed and the above identification process is executed (i.e., the correction process of step S130 is also executed without omitting it) is referred to as Example 2.

[0046] As a comparative example, a machine learning process was performed to learn the characteristics of the waveform of the P wave without paying attention to the periodicity of the P wave, and the timing of occurrence of the P wave was identified using the learned DNN. FIG. 8 shows the difference between the comparative example and this embodiment (Examples 1 and 2). As shown in the lower diagram of FIG. 8, in this embodiment, the occurrence period of the P wave of the electrocardiogram is learned by machine learning, the occurrence period of the P wave is generated from the electrocardiogram of the subject by the learned DNN, and the occurrence timing of the P wave is identified from the occurrence period of the generated P wave. On the other hand, as shown in the upper diagram of FIG. 8, in the comparative example, the characteristics of the waveform of the P wave on the electrocardiogram is learned by machine learning, and the waveform of the P wave is directly identified from the electrocardiogram of the subject by the learned DNN, thereby identifying the occurrence timing of the P wave. In the comparative example, in order to minimize the difference in performance due to the difference in the architecture of the machine learning model, an architecture in which one linearly connected layer is further added after the output layer of the machine learning model of this embodiment shown in FIG. 2 was used. In addition, data for detecting a section of 120 ms from the occurrence timing of the P wave was created as learning data. Other than these points, the learning data and learning conditions were the same as those in this embodiment, and the machine learning process was executed in the comparative example.

[0047] In Example 1, Example 2, and Comparative Example, a specification process was performed to specify the occurrence timing of P waves using the same electrocardiogram. In the specification process, an electrocardiogram of a subject other than the data for machine learning was used. The electrocardiogram of the subject was clinical electrocardiogram data of a patient with second degree atrioventricular block, including 218 P waves. Clinical electrocardiogram data of 20 cases from two patients was used.

[0048] In order to compare Example 1, Example 2, and Comparative Example, the following processing was further performed on the data after the identification processing. First, the timing (time) of the occurrence of the identified P wave was converted into a detection result of a segment with 120 ms from the occurrence as a P wave section. For each of Example 1, Example 2, and Comparative Example, the segment of the detected P wave section (detected segment) was compared with the segment of the correct label (GT) (correct segment). Between the detected segment and the correct segment, the segment with the closest time distance to the section center was identified. The detected segment with a time distance from the correct segment with a maximum allowable error of 500 ms was classified as a true positive, and the detected segments other than the true positive were classified as either false negative or false positive. In addition, taking into consideration the stability of behavior, errors that occurred within 1 second on either end of the approximately 10-second electrocardiogram recording were excluded.

[0049] Next, as shown in FIG. 9, the recall rate, precision rate, and F value were calculated from the obtained detection results. Furthermore, the distance between the section center of the detected segment classified as true positive and the section center of the correct segment was calculated to evaluate the time accuracy of P wave detection. The average and standard deviation of the estimated time error of the true positive P wave were 23±20 ms (n=196) in the comparative example, 24±26 ms (n=212) in the example 1, and 31±42 ms (n=218) in the example 2. From the above, it was confirmed that the precision rate was slightly lower in the example 1 compared to the comparative example, but the recall rate and F value were significantly improved. Furthermore, in the example 2, the precision rate and time estimation accuracy were slightly lower than in the comparative example, but the recall rate and F value were significantly improved, and it was confirmed that the recall rate and F value were improved even compared to the example 1.

[0050] FIG. 10 illustrates the verification results of the comparative example, Example 1, and Example 2. FIG. 10(c) and FIG. 10(d) show the phase trajectory of the phase θ calculated from the sine and cosine values ​​representing the P-wave generation period for each of Examples 1 and 2. As shown in FIG. 10(b) to FIG. 10(d), the phase trajectory of the P-wave generation period of Examples 1 and 2 is not generated accurately around 8 seconds compared to the phase trajectory of the P-wave generation period of the correct label, but is generated accurately in other parts. Therefore, it was found that the phase trajectory of the phase θ calculated from the sine and cosine values ​​representing the P-wave generation period can be generated with high accuracy by using the DNN trained by the machine learning process of this embodiment.

[0051] Also, Fig. 10(f) to Fig. 10(h) show the timing of occurrence of P waves identified for each of the comparative example, Example 1, and Example 2. As shown in Fig. 10(e) and Fig. 10(f), in the comparative example, compared to the correct answer label, the occurrence of P waves was not detected a total of three times. Also, as shown in Fig. 10(g), in Example 1, of the three occurrences of P waves not detected in the comparative example, the occurrence of two P waves was detected, excluding the occurrence around 8 seconds. Therefore, it was found that the number of false positives was reduced in Example 1 compared to the comparative example, and false negatives were improved compared to the comparative example.

[0052] Furthermore, as shown in FIG. 10(h), in Example 2, out of the three occurrences of P waves not detected in the comparative example, two occurrences of P waves were detected similarly to those in Example 1, and the occurrence of a P wave around 8 seconds not detected in Example 1 was also detected. Therefore, in Example 2, it was found that the false negatives were improved, and the orbit of the phase was corrected even for positions where the phase orbit of the phase θ was not accurately generated by executing the correction process of step S130, thereby improving the detection accuracy of the occurrence of P waves. From the above, it was confirmed that the occurrence timing of P waves can be accurately identified from the electrocardiogram of the subject by using the DNN trained by the machine learning process of this embodiment.

