State prediction device, classification model creation device, state prediction method, classification model creation method, and program

Abstract

This state prediction device predicts, in a flow cytometry method, the state of an inspection specimen provider from a plurality of observation objects for inspection that, as observation objects included in a specimen acquired from a specimen provider, are included in an inspection specimen acquired from the inspection specimen provider. The state prediction device comprises: a signal information acquisition unit that, due to the configuration of structured illumination or the configuration of structured detection, acquires, for each of the plurality of observation objects, signal information indicating a temporal change in the intensity of an electromagnetic wave, which is generated by a light detector that has received an electromagnetic wave modulated by the observation objects existing in a light irradiation area irradiated with illumination light from a light source; and a state prediction unit that, on the basis of a classification model, which has been trained on the relationship between training signal information including signal information for each of training observation objects and the state of a training specimen provider, predicts the state of the inspection specimen provider from inspection signal information including the signal information acquired for each of the plurality of observation objects for inspection.

Description

State prediction device, classification model creation device, state prediction method, classification model creation method, and program

[0001] This application claims priority to U.S. Provisional Application No. 63 / 570,273, filed March 27, 2024, the contents of which are incorporated herein by reference.

[0002] Traditionally, the disease status or risk of a patient or subject has been assessed. For example, in the diagnosis of leukemia, blood tests, bone marrow tests, chromosome and genetic tests, imaging tests, and cell surface marker tests are performed. Imaging tests examine the condition of organs and bone marrow. While imaging tests can be non-invasive, invasive tests such as CT scans and ultrasound scans using contrast agents can also be performed. In cell surface marker tests, for example, fluorescently labeled antibodies are added to cells in a sample, and the type of leukemia cell is identified by flow cytometry. Flow cytometry is a widely used cell biology technique that uses lasers to count, differentiate, and characterize cells in a heterogeneous mixture.

[0003] Ghost cytometry (GC) technology is known as a technology that can acquire more abundant and higher-resolution morphological information of cells than conventional flow cytometry (Patent Documents 1 and 2). In GC technology, for example, an optical signal emitted by an observation object, such as a cell, when the object is irradiated with structured illumination light is used to acquire morphological information of the observation object. The optical signal acquired as time-series waveform data by the structured illumination configuration contains compressed morphological information of the observation object irradiated with the structured illumination light.

[0004] International Publication No. WO 2016 / 136801 International Publication No. WO 2017 / 073737

[0005] When using flow cytometry that is not based on GC technology, it has been impossible to obtain morphological information of an object of observation with high accuracy in an unlabeled (label-free) state. Furthermore, conventional invasive tests such as imaging tests require a long testing time, and it is sometimes impossible to obtain morphological information of the object of observation in a label-free state. Therefore, there is a need for a method and device that can predict disease or morbidity risk quickly and accurately using a simple, non-invasive method.

[0006] The present invention has been made in consideration of the above points, and provides a condition prediction device, a classification model creation device, a condition prediction method, a classification model creation method, and a program that can predict disease or the risk of morbidity quickly and accurately using a simple configuration and a non-invasive method.

[0007] One aspect of the present invention is a state prediction device for predicting the state of a test specimen provider from a plurality of test observation objects contained in a test specimen obtained from the test specimen provider in a flow cytometry method, the state prediction device comprising one or more of a structured illumination configuration in which illumination light from a light source is converted into structured illumination light and irradiated onto the test specimen, and a structured detection configuration in which modulated electromagnetic waves from the test specimen are structured and detected, and modulated electromagnetic waves from the test specimen present in a light irradiation area irradiated with illumination light from the light source are received by a photodetector. and a state prediction unit that predicts the state of the test specimen provider from test signal information including the signal information acquired by the signal information acquisition unit for each of the plurality of test observation objects, based on a classification model that is a trained model that has learned the relationship between training signal information including the signal information for each of the plurality of training observation objects, which are the observation objects contained in the training sample acquired from the training specimen provider, and the state of the training specimen provider.

[0008] One aspect of the present invention is a classification model creation device for use in a flow cytometry method, which creates a classification model, which is a trained model that predicts the state of a test specimen provider, from a plurality of test observation objects contained in a test specimen obtained from the test specimen provider, the test specimen being an observation object contained in the test specimen provided from the test specimen provider. The classification model creation device includes: a signal information acquisition unit that acquires, for each of the plurality of observation objects, signal information indicating a time change in intensity of the electromagnetic waves generated when modulated electromagnetic waves from the observation objects present in a light irradiation area irradiated with illumination light from a light source are received by a photodetector using one or more of a structured illumination configuration in which illumination light from a light source is converted into structured illumination light and irradiated onto the observation objects, or a structured detection configuration in which modulated electromagnetic waves from the observation objects are structured and detected; and a learning unit that creates the classification model based on machine learning from the relationship between training signal information, which includes the signal information for each of the plurality of training observation objects that are the observation objects contained in the training specimen obtained from the training specimen provider, and the state of the training specimen provider.

[0009] One aspect of the present invention is a state prediction method for predicting the state of a test specimen provider from a plurality of test observation objects contained in a test specimen obtained from the test specimen provider in a flow cytometry method, the method comprising: a structured illumination configuration in which illumination light from a light source is converted into structured illumination light and irradiated onto the test specimen; and a structured detection configuration in which modulated electromagnetic waves from the test specimen are structured and detected, and the modulated electromagnetic waves from the test specimen present in a light irradiation area irradiated with illumination light from the light source are received by a photodetector to generate a signal. and a state prediction step of predicting the state of the test specimen provider from test signal information including the signal information acquired in the signal information acquisition step for each of the plurality of test observation objects, based on a classification model which is a trained model that has learned the relationship between training signal information including the signal information for each of the plurality of training observation objects, which are the observation objects contained in the training specimen acquired from the training specimen provider, and the state of the training specimen provider.

[0010] In one aspect of the present invention, in a flow cytometry method, a computer of a state prediction device that predicts the state of a test specimen provider from a plurality of test specimens contained in a test specimen obtained from the test specimen provider is provided with one or more of a structured illumination configuration that converts illumination light from a light source into structured illumination light and irradiates the test specimens, or a structured detection configuration that structures and detects modulated electromagnetic waves from the test specimens, and receives modulated electromagnetic waves from the test specimens present in a light irradiation area irradiated with illumination light from the light source, and generates a signal. and a state prediction step of predicting the state of the test specimen provider from test signal information including the signal information acquired in the signal information acquisition step for each of the plurality of test observation objects, based on a classification model that is a trained model that has learned the relationship between training signal information including the signal information for each of the plurality of training observation objects, which are the observation objects contained in the training sample acquired from the training specimen provider, and the state of the training specimen provider.

[0011] One aspect of the present invention is a classification model creation method for creating a classification model, which is a trained model that predicts the state of a test specimen provider, from a plurality of test observation objects that are observation objects contained in a sample obtained from the test specimen provider in a flow cytometry method, the plurality of test observation objects being observation objects contained in the test specimen obtained from the test specimen provider. The classification model creation method includes: a signal information acquisition step of acquiring, for each of the plurality of observation objects, signal information indicating a time change in intensity of the electromagnetic waves that are generated when modulated electromagnetic waves from the observation objects present in a light irradiation area that is irradiated with illumination light from a light source are received by a photodetector using one or more of a structured illumination configuration that converts illumination light from a light source into structured illumination light and irradiates the observation objects, or a structured detection configuration that structures and detects modulated electromagnetic waves from the observation objects; and a learning step of creating the classification model based on machine learning from the relationship between training signal information that includes the signal information for each of the plurality of training observation objects that are observation objects contained in the training specimen obtained from the training specimen provider, and the state of the training specimen provider.

[0012] One aspect of the present invention is a program for causing a computer of a classification model creation device that creates a classification model, which is a trained model that predicts the state of a test specimen provider, from a plurality of test observation objects that are observation objects contained in a test specimen obtained from the test specimen provider in a flow cytometry method, to execute: a signal information acquisition step of acquiring, for each of the plurality of observation objects, signal information indicating a time change in intensity of the electromagnetic waves that are generated when modulated electromagnetic waves from the observation objects present in a light irradiation area that is irradiated with illumination light from a light source and are received by a photodetector using one or more of a structured illumination configuration that converts illumination light from a light source into structured illumination light and irradiates the observation objects, or a structured detection configuration that structures and detects modulated electromagnetic waves from the observation objects; and a learning step of creating the classification model based on machine learning from the relationship between training signal information that includes the signal information for each of the plurality of training observation objects that are the observation objects contained in the training specimen obtained from the training specimen provider, and the state of the training specimen provider.

[0013] According to the present invention, a disease or risk of developing a disease can be predicted quickly and accurately using a simple and non-invasive method.

