JP7455295B2 - Biological information processing device and its control method - Google Patents

Description

Application of Article 30, Paragraph 2 of the Patent Act (1) Publication date: January 15, 2019 Publication: Journal of the Japan Medical AI Society Volume 1 “1st Japan Medical AI Society Academic Meeting Program/Abstract Collection” (2 ) Date: January 25, 2019 Meeting name: 1st Japanese Society of Medical AI Academic Meeting (3) Program app for the 66th Japanese Society of Arrhythmia and Electrocardiography Academic Meeting Publication date: July 12, 2021 Website address: http://new. jhrs. or. jp/contents_web/jhrs66/ (4) Date: July 25, 2020 Meeting name: 66th Japanese Arrhythmia Electrocardiography Society Academic Conference (5) Date: June 8, 2020 Meeting name: 39th Japan Holter Non-Invasive Electrocardiology Study Group

The present invention relates to a biological information processing device and a control method thereof.

Atrial fibrillation is known as one of the heart diseases. Atrial fibrillation begins paroxysmally and then progresses to chronicity. Since atrial fibrillation is involved in the formation of blood clots that cause cerebral infarction, it is desirable to detect atrial fibrillation at the paroxysmal stage. As a method for diagnosing the presence or absence of paroxysmal atrial fibrillation, Non-Patent Document 1 proposes a method using a P-wave summed electrocardiogram during non-paroxysmal times (during sinus rhythm).

Takahisa Yamada, Masatake Fukunami, "Clinical significance of P-wave averaged electrocardiogram", Heart, February 2006, Volume 38, Supplement No. 1, p. 28-32

Diagnostic methods using P-wave averaged electrocardiograms do not provide sufficient diagnostic performance due to examiner bias and noise. One aspect of the present invention aims to provide a technique for accurately determining whether a subject has a target disease based on electrocardiogram data.

In view of the above issues, we have developed a biological information processing device that provides electrocardiogram data during sinus rhythm that includes waveform data of a plurality of leads for each of a plurality of people, including those with a target disease and those without the target disease. and machine learning using the waveform data of one lead included in the acquired electrocardiogram data, based on the waveform data of one lead included in the electrocardiogram of the subject. and determining means for determining a model for determining whether or not the subject has the target disease.

With the above means, it is possible to accurately determine whether a subject has a target disease based on electrocardiogram data.

FIG. 1 is a block diagram illustrating the configuration of a biological information processing device according to some embodiments. FIG. 3 is a flow diagram illustrating operation of a learning phase according to some embodiments. FIG. 2 is a block diagram illustrating the structure of a model according to some embodiments. FIG. 13 is a flow diagram illustrating the operation of an inference phase according to some embodiments. FIG. 7 is a flow diagram illustrating operations in a learning phase according to another embodiment. FIG. 7 is a block diagram illustrating the structure of a model according to another embodiment. FIG. 3 is a diagram illustrating a decision tree obtained by learning according to some embodiments.

Hereinafter, embodiments will be described in detail with reference to the accompanying drawings. Note that the following embodiments do not limit the claimed invention, and not all combinations of features described in the embodiments are essential to the invention. Two or more features among the plurality of features described in the embodiments may be arbitrarily combined. In addition, the same or similar configurations are given the same reference numerals, and duplicate explanations will be omitted.

With reference to FIG. 1, the configuration of a biological information processing device 100 according to some embodiments will be described. The biological information processing device 100 determines whether the subject has a target disease based on the subject's electrocardiogram data. The target disease is a disease to be determined. The target disease may be, for example, heart disease. Furthermore, the target disease may be arrhythmia, such as paroxysmal atrial fibrillation (PAF). The embodiment described below is not limited to heart diseases such as PAF, but can determine any disease that has a correlation with electrocardiogram data.

