WO2022034801A1

WO2022034801A1 - Method, computer system, and program for creating histogram image, and method, computer system, and program for predicting state of object by using histogram image

Info

Publication number: WO2022034801A1
Application number: PCT/JP2021/028155
Authority: WO
Inventors: 郁郎鈴木; 直毅松田
Original assignee: 学校法人東北工業大学
Priority date: 2020-08-13
Filing date: 2021-07-29
Publication date: 2022-02-17
Also published as: JPWO2022034801A1; JP7099777B1

Abstract

The present invention provides a method for creating a histogram image that can be utilized to predict the state of an object. This method for creating a histogram image includes: a step (S501) for acquiring wavelet images; a step (S502) for dividing the wavelet images into a plurality of frequency bands; a step (S503) for creating a histogram image of spectral intensity for each of the plurality of divided wavelet images; and a step (S504) for combining the created plurality of histograms.

Description

How to create a histogram image, computer system, program, and how to predict the state of an object using a histogram image, computer system, program

The present invention relates to a method of creating a histogram image, a computer system, a program, and a method of predicting the state of an object using the histogram image, a computer system, and a program.

In non-clinical studies, research is being conducted to investigate the effects of pharmaceuticals by acquiring neural network activity such as human-derived neurons with a microelectrode array (MEA) or the like (Non-Patent Document 1).

An object of the present invention is to provide a novel method for predicting unknown properties of a target compound.

In one embodiment, the present invention provides, for example, the following items.
(Item 1)
How to create a histogram image
The process of acquiring a wavelet image and
The process of dividing the wavelet image into a plurality of frequency bands and
The process of creating a histogram of the spectral intensity for each of the multiple divided wavelet images,
A method that includes the process of combining multiple histograms created.
(Item 2)
The step of combining the plurality of histograms is
A step of converting each of the plurality of histograms into a color map, wherein the color of the color map represents a distribution ratio.
The method according to item 1, comprising the step of combining a plurality of converted color maps.
(Item 3)
The step of acquiring the wavelet image is
The process of acquiring waveform data and
The method according to item 1 or item 2, comprising the step of converting the waveform data into the wavelet image.
(Item 4)
The method according to item 3, wherein the waveform data includes an electroencephalogram waveform data.
(Item 5)
The method according to any one of items 1 to 4, wherein the plurality of frequency bands include at least 6 frequency bands.
(Item 6)
The plurality of frequency bands include at least a frequency band of about 1 Hz to about 4 Hz, a frequency band of about 4 Hz to about 8 Hz, a frequency band of about 4 Hz to about 8 Hz, a frequency band of about 8 Hz to about 12 Hz, and a frequency band of about 12 Hz to about 30 Hz. The method according to any one of items 1 to 5, comprising the frequency band of about 30 Hz to about 80 Hz, and the frequency band of about 100 Hz to about 200 Hz.
(Item 7)
It is a method of predicting the state of the target.
The step of receiving data indicating the biological signal of the target and
A step of creating a histogram image from the data showing the biological signal according to the method according to any one of items 1 to 6.
In the step of inputting the histogram image into an image recognition model trained by the training data set, the training data set is a plurality of known objects of the subject according to the method according to any one of items 1 to 6. A process, including multiple training images created from data showing biometric signals in the state of
A method comprising the steps of processing the histogram image in the image recognition model and outputting the state of the object.
(Item 8)
The method according to item 7, wherein the plurality of known states of the subject include a state having convulsive symptoms, a state not having convulsive symptoms, and a state having a precursor of convulsive symptoms. ..
(Item 9)
It is a method of constructing an image recognition model for predicting the state of an object.
, A step of creating a plurality of training images from data showing biological signals in a plurality of known states of a target according to the method according to any one of items 1 to 6.
A method comprising learning a training data set comprising the plurality of training images.
(Item 10)
It is a method of predicting the state of the target.
The step of receiving data indicating the biological signal of the target and
In the step of creating a histogram image from the data showing the biological signal, the histogram image is a color map in which the first axis represents the spectral intensity, the second axis represents the frequency, and the color represents the distribution ratio. There is a process and
A step of inputting the histogram image into an image recognition model trained by a training dataset, wherein the training dataset is a plurality of data created from data showing biometric signals in a plurality of known states of the subject. Processes and processes, including training histogram images,
A method comprising the steps of processing the histogram image in the image recognition model and outputting the state of the object.
(Item 11)
A computer system for creating histogram images,
How to get a wavelet image,
A means for dividing the wavelet image into a plurality of frequency bands,
A means of creating a histogram of spectral intensity for each of multiple post-split wavelet images,
A computer system with a means of combining multiple histograms created.
(Item 11A)
The computer system according to item 11, which comprises the features according to one or more of the above items.
(Item 12)
There is a program for creating a histogram image, the program is executed in a computer system equipped with a processor, and the program is
The process of acquiring a wavelet image and
The process of dividing the wavelet image into a plurality of frequency bands and
The process of creating a histogram of the spectral intensity for each of the multiple divided wavelet images,
A program that causes the processor to perform a process including a step of combining a plurality of created histograms.
(Item 12A)
The program according to item 12, which comprises the features described in one or more of the above items.
(Item 12B)
A storage medium for storing the program according to item 12 or item 12A.
(Item 13)
A computer system for predicting the state of an object,
A receiving means for receiving data indicating the biological signal of the target, and
A means for creating a histogram image from data showing the biological signal. In the histogram image, the first axis represents the spectral intensity, the second axis represents the frequency, and the color represents the distribution ratio. The means of creation and
An image recognition model trained by a training dataset, wherein the training dataset includes image recognition including a plurality of training histogram images created from data showing biometric signals in a plurality of known states of the subject. With the model
A computer system including an output means for outputting the state of the object.
(Item 13A)
13. The computer system of item 13, comprising the features described in one or more of the above items.
(Item 14)
A program for predicting the state of a target, wherein the program is executed in a computer system including a processor, and the program is
The step of receiving data indicating the biological signal of the target and
In the step of creating a histogram image from the data showing the biological signal, the histogram image is a color map in which the first axis represents the spectral intensity, the second axis represents the frequency, and the color represents the distribution ratio. There is a process and
A step of inputting the histogram image into an image recognition model trained by a training dataset, wherein the training dataset is a plurality of data created from data showing biometric signals in a plurality of known states of the subject. Processes and processes, including training histogram images,
A program for causing the processor to perform a process including a step of processing the histogram image in the image recognition model and outputting the state of the target.
(Item 14A)
The program of item 14, comprising the features described in one or more of the above items.
(Item 14B)
A storage medium for storing the program according to item 14 or item 14A.
(Item 15)
It is a method for predicting the characteristics of the target compound.
The step of receiving post-administration data indicating the biological signal of the subject after the subject compound is administered to the subject, and the step of receiving the post-administration data.
A step of creating a histogram image from the post-administration data according to the method according to any one of items 1 to 6.
The process of extracting the feature vector from the histogram image and
In the step of comparing the feature quantity vector with the plurality of reference feature quantity vectors, each of the plurality of reference feature quantity vectors is a plurality of known compounds according to the method according to any one of items 1 to 6. Includes a feature vector extracted from a histogram image created from baseline post-dose data showing each biometric signal after administration to the subject.
A method comprising the step of predicting the properties of the target compound based on the result of the comparison.
(Item 16)
The step of comparing the feature quantity vector with the plurality of reference feature quantity vectors is
The process of creating a feature map by mapping the feature vector, and
In the step of comparing the feature amount map with the plurality of reference feature amount maps, each of the plurality of reference feature amount maps is a map created by mapping each of the reference feature amount vectors. , The method of item 15, comprising the process.
(Item 17)
The method according to item 16, wherein the step of comparing the feature amount map with the plurality of reference feature amount maps comprises identifying at least one reference feature amount map similar to the feature amount map.
(Item 18)
The method according to item 16, wherein the step of comparing the feature amount map with the plurality of reference feature amount maps includes ranking the plurality of reference feature amount maps in an order similar to the feature amount map.

According to the present invention, it is possible to provide a method of creating a histogram image that can be used to predict the state of an object. Further, according to the present invention, it is possible to provide a method of predicting the state of an object using a histogram image and the like. This makes it possible to predict the unknown properties of the target compound.

The figure which shows an example of the configuration of the computer system 100 for predicting the state of an object. The figure which shows an example of the configuration of a processor 120 The figure which shows an example of the structure of the image creating means 121 The figure which shows an example of the structure of the processor 120'. Diagram showing an example of a histogram image Diagram showing another example of a histogram image The figure which shows an example of the feature amount vector extracted by the extraction means 124. A diagram showing an example of a feature map created by the comparison means 125. A diagram showing an example of one standard feature map that combines multiple reference feature maps. The figure which shows an example of the structure of the neural network 300 used for constructing an image recognition model 122. A flowchart showing an example of processing by the computer system 100 for predicting the state of the target. The figure which shows the specific example which creates a histogram image by process 500 A flowchart showing an example of processing by the computer system 100 for predicting the state of the target. A flowchart showing an example of processing by the computer system 100 for predicting the state of the target. A flowchart showing an example of processing by the computer system 100 for predicting the state of the target. The figure which shows the result of the experiment 1 of Example 1. The figure which shows the result of the experiment 2 of Example 1. The figure which shows the result of Example 2. The figure which shows the result of Example 2. The figure which shows the result of Example 2.

Hereinafter, the present invention will be described. It should be understood that the terms used herein are used in the meaning commonly used in the art unless otherwise noted. Accordingly, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. In case of conflict, this specification (including definitions) takes precedence.

1. 1. Definition In the present specification, the "object" means a living body for which a state is predicted. The subject may be a human, a non-human animal, or a human and an animal.

In the present specification, the "target compound" means a target compound whose characteristics are predicted. The target compound may be an unknown compound or a known compound. The properties of the target compound include, but are not limited to, for example, efficacy, toxicity and mechanism of action.