[0014] 1 is a diagram showing an overview of a state prediction method using a flow cytometer 1 according to a first embodiment of the present invention. FIG. 1 is a diagram showing an overview of a method for acquiring GMI waveform information from an observation object contained in a specimen according to a first embodiment of the present invention. FIG. 2 is a diagram showing an overview of a process for aggregating data obtained by performing dimensional reduction on GMI waveform information according to a first embodiment of the present invention and using the data as learning data. FIG. 2 is a diagram showing an overview of a learning process for creating a classification model according to a first embodiment of the present invention. FIG. 3 is a diagram showing an overview of a prediction process for a subject's state according to a first embodiment of the present invention. FIG. 4 is a diagram showing an example of the configuration of a flow cytometer 1 and a spatial light modulation unit 4 included in the flow cytometer 1 according to a first embodiment of the present invention. FIG. 5 is a diagram showing an example of the configuration of a state prediction device 10 according to a first embodiment of the present invention. FIG. 6 is a diagram showing an example of the flow of a learning process according to the first embodiment of the present invention. FIG. 7 is a diagram showing an example of the flow of a prediction process according to the first embodiment of the present invention. FIG. 8 is a diagram showing an overview of a learning process for creating a classification model according to a modified example of the first embodiment of the present invention. FIG. 9 is a diagram showing an overview of a prediction process for a subject's future state according to a modified example of the first embodiment of the present invention. FIG. 10 is a diagram showing an overview of a cell-level classification model according to a second embodiment of the present invention. FIG. 11 is a diagram showing an overview of a process for predicting a subject's state from a classification result by a cell-level classification model according to a second embodiment of the present invention. FIG. 11 is a diagram showing an example of the configuration of a state prediction device 10a according to a second embodiment of the present invention. FIG. 12 is a diagram showing an example of the flow of a learning process for an observation object classification model according to a second embodiment FIG. 1 is a diagram showing an example of the flow of a learning process of a classification model according to a second embodiment of the present invention. FIG. 2 is a diagram showing an example of the flow of a prediction process according to the second embodiment of the present invention. FIG. 3 is a diagram showing an overview of a classification model and prediction process according to a third embodiment of the present invention. FIG. 4 is a diagram showing an example of the configuration of a state prediction device 10b according to a third embodiment of the present invention. FIG. 5 is a diagram showing an example of the flow of a learning process according to the third embodiment of the present invention. FIG. 6 is a diagram showing an example of the flow of a prediction process according to the third embodiment of the present invention.

[0015] First Embodiment Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings. FIG. 1 is a diagram illustrating an overview of a condition prediction method using a flow cytometer 1 according to this embodiment. The flow cytometer 1 predicts the condition of a donor from multiple unlabeled observation objects contained in a sample using flow cytometry. The donor is, for example, an individual such as a donor who provides the sample whose condition is to be predicted. When the donor is an individual such as a human or animal, the donor's condition includes, for example, the type of disease or the state of the disease (early stage, late stage), and differences in responsiveness or resistance to drugs. The donor's condition also includes a combination of one or more of the type of disease, the state of the disease, and differences in responsiveness or resistance to drugs.

[0016] The provider may also be something like a unit of production, preparation, or storage of an observation object. A unit of production, preparation, or storage of an observation object is a certain amount of observation objects managed by the same production, preparation, or storage method. For example, in cell preparation, a certain amount of cell culture medium called a lot prepared by the same operation or processing at the same time is also included in the provider in this embodiment.

[0017] In this way, the donor whose state is predicted by the flow cytometer 1 includes an individual such as a human. The donor whose state is predicted by the flow cytometer 1 is also referred to as a test specimen donor. When the donor whose state is predicted by the flow cytometer 1 is a human, the test specimen donor is also referred to as a subject. The donor whose state is predicted by the flow cytometer 1 includes a unit of manufacture, preparation, or storage of an observation object such as a cell culture medium prepared under specified conditions or by a specified method, but the following explanation will focus on the case where the test specimen donor is a subject.

[0018] As will be described later as a modified example of this embodiment, the donor's condition predicted by the flow cytometer 1 is not limited to the current condition, but may also include a future condition. A classification model for predicting the donor's condition is used to predict the donor's condition by the flow cytometer 1. The flow cytometer 1 is also used not only to predict the donor's condition but also to create the classification model by machine learning. The following description will mainly focus on the case where the subject's condition predicted by the flow cytometer 1 is the subject's disease state.

[0019] To predict the condition of a subject using a classification model, for example, a time-series signal acquired for each subject using ghost cytometry (GC) technology is used. The time-series signal is also referred to as GMI (Ghost Motion Imaging) waveform information. The GMI waveform information is waveform data indicating the intensity of the time-series signal. The GMI waveform information contains compressed morphological information of the observed object. To predict the condition of a subject using a classification model, the time-series signal that is GMI waveform information can also be used in combination with a time-series signal acquired without using GC technology.

[0020] The observation object may be, for example, a cell. The observation object may also be a microorganism such as a bacterium, a fungus, a microalgae, or a protozoan. Therefore, the observation object may include one or more of a cell and a microorganism. The observation object may also be a spheroid (cell mass) consisting of multiple cells.

[0021] When the donor is a human, the specimen is, for example, a body fluid obtained from the individual, or a measurement sample obtained by diluting the body fluid with a buffer solution or the like as appropriate depending on the measurement. When the donor is a non-human individual (e.g., an animal), a measurement sample derived from body fluid can also be used as the specimen. Body fluid is a liquid that an individual has in some form within their body. Examples of body fluid include blood, lymph, tissue fluid (interstitial fluid, intercellular fluid, interstitial fluid), body cavity fluid, pleural effusion, ascites, cerebrospinal fluid (spinal fluid), synovial fluid (synovial fluid), and aqueous humor. In this embodiment, body fluid also includes various liquids secreted and excreted inside and outside the body, such as saliva, sweat, semen, and urine.

[0022] In this embodiment, the specimen also includes a measurement sample obtained at least in part from a unit of manufacture, preparation, or storage of the object to be observed (e.g., a production lot). In the following description, a specimen used by the flow cytometer 1 when predicting the condition of a donor will also be referred to as a test specimen. A test specimen is a specimen obtained from a donor (e.g., a subject) whose condition is to be predicted. A specimen obtained from a subject will also be referred to as a subject specimen. A specimen used by the flow cytometer 1 when training a classification model will also be referred to as a training specimen. A training specimen is a specimen obtained from a donor (e.g., a patient) to obtain training data to be used in creating a classification model.

[0023] A specimen contains a plurality of observation objects. For example, a specimen contains a plurality of cells as observation objects. Cells are an example of observation objects contained in a specimen. A plurality of objects means, for example, several thousand. There is no limit to the number of observation objects contained in a specimen, but it is desirable that the specimen contains 1,000 or more cells. It is desirable that the specimen contains more cells during learning than during evaluation. The number of cells contained in a test specimen may be fewer than that of a learning specimen, preferably 500 or more, but may also be around 100. In the following description, the observation object contained in a learning specimen will also be referred to as a learning observation object. Furthermore, the observation object contained in a test specimen will also be referred to as a test observation object.

[0024] A specimen may contain multiple cells. The multiple cells may be composed of, for example, different types of cells (e.g., different types of white blood cells, such as neutrophils and eosinophils). A specimen may also be composed of cells in different states (e.g., healthy cells and diseased cells). A specimen may also be composed of cells of the same type. For example, when samples of the same type of cells with different manufacturing conditions (e.g., culture period or culture method) or lots are used as specimens, the specimen may contain cells with different states, such as differentiation, metabolism, activation, and growth. Examples of cells with different states include cells with different efficacy or responsiveness to drugs, or cells with different productivity of target substances (antibodies or specific active ingredients). In this way, a specimen may contain cells of the same type but in different states. A specimen may contain cells with different types and / or states.

[0025] FIG. 2 shows an overview of a method for acquiring GMI waveform information from an observation object contained in a specimen. The GMI waveform information contains compressed morphological information of the observation object. FIG. 2 shows an overview of a method for acquiring GMI waveform information necessary for machine learning to create a classification model, using as an example a case where the specimen is a measurement sample containing blood collected from a patient suffering from acute myeloid leukemia (AML), and the observation object contained in the specimen is a cell. A specimen from which training data is acquired to create a classification model is called a training specimen. The provider who acquires the training specimen will also be referred to as the training specimen provider.

[0026] The same type of specimen as the subject specimen is used for the training specimen. Furthermore, the observation object contained in the training specimen includes the same type of observation object as the observation object contained in the subject specimen. As described above, the observation object contained in the training specimen is also referred to as the training observation object. The observation object contained in the subject specimen is also referred to as the subject observation object. The subject specimen is a specimen obtained from the subject whose condition is to be predicted. For example, if the subject specimen is blood and the subject observation object is cells, the training specimen is a measurement sample containing blood, and the training observation object is also cells.

[0027] GMI waveform information is acquired from the learning observation object by GC technology. The GMI waveform information is waveform information including cellular morphology information acquired by GC technology. Similarly to the learning observation object, GMI waveform information is also acquired from the subject observation object contained in the subject sample by GC technology. The learning sample differs from the subject sample in that it is pre-labeled with information regarding the condition of the donor (learning sample donor) from whom the sample was obtained. As a result, for example, if the learning sample donor is a patient suffering from AML, the GMI waveform information acquired from the learning observation object by GC technology is labeled as indicating that the donor is a patient suffering from AML, regardless of the type of cells in the observation object.

[0028] 2 illustrates an example in which the training observation object is a cell contained in a training sample obtained from an AML patient. When obtaining GMI waveform information from the training observation object to be used as training data for creating a classification model, multiple types of training samples provided by multiple types of training sample providers (e.g., AML patients, CML patients, and healthy individuals) are prepared, and GMI waveform information is obtained from the training observation object contained in these multiple types of training samples. The training sample is a sample obtained from the training sample provider.

[0029] FIG. 2 illustrates an example in which samples collected from n (n is a natural number) subjects suffering from AML are measured. Each sample contains m (m is a natural number) cells. Therefore, GMI waveform information is acquired for the number obtained by multiplying n by m. Note that m, the number of cells contained in each sample, may vary depending on the sample.

[0030] The cells contained in a single specimen do not necessarily have the same morphology. For example, the cells contained in a specimen collected from a donor suffering from AML may include normal cells unrelated to the disease and cells specific to the disease. Thus, the cells contained in a single specimen may include cells with different morphologies.