The biological information processing device 100 may include the components shown in FIG. The processor 101 is a device that controls the overall operation of the biological information processing device 100. The processor 101 functions, for example, as a CPU (central processing unit). The memory 102 is a device that stores programs and temporary data necessary for the operation of the biological information processing device 100. The memory 102 is configured of, for example, RAM (Random Access Memory) or ROM (Read Only Memory). The operation of the biological information processing device 100 may be performed by the processor 101 executing a program stored in the memory 102, for example. Alternatively, part or all of the operations of the biological information processing device 100 may be performed by a dedicated circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array).

The input device 103 is a device for acquiring input from the user of the biological information processing device 100. The input device 103 includes, for example, a keyboard and a mouse. The output device 104 is a device for outputting information to the user of the biological information processing device 100. The output device 104 is configured by, for example, a display or a speaker. The communication device 105 is a device for the biological information processing device 100 to communicate with other devices. The other device may be a computer connected to a network (eg, the Internet or a local area network).

The storage device 106 is a device that stores data used for the operation of the biological information processing device 100. The storage device 106 is configured by a storage medium such as a hard disk drive (HDD), solid state drive (SSD), or DVD (digital versatile disc).

The operation of the biological information processing device 100 according to one embodiment will be described with reference to FIGS. 2 to 4. The biological information processing device 100 determines a model for determining whether a subject has a target disease by performing machine learning, and performs determination using this model. The operation by the biological information processing device 100 is divided into a learning phase and an inference phase. The biological information processing device 100 determines a model based on a sample set of electrocardiogram data in a learning phase, and determines whether a subject has a target disease by applying the subject's electrocardiogram data to the model in an inference phase. judge whether In the following description, the biological information processing device 100 executes both the learning phase and the inference phase. Alternatively, the learning phase and the inference phase may be executed by separate biological information processing devices.

FIG. 2 explains the operation of the biological information processing device 100 in the learning phase. In step S201, the biological information processing apparatus 100 acquires electrocardiogram data during sinus rhythm for each of a plurality of people, including those with the target disease and those without the target disease. A plurality of acquired electrocardiogram data is called a sample set. Whether the person has the target disease may be determined based on whether the person has been diagnosed with the target disease by a doctor. Each electrocardiogram data in the sample set is attached with a label indicating the presence or absence of the target disease.

The sample set may be read from a server (for example, a data server) external to the biological information processing device 100 via the communication device 105, for example. Alternatively, the sample set may be read directly from the electrocardiograph. Each electrocardiogram data included in the sample set includes waveform data of a plurality of leads. The plurality of leads may be standard 12 leads, derived 18 leads, or any other number of leads. The sample set is divided into learning data, evaluation data, and verification data.

In step S202, the biological information processing device 100 performs machine learning by individually using the waveform data of each of the plurality of leads included in the electrocardiogram data of the learning data, thereby determining the presence or absence of the target disease for each lead. Determine the model for judgment. This machine learning may be performed using, for example, a convolutional neural network (CNN).

FIG. 3 describes a model when electrocardiogram data has standard 12 leads. Separate models are determined for each of the 12 leads. The model 301a is a model (function) that inputs the waveform data of lead I and outputs a determination as to whether the subject having the waveform data has the target disease. The output of the model 301a may include a probability that the subject has the target disease and a probability that the subject does not have the target disease. Since the sum of both probabilities is 1, the output of the model 301a may include only the probability that the subject has the target disease. Similarly, the models 301b to 301l are models that input waveform data from any one of a plurality of leads and output determination results. By performing individual machine learning on the models 301a to 301l, individual parameters are determined for each model. The structures (types of hyperparameters and activation functions) of the models 301a to 301l may be the same or different. In this way, a model (for example, model 301a) for determining whether a subject has a target disease is determined for each of a plurality of leads based on the waveform data of one lead (for example, lead I). be done.