In the present specification, the "medicinal effect" is the effect that occurs when a drug is applied to a target. For example, when the drug is an anticancer drug, the efficacy is directly affected by the reduction of the cancer area under X-ray observation, the delay of the progression of the cancer, and the extension of the survival time of the cancer patient. It may be an effect or an indirect effect such as a decrease in biomarkers that correlate with the progression of the cancer. As used herein, "medicinal efficacy" is intended to be effective under any applicable conditions. For example, if the drug is an anti-cancer drug, the efficacy may be an effect on a particular subject (eg, a man over 80 years of age) or a particular application condition (eg, other anti-cancer therapies). It may be the effect in combination with). In one embodiment, the agent may have a single efficacy or may have multiple efficacy. In one embodiment, the agent may have different efficacy under different application conditions. In general, medicinal effect refers to the effect aimed at achieving.

As used herein, "toxicity" is an unfavorable effect that occurs when a drug is applied to a subject. In general, toxicity is an effect that is different from the intended effect of the drug. Toxicity may occur by a different mechanism of action than the medicinal effect, or it may occur by the same mechanism of action as the medicinal effect. For example, when the drug is an anticancer drug, hepatotoxicity due to normal hepatocyte killing may occur at the same time as the medicinal effect of cancer cell killing through the mechanism of action of cell proliferation suppression, and cell proliferation. At the same time as the medicinal effect of cancer cell killing through the mechanism of action of inhibition, toxicity of neurological dysfunction may occur through the mechanism of action of membrane stabilization.

In the present specification, the "mechanism of action" is a mode in which a drug interacts with a biological mechanism. For example, if the drug was an anticancer drug, the mechanism of action would be to activate the immune system, kill fast-growing cells, block proliferative signaling, block specific receptors, or block specific genes. It can be an event of various levels, such as transcriptional inhibition. Once the mechanism of action has been identified, the efficacy, toxicity and / or appropriate mode of use can be predicted based on the accumulated information.

In the present specification, the "wavelet image" means a spectrogram obtained by wavelet transforming the data in a certain time window. The wavelet image is a color map in which the first axis represents time, the second axis represents frequency, and the color represents spectral intensity.

In the present specification, "about" means ± 10% of the value that follows.

2. 2. Prediction of Target State The inventor of the present invention uses artificial intelligence, so-called AI (Artificial Intelligence), to predict the state of a subject to which a target compound is administered in order to predict unknown properties of the target compound. Was developed. This artificial intelligence learns the relationship between an image created from data acquired when a plurality of known compounds are administered to a subject and the state of the subject by the known compound. When an image created from the data acquired when the target compound is administered to the subject is input to this artificial intelligence, the artificial intelligence can predict and output the state of the target. As a result, the characteristics of the target compound can be predicted based on this state.

This artificial intelligence can predict the state of the subject from the viewpoint of, for example, a neurological disease. The subject's condition in terms of a neurological disorder includes, for example, a condition with a neurological disorder, a condition without a neurological disorder, and a condition with a precursory state of the neurological disorder. Neurological disorders include, but are not limited to, convulsions, epilepsy, ADHD, dementia, autism, schizophrenia, depression and the like. This artificial intelligence is, for example, a condition having at least one of spasm, epilepsy, ADHD, dementia, aura, schizophrenia, depression, spasm, epilepsy, ADHD, dementia. , Aura, schizophrenia, depression, or at least of spasm, epilepsy, ADHD, dementia, aura, schizophrenia, depression It is possible to predict whether or not the patient has one precursor. For example, whether or not a person has a neurological disorder appears in the data obtained from the subject as a characteristic of frequency. For example, the component of the gamma wave band (about 30 to about 50 Hz) of the brain wave data acquired from the subject becomes stronger in the case of the subject having the symptom of epilepsy and the subject having the symptom of ADHD, and the subject having the symptom of dementia. And weakened in subjects with symptoms of posting ataxia. This artificial intelligence makes it possible to predict the state of an object by learning the characteristics peculiar to such a specific frequency band. This prediction can be used for the diagnosis of neurological disorders or as an index for the diagnosis of neurological disorders.

Based on this prediction, whether or not the target compound has the property of inducing a neurological disease (drug efficacy, toxicity, or mechanism of action) or whether or not it has the property of treating a neurological disease. It is possible to predict. This prediction can be applied, for example, to drug discovery of neurological disorders and evaluation of neurotoxicity.

For example, this artificial intelligence relates to a relationship between an image created from data acquired when a drug that induces convulsions (eg, 4-aminopyridine (4-AP)) is administered and the state of the subject (convulsive state). , The relationship between the image created from the data acquired when the drug was not administered and the state of the subject (non-convulsive state), acquired immediately after administration of the drug that induces convulsions (eg 4-AP) We are learning the relationship between the image created from the data and the state of the subject (convulsive precursor state). When an image created from the data obtained when the target compound is administered to the subject is input to this artificial intelligence, this artificial intelligence is in a state where the subject has convulsive symptoms or has convulsive symptoms. It is possible to predict whether the patient does not have the condition or has a precursor of convulsive symptoms. Based on this prediction, it is possible to predict whether or not the target compound has a convulsive-inducing property (medicinal efficacy, toxicity, or mechanism of action).

By learning the state of the subject when the known compound is administered for various medicinal effects, toxicity, and mechanism of action, what kind of medicinal effect or what kind of toxicity the target compound has. Or, it becomes possible to predict what kind of mechanism of action it has. This makes it possible to predict the efficacy, toxicity, and mechanism of action of the target compound, and for example, it becomes easy to classify which known compound the target compound is similar to. In addition, by predicting the efficacy, toxicity, and mechanism of action of the target compound, it is possible to promote the development of safe and secure compounds.

Such artificial intelligence can be realized by a computer system for predicting the state of an object described below.
Hereinafter, embodiments of the present invention will be described with reference to the drawings.

3. 3. Configuration of a Computer System for Predicting the State of a Target FIG. 1 shows an example of a configuration of a computer system 100 for predicting the state of a target according to an embodiment of the present invention.

The computer system 100 includes a receiving means 110, a processor 120, a memory 130, and an output means 140. The computer system 100 may be connected to the database unit 200.

The receiving means 110 is configured to be able to receive data from the outside of the computer system 100. The receiving means 110 may receive data from the outside of the computer system 100 via a network, for example, or data from a storage medium (for example, a USB memory, an optical disk, etc.) connected to the computer system 100 or a database unit 200. May be received. When receiving data over a network, the type of network does not matter. The receiving means 110 may receive the data using, for example, a wireless LAN such as Wi-fi, or may receive the data via the Internet.

The receiving means 110 is configured to receive, for example, the data acquired from the target. For example, the receiving means 110 can receive the data acquired from the subject when the known compound or the subject compound is administered. For example, the receiving means 110 can receive the data acquired from the subject when the compound is not administered. The data acquired from the subject is data indicating the biological signal of the subject, which can be acquired by any known method. The data indicating the biometric signal of interest can be any data as long as it is time series data or wave signal, for example, brain wave data, cultured nerve cell data, brain slice data, myocardial cell data, impedance measurement. Includes, but is not limited to, data, electrocardiogram data, myocardiogram data, brain surface blood flow data, and cerebral magnetic field data.

The electroencephalogram data is, for example, time-series potential data measured using an electroencephalograph. Specifically, the scalp electroencephalogram (EEG) data is potential time-series data of the body surface due to nerve activity measured using electrodes attached to the scalp. The cortical electroencephalogram (ECoG) data is time-series data of cortical potential due to neural activity measured using electrodes placed on the cortex in the skull.

Data on cultured nerve cells and data on cardiomyocytes can be obtained using, for example, a microelectrode array (MEA). This data provides, for example, the action potentials and synaptic current components of nerve cells or myocardial cells when a nerve cell or myocardial cell is cultured on MEA and a known compound or target compound is administered to the cultured nerve cell or myocardial cell. Obtained by measuring. For example, CMOS-MEA that utilizes CMOS may be used as MEA. By using CMOS-MEA, it becomes possible to acquire relatively high resolution data.

The electrocardiogram data can be acquired using, for example, an electrocardiograph. The electromyogram data can be acquired using, for example, an electromyogram. Blood flow data on the brain surface can be obtained using, for example, functional near-infrared spectroscopy (fNIRS: functional Near-Infrared Spectroscopy). Brain magnetic field data can be obtained, for example, using a magnetoencephalogram (MEG).

The receiving means 110 can receive the data acquired from the target in any format. The receiving means 110 may receive the data acquired from the target as raw data. For example, the receiving means 110 may receive the data acquired from the target as a wavelet image.

The receiving means 110 can receive arbitrary time-series data or wave signals other than the data acquired from the target. Arbitrary time-series data includes, but is not limited to, for example, data indicating sound waves, data indicating seismic waves, data indicating fluctuations in stock prices, and the like.

The data received by the receiving means 110 is passed to the processor 120 for subsequent processing.

The processor 120 controls the operation of the entire computer system 100. The processor 120 reads a program stored in the memory 130 and executes the program. This makes it possible to make the computer system 100 function as a device that performs a desired step. The processor 120 may be implemented by a single processor or by a plurality of processors. The data processed by the processor 120 is passed to the output means 140 for output.

The memory 130 stores a program for executing processing in the computer system 100, data required for executing the program, and the like. In the memory 130, for example, for constructing a program for creating a histogram image (for example, a program for realizing the process shown in FIG. 5 described later) or an image recognition model 122 for predicting the state of the target. (For example, a program that realizes the process shown in FIG. 7 described later) or a program for predicting the target state (for example, a program that realizes the process shown in FIG. 8 described later) is stored. There is. An application that implements an arbitrary function may be stored in the memory 130. Here, it does not matter how the program is stored in the memory 130. For example, the program may be pre-installed in memory 130. Alternatively, the program may be installed in memory 130 by being downloaded over the network. Alternatively, the program may be stored on a machine-readable non-transient storage medium. The memory 130 may be implemented by any storage means.

The output means 140 is configured to be able to output data to the outside of the computer system 100. It does not matter in what manner the output means 140 enables the information to be output from the computer system 100. For example, when the output means 140 is a display screen, information may be output to the display screen. Alternatively, when the output means 140 is a speaker, information may be output by voice from the speaker. Alternatively, when the output means 140 is a data writing device, the information may be output by writing the information to the storage medium or the database unit 200 connected to the computer system 100. Alternatively, when the output means 140 is a transmitter, the transmitter may output information by transmitting information to the outside of the computer system 100 via a network. In this case, the type of network does not matter. For example, the transmitter may transmit information via the Internet or may transmit information via LAN. For example, the output means 140 converts the data into a format that can be handled by the hardware or software to which the data is output, or adjusts the response speed to a response speed that can be handled by the hardware or software to which the data is output, and outputs the data. You may try to do it.