[0031] Multiple cells contained in samples collected from training sample donors in the same state are assigned the same label indicating the state of the donor. Figure 2 shows an example in which GMI waveform information obtained from each of m cells contained in five training samples is assigned a label indicating that the cells are suffering from AML.

[0032] Here, a label is information assigned to a training sample that identifies the condition of the donor from whom the sample was obtained. Labeling refers to linking information about the condition of the training sample donor with GMI waveform information acquired from the training sample. For example, when a body fluid acquired from a patient suffering from a specific disease (e.g., AML) is used as a training sample, labeling refers to linking and handling the individual GMI waveform information acquired from the sample with information about the specific disease (e.g., AML) assigned to the training sample.

[0033] Labels assigned to training samples can be assigned based on, for example, information regarding the expression levels of markers identifying unique cells contained in the sample. Unique cells include, but are not limited to, cells that have mutated from a normal state, such as cancerous cells or cells affected by a certain disease. In this case, the labels assigned to training samples can be obtained as measurements of the unique cells contained in the training sample. Furthermore, labels assigned to training samples can also be obtained from information regarding the results of other tests undergone by the training sample provider. Examples of other tests undergone by the training sample provider include imaging diagnostics such as MRI (Magnetic Resonance Imaging), CT (Computed Tomography), X-ray, nuclear medicine, and ultrasound, as well as blood pressure measurements, electrocardiograms, and blood tests. Furthermore, labels may be assigned to training samples based on genetic testing information for genes obtained from bodily fluids or tissue sections collected from the training sample provider.

[0034] In this embodiment, dimension reduction processing is performed on the GMI waveform information acquired for each sample. Next, data obtained by dimension reduction for training samples with the same label is accumulated and used as training data for machine learning. Figure 3 shows an overview of the process of accumulating data obtained by performing dimension reduction on GMI waveform information and using it as training data.

[0035] First, an overview of the dimensionality reduction process for GMI waveform information will be described. Dimensionality reduction is performed on multiple GMI waveform information acquired from one training sample. In the example of FIG. 3, the dimensionality reduction process visualizes the morphological information contained in the GMI waveform information as a two-dimensional scatter plot. Any dimensionality reduction method may be used. Examples of dimensionality reduction methods include uniform manifold approximation and projection (UMAP), principal component analysis (PCA), t-distributed stochastic neighbor embedding (t-SNE), and autoencoders. A neural network can also be used as a dimensionality reduction method. In the following description, data obtained by performing dimensional reduction on GMI waveform information will also be referred to as dimensionally reduced data.

[0036] The GMI waveform information acquired for multiple cells contained in one training sample is treated as a single population for each training sample. That is, the morphological information contained in the GMI waveform information for multiple cells contained in one training sample is collected and included in a single dimension-reduced data set. In FIG. 3 , as an example, the training sample is labeled with AML, which indicates the donor's condition. There are n AML training samples, and the ith training sample (i = 1, 2, ..., n: n is the number of training samples) among the n training samples contains mi cells. Because training samples do not necessarily contain the same number of cells, mi may vary depending on i.

[0037] Although the GMI waveform information contains compressed morphological information of the measured cells, the measurer cannot directly recognize it, making it difficult to utilize the abundant morphological information about cells contained in the GMI waveform information as it is. Furthermore, considering the entire i-th training sample, the mi pieces of GMI waveform information contain morphological information about cells derived from mi cells, but the amount of information is so large that it is difficult to comprehensively grasp the morphological characteristics of multiple cells contained in one training sample from the mi pieces of GMI waveform information. The dimension-reduced data obtained by the dimension reduction process displays the morphological information about multiple cells contained in one training sample in a more easily understandable manner.

[0038] In Fig. 3, mi GMI waveform information acquired from mi cells contained in the ith training sample is visualized as a single two-dimensional scatter plot through dimension reduction processing, and the distribution based on the morphological characteristics of multiple cells contained in the sample is displayed in a form that is easier to understand. In Fig. 3, GMI waveform information related to multiple cells acquired from one training sample is visualized as a single two-dimensional scatter plot through dimension reduction processing, but in the dimension reduction processing of GMI waveform information according to this embodiment, the GMI waveform information may be converted into dimension-reduced data other than a two-dimensional scatter plot.

[0039] As described above, the dimension-reduced data (e.g., the two-dimensional scatter plot in FIG. 3 ) obtained by dimension reduction processing from multiple pieces of GMI waveform information acquired from one training sample expresses morphological information about multiple cells contained in the training sample, which is included in the GMI waveform information. Furthermore, the dimension-reduced data (e.g., the two-dimensional scatter plot in FIG. 3 ) obtained by dimension reduction processing is also labeled with information about the condition of the training sample donor that is assigned to the training sample.

[0040] By similarly applying the above-described dimension reduction method to the m i GMI waveform information contained in the i-th training sample, dimension-reduced data for the i-th training sample is obtained. By repeatedly performing the same process on the first through n-th training samples, dimension-reduced data for each of the n training samples is obtained. As described above, in the example of FIG. 3 , the n training samples are each collected from a donor suffering from AML. Therefore, the dimension-reduced data for each of the n training samples is assigned the same label indicating the donor's condition, AML. The dimension-reduced data thus collected for the n training samples is used as training data for creating a classification model.

[0041] The state prediction device 10 according to the embodiment focuses on the distribution state in the dimension-reduced data, i.e., how the plurality of cells contained in the sample are distributed as a group when viewed in terms of a certain morphological characteristic, based on the morphological information of the plurality of cells contained in the sample, rather than on identifying the individual cells contained in the sample. The state prediction device 10 according to the present embodiment uses machine learning to learn the relationship between the sample and the label assigned to the sample that indicates the state of the training sample provider, based on information on how the plurality of cells contained in the sample are arranged in the dimension-reduced data when viewed in terms of a specific morphological characteristic.

[0042] For example, in Figure 3, a two-dimensional scatter plot of n training samples, each labeled with the same AML label, is used as training data for a donor with AML. This training data is used as training data for the condition of a specific donor (in this case, a donor with AML) in machine learning. Similarly, training data is created to be used as training data for the conditions of other donors (e.g., donors with CML and healthy donors).

[0043] Next, the creation of a classification model and the prediction of the subject's condition will be described. FIG. 4 shows an overview of the learning process for creating a classification model. The creation of the training data described above has mainly been described for the case where the label indicates AML (i.e., the state of the donor from whom the sample was collected is AML). Similarly, training data can be created by assigning a label indicating chronic myelogenous leukemia (CML) to dimension-reduced data obtained from training samples collected from donors (CML patients) suffering from CML. Furthermore, training data can be created by assigning a label indicating HS (Healthy Subject) to dimension-reduced data obtained from training samples collected from healthy donors (healthy individuals).

[0044] In this way, training data serving as training data for each condition (pathological condition) is created. As described above, the condition may include not only a diseased state but also a healthy state. Furthermore, multiple sets of training data may be created for one type of condition. For example, a label indicating AML may be assigned to each of the dimension-reduced data obtained for each of multiple training samples collected from k1 sets of donors (AML patients) suffering from AML, thereby creating k1 sets of training data for AML patients. Similarly, for example, training data for k2 sets of CML patients labeled with CML and training data for k3 sets of healthy individuals labeled with HS may be created. Note that k1, k2, and k3 are each a natural number.

[0045] A classification model is created by learning the model through machine learning using the created learning data. That is, the classification model is a trained model that has been trained in advance using the created learning data. As will be described later, a feature vector extracted from the dimensionality-reduced data is input to the classification model. Any supervised learning method may be used as the machine learning method, and for example, algorithms such as kNN (k-Nearest Neighbor) classifier, logistic regression, random forest, and support vector machine (SVM) may be used.

[0046] Figure 5 shows an overview of the process for predicting the state of a subject. In this process, dimension-reduced data is first generated from GMI waveform information obtained from each of multiple cells contained in a subject sample. A subject sample is a sample collected from a subject whose state is the target of prediction and whose state is unknown. Next, the dimension-reduced data generated for the subject is input into a classification model, which outputs the predicted state of the subject.

[0047] [Configuration of Flow Cytometer 1] FIG. 6 shows an example of the configuration of a flow cytometer 1 according to this embodiment and a spatial light modulation unit 4 included in the flow cytometer 1. FIG. 6(A) shows an example of the configuration of the flow cytometer 1. A flow cytometer is a measuring instrument that includes at least a microfluidic device having a flow path through which an observation object contained in a sample can flow together with a fluid, a light source that irradiates the flow path with illumination light, and a photodetector that detects signal light emitted from the observation object when the illumination light is irradiated onto the observation object flowing through the flow path. In the flow cytometer, the observation object flowing together with the fluid through the flow path is optically measured while moving within the flow path. The flow cytometer 1 according to this embodiment includes a microfluidic device 2, a light source 3, a spatial light modulation unit 4, a light detection optical system 5, a photodetector 6, a DAQ (Data Acquisition) device 7, and a personal computer (PC) 8.

[0048] The microfluidic device 2 includes a flow channel 20 through which an object to be observed can flow together with a fluid. The object to be observed is, for example, contained in a specimen collected from the bodily fluid of a donor. One example of the object to be observed is a cell. In FIG. 6, cell C1 is shown as an example of a cell. The microfluidic device 2 sequentially flows a plurality of objects to be observed through the flow channel 20, but it is preferable that only one object to be observed pass through the illumination light irradiation position of the flow channel 20 at a time.

[0049] The light source 3 and the spatial light modulation unit 4 are an optical configuration that irradiates the observation object with structured illumination light. As will be described below, the light source 3 and the spatial light modulation unit 4 irradiate the observation object passing through the flow path 20 with structured illumination light SLE1.