In step S203, the biological information processing device 100 uses the evaluation data to compare the determination performance of the models determined for each guidance, thereby identifying the guidance to be used in the inference phase. The determination performance may be evaluated using any one of accuracy, sensitivity, and specificity, or any combination thereof. The determination results from the model are: true positives where people with the target disease are determined to have the target disease, false positives where people without the target disease are determined to have the target disease, and people with the target disease are the target. They are classified into false negatives, in which a person is determined not to have the disease, and true negatives, in which a person who does not have the target disease is determined not to have the target disease. Accuracy is given by (number of true positives + number of true negatives)/total number. Sensitivity is given by number of true positives/(number of true positives + number of false negatives). Specificity is given by number of true negatives/(number of false positives + number of true negatives). For example, the biological information processing device 100 may specify the guidance with the highest accuracy, sensitivity, specificity, and average value as the guidance to be used in the inference phase. Since the guidance model specified in step S203 has already been determined in step S202, the biological information processing device 100 does not need to newly learn it.

By previously executing steps S201 to S203, effective leads for determining the target disease have been identified, and using electrocardiogram data from another sample set, a new model for determining the presence or absence of the target disease is created. Consider the case where the decision is made. In this case, the biological information processing device 100 performs learning in step S202 using only the waveform data of this valid lead (that is, omit learning for leads other than the valid lead), and performs learning in step S203. May be omitted. When effective guidance is a base, the biological information processing device 100 can determine the model to be used in the inference phase even if some processing is omitted in this way.

In the method of FIG. 2, step S203 was executed by the biological information processing device 100. Alternatively, step S203 may be performed by a human. In this case, the biological information processing device 100 displays the determination performance of the model determined for each guidance through the output device 104. The user (for example, a doctor) of the biological information processing device 100 specifies the guidance to be used in the inference phase based on this determination performance.

FIG. 4 explains the processing of the biological information processing device 100 in the inference phase. In step S401, the biological information processing apparatus 100 acquires electrocardiogram data of the subject. The electrocardiogram data may be read from a server (for example, a data server) external to the biological information processing device 100 via the communication device 105, for example. Alternatively, the electrocardiogram data may be read directly from the electrocardiograph. The electrocardiogram data includes at least waveform data of the lead identified in step S203.

In step S402, the biological information processing apparatus 100 determines whether the subject has the target disease by applying the subject's electrocardiogram data to the lead model specified in the learning phase. Since an effective guidance model is used, judgments can be made with high accuracy.

The operation of the biological information processing device 100 according to another embodiment will be described with reference to FIGS. 5 and 6. This embodiment differs from the embodiments of FIGS. 2-4 in that it uses ensemble learning. In step S501, similar to step S201, the biological information processing device 100 acquires a sample set of electrocardiogram data.

In step S502, the biological information processing apparatus 100 performs machine learning using the waveform data of a plurality of leads included in the electrocardiogram data of the sample set. A model for determining whether a subject has the target disease is determined based on the following. In this step, the biological information processing device 100 performs ensemble learning.

With reference to FIG. 6, a model 600 for performing ensemble learning will be described. The model 600 has a plurality of small models 601a-601l and a synthesizer 602. The number of small models 601a-601l corresponds to the number of leads of an electrocardiogram. Specifically, the small models 601a-601l correspond one-to-one to the leads of an electrocardiogram. The small models 601a-601l correspond to the models 301a-301l in FIG. 3. That is, each of the small models 601a-601l determines whether or not a subject has a target disease based on waveform data of individual leads included in the subject's electrocardiogram. For example, the small model 601a determines whether or not a subject has a target disease based on waveform data of lead I included in the subject's electrocardiogram.

A synthesizer 602 outputs a judgment result obtained by synthesizing judgment results from a plurality of small models 601a to 601l. The output from synthesizer 602 becomes the output from model 600. The synthesizer 602 may synthesize a plurality of determination results by majority vote.

When the model 600 is determined in the learning phase, the biological information processing device 100 uses the model 600 in the inference phase to determine whether the subject has the target disease. The operation of the biological information processing device 100 in the inference phase may be similar to the operation described in FIG. 4. Random forest may be used in the ensemble learning described above.

Experimental results for each of the above embodiments will be described below. Standard 12-lead electrocardiograms were obtained at a sampling frequency of 500 Hz and a sampling number of 5000 during sinus rhythm for a total of 600 people, 300 people diagnosed with PAF and 300 people not diagnosed with PAF. These 600 pieces of electrocardiogram data were divided into 480 pieces of learning data and 120 pieces of evaluation data and cross-validation was performed.