The database unit 200 connected to the computer system 100 may store, for example, data indicating a plurality of target biological signals in known states. Data showing biological signals in a plurality of known states of the subject may be stored in the database unit 200 in association with, for example, a known compound that induces the state. The database unit 200 may store, for example, data output by the computer system 100 (for example, a predicted target state, a created histogram image).

In the example shown in FIG. 1, the database unit 200 is provided outside the computer system 100, but the present invention is not limited thereto. It is also possible to provide the database unit 200 inside the computer system 100. At this time, the database unit 200 may be mounted by the same storage means as the storage means for mounting the memory 130, or may be mounted by a storage means different from the storage means for mounting the memory 130. In any case, the database unit 200 is configured as a storage unit for the computer system 100. The configuration of the database unit 200 is not limited to a specific hardware configuration. For example, the database unit 200 may be composed of a single hardware component or may be composed of a plurality of hardware components. For example, the database unit 200 may be configured as an external hard disk device of the computer system 100, or may be configured as a storage on the cloud connected via a network.

FIG. 2A shows an example of the configuration of the processor 120.

The processor 120 includes at least an image creating means 121 and an image recognition model 122.

The image creating means 121 is configured to create a histogram image from the data received by the receiving means 110. The data received by the receiving means 110 is a wavelet image. When the data received by the receiving means 110 is not a wavelet image, the image creating means 121 creates a wavelet image from the data received by the receiving means 110, and then creates a histogram image from the created wavelet image. You may do so. Here, the histogram image is an image showing the relationship between the spectral intensity, the frequency, and the distribution ratio. In one example, the histogram image is a color map in which the first axis represents the spectral intensity, the second axis represents the frequency, and the colors represent the distribution ratio. In another example, the histogram image is a three-dimensional graph in which the first axis represents the spectral intensity, the second axis represents the frequency, and the third axis represents the distribution ratio.

FIG. 3A shows an example of a histogram image, and FIG. 3B shows another example of a histogram image.

In the example shown in FIG. 3A, the histogram image is a color map. The horizontal axis represents the spectral intensity, and the larger the value, the stronger the spectral intensity. The vertical axis represents the frequency, and the value between 4 Hz and 250 Hz is shown. The lightness and darkness of the color represents the distribution ratio. The brighter the color, the higher the distribution ratio, and the darker the color, the lower the distribution ratio.

In the example shown in FIG. 3B, the histogram image is a three-dimensional graph. The first axis represents the spectral intensity, the second axis represents the frequency, and the third axis represents the distribution ratio.

The image recognition model 122 is configured to output the target state by processing the input image. The image recognition model 122 is trained by the training data set. When the histogram image created by the image creating means 121 is input to the image recognition model 122, the image recognition model 122 predicts and outputs the state of the target. The state of the subject predicted by the image recognition model 122 includes, for example, a state having a neurological disease, a state not having a neurological disease, and a state having a precursory state of the neurological disease. .. Neurological disorders include, but are not limited to, convulsions, epilepsy, ADHD, dementia, autism, schizophrenia, depression and the like. The subject's condition predicted by the image recognition model 122 is, for example, convulsions, epilepsy, ADHD, dementia, aura, schizophrenia, a condition having at least one of depression, convulsions. , Epilepsy, ADHD, dementia, aura, schizophrenia, depression, convulsions, epilepsy, ADHD, dementia, autism, schizophrenia, depression Includes a condition having at least one precursor of. In certain embodiments, the subject's condition predicted by the image recognition model 122 includes a condition with convulsive symptoms, a condition without convulsive symptoms, and a condition with a precursor to convulsive symptoms.

The processor 120 further includes a learning means 123.

The learning means 123 is configured to train the image recognition model 122 by learning the training data set. The training data set includes, for example, a plurality of training images created from data showing biological signals in a plurality of known states of the subject. The plurality of training images may be histogram images, and the histogram images may be those created by the image creating means 121 or may be received from outside the computer system 100.

Multiple known states of the subject used for the training dataset are achieved, for example, by administering to the subject multiple known compounds. A plurality of known compounds have known properties, and it is known what kind of state the subject will be when administered to the subject. The learning means 123 learns, for example, the relationship between a plurality of training images included in a training data set and known states corresponding to the plurality of training images. For example, the learning means 123 can train the image recognition model 122 by learning a plurality of training images as input teacher data and learning the corresponding states as output teacher data. As a result, the image recognition model 122 can predict and output the target state.

FIG. 4 shows an example of the structure of the neural network 300 used for constructing the image recognition model 122.

The neural network 300 has an input layer, at least one hidden layer, and an output layer. The number of nodes in the input layer of the neural network 300 corresponds to the number of dimensions of the input data. The number of nodes in the output layer of the neural network 300 corresponds to the number of dimensions of the output data. The hidden layer of the neural network 300 can contain any number of nodes. The neural network can be, for example, a convolutional neural network (CNN).

The weighting factor of each node of the hidden layer of the neural network 300 can be calculated based on the training data set. The process of calculating this weighting factor is the process of learning. For example, the value of the output layer when a plurality of training images created from the data showing the biological signals in a plurality of known states of the target are input to the input layer becomes the value indicating the corresponding known state. , The weighting factor of each node can be calculated. This can be done, for example, by backpropagation (backpropagation). A large amount of training data set may be preferable for training, but if it is too large, overfitting is likely to occur.

For example, in order to construct an image recognition model 122 that can predict a state with convulsive symptoms, a state without convulsive symptoms, and a state with aura of convulsive symptoms as known states. When training, the learning process utilizes multiple training images created from data showing biological signals obtained from the subject when a known compound known to induce convulsions is administered to the subject. For example, the set of (training image input to the input layer, value of the output layer) is (training image created from data showing biological signals in a state of having convulsive symptoms, [1]).
(Training image created from data showing biological signals in the absence of convulsive symptoms, [0]),
(Training image created from data showing biological signals with signs of convulsive symptoms, [0.5])
(1 of the output layer is a value indicating a state of having convulsive symptoms, 0 is a value indicating a state of not having convulsive symptoms, and 0.5 is a value indicating a sign of convulsive symptoms. It is a value indicating the state of being in.). In the learning process, the weighting factor of each node is calculated so as to satisfy these pairs. The ideal output of the neural network 300 in which the weighting coefficient of each node is calculated in this way is, for example, an output layer when an image created from data showing a biological signal in a state of having convulsive symptoms is input. Node is to output 1. However, in reality, it is difficult to obtain an ideal output due to the influence of noise and the like mixed in the data indicating the biological signal.

Referring again to FIG. 2A, the image recognition model 122 may preferably include, for example, a feature quantity extraction model 1221 and at least one state prediction model 1222, as shown in FIG. 2A. This is because the accuracy of prediction by the image recognition model 122 is improved.

The feature amount extraction model 1221 is configured to extract the feature amount of the input image by processing the input image. The feature extraction model 1221 is trained by the training data set. For example, when the histogram image created by the image creating means 121 is input to the feature amount extraction model 1221, the feature amount extraction model 1221 extracts the feature amount of the histogram image. The feature amount extracted by the feature amount extraction model is a numerical value of what kind of features the input image has.

As the feature amount extraction model 1221, for example, an existing trained image recognition model (for example, AlexNet, VGG-16, etc.) may be used, or a model constructed by further training an existing trained image recognition model. It may be a model constructed by training a neural network as shown in FIG. For example, with the existing trained image recognition model AlexNet, it is possible to extract 4096-dimensional features from the input image. The number of dimensions of the feature amount that can be extracted by the feature amount extraction model can be any number of 2 or more. As the number of dimensions increases, the accuracy of prediction by the image recognition model 122 improves, but as the number of dimensions increases, the processing load increases.

When the feature quantity extraction model 1221 is a model constructed by further training an existing trained image recognition model or a model constructed by training a neural network as shown in FIG. 4, the feature quantity extraction model 1221 Is trained by learning means 123. The feature extraction model 1221 is trained to be able to output a feature that captures the features of a point or wave well. For example, the weighting factor of each node of the hidden layer is a value in which the value of the output layer when a point image or a wave image is input to the input layer as a training image indicates the characteristics of the corresponding point image or the wave image. The weighting factor of each node can be calculated so as to be.

For example, the waveform image used as the training image is a sine wave image, various noise waveform images, and waveform images of different brain regions, and the corresponding output value is the name of the corresponding waveform. The feature amount extraction model 1221 trained in this way is compared with the feature amount extracted by the existing trained image recognition model, and the target brain wave data (for example, a histogram image created from the brain wave data, the brain wave data). It will be possible to extract features that capture the features of the waveform image).

The state prediction model 1222 is configured to receive a part or all of the feature amount extracted by the feature amount extraction model 1221 and process the received feature amount to output the target state. The state prediction model 1222 is trained by the training data set. When the feature amount extracted by the feature amount extraction model 1221 is input to the state prediction model 1222 for the image created from the data showing the biological signals in a plurality of known states of the target, the state prediction model 1222 displays the state of the target. Predict and output. The state of the subject predicted by the state prediction model 1222 is, for example, a state in which the subject has convulsive symptoms, a state in which the subject does not have convulsive states, and a state in which the subject has signs of convulsive symptoms. include. This, in turn, makes it possible to predict whether or not the target compound has the property of inducing convulsions.

The state prediction model 1222 is, for example, a model constructed by training a neural network as shown in FIG. 4, and the state prediction model 1222 is trained by the learning means 123. For example, the value of the output layer when the feature amount extracted by the feature amount extraction model 1221 from a plurality of training images created from the data showing the biological signals in a plurality of target states is input to the input layer corresponds to the value of the output layer. The weighting factor for each node can be calculated to be a value indicating the state.
For example, in order to construct a state prediction model 1222 that can predict a state having convulsive symptoms, a state not having convulsive symptoms, and a state having a precursor of convulsive symptoms as known states. When training, the learning process extracts features from multiple training images created from data showing biological signals obtained from the subject when a known compound known to induce seizures is administered to the subject. The feature quantity extracted by the model 1221 is used. For example, the set of (features input to the input layer, values of the output layer) is
(Features extracted by the feature extraction model 1221 from the training image created from the data showing the biological signals in the state of having convulsive symptoms, [1]),
(Features extracted by feature extraction model 1221 from training images created from data showing biological signals without convulsive symptoms, [0]),
(Features extracted by the feature extraction model 1221 from the training image created from the data showing the biological signals in the state of having aura of convulsive symptoms, [0.5])
(1 of the output layer is a value indicating a state of having convulsive symptoms, 0 is a value indicating a state of not having convulsive symptoms, and 0.5 is a value indicating a sign of convulsive symptoms. It is a value indicating the state of being in.). In the learning process, the weighting factor of each node is calculated so as to satisfy these pairs. The ideal output of the neural network 300 in which the weight coefficient of each node is calculated in this way is, for example, a feature quantity extraction model 1221 from an image created from data showing a biological signal in a state of having a spasm symptom. The node of the output layer outputs 1 when the extracted feature amount is input. However, in reality, it is difficult to obtain an ideal output due to the influence of noise and the like mixed in the data indicating the biological signal.

The image recognition model 122 can include a plurality of state prediction models 1222. When the image recognition model 122 includes a plurality of state prediction models 1222, each of the plurality of state prediction models 1222 can be trained, for example, by performing the same processing as the above-mentioned learning process. For example, by making the state of the target to be predicted different from each other by the plurality of state prediction models 1222, it becomes possible to selectively use the state prediction model 1222 according to the state of the target to be predicted. As a result, the processing load in learning each of the plurality of state prediction models 1222 can be reduced, and the accuracy in predicting various states can be improved.

FIG. 2B shows an example of the configuration of the image creating means 121.

The image creating means 121 includes a dividing means 1211, a histogram creating means 1212, and a joining means 1213.

The dividing means 1211 is configured to acquire a wavelet image and divide the acquired wavelet image into a plurality of frequency bands. The wavelet image may be one received by the receiving means 110, or may be obtained by wavelet transforming the data received by the receiving means 110. The wavelet image may be, for example, a wavelet image created by wavelet transforming the brain wave data received by the receiving means 110.

The dividing means 1211 can divide the wavelet image into a plurality of frequency bands by any known image processing technique.

The dividing means 1211 can divide the wavelet image in any unit. For example, the dividing means 1211 may divide the wavelet image for each pixel unit (for example, 1 μHz, 1 MHz, 1 Hz, 1 kHz, etc.) in the frequency axis direction of the wavelet image. Alternatively, the dividing means 1211 may divide the wavelet image into a plurality of pixels in the frequency axis direction of the wavelet image. For example, when dividing brain wave data, the dividing means 1211 has at least 6 frequency bands (for example, a delta wave band (about 1 Hz to about 4 Hz), a θ wave band (about 4 Hz to about 8 Hz), and an α wave band (about). 8Hz to about 12Hz), β wave band (about 12Hz to about 30Hz), γ wave band (about 30Hz to about 80Hz), ripple wave band (about 80Hz to about 200Hz)
)), And in a preferred embodiment, it can be divided into at least 10 frequency bands. Further, when dividing the brain wave data, it is preferable to divide the brain wave data so as to include a frequency band of about 1 Hz to about 250 Hz, and in particular, a frequency band of about 1 Hz to about 4 Hz, a frequency band of about 4 Hz to about 8 Hz, and so on. Divide to include a frequency band of 4 Hz to about 8 Hz, a frequency band of about 8 Hz to about 12 Hz, a frequency band of about 12 Hz to about 30 Hz, a frequency band of about 30 Hz to about 80 Hz, and a frequency band of about 100 Hz to about 200 Hz. Is preferable. By including the frequency bands divided in this way, it is possible to capture the characteristics peculiar to a specific state (or disease) that may appear in at least one of these frequency bands, and by using this for learning, it is possible to capture the characteristics. You will be able to predict that you are in a particular condition (or disease).

The number of divisions by the division means 1211 determines the number of pixels in the frequency axis direction of the final histogram image. The larger the number of divisions by the division means 1211, the larger the number of pixels in the frequency axis direction of the final histogram image, and the smaller the number of divisions, the smaller the number of pixels in the frequency axis direction of the final histogram image.

The wavelet image after being divided becomes a wavelet image from which the frequency component has been removed. That is, the output by the dividing means 1211 is a one-dimensional color map having a time axis and the colors representing the spectral intensities. For example, when the wavelet image is divided into a plurality of pixels in the frequency axis direction of the wavelet image, the frequency component can be removed by taking the average value of the spectral intensities of the plurality of pixels at each time. Here, instead of taking the average value, the frequency component may be removed by performing any other operation (for example, taking the maximum value, taking the minimum value, taking the intermediate value, etc.).

The histogram creating means 1212 is configured to create a histogram of the spectral intensity for each of the plurality of divided wavelet images after being divided by the dividing means 1211. A histogram is a graph having a first axis representing spectral intensity and a second axis representing frequency. The histogram creating means 1212 can create a histogram by counting the number of appearances of each spectral intensity in the divided wavelet image.

The joining means 1213 is configured to join a plurality of created histograms. The coupling means 1213 combines a plurality of histograms in the order of frequency, whereby a histogram image is created.

The coupling means 1213 can create a histogram image which is a three-dimensional graph by, for example, combining a plurality of two-dimensional histograms three-dimensionally in the frequency axis direction.

The combining means 1213 can create a histogram image by, for example, converting each of a plurality of histograms into a color map and combining the plurality of color maps. Here, in the color map converted from the plurality of histograms, the colors represent the distribution ratio. That is, by converting a two-dimensional histogram into a color map, it is possible to generate a one-dimensional color map having a time axis and representing a distribution ratio of colors. By combining a plurality of one-dimensional color maps two-dimensionally in the frequency axis direction, it is possible to create a histogram image which is a two-dimensional color map.

FIG. 2C shows an example of the configuration of the processor 120'. The processor unit 120'may be a processor included in the computer system 100 instead of the processor 120, or may be a processor included in the computer system 100 in addition to the processor 120. Hereinafter, a processor included in the computer system 100 instead of the processor 120 will be described. The same components as those in the above-mentioned example with reference to FIG. 2A are designated by the same reference numbers, and detailed description thereof will be omitted here.

The processor 120'includes at least an image creating means 121, an extracting means 124, and a comparison means 125.

As described above, the image creating means 121 is configured to create a histogram image from the data received by the receiving means 110. The created histogram image is passed to the extraction means 124.

The extraction means 124 is configured to extract a feature amount vector from a histogram image. Extraction means 124 can be trained by the training data set. For example, when the histogram image created by the image creating means 121 is input to the extracting means 124, the extracting means 124 extracts a plurality of feature quantities (that is, feature quantity vectors) of the histogram image. The feature amount extracted by the extraction means 124 is a numerical value of what kind of features the input histogram image has.

The extraction means 124 may have the same configuration as, for example, the feature amount extraction model 1221. The extraction means 124 may use, for example, an existing trained image recognition model (for example, AlexNet, VGG-16, etc.), or is a model constructed by further training an existing trained image recognition model. It may be a model constructed by training a neural network as shown in FIG. For example, with the existing trained image recognition model AlexNet, it is possible to extract a feature amount vector having a component of 4096 dimensions from the input image. The number of dimensions of the feature amount that can be extracted by the extraction means 124 can be any number of 2 or more. As the number of dimensions increases, the accuracy of prediction improves, but as the number of dimensions increases, the processing load increases.

For example, the image creating means 121 creates a histogram image from the data (post-administration data) acquired from the subject when the known compound or the target compound is administered, and the extraction means 124 creates a feature amount vector (feature amount vector (post-administration data) from the histogram image. When the post-administration feature amount vector) is extracted, the extraction means 124 extracts the post-administration feature amount vector from the histogram image created from the data (pre-administration data) acquired from the subject when the compound is not administered. It is preferable to normalize with the obtained feature amount vector (feature amount vector before administration). This is because the individual difference that can appear in the feature amount vector can be reduced. The normalization may be, for example, dividing the value of each component of the post-administration feature amount vector by the value of the corresponding component of the pre-dose feature amount vector, or the value of each component of the post-administration feature amount vector. May be divided by the average value of each component of the pre-administration feature amount vector.

The feature amount vector extracted by the extraction means 124 or the feature amount vector extracted and normalized is passed to the comparison means 125.

FIG. 3C shows an example of the feature amount vector extracted by the extraction means 124.

FIG. 3C is created from brain wave data obtained by administering feature 5 ml / kg, 4-AP 6 mg / kg, Strichnin 3 mg / kg, Aspirin 3000 mg / kg, pilocarpine 400 mg / kg, and Tramadol 150 mg / kg to rats. The 4096-dimensional feature amount vector extracted from the histogram image is shown. The horizontal axis corresponds to the dimension of the feature amount, and the vertical axis corresponds to the value of the feature amount. 4-AP, Strychnine, Pilocarpine, and Tramadol are known as convulsive-positive compounds, and Aspirin is known as a convulsive-negative compound. In this example, each feature vector is normalized by the feature vector extracted from the histogram image created from the data (pre-dose data) obtained from the rat when the compound was not administered. The component of the feature amount that does not change before and after administration is 1.

Referring to FIG. 2C again, the comparison means 125 is configured to compare the feature amount vector with the plurality of reference feature amount vectors. Here, the reference feature amount vector may include a feature amount vector derived from the data (post-administration data) acquired from the subject when the known compound is administered. That is, the reference feature amount vector may include the feature amount vector extracted by the extraction means 124 from the histogram image created by the image creation means 121 from the data acquired from the subject when the known compound is administered. The reference feature amount vector was obtained, for example, when (b) 4-AP 6 mg / kg was administered, and (c) Strychnin 3 mg / kg, which is shown in FIG. 3C. It contains at least one of a feature amount vector, (e) a feature amount vector obtained when Pilocarpine 400 mg / kg is administered, and (f) a feature amount vector obtained when Tramadol 150 mg / kg is administered. The comparison by the comparison means 125 includes a feature amount vector derived from the data obtained from the subject when the target compound was administered and a feature amount vector derived from the data obtained from the subject when the known compound was administered. Can be a comparison of.

The comparative means 125 is derived from, for example, the value of each component of the feature amount vector derived from the data obtained from the subject when the target compound is administered, and the data acquired from the subject when the known compound is administered. It is possible to compare the values of the corresponding components of the feature vector to be value-based. Alternatively, the comparison means 125 is obtained, for example, from a feature amount map created from a feature amount vector derived from data obtained from the subject when the target compound is administered, and from the subject when the known compound is administered. It is possible to compare the feature amount map created from the feature amount vector derived from the above data on an image basis. Specifically, the comparison means 125 can create a feature amount map by mapping the feature amount vector, and can compare the created feature amount maps with each other. Here, mapping can mean assigning and imaging corresponding colors or shades to the values of each component of the feature vector. For example, each pixel of the feature amount map corresponds to each component of the feature amount vector, and the pixel value of each pixel may correspond to the value of each component. The feature map can have any size depending on the number of dimensions of the feature vector. In one example, in the case of a feature vector having 409 6-dimensional components, the feature map has 1 × 4096 pixels, 2 × 2048 pixels, 4 × 1024 pixels, 8 × 512 pixels, 16 × 256 pixels, 32 × 128 pixels, It may have a size such as 64 × 64 pixels.

The comparison means 125 can assign a corresponding color or shade to the value of each component of the feature amount vector, for example, depending on whether or not there is a significant difference with respect to the reference data according to the object to be analyzed. For example, when analyzing convulsive characteristics, the reference data can be a feature vector obtained when a seizure-negative compound is administered, or a feature vector obtained when vehicl is administered. For example, when the comparison means 125 has a significant difference from the corresponding component of the feature amount vector of the reference data for a certain component of the feature amount vector, a specific color is assigned to the pixel corresponding to the component, and the significant difference. If not, another particular color is assigned to the pixel corresponding to that component. By doing this for all the components of the feature vector, a feature map can be created.

FIG. 3D shows an example of a feature amount map created by the comparison means 125.

FIG. 3D shows a feature map created from each of the four feature vectors shown in FIG. 3C. (A) shows a feature amount map created from the feature amount vector obtained when 4-AP 6 mg / kg was administered, and (b) is a feature obtained when Strychnin 3 mg / kg was administered. A feature amount map created from a quantity vector is shown, (c) shows a feature amount map created from a feature amount vector obtained when Pilocarpine 400 mg / kg is administered, and (d) shows a Tramadol 150 mg / kg. The feature amount map created from the feature amount vector obtained when kg was administered is shown. As described above, since these feature vector can be a reference feature vector, these feature maps can be a reference feature map. In this example, the 4096-dimensional component is represented by an image of 64 × 64 pixels. Of the feature quantities of the graph shown in FIG. 3C, the first to 64th feature quantities correspond to the first row, the 65th to 128th feature quantities correspond to the second row, ... 4033. The feature amount of the 4096th corresponds to the 64th row. Pixels corresponding to the components having a significant difference between the feature amount vector obtained when the convulsive negative compound (Aspirin 3000 mg / kg) was administered and the feature amount vector obtained when the vegetable was administered. Pixels corresponding to the components shown in black and having no significant difference are shown in white. The component corresponding to the pixel shown in black can be a useful feature quantity in analyzing the convulsive characteristics.

The comparison means 125 can compare the feature amount map created from the feature amount vector derived from the data acquired from the subject when the target compound is administered with a plurality of reference feature amount maps. For example, in the comparison means 125, the feature amount map created from the feature amount vector derived from the data acquired from the subject when the target compound is administered is a reference feature amount map among a plurality of reference feature amount maps. The feature map and the plurality of reference feature maps can be compared so as to identify whether they are similar to. For example, the comparative means 125 ranks a plurality of reference feature maps in an order similar to the feature map created from the feature vector derived from the data obtained from the subject when the subject compound is administered. It is possible to compare a feature map with a plurality of reference feature maps.

For example, the comparison means 125 performs pattern matching between a feature amount map created from a feature amount vector derived from data acquired from a subject when the target compound is administered and a plurality of reference feature amount maps. be able to. For example, the comparison means 125 can perform pattern matching using a trained model in which a plurality of reference feature amount maps are trained. The trained model trains multiple reference feature maps and their labels, and when an unlearned reference feature map is input, which of the learned reference feature maps is similar? Alternatively, it is possible to output the degree of similarity with each of the already learned reference feature amount maps. This output can identify which of the multiple reference feature maps the feature map is similar to, or the order in which the multiple reference feature maps are similar to the feature map. Can be identified.

In one example, the comparison means 125 can compare the feature amount map with one reference feature amount map in which a plurality of reference feature amount maps are combined. This is preferable in that comparisons with a plurality of reference feature map can be performed at one time.

FIG. 3E shows an example of one reference feature amount map in which a plurality of reference feature amount maps are combined.

FIG. 3E shows one reference feature amount map that combines the four reference feature amount maps shown in FIG. 3D. In this reference feature amount map, the pixels corresponding to the components having no significant difference are shown in white (0), and the pixels corresponding to the components having a significant difference are shown in colors other than white. In particular, the components having a significant difference are shown in colors according to how many of the four reference feature maps are common. In FIG. 3E, the pixels corresponding to the components having a significant difference in common in all four reference feature amount maps are shown in black (4), and the significant difference is commonly shown in the three reference feature amount maps. The pixels corresponding to the components having are shown in the darkest gray (3), and the pixels corresponding to the components having a significant difference in common in the two reference feature maps are shown in the next dark gray (2). Has been done. Pixels corresponding to the components having a significant difference existing only in the 4-AP reference feature map are shown in the next dark gray (1), and the significant difference existing only in the Trychnin reference feature map is shown. The pixel corresponding to the component having is shown in the next dark gray (1), and the pixel corresponding to the component having a significant difference present only in the reference feature amount map of Pilocarpine is the next dark gray (1). ), And the components with significant differences that exist only in Tramadol's reference feature map are shown in the lightest gray.

In the comparison means 125, for example, the pixel corresponding to the component having a significant difference in the feature amount map created from the feature amount vector derived from the data acquired from the subject when the target compound is administered is one reference. By calculating which color of the pixel in the feature map corresponds to the most, it is possible to identify which of the plurality of reference feature maps the feature map is similar to. Alternatively, in the comparison means 125, for example, the pixel corresponding to the component having a significant difference in the feature amount map created from the feature amount vector derived from the data acquired from the subject when the target compound is administered is 1. By calculating how many pixels of each color correspond to one reference feature amount map, it is possible to specify the order in which a plurality of reference feature amount maps are similar to the feature amount map.

For example, if it is identified that the feature map created from the feature vector derived from the data obtained from the subject when the subject compound is administered is similar to the 4-AP reference feature map, the subject. The compound can be expected to be 4-AP or a compound having properties similar to 4-AP. For example, if it is identified that the feature map created from the feature vector derived from the data obtained from the subject when the target compound is administered is similar to the Strychnine reference feature map, the target compound is determined. , Strychnine, or a compound having properties similar to Strychnine.

For example, the feature map created from the feature vector derived from the data obtained from the subject when the target compound was administered is similar to both the Pilocarpine reference feature map and the Tramadol reference feature map. If is identified, the target compound can be expected to be a compound having properties common to Pilocarpine and Tramadol.

For example, a feature map created from a feature vector derived from data obtained from a subject when a target compound is administered is a Pilocarpine reference feature map, a Tramadol reference feature map, and a Tramadol reference feature. If similar to the quantity map is identified, the compound of interest can be expected to be a compound having properties common to Pilocarpine, Tramadol, Tramadol. Such target compounds are expected to have, for example, at least convulsive toxicity.

For example, a feature map created from a feature vector derived from data obtained from a subject when a subject compound is administered is most similar to the Pilocarpine reference feature map, followed by Tramadol's reference feature map. When it is identified that it is similar to Pilocarpine, it can be predicted that the target compound is a compound having mainly properties similar to Pilocarpine and also having properties similar to Tramadol.

In this way, the processor 120'can predict which known compound the target compound resembles, and further predict the characteristics of the target compound and the ranking of the characteristics of the target compound.

In the above example, each component of the computer system 100 is provided in the computer system 100, but the present invention is not limited to this. It is also possible that any of the components of the computer system 100 is provided outside the computer system 100. For example, when each of the processor 120 and the memory 130 is composed of separate hardware components, each hardware component may be connected via an arbitrary network. At this time, the type of network does not matter. Each hardware component may be connected via a LAN, may be wirelessly connected, or may be connected by wire, for example. The computer system 100 is not limited to a specific hardware configuration. For example, it is also within the scope of the present invention to configure the processor 120 with an analog circuit instead of a digital circuit. The configuration of the computer system 100 is not limited to the above-mentioned one as long as the function can be realized.

In the above example, the components of the processors 120 and 120'are provided in the same processor 120 and 120', but the present invention is not limited to this. It is also within the scope of the present invention that the components of the processors 120 and 120'are distributed to a plurality of processor units.

In the above-mentioned example, it has been described that the image creating means 121 creates a histogram image based on the data received by the receiving means 110, but the present invention is not limited to this. For example, the receiving means 110 may receive the histogram image created outside the computer system 100. At this time, the image creating means 121 may be omitted, and the image recognition model 122 can directly receive the histogram image from the receiving means 110.

4. Processing by the computer system for predicting the state of the target FIG. 5 is a flowchart showing an example of processing by the computer system 100 for predicting the state of the target. In the example shown in FIG. 5, the process 500 for creating a histogram image used for predicting the state of the target will be described. FIG. 6 shows a specific example of creating a histogram image by the process 500. In the examples described below, the process 500 will be described as being executed in the processor 120, but it is understood that the process 500 is also executed in the processor 120'.

In step S501, the image creating means 121 of the processor unit 120 acquires the wavelet image. The wavelet image is an image obtained by wavelet transforming the data acquired from the target. The acquired wavelet image may be one received by the receiving means 110 from the outside of the computer system 100, or may be obtained by the processor unit 120 performing wavelet transform on the data received by the receiving means 110. There may be.

For example, in step S501, a wavelet image as shown in FIG. 6A is acquired. The wavelet image shown in FIG. 6A is a spectrogram obtained by wavelet transforming a waveform of an electroencephalogram acquired from a subject for a certain 60 seconds, and is a spectrum of 4 Hz to 250 Hz in 1 Hz units. The strength is shown in chronological order.