[0050] The illumination light LE1 emitted from the light source 3 is converted into structured illumination light SLE1 by the spatial light modulation unit 4 and irradiated onto the irradiation position of the flow channel 20. The light source 3 is, for example, a laser light source, a semiconductor laser light source, or an LED (Light Emitting Diode) light source. The light source 3 may be a fiber laser light source. The illumination light LE1 emitted by the light source 3 may be continuous light or pulsed light, but continuous light is preferable. The illumination light LE1 emitted by the light source 3 may be spatially incoherent light, but spatially coherent light is preferable. In this embodiment, the illumination light LE1 emitted by the light source 3 is, for example, spatially coherent light. Furthermore, the illumination light LE1 emitted by the light source 3 is preferably light of a single wavelength, but may also be multi-wavelength illumination light having a discrete spectrum by mixing light of multiple specific wavelengths.

[0051] In this embodiment, the spatial light modulation unit 4 is disposed on the optical path between the light source 3 and the flow path 20. This arrangement is also referred to as a structured illumination configuration. Illumination light LE1 emitted from the light source 3 is converted into structured illumination light SLE1 by the spatial light modulation unit 4, and the structured illumination light SLE1 is irradiated onto the flow path 20. The structured illumination light SLE1 is illumination light having a structured illumination pattern described below, and is irradiated onto the observation object so that the structured illumination pattern is imaged at the irradiation position on the flow path 20.

[0052] Here, the spatial light modulation unit 4 will be described with reference to FIG. 6(B). The spatial light modulation unit 4 is an optical system including a spatial light modulator 40. FIG. 6(B) is a diagram showing an example of the spatial light modulation unit 4 according to this embodiment. In FIG. 6(B), the spatial light modulation unit 4 includes the spatial light modulator 40, a first lens 41, a spatial filter 42, a second lens 43, and an objective lens 44. In the spatial light modulation unit 4, the spatial light modulator 40, the first lens 41, the spatial filter 42, the second lens 43, and the objective lens 44 are arranged on the optical path between the light source 3 and the flow path 20 in this order from the side closest to the light source 3.

[0053] The spatial light modulator 40 changes the optical characteristics of the incident light for each of multiple regions included in the incident surface of the incident light. The spatial light modulator 40 is an optical element that changes the optical characteristics of incident light, structuring the illumination light LE1 and converting it into structured illumination light SLE1. The surface onto which the light of the spatial light modulator 40 is incident has multiple regions, and the optical characteristics of the illumination light LE1 are individually converted in each of the multiple regions through which the light passes. In other words, the optical characteristics of the light that passes through the spatial light modulator 40 are changed so that they differ from each other in the multiple regions compared to the optical characteristics of the incident light. Here, the optical characteristics refer to, for example, one or more of the optical characteristics of intensity, wavelength, phase, and polarization state, but are not limited to these. The spatial light modulator 40 changes the optical characteristics of the incident illumination light LE1 for each of the multiple regions included in the spatial light modulator 40, enabling the structured illumination light SLE1 to be irradiated onto an observation object. The structured illumination light SLE1 is provided with a spatial distribution pattern of light characteristics (e.g., a spatial distribution of light intensity) based on changes in light characteristics occurring in each of the multiple regions. Examples of the distribution pattern of light characteristics include a random pattern, a regular pattern, or a raster pattern.

[0054] The spatial distribution pattern of light characteristics imparted to the structured illumination light SLE1 based on the changes in light characteristics occurring in each of a plurality of regions is also referred to as a structured illumination pattern. Also, changing the light characteristics of the incident light for each of a plurality of regions included in the incident surface of the incident light and imparting a spatial distribution pattern of light characteristics to the illumination light based on the changes in light characteristics occurring in each of the regions is also referred to as structuring the illumination light.

[0055] In this embodiment, the spatial light modulator 40 is, for example, a diffractive optical element (DOE), which is an optical element that controls the diffraction of light using a microscopic shape formed thereon. Other examples of the spatial light modulator 40 include a spatial light modulator (SLM) and a digital micromirror device (DMD). Further examples of the spatial light modulator 40 include a film on whose surface multiple regions with different optical properties are printed, or a structured detection mask or optical filter in which light-transmitting portions and light-blocking portions are arranged to form a binary pattern. Note that when the illumination light LE1 emitted by the light source 3 is incoherent light, the spatial light modulator 40 includes a DMD.

[0056] As the object of observation passes through the illumination region illuminated with the structured illumination light SLE1, an optical signal LS1 is generated in which the structured illumination light SLE1 is modulated by the object of observation. The optical signal LS1 is an electromagnetic wave generated by modulating the structured illumination light SLE1 by the object of observation. The optical signal LS1 is, for example, scattered light. The scattered light is an example of an electromagnetic wave modulated by the object of observation present in the illumination region illuminated with the structured illumination light SLE1.

[0057] Other examples of the optical signal LS1 include fluorescence emitted when fluorescent molecules are excited by the structured illumination light SLE1 while the object of observation passes through the light irradiation area illuminated with the structured illumination light SLE1, and transmitted light when the structured illumination light SLE1 passes through the object of observation. The optical signal LS1 also includes diffracted light and interference light generated while the object of observation passes through the light irradiation area illuminated with the structured illumination light SLE1.

[0058] 6 illustrates an example in which the optical signal LS1 is detected as scattered light by a photodetector 6 installed in the forward direction. Therefore, the following description will be directed to a case in which the optical signal LS1 is scattered light. However, the method described below can also be applied to cases in which the optical signal LS1 is a type of electromagnetic wave other than scattered light, such as fluorescence, transmitted light, or diffracted light. The method described below can also be applied simultaneously to multiple different types of optical signals LS1. For example, the phase difference between cell-derived fluorescence (autofluorescence) or transmitted light and diffracted light can be detected as GMI waveform information. Dark-field light can also be acquired as GMI waveform information.

[0059] The light detection optical system 5 is an optical mechanism for focusing the light signal LS1, which is light scattered from the object of observation, onto the photodetector 6, and includes an imaging lens (not shown in FIG. 6 ). The imaging lens (not shown in FIG. 6 ) included in the light detection optical system 5 focuses the light signal LS1, which is light scattered from the object of observation, onto the position of the photodetector 6. Furthermore, the light detection optical system 5 may further include a dichroic mirror or a wavelength-selective optical filter in addition to the imaging lens.

[0060] The optical signal LS1 acquired using the GC technology contains compressed morphological information of the object being observed. The morphological information of the object being observed acquired using the GC technology includes information on the internal structure of the object being observed (such as the structure and arrangement of organelles in the case where the object being observed is a cell) in addition to the shape of the object being observed.

[0061] The photodetector 6 detects the optical signal LS1 collected by the imaging lens (not shown in FIG. 6 ). The photodetector 6 detects the optical signal LS1 and converts it into an electrical signal. An example of the photodetector 6 is a photomultiplier tube (PMT). The photodetector 6 detects the intensity of the optical signal LS1 collected by the imaging lens (not shown in FIG. 6 ) in a time series. The photodetector 6 may be a single sensor composed of a single light-receiving element, or a multi-sensor composed of multiple light-receiving elements. As another example of the photodetector 6, a photodiode (PD), an avalanche photodiode (APD), or a line-type PMT may be used.

[0062] The DAQ device 7 converts the electrical signal pulses output by the photodetector 6 into electronic data for each pulse. The electronic data includes a pair of time and the intensity of the electrical signal pulse. The electronic data is time-series waveform data. One example of the DAQ device 7 is an oscilloscope.

[0063] The PC 8 is a device provided in the flow cytometer 1 that performs information processing. This information processing includes processing for predicting the state of the subject. The electronic data output from the DAQ device 7 is called signal information D1. The signal information D1 is information indicating the temporal change in the intensity of the optical signal LS1. In other words, the signal information D1 is electronic data indicating the temporal change in the intensity of the optical signal LS1 (for example, scattered light) emitted from an observation object present in a light irradiation area irradiated with the structured illumination light SLE1. The PC 8 is an example of a state prediction device. This state prediction device predicts the state of the subject based on signal information D1 acquired from multiple unlabeled subject observation objects contained in the subject's sample.

[0064] The PC 8 acquires the signal information D1 from the DAQ device 7 and generates waveform data based on the acquired signal information D1. The waveform data indicates changes in the intensity of the electromagnetic wave over time. The signal information D1 includes the GMI waveform information described above.

[0065] In this embodiment, an example will be described in which the optical signal LS1 is detected using a structured illumination configuration in the flow cytometer 1, but the configuration for acquiring the optical signal LS1 using GC technology is not limited to this. The optical signal LS1 may also be detected using a structured detection configuration in the flow cytometer 1. In the structured detection configuration, the spatial light modulation unit 4 is disposed on the optical path between the flow path 20 and the photodetector 6.

[0066] Furthermore, the flow cytometer 1 may be provided with an optical configuration that detects the intensity of electromagnetic waves (e.g., scattered light) from an object of observation over time without using GC technology, in addition to the configuration that acquires the optical signal LS1 using the above-described GC technology. In this case, the optical configuration that detects electromagnetic waves from an object of observation without using GC technology (hereinafter referred to as the conventional optical configuration) detects the intensity of electromagnetic waves (e.g., scattered light) from the object of observation over time using an optical configuration that does not include the spatial light modulation unit 4 from the configuration shown in FIG. 6A. The conventional optical configuration is composed of optical elements such as a light source, lenses, optical filters, and mirrors, but some of the optical elements can also be used as optical elements included in the configuration that acquires the optical signal LS1 using GC technology.