In the first scenario, learning was performed using CNN using 12 leads as one data of 12 channels. In the second scenario, according to the method shown in FIG. 2 described above, learning was performed using CNN using 12 leads as individual data of one channel. In the third scenario, ensemble learning was performed using 12 leads as individual data of one channel according to the method shown in FIG. 5 described above. All CNNs had an input length of 6 seconds, 8 convolution blocks, residual blocks and shortcut connections, no random noise, no random scaling, and no random translation. In the third scenario, a decision tree 700 shown in FIG. 7 was generated based on the learning results of the model 600 shown in FIG. Internal nodes of the decision tree 700 represent branching conditions for each induction, and external nodes represent estimation results (with or without PAF). The following judgment performance was obtained for the first to third scenarios.

It was found that the judgment performance (average values of accuracy, sensitivity, and specificity) was higher in the second scenario than in the first scenario, and higher in the third scenario than in the second scenario. Furthermore, in the second scenario, it was found that the model using aVR guidance had the highest determination performance among the multiple guidances.

The invention is not limited to the above-described embodiments, and various modifications and changes can be made within the scope of the invention.

100 biological information processing device, 301a to 301l model, 600 model

Claims (10)

A biological information processing device,
acquisition means for acquiring electrocardiogram data during sinus rhythm, including waveform data of a plurality of leads, for each of a plurality of people, including a person with a target disease and a person without the target disease;
By performing machine learning using the waveform data of one lead included in the acquired electrocardiogram data, the subject can a determining means for determining a model for determining whether or not the patient has a target disease;
A biological information processing device comprising: The determining means compares the determination performance of a plurality of models obtained by performing machine learning using individually the waveform data of each of the plurality of leads included in the acquired electrocardiogram data, The biological information processing device according to claim 1, wherein the one guidance is specified. Further comprising determining means for determining whether the subject has the target disease using a model obtained using waveform data of the identified one lead among the plurality of models. The biological information processing device according to claim 2. A biological information processing device,
acquisition means for acquiring electrocardiogram data during sinus rhythm, including waveform data of a plurality of leads, for each of a plurality of people, including a person with a target disease and a person without the target disease;
By performing machine learning using the waveform data of the plurality of leads included in the acquired electrocardiogram data, the subject can Determining means for determining a model for determining whether or not the patient has the target disease;
Equipped with
The model is
a plurality of small models for each determining whether the subject has a target disease based on waveform data of individual leads included in the electrocardiogram of the subject;
a synthesizer that synthesizes determination results from the plurality of small models;
A biological information processing device including. The biological information processing device according to claim 4 , wherein the synthesizer synthesizes the determination results from the plurality of small models by majority vote. The biometric information processing device according to claim 1 , wherein the machine learning uses a convolutional neural network. The biological information processing device according to any one of claims 1 to 6 , wherein the target disease includes paroxysmal atrial fibrillation. A program for causing a computer to function as each means of the biological information processing device according to any one of claims 1 to 7 . A method for controlling a biological information processing device, the method comprising:
an acquisition step of acquiring electrocardiogram data during sinus rhythm, including waveform data of a plurality of leads, for each of a plurality of people, including a person with a target disease and a person without the target disease;
By performing machine learning using the waveform data of one lead included in the acquired electrocardiogram data, the subject can a determination step of determining a model for determining whether or not the patient has the target disease;
A control method having. A method for controlling a biological information processing device, comprising:
An acquisition step of acquiring electrocardiogram data during sinus rhythm including waveform data of a plurality of leads for each of a plurality of people including people with a target disease and people without the target disease;
A determination step of determining a model for determining whether or not the subject has the target disease based on the waveform data of the multiple leads included in the electrocardiogram of the subject by performing machine learning using the waveform data of the multiple leads included in the acquired electrocardiogram data;
having
The model is
A plurality of small models for determining whether the subject has a target disease based on waveform data of individual leads included in an electrocardiogram of the subject;
a synthesizer for synthesizing the judgment results of the plurality of small models;
A control method comprising: