JP7022957B1 - Atrial fibrillation detection program, atrial fibrillation detection device, atrial fibrillation detection method and atrial fibrillation detection system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an atrial fibrillation detection program, an atrial fibrillation detection device and an atrial fibrillation detection method capable of automatically detecting the onset of atrial fibrillation. SOLUTION: The waveform data corresponding to the first time corresponding to the waveform data required for detecting atrial fibrillation is divided into a plurality of first waveform data corresponding to the second time shorter than the first time. , A learning model by performing machine learning using a plurality of training data including a plurality of divided first waveform data and information indicating whether or not each of the plurality of first waveform data corresponds to atrial fibrillation. To generate. [Selection diagram] FIG. 12

Description

The present invention relates to an atrial fibrillation detection program, an atrial fibrillation detection device, an atrial fibrillation detection method, and an atrial fibrillation detection system.

Atrial fibrillation (AF) is a type of arrhythmia caused by abnormal movements in the atrium. Such atrial fibrillation is a disease that causes, for example, a thrombus in the atrium and may further cause a cerebral infarction or the like.

In the detection of such atrial fibrillation, the doctor generally detects the waveform pattern (abnormal waveform pattern) related to atrial fibrillation from the pulse wave or electrocardiogram collected from the patient based on his knowledge and experience. (See Patent Documents 1 to 5).

Japanese Unexamined Patent Publication No. 2018-086278 Japanese Unexamined Patent Publication No. 2018-153487 Japanese Unexamined Patent Publication No. 2019-201886 Japanese Unexamined Patent Publication No. 2018-102518 Japanese Unexamined Patent Publication No. 2018-149318

Here, when the above-mentioned atrial fibrillation is detected at an early stage, the doctor can completely cure the atrial fibrillation by utilizing ablation or the like.

However, the above-mentioned detection by a doctor is available only in hospitals and the like, and may not be able to detect atrial fibrillation at an early stage. Therefore, in recent years, there has been a demand for a method that allows patients to easily detect atrial fibrillation by themselves.

Therefore, an object of the present invention is to provide an atrial fibrillation detection program, an atrial fibrillation detection device, and an atrial fibrillation detection method that can automatically detect the onset of atrial fibrillation.

The atrial fibrillation detection program in the present invention for achieving the above object causes a computer to execute an atrial fibrillation detection process that detects atrial fibrillation by performing machine learning using waveform data of an electrocardiographic signal. In the fibrillation detection program, the waveform data corresponding to the first time equal to or longer than the time of the waveform data required to detect the atrioventricular fibrillation is the second time less than the time of the waveform data required to detect the atrial fibrillation. It is divided into a plurality of first waveform data corresponding to the above, and includes the plurality of divided first waveform data and information indicating whether or not each of the plurality of first waveform data corresponds to the atrial fibrillation. It is characterized by having a computer execute a process of generating a learning model by performing machine learning using a plurality of training data.

Further, in one embodiment, the atrial fibrillation detection program in the present invention for achieving the above object generates the second waveform data by performing trend removal on the waveform data corresponding to the first time. The third waveform data is generated by performing motion artifact removal on the generated second waveform data, and the generated third waveform data is converted into the plurality of first waveform data corresponding to the second time. It is characterized by being divided.

Further, in one embodiment, the atrial fibrillation detection program in the present invention for achieving the above object uses new waveform data corresponding to the first time and a plurality of fourth waveform data corresponding to the second time. By inputting each of the divided fourth waveform data into the learning model, it is determined whether or not each of the plurality of fourth waveform data is the waveform data corresponding to the atrial fibrillation. A process of acquiring the information to be shown, determining whether or not the new waveform data is waveform data corresponding to the atrial fibrillation, and outputting the result of the determination based on the acquired information. It is characterized by having a computer execute it.

Further, in one embodiment, the atrial fibrillation detection program in the present invention for achieving the above object obtains the fifth waveform data by performing trend removal on the new waveform data corresponding to the first time. The sixth waveform data is generated by performing motion artifact removal on the generated fifth waveform data, and the generated sixth waveform data is used as the plurality of fourth waveforms corresponding to the second time. It is characterized by dividing it into data.

Further, in one embodiment, the atrial fibrillation detection program in the present invention for achieving the above object corresponds to the atrial fibrillation in each of a predetermined number of continuous waveform data among the plurality of fourth waveform data. When it is determined that the waveform data is to be used, it is determined that the new waveform data is the waveform data corresponding to the atrial fibrillation.

Further, in one embodiment, the atrial fibrillation detection program in the present invention for achieving the above object corresponds to the waveform data required for the detection of the atrial fibrillation during the time corresponding to the continuous predetermined number of waveform data. It is characterized by the fact that it is more than the time to do.

Further, in one embodiment, the atrial fibrillation detection program in the present invention for achieving the above object determines that the heart rate of a person corresponding to the new waveform data is not equal to or higher than a predetermined value, the new waveform data. Is determined to be waveform data corresponding to the atrial fibrillation.

Further, in one embodiment, the atrial fibrillation detection program in the present invention for achieving the above object determines that the heart rate of a person corresponding to the new waveform data is equal to or higher than a predetermined value, the new waveform. When it is determined whether or not the standard deviation of the heart rate of the person corresponding to the data is equal to or less than the predetermined value, and when it is determined that the standard deviation of the heart rate of the person corresponding to the new waveform data is not equal to or less than the predetermined value, the new new waveform data is determined. It is characterized in that it is determined that the waveform data is the waveform data corresponding to the atrial fibrillation.

Further, the atrial fibrillation detection device in the present invention for achieving the above object is an atrial fibrillation detection device that detects atrial fibrillation by performing machine learning using waveform data of an electrocardiographic signal. The waveform data corresponding to the first time equal to or longer than the time of the waveform data required for detecting atrial fibrillation is the plurality of first waveforms corresponding to the second time less than the time of the waveform data required to detect atrial fibrillation. A plurality of training data including a data dividing unit for dividing into data, the plurality of divided first waveform data, and information indicating whether or not each of the plurality of first waveform data corresponds to the atrial fibrillation. It is characterized by having a model generation unit that generates a learning model by performing machine learning using the above.