The acquired wavelet image is passed to the dividing means 1211.

In step S502, the image segmentation means 1211 divides the wavelet image acquired in step S501 into a plurality of frequency bands. The dividing means 1211 can divide the wavelet image into a plurality of frequency bands by any known image processing technique. Further, the dividing means 1211 can divide the wavelet image in any unit.

For example, in step S502, as shown in FIG. 6B, the wavelet image shown in FIG. 6A is divided into a plurality of frequency bands. In the example shown in FIG. 6B, for the sake of brevity, only the 4Hz, 10Hz, 50Hz, 100Hz, and 250Hz frequency bands are shown, but the wavelet image is between these frequency bands. It is also divided into intervening frequency bands. That is, the wavelet image is divided into a total of 247 images from 4 Hz to 250 Hz.

The divided wavelet image has a time axis and is a one-dimensional color map in which colors represent spectral intensities.

In step S503, the histogram creating means 1212 creates a histogram of the spectral intensity for each of the plurality of divided wavelet images after being divided in step S502. The histogram creating means 1212 can create a histogram by counting the number of appearances of each spectral intensity in the divided wavelet image. In the histogram, the horizontal axis represents the spectral intensity, and the vertical axis represents the number of occurrences (or distribution ratio) of each spectral intensity.

For example, in step S503, as shown in FIG. 6 (c), a histogram is created from a plurality of divided wavelet images after being divided in FIG. 6 (b). In the example shown in FIG. 6 (c), for simplification of the description, only the histogram created from the divided wavelet image of the frequency bands of 4 Hz, 10 Hz, 50 Hz, 100 Hz, and 250 Hz is shown, but the histogram is shown. Is also created in the frequency bands that intervene between these frequency bands. That is, a total of 247 histograms from 4 Hz to 250 Hz are created. The histogram is a two-dimensional histogram in which the horizontal axis represents the spectral intensity and the vertical axis represents the distribution ratio of each spectral intensity. The vertical axis is the distribution ratio when the maximum number of occurrences is 100%.

In step S504, the joining means 1213 joins a plurality of histograms created in step S503. The coupling means 1213 combines a plurality of histograms in the order of frequency, whereby a histogram image is created.

For example, in step S504, as shown in FIG. 6D, 247 color maps converted from the 247 histograms created in FIG. 6C are two-dimensionally combined in the frequency axis direction. This creates a histogram image.

The histogram image created in this way is suitable for predicting the state of the target using the image recognition model. First, in the histogram image, the time information contained in the time series data (for example, wavelet image) used to create the histogram image is deleted, so that there is an advantage that it is easy to learn the characteristics of the time series data as an image. be. Secondly, it is possible to create a histogram image having the same characteristics even when the time window of the time series data used for creating the histogram image is shifted, and it is possible to increase the number of images used for learning. There is. Thirdly, since the characteristics of the frequency intensity distribution can be detected independently of the difference between the samples, there is an advantage that the prediction accuracy can be improved.

FIG. 7 is a flowchart showing an example of processing by the computer system 100 for predicting the state of the target. In the example shown in FIG. 7, the process 700 for constructing the image recognition model 122 for predicting the state of the target will be described.

When the computer system 100 receives data indicating biological signals in a plurality of known states of the target via the receiving means 110, the received data is passed to the processor 120.

In step S701, the image creating means 121 of the processor 120 creates a plurality of training images from data showing biological signals in a plurality of known states of the target. The plurality of training images are histogram images, and the image creating means 121 can create a plurality of histogram images by the above-mentioned process 500 with reference to FIG.

In step S702, the learning means 123 of the processor 120 learns a training data set including a plurality of training images. For example, the learning means 123 learns the relationship between the plurality of training images and the known states corresponding to the plurality of training images. This means that, for example, as described above with reference to FIG. 4, in the neural network, the value of the output layer when a plurality of training images are input to the input layer becomes a value indicating the state of the corresponding target. , By calculating the weighting factor of each node.

The image recognition model 122 for predicting the state of the target constructed in this way can be used in the process for predicting the state of the target, which will be described later.

FIG. 8A is a flowchart showing an example of processing by the computer system 100 for predicting the state of the target. In the example shown in FIG. 8A, the process 800 for predicting the state of the target will be described.

In step S801, the computer system 100 receives data indicating a target biological signal via the receiving means 110. The receiving means 110 may receive data indicating a target biological signal as a wavelet image. Alternatively, the receiving means 110 may receive the data acquired from the target in a format other than the wavelet image.

The received data is passed to the processor 120, and the processor 120 receives this.

In step S802, the image creating means 121 of the processor 120 creates a histogram image from the data indicating the biological signal received by the receiving means 110. The image creating means 121 can create a histogram image from the wavelet image received by the receiving means 110. When the data received by the receiving means 110 is not a wavelet image, the image creating means 121 creates a wavelet image from the data received by the receiving means 110, and then creates a histogram image from the created wavelet image. You may do so.

The image creating means 121 can create a plurality of histogram images by, for example, the process 500 described above with reference to FIG.

In step S803, the processor 120 inputs the histogram image created in step S802 into the image recognition model 122. The image recognition model 122 is trained by the process 700 described above with reference to FIG.

In step S804, the processor 120 processes the image in the image recognition model 122 and outputs the target state. In this way, the state of the target can be predicted.

FIG. 8B is a flowchart showing an example of processing by the computer system 100 for predicting the state of the target. In the example shown in FIG. 8B, the process 810 for predicting the characteristics of the target compound will be described.

In step S811, the computer system 100 receives the post-administration data indicating the biological signal of the target after the target compound is administered to the target via the receiving means 110. The receiving means 110 may receive the post-administration data as a wavelet image. Alternatively, the receiving means 110 may receive the post-administration data in a format other than the wavelet image. The receiving means 110 may also receive pre-dose data indicating the biological signal of the subject before the subject compound is administered to the subject.

The received data is passed to the processor 120', and the processor 120' receives the data.

In step S812, the image creating means 121 of the processor 120'creates a histogram image from the post-administration data received in step S811. In step S812, a histogram image is created by the same processing as in step S802. When the pre-administration data is also received in step S811, the image creating means 121 can also create a histogram image from the pre-administration data.

In step S813, the extraction means 124 of the processor 120'extracts the feature amount vector from the histogram image created in step S812. The extraction means 124 can extract a feature amount vector by using, for example, a trained image recognition model. For example, a feature vector as shown in FIG. 3C is extracted.

When the histogram image is created from the pre-administration data in step S812, the extraction means 124 can also extract the feature amount vector from the histogram image created from the pre-administration data. Thereby, the extraction means 124 can normalize the feature amount vector (post-administration feature amount vector) from the post-administration data with the feature amount vector (pre-dose feature amount vector) from the pre-administration data. This is preferable in that individual differences that may appear in the feature amount vector can be reduced. The normalization may be, for example, dividing the value of each component of the post-administration feature amount vector by the value of the corresponding component of the pre-dose feature amount vector, or the value of each component of the post-administration feature amount vector. May be divided by the average value of each component of the pre-administration feature amount vector.

In step S814, the comparison means 125 of the processor 120'compares the feature amount vector extracted in step S813 with a plurality of reference feature amount vectors. When the feature quantity vector is normalized in step S813, the comparison means 125 compares the normalized feature quantity vector with the plurality of reference feature quantity vectors.

In the computer system 100, a process of extracting a reference feature quantity vector is performed before starting the process 810. The process for extracting the reference feature amount vector may be the process of steps S811 to S813 for the post-administration data showing the biological signal of the subject after the known compound is administered to the subject. That is, the process of extracting the reference feature amount vector is to receive post-administration data showing the biological signal of the target after administering the known compound to the subject, and to create a histogram image from the post-administration data of the known compound. , Extracting a reference feature vector from a histogram image may be included. The process of extracting the reference feature vector is to receive pre-dose data showing the biometric signal of the subject before administering the known compound to the subject, to create a histogram image from the pre-administration data of the known compound, and to create a histogram. It is preferable to further include extracting the pre-dose reference feature amount vector from the image and normalizing the post-dose reference feature amount vector with the pre-dose reference feature amount vector.

The comparison means 125 can, for example, compare the value of each component of the feature amount vector with the value of the corresponding component of the reference feature amount vector extracted in step S813 on a value basis. Alternatively, the comparison means 125 can, for example, compare the feature amount map created from the feature amount vector extracted in step S813 with the reference feature amount map created from the reference feature amount vector on an image basis. The comparison means 125 creates, for example, a feature map as shown in FIG. 3D or FIG. 3E for both the feature vector and the reference feature vector, and compares using the created feature map. be able to.

The comparison means 125 can perform pattern matching between the feature amount map and each of the plurality of reference feature amount maps (for example, the reference feature amount map shown in FIG. 3D). Alternatively, the comparison means 125 can perform pattern matching between, for example, a feature amount map and one feature amount map (for example, the reference feature amount map shown in FIG. 3E) in which a plurality of reference feature amount maps are combined. .. This makes it possible to identify which of the multiple reference feature maps the feature map is similar to, or the order in which the multiple reference feature maps are similar to the feature map. Can be identified.

In step S815, the processor 120'predicts the characteristics of the target compound based on the result of step S814. The processor 120'can predict, for example, the characteristics of the known compound corresponding to the reference feature amount vector to which the feature amount vector is most similar as the characteristic of the target compound. Alternatively, the processor 120'can predict, for example, the properties of some known compounds corresponding to some reference feature vectors with similar feature quantity vectors as the properties of the target compound. Alternatively, the processor 120'is likely to have, for example, the properties of some known compounds corresponding to some reference feature vectors that are said to be similar, in order of similarity. Can be predicted as.

The processor 120'can predict, for example, the characteristics of a known compound corresponding to the reference feature amount map to which the feature amount map is most similar as the characteristics of the target compound. Alternatively, the processor 120'can predict, for example, the properties of some known compounds corresponding to some reference feature maps to which the feature maps are similar as the properties of the target compound. Alternatively, the processor 120'is likely, for example, the properties of some known compounds corresponding to some reference feature maps to which the feature maps are similar, in order of similarity. Can be predicted as.

In this way, the treatment 810 can predict which known compound the target compound resembles, and further predict the characteristics of the target compound and the ranking of the characteristics of the target compound.