[0053] In this embodiment, the sine value and the cosine value representing the occurrence period of the P wave of the electrocardiogram are learned by machine learning processing, but the present invention is not limited to such a configuration. For example, either the sine value or the cosine value representing the occurrence period of the P wave of the electrocardiogram may be learned by machine learning. In this case, the occurrence period of the P wave of the electrocardiogram can be learned, so that the accuracy of identifying the occurrence timing of the P wave from the electrocardiogram of the subject can be improved compared to the case where the characteristics of the waveform of the P wave are learned without focusing on the periodicity of the P wave (the above comparative example).

[0054] In addition, in this embodiment, the machine learning process is performed using one identifying device 10, and the identifying process is also performed, but this is not limited to the configuration. The machine learning process may be performed by another device, and the identifying device may perform the identifying process using the DNN learned by the machine learning process performed by the other device. That is, the calculation unit 12 of the identifying device may include a learned DNN storage unit 18, a subject information storage unit 20, a subject information storage unit 20, and an identifying unit 34, and the other device may include a learning data storage unit 16, a learned DNN storage unit 18, and a learning processing unit 32. In this case, the other device may perform the machine learning process, and the identifying device may acquire the DNN learned by the other device from the other device via the communication unit 40, and the timing of occurrence of the P wave may be identified using the acquired DNN. Also, another device may execute a process (step S120) of calculating sine and cosine values ​​(i.e., phase information over time) representing the occurrence period of P waves in the electrocardiogram of the subject using the trained DNN, and the identifying device 10 may execute a process (step S130) of identifying the occurrence timing of P waves from the phase information. In this case, the identifying device 10 may acquire phase information via the communication unit 40 from the device that executed the process of step S120, and identify the occurrence timing of P waves using the acquired phase information. Also, the identifying device 10 may execute the machine learning process and the identifying process by utilizing an environment provided by a cloud computing service, rather than its own hardware.

[0055] In the present embodiment, the identification device 10 identifies the occurrence timing of the P wave, but is not limited to such a configuration. If the wave occurs periodically on the electrocardiogram, the occurrence timing of the wave can be identified using the identification device. For example, the identification device can also identify the occurrence timing of the QRS wave and the T wave using a method similar to that for identifying the occurrence timing of the P wave focusing on the periodicity described above.

[0056] Although specific examples of the technology disclosed in this specification have been described in detail above, these are merely examples and do not limit the scope of the claims. The technology described in the claims includes various modifications and variations of the specific examples exemplified above. Furthermore, the technical elements described in this specification or drawings exhibit technical utility alone or in various combinations, and are not limited to the combinations described in the claims at the time of filing. Furthermore, the technology exemplified in this specification or drawings achieves multiple objectives simultaneously, and achieving one of the objectives itself has technical utility. [Explanation of symbols]

[0057] 10:Specific equipment 12: Arithmetic section 14: Memory 16: Learning data storage unit 18: Trained DNN memory 20: Subject information storage unit 30: CPU 32: Learning processing unit 34: Specific part 40: Communications Department

Claims

1. An identification device for identifying information related to the occurrence timing of a specific wave on an electrocardiogram, comprising: An acquisition unit for acquiring a specific electrocardiogram of a subject; A calculation unit, The calculation unit is acquiring time-varying phase information representing periodic occurrence of a specific wave of the specific electrocardiogram based on machine learning using learning data including a learning electrocardiogram and occurrence timing of a specific wave of the learning electrocardiogram; An identification device that identifies timing information representing the occurrence timing of the specific wave contained in the specific electrocardiogram from the acquired phase information.

2. The identifying device according to claim 1 , wherein the phase information includes at least one of a sine wave and a cosine wave, the phase of which is 0 at the occurrence timing of the specific wave.

3. The calculation unit is Using the phase information, a locus of change in sine and cosine values ​​due to a phase transition indicating 0 at a timing of occurrence of a specific wave of the specific electrocardiogram is identified; The identification device according to claim 2 , wherein timing information representing a generation timing of the specific wave is identified by correcting an already identified trajectory.

4. the learning electrocardiogram includes a first pseudo learning electrocardiogram including a specific wave and a second learning electrocardiogram previously measured from a person; The calculation unit, in the machine learning, a first learning step using the first learning electrocardiogram and occurrence timing of a specific wave of the first learning electrocardiogram; 2. The identification device according to claim 1, further comprising: a second learning step using the second learning electrocardiogram and occurrence timing of a specific wave of the second learning electrocardiogram.

5. The identification device according to any one of claims 1 to 4, wherein the specific wave is a P wave.

6. 1. A computer program for identifying information relating to the timing of occurrence of specific waves on an electrocardiogram, comprising: Computer, an acquisition unit that acquires, based on machine learning using learning data including a learning electrocardiogram and occurrence timings of specific waves of the learning electrocardiogram, time-series phase information that represents periodic occurrences of specific waves of the specific electrocardiogram; A computer program that functions as an identification unit that identifies timing information representing the occurrence timing of the specific wave contained in the specific electrocardiogram from the acquired phase information.

Images