Further, in the atrial fibrillation detection method in the present invention for achieving the above object, the atrial fibrillation detection process for detecting atrial fibrillation by performing machine learning using the waveform data of the electrocardiographic signal is executed on the computer. A method of detecting atrial fibrillation, wherein the waveform data corresponding to the first time equal to or longer than the time of the waveform data required for detecting the atrial fibrillation is less than the time of the waveform data required for detecting the atherosclerotic fibrillation. It is divided into a plurality of first waveform data corresponding to 2 hours, and the divided plurality of first waveform data and information indicating whether or not each of the plurality of first waveform data corresponds to the atrial fibrillation. It is characterized in that a computer is made to execute a process of generating a learning model by performing machine learning using a plurality of training data including.

Further, the atrial fibrillation detection system in the present invention for achieving the above object detects atrial fibrillation by performing machine learning using the waveform data and an operation terminal that collects waveform data of an electrocardiographic signal. An atrial fibrillation detection system having an atrial fibrillation detecting device for performing the atrial fibrillation, wherein the atrial fibrillation detecting device obtains the waveform data corresponding to a first time equal to or longer than the time of the waveform data required for detecting the atrial fibrillation. , The plurality of first waveform data corresponding to the second time less than the time of the waveform data required for the detection of atrial fibrillation, and the divided first waveform data and the plurality of first waveform data. It is characterized in that a learning model is generated by performing machine learning using a plurality of training data including information indicating whether or not each corresponds to the atrial fibrillation.

Further, in the atrial fibrillation detection system of the present invention for achieving the above object, in one embodiment, the operation terminal collects new waveform data corresponding to the first time from the subject and collects the data. When new waveform data is transmitted to the atrial fibrillation detection device, and the atrial fibrillation detection device receives the new waveform data transmitted by the operation terminal, the new waveform data received is transmitted to the first. By dividing into a plurality of fourth waveform data corresponding to two hours and inputting each of the divided fourth waveform data into the learning model, each of the plurality of fourth waveform data is said to have atrial fibrillation. Information indicating whether or not the waveform data corresponds to the above is acquired, and based on the acquired information, it is determined whether or not the new waveform data is the waveform data corresponding to the atrial fibrillation. , The result of the determination is output.

According to the atrial fibrillation detection program, the atrial fibrillation detection device, the atrial fibrillation detection method, and the atrial fibrillation detection system in the present invention, the onset of atrial fibrillation can be automatically detected.

FIG. 1 is a diagram showing a configuration example of the information processing apparatus 1 according to the first embodiment. FIG. 2 is a diagram showing a configuration example of the patient terminal 2 according to the first embodiment. FIG. 3 is a diagram showing a configuration example of the information processing apparatus 1 according to the first embodiment. FIG. 4 is a diagram showing a configuration example of the information processing apparatus 1 according to the first embodiment. FIG. 5 is a diagram illustrating an outline of the atrial fibrillation detection process according to the first embodiment. FIG. 6 is a diagram illustrating an outline of the atrial fibrillation detection process according to the first embodiment. FIG. 7 is a flowchart illustrating the details of the atrial fibrillation detection process according to the first embodiment. FIG. 8 is a flowchart illustrating the details of the atrial fibrillation detection process according to the first embodiment. FIG. 9 is a flowchart illustrating the details of the atrial fibrillation detection process according to the first embodiment. FIG. 10 is a flowchart illustrating the details of the atrial fibrillation detection process according to the first embodiment. FIG. 11 is a flowchart illustrating the details of the atrial fibrillation detection process according to the first embodiment. FIG. 12 is a diagram illustrating details of the atrial fibrillation detection process according to the first embodiment. FIG. 13 is a diagram illustrating details of the atrial fibrillation detection process according to the first embodiment. FIG. 14 is a diagram showing a specific example of a plurality of divided waveform data divided by the process of S14. FIG. 15 is a diagram illustrating details of the atrial fibrillation detection process according to the first embodiment. FIG. 16 is a diagram showing a specific example of one of the plurality of divided waveform data divided by the process of S34. FIG. 17 is a diagram showing a performance evaluation result of the atrial fibrillation detection process in the first embodiment.

Hereinafter, embodiments of the present invention will be described with reference to the drawings. However, such embodiments do not limit the technical scope of the invention.

First, a configuration example of the information processing apparatus 1 according to the first embodiment will be described. FIG. 1 is a diagram showing a configuration example of the information processing apparatus 1 according to the first embodiment.

The information processing device 1 (hereinafter, also referred to as an atrial fibrillation detection device 1) is a computer device, for example, a general-purpose PC (Personal Computer). Then, the information processing apparatus 1 obtains, for example, time-series data (hereinafter, also referred to as electrocardiographic waveform data or waveform data) about the patient's electrocardiographic signal transmitted from the patient terminal 2 (hereinafter, also referred to as an operation terminal 2). By using it, a process for detecting atrial fibrillation (hereinafter, also referred to as an atrial fibrillation detection process) is performed.

The information processing apparatus 1 has a hardware configuration of a general-purpose computer apparatus, and has, for example, as shown in FIG. 1, a CPU 101 which is a processor, a memory 102, a communication interface 103, and a storage medium 104. The parts are connected to each other via the bus 105.

The storage medium 104 has, for example, a program storage area (not shown) for storing a program (not shown) for performing atrial fibrillation detection processing.

Further, the storage medium 104 has, for example, a storage unit 110 (hereinafter, also referred to as a storage area 110) for storing information used when performing atrial fibrillation detection processing. The storage medium 104 may be, for example, an HDD (Hard Disk Drive) or an SSD (Solid State Drive).

The CPU 101 executes a program loaded from the storage medium 104 into the memory 102 to perform atrial fibrillation detection processing.

The communication interface 103 performs wired communication with the access point AP via a network NW such as an Internet network, for example. Then, the access point AP wirelessly communicates with the patient terminal 2 by using, for example, WiFi. That is, the information processing device 1 may be, for example, a computer device (cloud server) provided on the cloud.