By combining the treatment 810 with the treatment 800, it is possible to predict what kind of characteristics of the target compound the predicted state of the target is due to. This prediction can be applied, for example, to drug discovery of neurological disorders and evaluation of neurotoxicity.

In the above example, it was explained that the processes are performed in a specific order, but it should be noted that the order of each process is not limited to the one described and is performed in any logically possible order.

In the example described above with reference to FIGS. 5, 7, 8A, and 8B, the processing of each step shown in FIGS. 5, 7, 8A, and 8B is a program stored in the processor 120 and the memory 130. However, the present invention is not limited to this. At least one of the processes of each step shown in FIGS. 5, 7, 8A, and 8B may be realized by a hardware configuration such as a control circuit.

(Example 1)
After obtaining cortical EEG for 15 minutes before administration, 5 convulsive positive drugs (4-AP, Strychnine, Pilocarpine, Isoniazid, PTZ) were administered at 3 mg / kg, 1 mg / kg, 150 mg / kg, 150 mg / kg, 30 mg /, respectively. Cortical electroencephalograms when administered to kg specimens were obtained for 2 hours or until death. Here, a rat was used as a sample. The following three states were set from the general condition observation records before and after drug administration.
(1) Before administration (2) Precursor state (after twitching, which is a prodromal symptom of seizure, to 1 hour)
(3) Convulsions (after convulsive seizures such as clonic convulsions until death)

The acquired cortical EEG was divided by a time window of 60 seconds, and each data with the time window shifted by 30 seconds was wavelet-converted to obtain a wavelet image. Then, the obtained wavelet image was subjected to the above-mentioned processing 500 with reference to FIG. 5, to obtain a plurality of histogram images in which the time window of 60 seconds was shifted by 30 seconds.

The feature amount of the histogram image was extracted for each of the above three states. AlexNet, which is an existing trained image recognition model, was used as the feature extraction model. The 4096-dimensional features extracted by the feature extraction model were input to the state prediction model and trained. In the state prediction model, the feature amounts of 263 images before administration, 220 images of aura, and 637 images of convulsions were input and trained. The state prediction model consists of an input layer of 4096 units, a hidden layer of 10 layers, and an output layer of 1 unit, and depending on the value of the output layer, whether it is a pre-administration state, a convulsive precursor state, or a convulsive state Trained to be predictable.

After training the state prediction model, when unlearned data was input to the state prediction model, we experimented with which of the three states to predict and verified the separation system of the three states (Experiment 1). The input data are histogram images of 118 images before administration, 85 images of aura, and 277 images of convulsions.

Furthermore, after training the state prediction model, a histogram image for about 2 hours immediately after drug administration containing unlearned data was input to the sample in which only precursors were seen, and the state transition was predicted (Experiment 2).

FIG. 9 is a diagram showing the results of Experiment 1.

The table of FIG. 9 shows which state the state prediction model predicted with respect to the histogram image of the actual state indicated by the actual label. The state prediction model predicted that 99 images were in the pre-administration state, 9 images were predicted to be in the precursory state, and 10 images were predicted to be in the convulsive state for 118 histogram images before administration. .. In addition, the state prediction model predicts that 16 images are in the pre-administration state, 49 images are predicted to be in the precursory state, and 20 images are in the convulsive state for 85 histogram images in the precursory state. I predicted. In addition, the state prediction model predicts that 7 images are in the pre-administration state, 14 images are predicted to be in the precursory state, and 256 images are in the convulsive state for 277 histogram images in the convulsive state. I predicted.

From this result, the accuracy ((true positive + true negative) / overall) was 84.2%. The specificity (true negative / (false positive + true negative)) was 83.9% (ie, the probability of false positive was 16.1%). The sensitivity of the precursor (true positive / (true positive + false negative)) was 57.6%. The accuracy of the precursor (true positive / (true positive + false positive)) was 68.1%. The sensitivity of the convulsions was 92.4%. The accuracy of the convulsions was 89.5%. From this result, it can be seen that the pre-administration state and the convulsive state can be predicted with high accuracy by the state prediction model. The sensitivity of the precursory state remains at 57.6%, but the waveform of the precursory state was classified by the researchers based on the observation of the general condition, and the precursory state includes both the pre-dose state and the convulsive state. It is thought that it is. Therefore, the convulsive risk is predicted at 81.2% when the precursory state and the convulsive state are combined, and it can be said that the convulsive risk prediction can be predicted with high accuracy. It can also be considered that the state prediction model correctly determines the waveform in the precursory state.

FIG. 10 is a diagram showing the results of Experiment 2.

In the graph of FIG. 10, the horizontal axis represents time and the vertical axis represents states, and the state prediction model predicts which state in chronological order. The state prediction model learns a histogram image of 1618 seconds to 4318 seconds as a precursor state.

From this result, it can be seen that the learning prediction model can predict the learned precursor range as a precursor state. In addition, it can be seen that the portion (600 to 1500 seconds) that was not set as the precursor state in the general state observation record is also predicted as the precursor state. This suggests that even precursory states that cannot be captured by general state observation records can be captured by the state prediction model. That is, it is considered that by predicting a precursory state using a state prediction model, it is possible to detect the precursory state at an early stage and use it for diagnosis, early treatment, and prevention.

(Example 2: Comparison with the conventional method)
We compared the conventional method of using FFT spectral intensity in EEG analysis to detect convulsive aura and seizure, and the method of predicting the state of a subject using the histogram image of the present invention.

EEGs were obtained when the vehicle and three convulsive-positive drugs (4-AP, Isoniazid, Pilocarpine) were administered to the sample. Here, a rat was used as a sample. Three convulsive-positive agents (4-AP, Isoniazid, Pilocarpine) were administered at 3 mg / kg, 150 mg / kg, and 150 mg / kg, respectively, to induce convulsive aura. Three seizure-positive agents (4-AP, Isoniazid, Pilocarpine) were administered at 6 mg / kg, 300 mg / kg, and 400 mg / kg, respectively, to induce seizure conditions. From the general condition observation records before and after drug administration, EEG data were classified into (1) before administration, (2) immediately after administration (before the effect of the drug appeared), (3) convulsive aura state, and (4) seizure state.

In the conventional method, FFT (Fast Fourier Transform) was performed on each of the classified EEG data, and the frequency spectrum was calculated. Of the calculated frequency spectra, the total value of the spectral intensities was calculated for the frequency spectra in the 4 to 200 Hz band. The total value of each spectral intensity was normalized, with the total value of the spectral intensities of the frequency spectra obtained from the data before drug administration as 100%.

In the method of predicting the state of a target using the histogram image of the present invention, a histogram image in the 4 to 250 Hz band was created from each of the classified EEG data. Specifically, for each of the classified electroencephalogram data, the acquired electroencephalogram was divided by a time window of 60 seconds, and each data in which the time window was shifted by 30 seconds was wavelet-transformed to obtain a wavelet image. Then, the obtained wavelet image was subjected to the above-mentioned processing 500 with reference to FIG. 5, to obtain a plurality of histogram images in which the time window of 60 seconds was shifted by 30 seconds.

The feature amounts of the histogram images were extracted for each of the three states of pre-administration, convulsive aura state, and convulsive seizure state. Alex Net, which is an existing trained image recognition model, was used as the feature extraction model. The 4096-dimensional features extracted by the feature extraction model were input to the state prediction model and trained. The state prediction model was trained to be able to output three probabilities: the probability of being in a pre-dose state, the probability of being in a convulsive aura state, and the probability of being in a seizure state. From the prediction results of the training data, an ROC curve for the pre-administration state was created, and the threshold value of the optimum pre-administration probability when determining using the probability of the pre-administration state was calculated from the optimum operating point of the ROC curve. The toxicity score was defined as toxicity score = 1-pre-dose probability. The toxicity score at the optimum operating point was 0.1308. Images with a toxicity score equal to or higher than the toxicity score threshold (0.1308) at the optimum operating point were judged to be toxic. Toxicity probability = Toxicity probability = number of images judged to be toxic / total number of images x 100 (%)
Was defined as.

After training the state prediction model, when unlearned data is input to the state prediction model, by calculating the toxicity probabilities of each of the convulsive aura state and the seizure state during vehicle administration, the convulsive aura state and the seizure state. Attempted to detect (Experiment 1).

Furthermore, by calculating the toxicity score and toxicity probability when the data immediately after the vehicle administration and the data 60 minutes after the vehicle administration are input to the state prediction model after the training of the state prediction model, the influence on the prediction by the vehicle administration can be obtained. Investigated (Experiment 2).

FIG. 11 is a diagram showing the results of the conventional method.

FIG. 11 (a) shows the results immediately after and 60 minutes after the administration of the vehicle, and the three convulsive-positive agents (4-AP, Isoniazid, and Pilocarpine) were administered at 3 mg / kg, 150 mg / kg, and 150 mg / kg, respectively. The results immediately after administration and at the time of aura of convulsions are shown, and FIG. 11 (b) shows the results immediately after administration of the vehicle and 60 minutes after administration, and three convulsive-positive drugs (4-AP, Isoniazid, and pilocarpine), respectively. The results immediately after administration and at the time of convulsive seizure when 6 mg / kg, 300 mg / kg, and 400 mg / kg were administered are shown. FIG. 11 (c) shows the results at the time of convulsive aura when three convulsive positive agents (4-AP, Isoniazid, pilocarpine) were administered at 3 mg / kg, 150 mg / kg, and 150 mg / kg, respectively. The results at the time of a seizure when the drugs (4-AP, Isoniazid, Pilocarpine) were administered at 6 mg / kg, 300 mg / kg, and 400 mg / kg, respectively, are shown. In each figure, the vertical axis shows the spectral intensity normalized by the spectral intensity before administration, and the horizontal axis shows the label. In FIGS. 11 (a) and 11 (b), whether there was a significant difference between the result immediately after administration and the result 60 minutes after administration, and between the result immediately after administration and the result at the time of aura of convulsions. It is shown. “*” Indicates that there was a significant difference, and “NS” indicates that there was no significant difference. 11 (c) and 11 (d) show whether there was a significant difference between the results at the time of aura of convulsions and the results immediately after administration of the vehicle. “*” Indicates that there was a significant difference, and “NS” indicates that there was no significant difference.