Next, a configuration example of the patient terminal 2 according to the first embodiment will be described. FIG. 2 is a diagram showing a configuration example of the patient terminal 2 according to the first embodiment.

The patient terminal 2 is, for example, a portable terminal such as an electrocardiograph, a wearable watch, or a smartphone owned by the patient himself, and is a terminal capable of collecting electrocardiographic waveform data of the patient. Specifically, the patient terminal 2 collects the patient's electrocardiographic waveform data and transmits it to the information processing apparatus 1 at a periodic timing, for example.

The patient terminal 2 has a hardware configuration of a general-purpose computer device, and has, for example, as shown in FIG. 2, a CPU 201 which is a processor, a memory 202, a communication interface 203, and a storage medium 204. The parts are connected to each other via the bus 205.

The storage medium 204 has, for example, a program storage area (not shown) for storing a program (not shown) for performing atrial fibrillation detection processing.

Further, the storage medium 204 has, for example, a storage unit 210 (hereinafter, also referred to as a storage area 210) for storing information used when performing atrial fibrillation detection processing. The storage medium 204 may be, for example, an HDD or SSD.

The CPU 201 executes a program loaded from the storage medium 204 into the memory 202 to perform atrial fibrillation detection processing.

The communication interface 203 performs wireless communication with the access point AP by using, for example, WiFi. Then, the access point AP performs wired communication with the information processing device 1 via a network NW such as an Internet network.

Further, the administrator terminal 3 shown in FIG. 1 is, for example, a PC in which an administrator of the information processing apparatus 1 (hereinafter, also simply referred to as an administrator) inputs necessary information. Specifically, the administrator uses, for example, the administrator terminal 3 to transmit a plurality of waveform data (waveform data for learning) used for generating a learning model to the information processing apparatus 1.

The information processing device 1 may be, for example, a mobile terminal such as a smartphone. In this case, as shown in FIG. 3, the communication interface 103 may directly perform wireless communication with, for example, the operation terminal 2.

Further, the information processing apparatus 1 may also have, for example, a function of the patient terminal 2 (a function of collecting electrocardiographic waveform data of a patient). In this case, as shown in FIG. 4, the information processing apparatus 1 does not need to communicate with the patient terminal 2.

Further, the patient terminal 2 may be a stationary device such as a 12-lead electrocardiograph or a bedside monitor.

[Outline of the first embodiment]
5 and 6 are diagrams illustrating an outline of the atrial fibrillation detection process according to the first embodiment. Specifically, FIG. 5 is a diagram illustrating a process at the learning stage among the atrial fibrillation detection processes according to the first embodiment. Further, FIG. 6 is a diagram illustrating a process at the determination stage among the atrial fibrillation detection processes according to the first embodiment.

First, among the atrial fibrillation detection processes in the first embodiment, the processes in the learning stage will be described.

As shown in FIG. 5, the data receiving unit 111 of the information processing apparatus 1 receives, for example, a plurality of waveform data (waveform data for learning) transmitted by the administrator via the administrator terminal 3. The plurality of waveform data is, for example, waveform data corresponding to a time equal to or longer than the time of the waveform data required for detecting atrial fibrillation (hereinafter, also referred to as a first time). That is, the first time is longer than the time of the waveform data required to generate a learning model (learning model described later) for detecting atrial fibrillation, and the patient can use the learning model to make the atrial atrial. The time is longer than the time of the waveform data required to determine whether or not it is fibrillation. Specifically, the time required to detect atrial fibrillation may be, for example, 30 seconds. Further, the first hour may be, for example, a time such as 10 minutes or 1 hour.

Subsequently, the data pre-processing unit 112 (hereinafter, also referred to as a data dividing unit 112) of the information processing apparatus 1 performs pre-processing on each of the plurality of waveform data received by the data receiving unit 111.

Specifically, the data pre-processing unit 112, for example, takes less than the time of the waveform data required for detecting atrial fibrillation for each of the waveform data received by the data receiving unit 111 (hereinafter, both the second time). It is divided into a plurality of divided waveform data (hereinafter, also referred to as first waveform data) corresponding to (referred to as). That is, the second time is less than the time of the waveform data required to generate the learning model (learning model described later) for detecting atrial fibrillation, and the patient can use the learning model to make the atrial atrial. The time is less than the time of the waveform data required to determine whether or not it is fibrillation. Specifically, the second time may be, for example, a time such as 10 seconds.

Then, the data generation unit 113 of the information processing apparatus 1 has each divided waveform data and each divided waveform data corresponding to atrial fibrillation for each of the plurality of divided waveform data divided by the data preprocessing unit 112. It generates training data including information indicating whether or not it exists (hereinafter, also referred to as a label).

After that, the model generation unit 114 of the information processing apparatus 1 generates a learning model by performing machine learning using a plurality of learning data generated by the data generation unit 113.

Specifically, the model generation unit 114 generates a CNN (Convolutional Neural Network), for example, by performing machine learning using a plurality of learning data generated by the data generation unit 113.

The data receiving unit 111 may receive waveform data continuously transmitted from the patient terminal 2 owned by the specific patient (that is, electrocardiographic waveform data for the specific patient) at any time. The data preprocessing unit 112 and the data generation unit 113 may generate learning data each time the data reception unit 111 receives waveform data. In this case, the data preprocessing unit 112 and the data generation unit 113 use, for example, only the waveform data received by the data receiving unit 111, which is determined by the doctor to be the waveform data corresponding to atrial fibrillation. By doing so, the training data may be generated. Further, the model generation unit 114 may be one that causes the learning model to sequentially learn the generated learning data each time the data generation unit 113 generates learning data.

Next, among the atrial fibrillation detection processes in the first embodiment, the process at the determination stage will be described.

As shown in FIG. 6, the data receiving unit 111 receives, for example, new waveform data (waveform data to be determined) transmitted by the patient via the patient terminal 2.

Specifically, the patient terminal 2 collects waveform data corresponding to the first time from the patient, for example. Then, the patient terminal 2 transmits the collected waveform data to the information processing device 1. After that, the data receiving unit 111 receives the waveform data transmitted from the patient terminal 2.

Subsequently, the data pre-processing unit 112 of the information processing apparatus 1 performs pre-processing on the new waveform data received by the data receiving unit 111.

Specifically, the data preprocessing unit 112 divides, for example, the new waveform data received by the data receiving unit 111 into a plurality of divided waveform data (hereinafter, also referred to as second waveform data) corresponding to the second time. ..

Then, the data determination unit 115 of the information processing apparatus 1 inputs each of the plurality of divided waveform data divided by the data preprocessing unit 112 into a learning model (a learning model generated by the model generation unit 114), whereby a plurality of divided waveform data are input. A determination value indicating whether or not each of the divided waveform data is waveform data corresponding to atrial fibrillation is calculated.

Specifically, the data determination unit 115 uses, for example, the model parameter set of the CNN generated by the model generation unit 114, so that each of the divided waveform data divided by the data preprocessing unit 112 is divided. Is determined whether or not is the waveform data corresponding to atrial fibrillation.

Further, the data determination unit 115 determines whether or not the new waveform data received by the data reception unit 111 is waveform data corresponding to atrial fibrillation, based on each of the acquired determination values.

After that, the result output unit 116 of the information processing apparatus 1 outputs (transmits) the result of the determination by the data determination unit 115 to the patient terminal 2, for example.

Specifically, the result output unit 116 outputs information indicating whether or not the new waveform data received by the data reception unit 111 is waveform data corresponding to atrial fibrillation to the patient terminal 2.

That is, when the information processing apparatus 1 in the present embodiment receives, for example, new waveform data transmitted from the patient terminal 2, the patient corresponding to the new waveform data can use the learning model generated in advance. Determine if you have atrial fibrillation.

This allows the patient to determine whether or not atrial fibrillation has developed without having to be examined by a doctor or the like. Therefore, the patient can perform early detection of atrial fibrillation.

Further, in the information processing apparatus 1 of the present embodiment, in the learning stage, the learning waveform data corresponding to a time longer than the time required for detecting atrial fibrillation (first time) is used as the learning data as it is. Instead of using it, a plurality of divided waveform data corresponding to a time shorter than the time required for detecting atrial fibrillation (second time) (a plurality of divided waveform data generated by dividing the learning waveform data). Is used as training data.

Then, in the determination stage, the information processing apparatus 1 in the present embodiment divides a plurality of divided waveform data (by dividing the waveform data to be determined) corresponding to a time shorter than the time required for detecting atrial fibrillation. By inputting the generated plurality of divided waveform data) into the training model, information indicating whether or not the waveform data to be determined is the waveform data corresponding to atrial fibrillation is output.

As a result, the information processing apparatus 1 in the present embodiment has a scale as compared with the case where the learning waveform data corresponding to the time longer than the time required for detecting atrial fibrillation is used as it is. It is possible to generate a small learning model (a learning model with a small number of convolutional layers and fully connected layers). Therefore, the information processing apparatus 1 can perform the learning model generation processing and the determination processing for new waveform data in a short time even when the processing capacity is weak.

[Details of the first embodiment]
Next, the details of the atrial fibrillation detection process in the first embodiment will be described. 7 to 11 are flowcharts illustrating the details of the atrial fibrillation detection process according to the first embodiment. Further, FIGS. 12 to 17 are diagrams illustrating details of the atrial fibrillation detection process according to the first embodiment.

[Processing in the learning stage]
First, among the details of the atrial fibrillation detection process in the first embodiment, the details of the process in the learning stage will be described. 7 and 8 are diagrams illustrating details of processing in the learning stage.

As shown in FIG. 7, the data receiving unit 111 is, for example, a plurality of waveform data (waveform data for learning) transmitted from the administrator terminal 3 or a plurality of waveform data transmitted from the patient terminal 2 of a specific patient. Wait until the waveform data (electrocardiographic waveform data of a specific patient) is received (NO in S11).

Then, when a plurality of waveform data are received (YES in S11), the data preprocessing unit 112 performs trend removal for each of the plurality of waveform data received in the process of S11 (S12).

Subsequently, in this case, the data preprocessing unit 112 performs motion artifact removal for each of the plurality of waveform data for which the trend has been removed in the processing of S12 (S13).

Further, the data preprocessing unit 112 divides each of the plurality of waveform data for which the motion artifacts have been removed in the processing of S13 into a plurality of divided waveform data (S14).

Specifically, the data preprocessing unit 112 has, for example, a plurality of waveform data (for example, 30 seconds or more) for which motion artifacts have been removed in the processing of S13 so that the time corresponding to each divided waveform data is 10 seconds. Divide each of multiple waveform data corresponding to time).

Then, the data preprocessing unit 112 performs downsampling for each of the plurality of divided waveform data divided by the processing of S14 (S15). Further, the data preprocessing unit 112 normalizes each of the plurality of divided waveform data downsampled in the processing of S15 (S16).

That is, the information processing apparatus 1 performs preprocessing so that the waveform data received in the processing of S11 has a form suitable for generating a learning model. Hereinafter, specific examples of the processes from S11 to S16 will be described.

[Specific example of processing from S11 to S16]
FIG. 12 is a diagram illustrating a specific example of the process from S11 to the process of S16. The horizontal axis and the vertical axis in each waveform data shown in FIG. 12 indicate data numbers and amplitudes, respectively.

The data receiving unit 111 receives, for example, the waveform data shown in FIG. 12A (S11). Specifically, the waveform data shown in FIG. 12A is electrocardiographic waveform data collected for 30 seconds.

Then, the data preprocessing unit 112 generates the waveform data shown in FIG. 12B by performing trend removal on the waveform data shown in FIG. 12A (S12).

Subsequently, the data preprocessing unit 112 generates the waveform data shown in FIG. 12 (C) by performing motion artifact removal on the waveform data shown in FIG. 12 (B) (S13).