As can be seen from FIGS. 11 (c) and 11 (d), there were some drugs that did not show a significant difference from the results immediately after the administration of the vehicle at both the aura of convulsions and the seizures. With drugs that did not show a significant difference, it was not possible to distinguish between the time of vehicle administration and the time of convulsive aura and seizure. Conventional methods have been found to be unsuitable for detecting convulsive aura and seizure conditions for a wide range of drugs.

As can be seen from FIGS. 11 (a) and 11 (b), a significant difference was observed between the result immediately after the vehicle administration and the result 60 minutes after the vehicle administration. It can be seen that the conventional method is not a stable evaluation system.

12A and 12B show the results of the method of predicting the state of the object using the histogram image of the present invention.

FIG. 12A (a) shows the ROC curve for the pre-administration state created from the prediction results of the training data. In the ROC curve, the vertical axis shows the ratio predicted to be the pre-dose state when the data of the pre-dose state is input, and the horizontal axis shows the data of the convulsive precursor state and the data of the convulsive attack state. The rate predicted to be in the pre-administration state is shown. As can be seen from FIG. 12 (A), the post-training state prediction model separates the pre-dose state data, the convulsive aura state data, and the seizure state data with an accuracy of 91.6% for the training data. We were able to.

FIG. 12A (b) shows the results for Experiment 1. The table shows the probabilities of toxicity of each of the three convulsive-positive drugs, when unlearned data is input to the state prediction model, when the vehicle is administered, the convulsive aura state, and the convulsive seizure state. The graph is an average of the tables and shows the predicted probabilities of toxicity at the time of vehicle administration, convulsive aura state, and seizure state. As can be seen from FIG. 12 (b), the post-training state prediction model predicts that the data at the time of vehicle administration is the pre-administration state with an average accuracy of 89.9 ± 5.2% for the unlearned data, and convulsions. The average toxicity probability of the aura state was determined to be 84.4 ± 9.0%, and the average toxicity probability of the seizure state was determined to be 98.8 ± 0.6%. Thus, according to the post-training state prediction model, it was possible to detect the convulsive aura state and the convulsive seizure state in each of the three convulsive positive drugs. That is, it can be said that the method of predicting the state of a subject using the histogram image of the present invention is more advantageous than the conventional method in that it can detect a convulsive aura state and a seizure state for a wide range of drugs. ..

Further, in the conventional method, the determination is made based on the spectral intensity of the entire measurement time, whereas in the method of predicting the target state using the histogram image of the present invention, the determination can be made for each image, so that the time information is also quantified. It is possible to identify precursory states that could not be captured by general state observation. Furthermore, in the method of predicting the state of the target using the histogram image of the present invention, since it can be determined using the 4096-dimensional feature amount extracted by the feature amount extraction model, the convulsive aura state and the seizure state. It can be said that the detection sensitivity of is excellent. Furthermore, in the method of predicting the state of a target using the histogram image of the present invention, by appropriately changing the setting of the label to be learned, not only the prediction of the convulsive state of the drug but also the mechanism of action of a wide range of target compounds can be determined. Prediction is also possible. In these respects as well, it can be said that it is more advantageous than the conventional method.

FIG. 12B shows the results for Experiment 2. FIG. 12B (a) shows the average toxicity score when the data immediately after the vehicle administration and the data 60 minutes after the vehicle administration are input to the state prediction model. The vertical axis shows the average toxicity score, and the horizontal axis shows the label. FIG. 12B (b) shows the average toxicity probability when the data immediately after the vehicle administration and the data 60 minutes after the vehicle administration are input to the state prediction model. The vertical axis shows the average toxicity probability, and the horizontal axis shows the label. In FIGS. 12B (a) and 12B (b), it is shown whether or not there was a significant difference between the result immediately after the vehicle administration and the result 60 minutes after the vehicle administration, and "NS" is significant. It shows that there was no difference. As can be seen from FIGS. 12B (a) and 12B (b), no significant difference was observed between the result immediately after the vehicle administration and the result 60 minutes after the vehicle administration. The method of predicting the state of an object using the histogram image of the present invention is a stable evaluation system, and can be said to be more advantageous than the conventional method in this respect as well.

The present invention is not limited to the above-described embodiment. It is understood that the invention should be construed only by the claims. It will be understood by those skilled in the art that from the description of the specific preferred embodiments of the present invention, the equivalent scope can be carried out based on the description of the present invention and common general technical knowledge.

The present invention provides a method for creating a histogram image that can be used to predict the state of an object, and is useful as a method for predicting the state of an object using a histogram image.

100 Computer system 110 Receiving means 120 Processor 130 Memory 140 Output means 200 Database section

Claims

How to create a histogram image
The process of acquiring a wavelet image and
The process of dividing the wavelet image into a plurality of frequency bands and
The process of creating a histogram of the spectral intensity for each of the multiple divided wavelet images,
A method that includes the process of combining multiple histograms created.
The step of combining the plurality of histograms is
A step of converting each of the plurality of histograms into a color map, wherein the color of the color map represents a distribution ratio.
The method of claim 1, comprising combining a plurality of converted colormaps.
The step of acquiring the wavelet image is
The process of acquiring waveform data and
The method according to claim 1 or 2, comprising the step of converting the waveform data into the wavelet image.
The method according to claim 3, wherein the waveform data includes an electroencephalogram waveform data.
The method according to any one of claims 1 to 4, wherein the plurality of frequency bands include at least 6 frequency bands.
The plurality of frequency bands include at least a frequency band of about 1 Hz to about 4 Hz, a frequency band of about 4 Hz to about 8 Hz, a frequency band of about 4 Hz to about 8 Hz, a frequency band of about 8 Hz to about 12 Hz, and a frequency band of about 12 Hz to about 30 Hz. The method according to any one of claims 1 to 5, comprising the frequency band of about 30 Hz to about 80 Hz, and the frequency band of about 100 Hz to about 200 Hz.
It is a method of predicting the state of the target.
The step of receiving data indicating the biological signal of the target and
A step of creating a histogram image from data showing the biological signal according to the method according to any one of claims 1 to 6.
A step of inputting the histogram image into an image recognition model trained by the training data set, wherein the training data set is a plurality of objects according to the method according to any one of claims 1 to 6. A process and a process that includes multiple training images created from data showing biometric signals in known conditions.
A method comprising the steps of processing the histogram image in the image recognition model and outputting the state of the object.
The seventh aspect of claim 7, wherein the plurality of known states of the subject include a state having a convulsive symptom, a state having no convulsive symptom, and a state having a precursor of the convulsive symptom. Method.
It is a method of constructing an image recognition model for predicting the state of an object.
A step of creating a plurality of training images from data showing biological signals in a plurality of known states of a target according to the method according to any one of claims 1 to 6.
A method comprising learning a training data set comprising the plurality of training images.
It is a method of predicting the state of the target.
The step of receiving data indicating the biological signal of the target and
In the step of creating a histogram image from the data showing the biological signal, the histogram image is a color map in which the first axis represents the spectral intensity, the second axis represents the frequency, and the color represents the distribution ratio. There is a process and
A step of inputting the histogram image into an image recognition model trained by a training data set, wherein the training data set is a plurality of data created from data showing biometric signals in a plurality of known states of the subject. Processes and processes, including training histogram images,
A method comprising the steps of processing the histogram image in the image recognition model and outputting the state of the object.
A computer system for creating histogram images,
How to get a wavelet image,
A means for dividing the wavelet image into a plurality of frequency bands,
A means of creating a histogram of spectral intensity for each of multiple post-split wavelet images,
A computer system with a means of combining multiple histograms created.
There is a program for creating a histogram image, the program is executed in a computer system equipped with a processor, and the program is
The process of acquiring a wavelet image and
The process of dividing the wavelet image into a plurality of frequency bands and
The process of creating a histogram of the spectral intensity for each of the multiple divided wavelet images,
A program that causes the processor to perform a process including a step of combining a plurality of created histograms.
A computer system for predicting the state of an object,
A receiving means for receiving data indicating the biological signal of the target, and
A means for creating a histogram image from data showing the biological signal. In the histogram image, the first axis represents the spectral intensity, the second axis represents the frequency, and the color represents the distribution ratio. The means of creation and
An image recognition model trained by a training dataset, wherein the training dataset includes image recognition including a plurality of training histogram images created from data showing biometric signals in a plurality of known states of the subject. With the model
A computer system including an output means for outputting the state of the object.
A program for predicting the state of a target, wherein the program is executed in a computer system including a processor, and the program is
The step of receiving data indicating the biological signal of the target and
In the step of creating a histogram image from the data showing the biological signal, the histogram image is a color map in which the first axis represents the spectral intensity, the second axis represents the frequency, and the color represents the distribution ratio. There is a process and
A step of inputting the histogram image into an image recognition model trained by a training data set, wherein the training data set is a plurality of data created from data showing biometric signals in a plurality of known states of the subject. Processes and processes, including training histogram images,
A program for causing the processor to perform a process including a step of processing the histogram image in the image recognition model and outputting the state of the target.
It is a method for predicting the characteristics of the target compound.
The step of receiving post-administration data indicating the biological signal of the subject after the subject compound is administered to the subject, and the step of receiving the post-administration data.
A step of creating a histogram image from the post-administration data according to the method according to any one of claims 1 to 6.
The process of extracting the feature vector from the histogram image and
In the step of comparing the feature quantity vector with the plurality of reference feature quantity vectors, each of the plurality of reference feature quantity vectors is known according to the method according to any one of claims 1 to 6. A step comprising a feature vector extracted from a histogram image created from baseline post-administration data showing each biometric signal after administration of the compound to a subject.
A method comprising the step of predicting the properties of the target compound based on the result of the comparison.
The step of comparing the feature quantity vector with the plurality of reference feature quantity vectors is
The process of creating a feature map by mapping the feature vector, and
In the step of comparing the feature amount map with the plurality of reference feature amount maps, each of the plurality of reference feature amount maps is a map created by mapping each of the reference feature amount vectors. , The method of claim 15, comprising the steps.
The method according to claim 16, wherein the step of comparing the feature amount map with the plurality of reference feature amount maps includes identifying at least one reference feature amount map similar to the feature amount map.
The method according to claim 16, wherein the step of comparing the feature amount map with the plurality of reference feature amount maps includes ranking the plurality of reference feature amount maps in an order similar to the feature amount map.