Next, the data preprocessing unit 112 divides the waveform data shown in FIG. 12 (C) to generate, for example, three divided waveform data corresponding to FIG. 12 (D) (S14). Specifically, each of the divided waveform data shown in FIG. 12 (D) is an electrocardiographic waveform data collected for 10 seconds.

Then, the data preprocessing unit 112 generates each of the divided waveform data shown in FIG. 12 (E) by downsampling each of the waveform data divided and shown in FIG. 12 (D) (S15).

After that, the data preprocessing unit 112 generates each of the divided waveform data shown in FIG. 12 (F) by normalizing each of the divided waveform data shown in FIG. 12 (E) (S16).

Returning to FIG. 8, the data generation unit 113 of the information processing apparatus 1 indicates whether or not the plurality of divided waveform data normalized in the process of S16 are the divided waveform data corresponding to atrial fibrillation (label). ) Is accepted (NO in S21).

Specifically, the person in charge of generating the training data (for example, a doctor) has the divided waveform data in which each divided waveform data corresponds to atrial fibrillation for each of the plurality of divided waveform data normalized by the processing of S16. It is determined whether or not it is. Then, the person in charge displays a label indicating a determination result of whether or not each of the plurality of divided waveform data normalized in the process of S16 is the divided waveform data corresponding to atrial fibrillation (illustrated). It is input to the information processing device 1 via (not).

Then, when the input of information (label) indicating whether or not the plurality of divided waveform data normalized in the processing of S16 is the divided waveform data corresponding to atrial fibrillation is accepted (YES in S21), the data. The generation unit 113 generates training data including each divided waveform data and a label corresponding to each divided waveform data for each divided waveform data normalized by the process of S16 (S22).

That is, the information processing apparatus 1 divides the waveform data received in the processing of S11 instead of using the waveform data (waveform data having a length capable of detecting atrial fibrillation) received in the processing of S11 as it is as training data. The divided waveform data generated in the above process is used as training data.

After that, the data generation unit 113 stores each of the learning data generated in the process of S22 in the storage area 110 (S23).

When the learning data used for generating the learning model is stored in the storage area 110 in advance, the information processing apparatus 1 may not perform the processing of S11 to S23.

Then, the model generation unit 114 of the information processing apparatus 1 determines whether or not the model generation timing has come (S24). The model generation timing may be, for example, the timing when the number of training data generated in the process of S22 reaches a predetermined number. That is, the model generation timing may be, for example, the timing at which the number of training data generated in the process of S22 reaches the number of learning data required to generate a learning model having sufficient determination accuracy.

As a result, when it is determined that the model generation timing has not been reached (NO in S24), the information processing apparatus 1 performs the processing after S11 again. That is, in this case, the information processing apparatus 1 repeats the processing after S11 until the model generation timing is reached.

After that, when the model generation timing comes (YES in S24), the model generation unit 114 generates a learning model by performing machine learning using each of the learning data stored in the storage area 110 (S25). Hereinafter, a specific example of the processing of S25 will be described.

[Specific example of processing of S25]
FIG. 13 is a diagram illustrating a specific example of the process of S25. Further, FIG. 14 is a diagram showing a specific example of a plurality of divided waveform data divided by the process of S14. Specifically, FIG. 13 is a specific example of generating a CNN as a learning model.

For example, as shown in FIG. 13, the model generation unit 114 inputs each of the plurality of divided waveform data (the plurality of divided waveform data shown in FIG. 14) included in the training data stored in the storage area 110 to each of the input layers. By sequentially inputting from the node, the convolution operation is performed in the convolution layer (Conv layer), and the feature pattern corresponding to each of the plurality of divided waveform data is automatically extracted. Then, the model generation unit 114 learns the optimum model parameter set by using each of the extracted feature patterns and each of the labels included in the training data.

Here, the training data used to generate the learning model is training data including the divided waveform data divided in the process of S14. Therefore, the model generation unit 114 generates a learning model having a relatively small scale (for example, a learning model in which the number of input nodes and the number of convolution layers is suppressed) in the processing of S25. Specifically, the model generation unit 114 generates, for example, a CNN having two convolutional layers and one fully connected layer in the processing of S25.

As a result, the information processing apparatus 1 can perform the learning model generation processing and the determination processing for new waveform data in a short time even when the processing capacity is weak.

Further, the information processing apparatus 1 can automate the extraction of the feature amount from the waveform data by using the divided waveform data divided in the processing of S14 as the learning data as it is.

[Processing at the judgment stage]
Next, among the details of the atrial fibrillation detection process in the first embodiment, the details of the process in the determination stage will be described. 9 to 11 are diagrams illustrating details of processing in the determination stage.

As shown in FIG. 9, the data receiving unit 111 waits until, for example, receives new waveform data (waveform data to be determined) transmitted from the patient terminal 2 (NO in S31).

Then, when new waveform data is received (YES in S31), the data preprocessing unit 112 performs trend removal on the new waveform data received in the processing of S31 (S32).

Subsequently, in this case, the data preprocessing unit 112 performs motion artifact removal on the new waveform data for which the trend has been removed in the processing of S32 (S33).

Further, the data preprocessing unit 112 divides the new waveform data from which the motion artifacts have been removed in the processing of S33 into a plurality of divided waveform data (S34).

Specifically, the data preprocessing unit 112 performs new waveform data (for example, 30 seconds or more) in which motion artifacts are removed by the processing of S33 so that the time corresponding to each divided waveform data becomes 10 seconds. (Waveform data corresponding to time) is divided.

Then, the data preprocessing unit 112 performs downsampling for each of the plurality of divided waveform data divided by the processing of S34 (S35). Further, the data preprocessing unit 112 normalizes each of the plurality of divided waveform data downsampled in the processing of S35 (S36).

That is, the information processing apparatus 1 performs preprocessing so that the new waveform data received in the processing of S31 becomes a form suitable for inputting the learning model.

Next, as shown in FIG. 10, the data determination unit 115 of the information processing apparatus 1 inputs each of the plurality of divided waveform data normalized by the processing of S36 into the learning model, so that the divided waveform data can be obtained. A determination value indicating whether or not each corresponds to atrial fibrillation is acquired (S41).

Specifically, the data determination unit 115 determines the determination value output from the model generation unit 114 when the divided waveform data is input for each of the plurality of divided waveform data normalized in the process of S36. It is acquired as information indicating whether or not each divided waveform data is waveform data corresponding to atrial fibrillation. Hereinafter, a specific example of the processing of S41 will be described.

[Specific example of processing of S41]
FIG. 15 is a diagram illustrating a specific example of the process of S41. Further, FIG. 16 is a diagram showing a specific example of one of the plurality of divided waveform data divided by the process of S34. Specifically, FIG. 15 is a specific example when determining one of the plurality of divided waveform data divided by the process of S34.

For example, as shown in FIG. 15, the model generation unit 114 has each of the input layers in the CNN generated by the processing of S25 the divided waveform data (divided waveform data shown in FIG. 16) normalized by the processing of S36. Enter from the node. Then, the model generation unit 114 acquires the determination value output from the output layer of the CNN.

Returning to FIG. 10, the data determination unit 115 identifies the divided waveform data corresponding to atrial fibrillation based on the determination value acquired in the process of S41 among the divided waveform data normalized in the process of S36. (S42).

Subsequently, the data determination unit 115 determines whether or not the divided waveform data specified in the process of S42 is continuous by a predetermined number or more in the divided waveform data normalized by the process of S36 (S43).

That is, each of the divided waveform data normalized by the processing of S36 is waveform data having no length required for detecting atrial fibrillation. Therefore, for example, when the data determination unit 115 has a plurality of continuous divided waveform data corresponding to the length required for detecting atrial fibrillation in the divided waveform data normalized by the processing of S36, the data determination unit 115 may use the divided waveform data. It is determined that the new waveform data received in the process of S31 is the waveform data corresponding to atrial fibrillation.

Specifically, for example, when the new waveform data received in the processing of S31 is the waveform data corresponding to 30 seconds, and the divided waveform data normalized in the processing of S36 is the waveform data corresponding to 10 seconds. The data determination unit 115 determines whether or not the divided waveform data specified in the process of S42 is continuous by 3 or more in the divided waveform data normalized by the process of S36.

As a result, when it is determined that the divided waveform data specified in the process of S42 is continuous by a predetermined number or more (YES in S44), the data determination unit 115 is the patient corresponding to the new waveform data received in the process of S31. It is determined whether or not the heart rate of is equal to or higher than a predetermined value (S45).

Specifically, the data determination unit 115 determines, for example, whether or not the heart rate of the patient corresponding to the new waveform data received in the process of S31 is 100 (bpm) or more.

Then, as shown in FIG. 11, when it is determined that the heart rate of the patient corresponding to the waveform data received in the process of S31 is equal to or higher than a predetermined value (YES in S51), the data determination unit 115 receives the data in the process of S31. It is determined whether or not the standard deviation of the patient's heart rate corresponding to the new waveform data is equal to or greater than a predetermined value (S52).

Specifically, the data determination unit 115 determines, for example, whether or not the standard deviation of the patient's heart rate corresponding to the new waveform data received in the process of S31 is 11 (bpm) or more.

As a result, when it is determined that the heart rate of the patient corresponding to the waveform data received in the processing of S31 is equal to or higher than a predetermined value (YES in S53), the data determination unit 115 receives new waveform data received in the processing of S31. It is determined that the waveform data corresponds to atrial fibrillation (S54).

Further, the data determination unit 115 also performs the processing of S54 in the same manner (NO of S51) when it is determined that the heart rate of the patient corresponding to the waveform data received in the processing of S31 is not equal to or higher than a predetermined value.

That is, the information processing apparatus 1 receives the divided waveform data normalized by the processing of S36 in the processing of S31 even when the divided waveform data specified by the processing of S42 are continuous by a predetermined number or more. When the heart rate of the patient corresponding to the new waveform data is equal to or higher than a predetermined value, for example, the new waveform data received in the processing of S31 is the waveform data corresponding to tachycardia or the like (symptoms other than atrial fibrillation). Judge that there is a possibility.

Then, the information processing apparatus 1 is, for example, when the heart rate of the patient corresponding to the new waveform data received in the processing of S31 is equal to or higher than a predetermined value, the information processing device 1 corresponds to the new waveform data received in the processing of S31. When the standard deviation (variation) of the heart rate is not more than a predetermined value, it is determined that the new waveform data received in the process of S31 is the waveform data corresponding to tachycardia or the like.

As a result, the information processing apparatus 1 can more accurately determine whether or not the new waveform data received in the process of S31 is the waveform data corresponding to atrial fibrillation.

Returning to FIG. 11, after the processing of S54, the result output unit 116 of the information processing apparatus 1 outputs the determination result in the processing of S54 to the patient terminal 2 (S55).

That is, when the processing of S54 is performed, the result output unit 116 provides information indicating that the new waveform data received in the processing of S31 is the waveform data corresponding to atrial fibrillation in the patient terminal 2 (processing of S31). The new waveform data received in the above is transmitted to the patient terminal 2) that has transmitted the data.

Further, when it is determined that the divided waveform data specified in the processing of S42 is not continuous by a predetermined number or more (NO in S43), the result output unit 116 causes the new waveform data received in the processing of S31 to cause atrial fibrillation. Information indicating that it is not the corresponding waveform data is transmitted to the patient terminal 2.

Further, similarly, when it is determined that the heart rate of the patient corresponding to the waveform data received in the processing of S31 is not equal to or higher than the predetermined value (NO in S53), the result output unit 116 receives a new waveform in the processing of S31. Information indicating that the data is not waveform data corresponding to atrial fibrillation is transmitted to the patient terminal 2.

[Performance evaluation results of atrial fibrillation detection processing]
Next, the performance evaluation result of the atrial fibrillation detection process will be described. FIG. 17 is a diagram showing the performance evaluation result of the atrial fibrillation detection process. Specifically, FIG. 17 shows the determination accuracy of a learning model (hereinafter, also referred to as a first learning model) generated by using learning data including waveform data (learning waveform data) stored in advance in the first DB. , A diagram illustrating information indicating the determination accuracy of a learning model (hereinafter, also referred to as a second learning model) generated by using learning data including waveform data (waveform data for learning) stored in advance in the second DB. Is.

Of the learning data used to generate the first and second learning models, the learning data corresponding to the abnormal label is the learning data including the waveform data corresponding to the atrial fibrillation, and the learning data corresponding to the normal label. Is learning data including waveform data corresponding to symptoms other than atrial fibrillation. That is, the training data corresponding to the normal label includes training data including waveform data corresponding to normal sinus rhythm, learning data including waveform data corresponding to atrial flutter, and waveform data corresponding to atrioventricular junction rhythm. Contains training data including.

The information shown in FIG. 17 includes TP (true positive rate) indicating the sensitivity of atrial fibrillation detection, TN (true negative rate) indicating the specificity of atrial fibrillation detection, and ACC (true negative rate) indicating the accuracy of atrial fibrillation detection. It has an item of accuracy). In addition, the information shown in FIG. 17 has PPV (possive positive value) indicating a positive predicted value for atrial fibrillation detection and NPV (negate positive value) indicating a negative predicted value for atrial fibrillation detection as items.

Specifically, the information shown in FIG. 17 is "92.07 (%)" and "93.53 (%), respectively, of TP, TN, ACC, PPV and NPV corresponding to the information corresponding to the first learning model. , "97.10 (%)", "73.90 (%)" and "93.16 (%)".

Further, the information shown in FIG. 17 is "94.51 (%)" and "87.10 (%)" for TP, TN, ACC, PPV and NPV corresponding to the information corresponding to the second learning model, respectively. It shows that it is "92.21 (%)", "20.12 (%)" and "96.82 (%)".

As described above, the information processing apparatus 1 in the present embodiment converts the waveform data corresponding to the first time, which is longer than the time corresponding to the waveform data required for detecting atrial fibrillation, into the second time, which is shorter than the first time. Multiple training data including a plurality of divided waveform data divided into a plurality of corresponding divided waveform data and information indicating whether or not each of the divided plurality of divided waveform data corresponds to atrial fibrillation. A learning model is generated by performing machine learning using.

That is, when the information processing apparatus 1 in the present embodiment receives, for example, new waveform data transmitted from the patient terminal 2, the patient corresponding to the new waveform data can use the learning model generated in advance. Determine if you have atrial fibrillation.

This enables the patient to determine whether or not atrial fibrillation has developed without consulting a doctor or the like. Therefore, the patient can perform early detection of atrial fibrillation.

Further, in the information processing apparatus 1 of the present embodiment, instead of using the learning waveform data corresponding to a time longer than the time required for detecting atrial fibrillation as the learning data as it is, the atriosphere is used in the learning stage. A plurality of divided waveform data (a plurality of divided waveform data generated by dividing the learning waveform data) corresponding to a time shorter than the time required for detecting fibrillation are used as training data.

Then, in the determination stage, the information processing apparatus 1 in the present embodiment divides a plurality of divided waveform data (by dividing the waveform data to be determined) corresponding to a time shorter than the time required for detecting atrial fibrillation. By inputting the generated plurality of divided waveform data) into the training model, information indicating whether or not the waveform data to be determined is the waveform data corresponding to atrial fibrillation is output.

As a result, the information processing apparatus 1 in the present embodiment has a scale as compared with the case where the learning waveform data corresponding to the time longer than the time required for detecting atrial fibrillation is used as it is. It is possible to generate a small learning model (for example, a learning model with a small number of convolutional layers and fully connected layers). Therefore, the information processing apparatus 1 can perform a learning model generation process and a new waveform data determination process by using the learning model in a short time even when the processing capacity is weak.

In the above example, the case where the information processing apparatus 1 performs the atrial fibrillation detection process by using the waveform data collected in the patient terminal 2 has been described, but the atrial fibrillation detection process is performed in the patient terminal 2. It may be done.

Specifically, the patient terminal 2 may have the same functions as the data preprocessing unit 112, the data generation unit 113, the model generation unit 114, the data determination unit 115, and the result output unit 116.

More specifically, the patient terminal 2 may generate a learning model by using, for example, learning data for learning received from the administrator terminal 3. Then, when new waveform data is collected from the patient, the patient terminal 2 uses a learning model to determine whether or not the new waveform data is waveform data corresponding to atrial fibrillation. There may be.

Further, the patient terminal 2 may store, for example, the learning model generated in the information processing apparatus 1 in the storage area 210. Then, when new waveform data is collected from the patient, the patient terminal 2 uses a learning model to determine whether or not the new waveform data is waveform data corresponding to atrial fibrillation. There may be.

In this respect, in the atrial fibrillation detection process in the present embodiment, a small-scale learning model is used as described above. Therefore, the patient terminal 2 can perform atrial fibrillation detection processing even when the processing capacity of the own terminal is not sufficient.

1: Atrial fibrillation detection device 2: Patient terminal 3: Administrator terminal 101: CPU
102: Memory 103: Communication interface 104: Storage medium 105: Bus

Claims (10)

It is an atrial fibrillation detection program that causes a computer to execute atrial fibrillation detection processing that detects atrial fibrillation by performing machine learning using waveform data of electrocardiographic signals.
A plurality of learning waveform data corresponding to the first time of a predetermined time or more, which is predetermined as the time of the waveform data required for detecting atrial fibrillation, and a plurality of waveform data corresponding to the second time of less than the predetermined time. Divide into the first waveform data
Performing machine learning using a plurality of learning data including the divided plurality of first waveform data and information indicating whether or not each of the plurality of first waveform data corresponds to the atrial fibrillation. Generates a learning model by
The new waveform data corresponding to the first time is divided into a plurality of fourth waveform data corresponding to the second time.
By inputting each of the plurality of divided fourth waveform data into the learning model, information indicating whether or not each of the plurality of fourth waveform data is waveform data corresponding to the atrial fibrillation is acquired. death,
When it is determined that each of the information corresponding to the plurality of fourth waveform data in succession is the waveform data corresponding to the atrial fibrillation, the new waveform data is the atrial fibrillation. Outputs information indicating that the waveform data corresponds to the motion.
An atrial fibrillation detection program characterized by having a computer perform processing. In claim 1,
In the process of dividing the learning waveform data,
The second waveform data is generated by performing trend removal on the learning waveform data corresponding to the first time.
The third waveform data is generated by performing motion artifact removal on the generated second waveform data.
The generated third waveform data is divided into the plurality of first waveform data corresponding to the second time.
Atrial fibrillation detection program characterized by this. In claim 1 ,
In the process of dividing the new waveform data,
The fifth waveform data is generated by performing trend removal on the new waveform data corresponding to the first time.
The sixth waveform data is generated by performing motion artifact removal on the generated fifth waveform data.
The generated sixth waveform data is divided into the plurality of fourth waveform data corresponding to the second time.
Atrial fibrillation detection program characterized by this. In claim 1 ,
The time corresponding to the continuous predetermined number of waveform data is a time equal to or longer than the predetermined time.
Atrial fibrillation detection program characterized by this. In claim 1 ,
In the output process,
When it is determined that the heart rate of the person corresponding to the new waveform data is not equal to or higher than a predetermined value, it is determined that the new waveform data is the waveform data corresponding to the atrial fibrillation.
Atrial fibrillation detection program characterized by this. In claim 5 ,
In the output process,
When it is determined that the heart rate of the person corresponding to the new waveform data is equal to or higher than the predetermined value, it is determined whether or not the standard deviation of the heart rate of the person corresponding to the new waveform data is equal to or lower than the predetermined value.
When it is determined that the standard deviation of the heart rate of the person corresponding to the new waveform data is not less than or equal to a predetermined value, it is determined that the new waveform data is the waveform data corresponding to the atrial fibrillation.
Atrial fibrillation detection program characterized by this. It is an atrial fibrillation detection device that detects atrial fibrillation by performing machine learning using waveform data of electrocardiographic signals.
A plurality of learning waveform data corresponding to the first time of a predetermined time or more, which is predetermined as the time of the waveform data required for detecting atrial fibrillation, and a plurality of waveform data corresponding to the second time of less than the predetermined time. The data division part that divides into the first waveform data,
Performing machine learning using a plurality of learning data including the divided plurality of first waveform data and information indicating whether or not each of the plurality of first waveform data corresponds to the atrial fibrillation. Has a model generator, which generates a learning model by
The data division unit divides the new waveform data corresponding to the first time into a plurality of fourth waveform data corresponding to the second time, and further divides the new waveform data into a plurality of fourth waveform data corresponding to the second time.
By inputting each of the plurality of divided fourth waveform data into the learning model, information indicating whether or not each of the plurality of fourth waveform data is waveform data corresponding to the atrial fibrillation is acquired. Data judgment unit and
When it is determined that each of the information corresponding to the plurality of fourth waveform data in succession is the waveform data corresponding to the atrial fibrillation, the new waveform data is the atrial fibrillation. It has a result output unit that outputs information indicating that it is waveform data corresponding to motion.
An atrial fibrillation detector characterized by this. It is an atrial fibrillation detection method that causes a computer to execute atrial fibrillation detection processing that detects atrial fibrillation by performing machine learning using waveform data of electrocardiographic signals.
A plurality of learning waveform data corresponding to the first time of a predetermined time or more, which is predetermined as the time of the waveform data required for detecting atrial fibrillation, and a plurality of waveform data corresponding to the second time of less than the predetermined time. Divide into the first waveform data
Performing machine learning using a plurality of learning data including the divided plurality of first waveform data and information indicating whether or not each of the plurality of first waveform data corresponds to the atrial fibrillation. Generates a learning model by
The new waveform data corresponding to the first time is divided into a plurality of fourth waveform data corresponding to the second time.
By inputting each of the plurality of divided fourth waveform data into the learning model, information indicating whether or not each of the plurality of fourth waveform data is waveform data corresponding to the atrial fibrillation is acquired. death,
When it is determined that each of the information corresponding to the plurality of fourth waveform data in succession is the waveform data corresponding to the atrial fibrillation, the new waveform data is the atrial fibrillation. Outputs information indicating that the waveform data corresponds to the motion.
An atrial fibrillation detection method characterized by having a computer perform processing. It is an atrial fibrillation detection system having an operation terminal that collects waveform data of an electrocardiogram and an atrial fibrillation detection device that detects atrial fibrillation by performing machine learning using the waveform data.
The atrial fibrillation detection device is
A plurality of learning waveform data corresponding to the first time of a predetermined time or more, which is predetermined as the time of the waveform data required for detecting atrial fibrillation, and a plurality of waveform data corresponding to the second time of less than the predetermined time. Divide into the first waveform data
Performing machine learning using a plurality of learning data including the divided plurality of first waveform data and information indicating whether or not each of the plurality of first waveform data corresponds to the atrial fibrillation. Generates a learning model by
The new waveform data corresponding to the first time is divided into a plurality of fourth waveform data corresponding to the second time.
By inputting each of the plurality of divided fourth waveform data into the learning model, information indicating whether or not each of the plurality of fourth waveform data is waveform data corresponding to the atrial fibrillation is acquired. death,
When it is determined that each of the information corresponding to the plurality of fourth waveform data in succession is the waveform data corresponding to the atrial fibrillation, the new waveform data is the atrial fibrillation. Outputs information indicating that the waveform data corresponds to the motion.
Atrial fibrillation detection system characterized by this. In claim 9 .
The operation terminal is
New waveform data corresponding to the first hour was collected from the subject, and
The collected new waveform data is transmitted to the atrial fibrillation detection device, and the data is transmitted to the atrial fibrillation detector.
The atrial fibrillation detection device is
When the new waveform data transmitted by the operation terminal is received, the received new waveform data is divided into a plurality of fourth waveform data corresponding to the second time.
Atrial fibrillation detection system characterized by this.

Images