[0067] [Configuration of State Prediction Device 10] FIG. 7 is a diagram showing an example of the configuration of the state prediction device 10 according to this embodiment. The state prediction device 10 is a device that predicts the state of a subject based on signal information D1 acquired by a flow cytometer 1 from multiple unlabeled subject observation objects contained in the subject's sample. The state prediction device 10 includes a control unit 11 and a storage unit 12. The control unit 11 includes, for example, a central processing unit (CPU), a graphics processing unit (GPU), a field-programmable gate array (FPGA), etc., and performs various calculations and information exchange. Each functional unit of the control unit 11 is realized by the CPU expanding a program from a read-only memory (ROM) to a random access memory (RAM), reading the program, and executing processing. Each functional unit of the control unit 11 can be realized by the CPU reading a program from a hard disk and executing the process. Also, the control unit 11 can read and execute all or part of the program from an external device (not shown).

[0068] The control unit 11 includes a signal information acquisition unit 110 , a dimension reduction unit 111 , a state prediction unit 112 , an output unit 113 , a learning unit 114 , and an operation unit 115 .

[0069] The signal information acquiring unit 110 acquires signal information D1 from the DAQ device 7. The signal information acquiring unit 110 acquires signal information D1 for each of the multiple observation objects. The signal information D1 includes GMI waveform information acquired using the above-described GC technology. Therefore, the signal information acquiring unit 110 acquires, for each of the multiple observation objects, signal information D1 indicating a temporal change in the intensity of electromagnetic waves generated when modulated electromagnetic waves from the observation object present in the light irradiation area irradiated with illumination light LE1 from the light source 3 are received by the photodetector 6 using one or more of the structured illumination configuration and the structured detection configuration. As described above, in the structured illumination configuration, the illumination light LE1 from the light source 3 is converted into structured illumination light SLE1 and irradiated onto the observation object included in the subject's sample.

[0070] The dimension reduction unit 111 performs dimension reduction on the signal information D1 for each of the multiple observation objects acquired by the signal information acquisition unit 110. The dimension reduction unit 111 generates dimension-reduced data as data obtained by performing dimension reduction on the signal information D1 for each of the multiple observation objects acquired by the signal information acquisition unit 110.

[0071] The state prediction unit 112 predicts the state of the subject from the dimension-reduced data based on the classification model L1.

[0072] The output unit 113 outputs the prediction result by the state prediction unit 112 to an external device (not shown). The external device is, for example, a display device or other display device, or a storage device provided in a server, etc. The output unit 113 may output the prediction result to its own display device or storage device. In this case, the state prediction device 10 has a display device and a storage device (not shown) in addition to the control unit 11.

[0073] The learning unit 114 learns the classification model L1. In the classification model L1, the relationship between dimension-reduced data obtained by dimension reduction of signal information D1 for each of a plurality of unlabeled cells (learning observation objects) contained in a sample obtained from a training sample provider and labels indicating the state of the training sample provider is learned, and a trained classification model L1 is created. Learning the classification model L1 is also referred to as creating the classification model L1.

[0074] The operation unit 115 accepts various operations from the person performing the measurement.

[0075] The storage unit 12 stores various types of information. The storage unit 12 is configured using a storage device such as a magnetic hard disk drive or a semiconductor storage device. The storage unit 12 stores, for example, a trained classification model L1. The storage unit 12 may also store signal information D1 acquired by the signal information acquisition unit 110.

[0076] The functional units of the control unit 11 may be distributed across multiple servers. For example, the learning unit 114 may be provided in a learning device that is a separate device from the state prediction device 10. In that case, the state prediction device 10 acquires the classification model L1 created by the learning device. The state prediction device 10 may also be configured not to include the memory unit 12. In that case, the memory unit 12 may be provided in an external server, and may communicate with the state prediction device 10 to exchange information as needed.

[0077] [Learning Process] The learning process, which is a process for learning the classification model L1, will be described. Fig. 8 is a diagram showing an example of the flow of the learning process according to this embodiment. The learning process is executed prior to the prediction process performed by the control unit 11 of the state prediction device 10.

[0078] Step S10: The signal information acquisition unit 110 acquires signal information D1 for each of the multiple observation objects from the DAQ device 7. The multiple observation objects are included in a training sample collected from a training sample provider. The signal information acquisition unit 110 supplies the acquired signal information D1 for each of the multiple observation objects to the dimension reduction unit 111. The signal information D1 acquired from the training observation objects is also referred to as training signal information. The training observation objects are observation objects included in the training sample.

[0079] Step S20: The dimension reduction unit 111 performs dimension reduction on the signal information D1 for each of the plurality of observation objects acquired by the signal information acquisition unit 110. The dimension reduction unit 111 generates dimension-reduced data by the dimension reduction.

[0080] Step S30: The learning unit 114 extracts feature vectors from the dimension-reduced data. As a method for extracting feature vectors from the dimension-reduced data, any of the first, second, and third extraction methods described below can be used, but other methods may also be used to extract feature vectors.

[0081] The first extraction method can be used when the signal information D1 is converted into two-dimensional dimension-reduced data by dimension reduction. The signal information D1 includes GMI waveform information. The two-dimensional dimension-reduced data includes a two-dimensional scatter plot. In the first extraction method, first, data related to individual observations included in the two-dimensional scatter plot generated by dimension reduction are classified into clusters designated by the operator. Next, for each two-dimensional scatter plot for each specimen, the number of data included in each cluster is counted. Each piece of data included in each cluster corresponds to one observation in the specimen. As a result, for each two-dimensional scatter plot for each specimen, the ratio of the data included in each cluster to the total data included in the dimension-reduced data can be calculated. Finally, a feature vector having the ratio as its components is generated. In the first extraction method, the dimension of the feature vector is equal to the number of clusters designated by the operator.

[0082] In the second extraction method, first, the space of the dimension-reduced data for each specimen (e.g., a two-dimensional scatter plot) is divided into grids of a predetermined size. Next, the density of the data contained in each grid is calculated. A feature vector is generated by using the density of the data contained in each grid as components. That is, in the second extraction method, the dimension of the feature vector is equal to the number of grids. In the second extraction method, the degree of data concentration is automatically extracted as the density of the data contained in each grid. The second extraction method is used when the dimension-reduced data is two-dimensional data, but may also be used when the dimension-reduced data is three or more dimensional data.

[0083] The third extraction method is used when the dimension of the signal information D1 is reduced to a dimension greater than two (referred to as low dimension) by dimensionality reduction. The dimension of the signal information D1 is, for example, about 1000 dimensions. The low dimension is, for example, 10 dimensions. First, a vector whose components are the coordinates of the data in the low-dimensional dimension-reduced data is generated for each data item of the observed object contained in each specimen. For each coordinate component of the vector, a feature vector for the specimen is generated by averaging the observed object contained in the specimen between the vectors.

[0084] In the first extraction method, the measurer must manually specify clusters, but it may be difficult to identify clusters in the dimension-reduced data. Furthermore, because the clusters are specified by the measurer, it may be difficult to maintain objectivity in the cluster specification. The second and third extraction methods are suitable, for example, when it is difficult to identify clusters in the dimension-reduced data or when there is variability in the cluster specification among the measurers. The measurer may select which of the first, second, and third extraction methods to use.

[0085] Step S40: The learning unit 114 trains the classification model L1 by machine learning. The learning unit 114 trains the classification model L1 using pairs of feature vectors and labels extracted from the dimension-reduced data acquired from the training samples as training data. Each piece of signal information D1 included in the training samples is pre-assigned a label indicating the state of the training sample provider associated with the sample from which the signal information D1 was acquired. Accordingly, the feature vectors extracted from the dimension-reduced data generated based on the signal information D1 acquired from multiple observation objects (training observation objects) included in the training samples are assigned the labels assigned to the signal information D1 used to generate the dimension-reduced data.

[0086] The classification model L1 is trained by machine learning so that when a feature vector extracted from dimension-reduced data generated based on signal information D1 acquired from the subject's sample is input, the classification model L1 outputs a label indicating the subject's state. The training unit 114 stores the trained classification model L1 created by training in the storage unit 12. The control unit 11 then ends the training process.

[0087] [Prediction Processing] The prediction processing, which is processing for predicting the state of a subject, will be described. Fig. 9 is a diagram showing an example of the flow of the prediction processing according to this embodiment. The prediction processing is executed by the control unit 11 of the state prediction device 10.

[0088] Step S110: The signal information acquisition unit 110 acquires signal information D1 for each of a plurality of observation objects (subject observation objects) contained in the subject sample from the DAQ device 7. The plurality of subject observation objects are cells C1 contained in the subject sample. The signal information acquisition unit 110 supplies the signal information D1 acquired for each of the plurality of subject observation objects contained in the subject sample to the dimension reduction unit 111. The signal information D1 acquired from the subject observation objects is also referred to as subject signal information. Furthermore, the signal information D1 acquired from the test observation object is also referred to as test signal information. The subject is an example of a test sample provider.

[0089] Step S120: The dimension reduction unit 111 performs dimension reduction on the signal information D1 for each of the plurality of subject observation objects acquired by the signal information acquisition unit 110. The dimension reduction unit 111 generates dimension-reduced data by the dimension reduction.

[0090] Step S130: The state prediction unit 112 extracts a feature vector from the dimension-reduced data. The method for extracting the feature vector from the dimension-reduced data is the same as the method described in step S30 of the learning process. Therefore, the dimension of the feature vector extracted in step S130 of the prediction process is equal to the dimension of the feature vector extracted in step S30 of the learning process.

[0091] Step S140: The state prediction unit 112 predicts the state of the subject based on the classification model L1. The state prediction unit 112 inputs the feature vector extracted in step S130 into the classification model L1 and causes the classification model L1 to output the predicted result. The state prediction unit 112 predicts the result output by the classification model L1 as the state of the subject. Therefore, the state prediction unit 112 predicts the state of the subject based on the classification model L1 from data obtained by performing dimensional reduction on the signal information D1 (subject signal information) for each of the multiple subject observation objects acquired by the signal information acquisition unit 110 using the dimensionality reduction unit 111.

[0092] In this embodiment, the classification model L1 is a trained model that has learned the relationship between the feature vectors extracted by performing dimension reduction processing on training signal information acquired from training observation objects contained in the training sample and the state of the training sample donor attached to the training sample. Cell morphological information contained in the training signal information is extracted using a dimension reduction processing technique and extracted as information that provides an overview of the morphological characteristics of all cells contained in the training sample, and the classification model L1 has learned the relationship between this information and the state of the training sample donor. Therefore, the classification model L1 is also a trained model that has learned the relationship between the training signal information acquired from multiple training observation objects contained in the training sample and the state of the training sample donor. The state prediction unit 112 predicts the state of the subject who provided the subject sample based on the trained model, using the subject signal information acquired by the signal information acquisition unit 110 for each of the multiple subject observation objects contained in the subject sample.

[0093] In the above description, the signal information D1 acquired from the subject observation object contained in the subject sample is referred to as subject signal information, and the signal information D1 acquired from the learning observation object contained in the learning sample acquired from the learning sample provider is referred to as learning signal information. The subject test signal information is test signal information when the test sample provider is the subject, and is one example of test signal information. The signal information D1 includes GMI waveform information.

[0094] Step S150: The output unit 113 outputs to an external device the prediction result obtained by the state prediction unit 112. With this, the control unit 11 ends the prediction process.

[0095] In this embodiment, the case where GMI waveform information is used as the signal information D1 has been described, but this is not limiting. In this embodiment, GMI waveform information acquired using GC technology can also be combined with information on an optical signal acquired from an electromagnetic wave (e.g., scattered signal light) from an object under observation using a conventional optical configuration without using GC technology, and used as the signal information D1. Furthermore, the GMI waveform information can also be converted into a frequency spectrum and used as the signal information D1.

[0096] Furthermore, in the present embodiment, the flow cytometer 1 has been described as irradiating the observation object with illumination light LE1, which is light of a single predetermined wavelength, and acquiring GMI waveform information using a structured illumination configuration or a structured detection configuration. However, this is not limited to this. The flow cytometer 1 can also be configured, for example, to irradiate the observation object with multiple illumination lights of different wavelengths using a structured illumination configuration to acquire multiple GMI waveform information. In this case, the number of types of feature quantities obtained from the GMI waveform information is the number of types of feature quantities obtained when using light of a single wavelength multiplied by the number of wavelengths, thereby enabling the acquisition of more abundant morphological information about the observation object. When light of multiple wavelengths is used as illumination light in the flow cytometer 1, for example, the light source 3 can be configured to include multiple light sources emitting illumination light of different wavelengths.

[0097] (Variation 1) In the above-described embodiment, a case where the current state of a subject is predicted has been described as an example, but this is not limiting. This embodiment also includes, as a variation, a case where the future state of a subject is predicted. Fig. 10 shows an overview of the learning process for creating a classification model according to this variation. This classification model is a classification model that predicts the future state of a subject.

[0098] In this modification, to generate training data, GMI waveform information obtained from samples collected at multiple past periods from a training sample donor who has reached a predetermined state is used. The sample donor may not have the predetermined state at the time the sample was collected. For example, the predetermined state may be a state of having AML. The training sample donor who provided the training sample may or may not have AML at the time the sample was collected, and may or may not have symptoms of AML. However, the donor is ultimately found to have AML at a time greater than a predetermined time after the sample was collected. It is preferable, but not limited to, that the training samples be collected at multiple past periods. The past times at which the training samples are collected can be adjusted depending on the predetermined state and the number of samples.

[0099] Signal information D1 is acquired from each of a plurality of cells, which are the learning observation objects contained in the learning sample, and dimension-reduced data is further generated from the acquired signal information D1. Here, the dimension-reduced data generated for learning is generated from each of the samples acquired at multiple time periods. A label indicating the future state of the donor who acquired the learning sample (e.g., a label indicating AML) is linked to the dimension-reduced data for each of the multiple time periods, and learning data that serves as training data for the future state is created. The training data is generated for each state. A model is trained by machine learning using the created training data, and a trained classification model is created.

[0100] In the example shown in Figure 10, there are three past periods. The multiple periods are, for example, monthly, three months, six months, or yearly. The multiple periods can be set appropriately depending on the condition to be predicted. For example, if the training sample provider is a patient with a specific disease, the multiple periods for obtaining the training sample can be set appropriately depending on appropriate intervals for observing the progression rate of the disease or the transition of the condition.

[0101] FIG. 11 shows an overview of the process for predicting the future state of a subject. In the prediction process, first, specimens that serve as subject specimens are collected at multiple time periods from the subject to be predicted. Next, signal information D1 is obtained from multiple cells that are subject observation objects contained in the subject specimen. Next, dimension-reduced data is generated for each time period when the subject specimen was collected from the signal information D1 obtained from the multiple cells contained in the subject specimen. Finally, the generated dimension-reduced data for the multiple time periods is input into a classification model, and the future state of the subject to be predicted is output.

[0102] The classification model learns time-series changes in the state. Accordingly, the time intervals and the number of time periods at which a subject sample is collected from a subject to be predicted match the time intervals and the number of time periods at which training samples are collected from training sample providers in creating the training data, for example. For example, assume that in creating the training data, training samples are collected at three monthly intervals, such as the time of collection of the initial training sample, one month after the initial collection, and two months after the initial collection. In this case, the time of collection of the subject sample from the subject to be predicted is also set to three monthly intervals, such as the time of collection of the initial subject sample, one month after the initial collection, and two months after the initial collection.

[0103] The time intervals and number of sample collection periods may be different between the training sample and the subject sample. Furthermore, the time intervals and number of sample collection periods may also be different between the training samples. Furthermore, even if there are gaps in the state time series in the classification model, the gaps are complemented and learned based on machine learning. For example, even if there are few sample collection periods for a single training sample, if there are many training samples learning the same state and the samples are collected with a variety of time intervals and periods overall, learning can be complemented and learned.

[0104] Furthermore, the time periods during which samples are collected from a training sample provider do not have to be multiple. In creating the training data, a sample collected at a single time in the past from a training sample provider who is known to have reached the state to be predicted may be used as the training sample. A label indicating the state is assigned to dimension-reduced data generated for the sample, and the training data is created. During prediction, dimension-reduced data is generated for a sample collected as a subject sample at a single time period from the subject to be predicted. The dimension-reduced data is input into a classification model, and the future state of the subject is output.

[0105] Even when only samples collected during a limited period in the past are used as training samples, by increasing the number of training sample donors and the variation in the timing of sample collection across all training samples, it is possible to train a classification model using dimension-reduced data with overall variation. On the other hand, when comparing the number of sample collections at the time of creating the training data and at the time of prediction, it is preferable that samples be collected from subjects to be predicted over a larger number of periods.

[0106] As described above, in this modification, the classification model has learned the relationship between the learning signal information acquired at a first time period and the state of the test specimen provider at a second time period after the first time period. The state prediction unit 112 predicts the state of the test specimen provider at a fourth time period after the third time period based on the classification model, using the testing signal information acquired at a third time period by the signal information acquisition unit 110. Because the fourth time period is a time period after the third time period, in this modification, the future state of the test specimen provider is predicted based on the classification model.

[0107] Second Embodiment A second embodiment of the present invention will be described in detail below with reference to the drawings. In the first embodiment, dimensionality reduction is performed on GMI waveform information acquired from multiple subject observation objects contained in a subject sample, and the subject's state is predicted from the resulting dimension-reduced data based on a trained classification model. In this embodiment, a case will be described in which GMI waveform information acquired from multiple subject observation objects contained in a subject sample is first classified at the observation object level, and the subject's state is then predicted based on the classification results. The state prediction method according to this embodiment does not require a dimensionality reduction process. The flow cytometer according to this embodiment will be referred to as a flow cytometer 1a, and the state prediction device will be referred to as a state prediction device 10a.

[0108] [Prediction of Subject's State Using Observation Object Classification Model] Hereinafter, prediction of a subject's state using the observation object classification model according to this embodiment will be described. The observation object classification model is a classification model that, when GMI waveform information acquired from an observation object is input, predicts and classifies the observation object as to which donor the specimen the observation object was acquired from. The cell level classification model is an observation object classification model when the observation object to be classified is a cell. The same algorithm as used in creating the classification model described above can be used to create the cell level classification model. A neural network can also be used to create the cell level classification model.

[0109] FIG. 12 shows an overview of the observation object classification model according to this embodiment. Hereinafter, an example will be described in which the observation object to be classified is a cell. When GMI waveform information is input, the cell-level classification model uses a trained model that has been trained using GMI waveform information acquired in advance from a training sample containing cells belonging to a specific class as training data. The model classifies the cell from which the GMI waveform was acquired as to which specific class the cell is likely to belong. In the following description, the training sample for training the observation object classification model will also be referred to as the observation object classification model training sample.

[0110] The class may not correspond to the state of the specimen donor to be predicted. In this case, the class does not correspond to the state of the subject on which the classification model used to predict the state of the subject is trained. The class may be, for example, cells identified by cell surface markers. In this case, the class may be a class corresponding to a cell type, such as T cells, B cells, or natural killer (NK) cells. The class may also be a class distinguished by differences in the responsiveness of specific cells to drugs. The class may also be a class distinguished by differences in treatment of cells in which a specific gene has been deleted or modified by gene editing technology. The class may also correspond to the state of the specimen donor to be predicted. For example, if a training sample is obtained from an AML patient, multiple cells contained in the training sample would be included in a class called cells derived from the AML patient.

[0111] First, a cell-level classification model is trained. The cell-level classification model is an example of an observation object classification model. A cell-level classification model training sample is prepared for training the cell-level classification model, and GMI waveform information is individually acquired for multiple cells contained in the training sample. Here, multiple samples are prepared for each of several classes as the cell-level classification model training sample. The GMI waveform information and information regarding the class to which the cell from which the GMI waveform information was acquired belongs are linked and included in the training data for the cell-level classification model. The cell-level classification model is trained based on machine learning using the training data. The cell-level classification model training sample is an example of a sample for training an observation object classification model.

[0112] Next, the classification model according to this embodiment is trained. The classification model is a trained model that predicts the state of a subject based on the classification results of the cell-level classification model. First, a training sample for training the classification model is collected from a training sample donor. The training sample donor is, for example, a patient suffering from AML. Multiple types of training samples to be used for training the classification model are prepared, one for each type of training sample.

[0113] Next, GMI waveform information is obtained from each of the multiple cells contained in the training sample. The GMI waveform information obtained from the multiple cells contained in a specific training sample is each assigned a label indicating the state of the training sample provider (e.g., AML patient). For each training sample, the GMI waveform information obtained from each of the multiple cells contained in the sample is input into a trained cell-level classification model, and the cell-level classification model classifies each cell as belonging to a specific class. Here, the term "class" refers to the classification of cells contained in a sample prepared in advance for training the cell-level classification model, as described above. As a result, classification results are obtained for the multiple cells contained in the training sample for training the classification model, based on the cell-level classification model training sample.

[0114] In a cell-level classification model, cells are classified into classes into which the training samples used to train the classification model are likely to belong. The classification results based on the cell-level classification model can also be output as feature vectors indicating the proportions of multiple cells contained in each sample that belong to each class. In this case, the feature vectors output from the cell-level classification model are used as training data to train the classification model.

[0115] Next, prediction of a subject's condition using the cell-level classification model according to this embodiment will be described. FIG. 13 shows an overview of the process of predicting a subject's condition from the classification results by the cell-level classification model according to this embodiment. Two trained models, a cell-level classification model and a classification model, are used in subject prediction according to this embodiment. The cell-level classification model is a classification model that, when GMI waveform information acquired from a subject sample is input, predicts and classifies the class of sample from which the cell is likely to be included. The classification model is a trained model that predicts the subject's condition based on the classification results by the cell-level classification model.

[0116] First, a test specimen is collected from a subject. Next, GMI waveform information is obtained from each of a plurality of cells contained in the test specimen collected from the subject. Next, the GMI waveform information obtained from the plurality of cells contained in the subject specimen is input into a trained cell-level classification model, and the cell-level classification model is caused to classify the plurality of cells contained in the subject specimen as belonging to a particular class, thereby obtaining a classification result based on the subject specimen. The classification result based on the subject specimen can also be output as a feature vector based on the class of the plurality of cells contained in the subject specimen.

[0117] Finally, the classification results from the cell-level classification model are input into a classification model to predict the state of the subject. For example, the proportions of cells contained in a subject's sample that indicate which class each cell belongs to can be calculated, and the proportions can be further input into a classification model as a feature vector to predict the state of the subject. In this case, the classification model is trained by supervised machine learning so that when the feature vector indicating the proportions is input, the classification model outputs the state of the subject.

[0118] [Configuration of State Prediction Device 10a] Fig. 14 is a diagram showing an example of the configuration of the state prediction device 10a according to this embodiment. The state prediction device 10a includes a signal information acquisition unit 110, an observed object classification unit 116a, a state prediction unit 112a, an output unit 113, a learning unit 114a, and an operation unit 115. Comparing the state prediction device 10a according to this embodiment (Fig. 14) with the state prediction device 10 according to the first embodiment (Fig. 9), the latter differs in that the former includes an observed object classification unit 116a, a state prediction unit 112a, and a learning unit 114a, and in that an observed object classification model L2a and a classification model L3a are stored in the storage unit 12. Here, the functions of the other components are the same as those of the first embodiment.

[0119] The observation object classification unit 116a classifies each of a plurality of observation objects from the signal information D1 based on the observation object classification model L2a. When the observation object is a cell, the observation object classification model is the cell level classification model described above. That is, the cell level classification model is an example of an observation object classification model.

[0120] The state prediction unit 112a predicts the state of the subject using a classification model L3a based on the results of classification of the multiple subject observation objects by the observation object classification unit 116a from the signal information D1 for each of the multiple observation objects (subject observation objects) contained in the subject sample acquired by the signal information acquisition unit 110.

[0121] The learning unit 114a performs machine learning to learn the observation object classification model L2a and the classification model L3a. In the observation object classification model L2a, the relationship between the observation object contained in a learning sample prepared for learning the observation object classification model (observation object for learning the observation object classification model) and information regarding the class assigned to the learning sample is learned. When the observation object is a cell, the observation object classification model L2a is the cell-level classification model described above. In the classification model L3a, the relationship between the results obtained by classifying each of the multiple learning observation objects contained in a learning sample obtained from a learning sample provider based on the observation object classification model L2a and the state of the learning sample provider is learned. The observation object classification model training sample used for learning the observation object classification model L2a may be different from the training sample used for learning the classification model L3a, but they can also be the same.

[0122] The learning unit 114a may be provided in a learning device that is a device separate from the state prediction device 10a. In that case, the state prediction device 10a acquires the observed object classification model L2a and the classification model L3a created by the learning device. The state prediction device 10a may also acquire either the observed object classification model L2a or the classification model L3a from a learning device separate from the state prediction device 10a.

[0123] [Learning Process] The following describes the learning process of the observed object classification model, which is a process for learning the observed object classification model L2a, and the learning process, which is a process for learning the classification model L3a. Fig. 15 is a diagram showing an example of the flow of the learning process of the observed object classification model according to this embodiment. The learning process of the observed object classification model is executed by the control unit 11a of the state prediction device 10a at a time before the learning process of the classification model.

[0124] Step S210: The signal information acquisition unit 110 acquires signal information D1 for each of a plurality of observation objects for learning an observation object classification model from the DAQ device 7. The plurality of observation objects are observation objects included in a specimen for learning an observation object classification model.

[0125] Step S220: The learning unit 114a learns the observation object classification model L2a. The learning unit 114a learns the observation object classification model L2a using, as learning data, a set of signal information D1 of an observation object for learning the observation object classification model as a unit and information on a class indicating the origin of the observation object. The signal information D1 of the observation object for learning the observation object classification model is pre-assigned information on the class indicating the origin of the observation object from which the signal information D1 was obtained. The learning unit 114a stores the observation object classification model L2a created as a result of learning in the storage unit 12. Here, the class refers to the classification of cells contained in the learning specimen of the cell-level classification model, as described above. With this, the control unit 11a ends the learning process of the observation object classification model.

[0126] 16 is a diagram showing an example of the flow of the classification model learning process according to this embodiment. The classification model learning process is executed by the control unit 11a of the state prediction device 10a prior to the prediction process.

[0127] Step S310: The signal information acquisition unit 110 acquires signal information D1 for each of a plurality of training observation objects from the DAQ device 7. The plurality of observation objects are observation objects included in a training sample. A training sample is a sample containing training observation objects for training a classification model that predicts the state of a subject. A training sample is, for example, a sample acquired from a training sample provider suffering from AML. A training sample is a sample containing training observation objects for training a classification model L3a.

[0128] Step S320: The observation object classification unit 116a classifies each of the plurality of observation objects included in the learning specimen from the signal information D1 for each learning observation object based on the observation object classification model L2a.

[0129] Step S330: The learning unit 114a acquires the classification result of the observation object by the observation object classification model L2a.

[0130] Step S340: The learning unit 114a trains the classification model L3a. The learning unit 114a trains the classification model L3a using, as training data, a pair of the classification result from the observation object classification model L2a and a label indicating the state of the training specimen donor. The classification result from the observation object classification model L2a is, for example, a proportion indicating which of the observation object classification model training specimens each of a plurality of cells contained in the training specimen is predicted to originate from. The classification result from the observation object classification model L2a is assigned in advance a label indicating the state of the training specimen donor from which the signal information D1 was acquired. The label indicating the state of the training specimen donor is, for example, a label indicating that the training specimen donor is suffering from AML.

[0131] When the classification result by the observed object classification model L2a is input, the classification model L3a is trained based on machine learning so as to output a label indicating the state of the provider who acquired the signal information D1. The learning unit 114 stores the classification model L3a created as a result of the training in the storage unit 12. With this, the control unit 11a ends the training process.

[0132] In addition, if the observation object classification model training specimen used to train the observation object classification model L2a and the training specimen used to train the classification model L3a are the same, the processing of steps S310 and S320 may be omitted.

[0133] 17 is a diagram showing an example of the flow of the prediction process according to this embodiment. The prediction process is executed by the control unit 11a of the state prediction device 10a.

[0134] Step S410: The signal information acquiring unit 110 acquires signal information D1 for each of a plurality of observation objects contained in the subject's sample from the DAQ device 7. The plurality of observation objects are cells C1 contained in the subject's sample collected from the subject. The signal information acquiring unit 110 supplies the signal information D1 acquired from each of the plurality of observation objects contained in the subject's sample to the observation object classification unit 116a.

[0135] Step S420: The observation object classifying unit 116a classifies each of the plurality of observation objects from the signal information D1 based on the observation object classification model L2a. The observation object classifying unit 116a supplies the classification result to the state predicting unit 112a.

[0136] Step S430: The state prediction unit 112a predicts the state of the subject from whom the subject sample was collected based on the classification model L3a from the results of classification of the multiple observation objects contained in the subject sample by the observation object classification unit 116a.

[0137] The state prediction unit 112a predicts the state of the subject based on the classification model L3a (an example of a trained model) from the classification results of the multiple observation objects included in the subject's sample, which are obtained by the observation object classification unit 116a, based on the test signal information for each of the multiple test observation objects acquired by the signal information acquisition unit 110. Here, the classification model L3a has learned the relationship between the classification results of the multiple observation objects included in the training sample based on the observation object classification model L2a (an example of a classification model) and the state of the training sample provider.

[0138] Step S440: The output unit 113 outputs the prediction result by the state prediction unit 112a to an external device. Then, the control unit 11a ends the prediction process.

[0139] In this embodiment, an example of classification based on a cell-level classification model has been described, but the present invention is not limited to this. Instead of using cells contained in a specimen as units, classification may be performed using subsets of cells contained in the specimen as units. In this case, for example, a Multiple-Instance Learning (MIL) method can be used.

[0140] In the above embodiment, an example in which the classification results from the cell-level classification model are input to the classification model as feature vectors has been described, but this is not limiting. A confusion matrix generated based on the classification results from the cell-level classification model may be used as input to the classification model. The confusion matrix is ​​a matrix whose elements are the number (or frequency) of combinations of predicted classes from the cell-level classification model and actual correct classes.

[0141] (Third Embodiment) A third embodiment of the present invention will be described below with reference to the drawings. In this embodiment, a case will be described in which GMI waveform information acquired from multiple subject observation objects contained in a subject sample is first clustered using a clustering algorithm, and the subject's state is then predicted based on the clustering results. Unlike the first or second embodiment, this embodiment does not require the above-described dimensionality reduction process or classification using a cell-level classification model for the GMI waveform information acquired from multiple subject observation objects contained in a subject sample. A flow cytometer according to this embodiment will be referred to as flow cytometer 1b, and a state prediction device will be referred to as state prediction device 10b.

[0142] FIG. 18 shows an overview of the classification model and prediction process according to this embodiment. First, GMI waveform information is acquired for each of multiple cells contained in a training sample using a flow cytometer 1b. The GMI waveform information of multiple cells contained in the training sample is classified into several clusters (groups) using a predetermined clustering algorithm. A clustering algorithm is an algorithm for dividing data into several groups (clusters). While a method such as the K-means method can be used as a clustering algorithm, clustering can also be performed using a neural network. As a result, multiple cells contained in the sample are classified into several clusters (groups) based on the morphological information contained in the GMI waveform information of each cell. Based on the classification results, a feature vector whose components are the proportion of cells contained in a specific cluster is calculated. For each training sample of the same type, feature vector data is accumulated, and training data for the classification model is created, linked to a label indicating the condition of the sample (e.g., AML donor). Finally, machine learning is performed using the training data created for each condition of the training sample donor as training data, and a trained classification model is created.

[0143] Next, an overview of the process for predicting the state of a subject from the calculation of feature vectors by cluster classification according to this embodiment will be described. First, GMI waveform information is obtained from each of multiple cells contained in a specimen of the subject (subject specimen). Next, the GMI waveform information of multiple cells contained in the subject specimen is cluster-classified using a predetermined clustering algorithm. Finally, the calculated feature vector data for the subject specimen is input into a trained classification model, and the classification model is allowed to predict the state of the subject.

[0144] [Configuration of State Prediction Device 10b] Fig. 19 is a diagram showing an example of the configuration of the state prediction device 10b according to this embodiment. The state prediction device 10b includes a signal information acquisition unit 110, a clustering unit 117b, a state prediction unit 112b, an output unit 113, a learning unit 114b, and an operation unit 115. Comparing the state prediction device 10b according to this embodiment (Fig. 19) with the state prediction device 10 according to the first embodiment (Fig. 9), the state prediction device 10b according to this embodiment differs in that it includes a clustering unit 117b, a state prediction unit 112b, and a learning unit 114b, and in that a classification model L4b is stored in a memory unit 12. Here, the functions of the other components are the same as those of the first embodiment.

[0145] The clustering unit 117b classifies a plurality of observation objects from the signal information D1 into clusters (groups) based on a clustering algorithm.

[0146] The state prediction unit 112b predicts the state of the subject using the classification model L4b based on the signal information D1 for each of the multiple observation objects (subject observation objects) contained in the subject sample acquired by the signal information acquisition unit 110, which is obtained by classifying the multiple subject observation objects into clusters by the clustering unit 117b.

[0147] The learning unit 114b performs machine learning to train the classification model L4b. The classification model L4b learns the relationship between the feature vectors calculated from the results of classifying, based on a clustering algorithm, each of the multiple training observation objects contained in the training sample obtained from the training sample provider, and the state of the training sample provider.

[0148] The learning unit 114b may be provided in a learning device that is separate from the state prediction device 10b. In this case, the state prediction device 10b acquires the classification model L4b created by the learning device.

[0149] [Learning Process] The learning process, which is a process for learning the classification model L4b, will be described. Fig. 20 is a diagram showing an example of the flow of the learning process according to this embodiment. The learning process is executed prior to the prediction process performed by the control unit 11b of the state prediction device 10b.

[0150] Step S510: The signal information acquisition unit 110 acquires signal information D1 for each of the multiple observation objects from the DAQ device 7. The multiple observation objects are included in a training sample collected from a training sample provider. The signal information acquisition unit 110 supplies the acquired signal information D1 for each of the multiple observation objects to the clustering unit 117b.

[0151] Step S520: The clustering unit 117b classifies the plurality of observation objects from the signal information D1 into clusters (groups) based on a clustering algorithm.

[0152] Step S530: The learning unit 114b extracts a feature vector from the clustering results obtained by the clustering unit 117b. The feature vector has components that represent the proportion of cells contained in a specific cluster. To extract the feature vector from the clustering results, for example, any of the first extraction method, second extraction method, or third extraction method described in the first embodiment above can be used. Methods other than these may also be used to extract the feature vector from the clustering results.

[0153] Step S540: The learning unit 114b trains the classification model L4b by machine learning. The learning unit 114 trains the classification model L4b using, as training data, pairs of feature vectors extracted from the clustering results obtained from the training samples and labels indicating the status of the training sample providers.

[0154] The classification model L4b is trained by machine learning so that when a feature vector extracted from the clustering result based on the signal information D1 acquired from the subject's sample is input, the classification model L4b outputs a label indicating the subject's condition. The training unit 114b stores the trained classification model L4b created by training in the storage unit 12. The control unit 11b then ends the training process.

[0155] [Prediction Process] Fig. 21 is a diagram showing an example of the flow of the prediction process according to this embodiment. The prediction process is executed by the control unit 11b of the state prediction device 10b.

[0156] Step S610: The signal information acquiring unit 110 acquires signal information D1 for each of a plurality of observation objects (subject observation objects) contained in the subject sample from the DAQ device 7. The plurality of subject observation objects are cells C1 contained in the subject sample. The signal information acquiring unit 110 supplies the acquired signal information D1 for each of the plurality of subject observation objects contained in the subject sample to the clustering unit 117b.

[0157] Step S620: The clustering unit 117b classifies the plurality of subject observation objects from the signal information D1 into clusters (groups) based on a clustering algorithm.

[0158] Step S630: The state prediction unit 112b extracts feature vectors from the clustering results. As in the learning process, the feature vectors can be extracted from the clustering results using any one of the first, second, or third extraction methods described in the first embodiment, but other methods may also be used.

[0159] Step S640: The state prediction unit 112b predicts the state of the subject based on the classification model L4b. The state prediction unit 112b inputs the feature vector extracted in step S630 to the classification model L4b and causes the classification model L4b to output the predicted result. The state prediction unit 112b predicts the result output by the classification model L4b as the state of the subject.

[0160] Step S650: The output unit 113 outputs the prediction result by the state prediction unit 112b to an external device. Then, the control unit 11b ends the prediction process.

[0161] As described above, the state prediction device according to each embodiment (state prediction device 10, state prediction device 10a, or state prediction device 10b) is a state prediction device that predicts the state of a subject from a plurality of unlabeled subject observation objects (cells C1 in this embodiment) contained in a sample acquired from the subject (a subject sample in this embodiment) in a flow cytometry method, and includes a signal information acquisition unit 110 and a state prediction unit (state prediction unit 112, state prediction unit 112a, or state prediction unit 112b). The signal information acquisition unit 110 acquires, as subject signal information for each of a plurality of subject observation objects (cells C1 in this embodiment), signal information (signal information D1 in this embodiment) indicating the time change in intensity of electromagnetic waves generated when modulated electromagnetic waves from the subject observation object present in the light irradiation area irradiated with the illumination light LE1 from the light source 3 are received by the photodetector 6, using one or more of a structured illumination configuration in which illumination light LE1 from the light source 3 is converted into structured illumination light SLE1 and irradiated onto the subject observation object, or a structured detection configuration in which modulated electromagnetic waves from the subject observation object are structured and detected. The state prediction unit (state prediction unit 112, state prediction unit 112a, or state prediction unit 112b) predicts the state of the subject from the subject signal information (in this embodiment, signal information D1 acquired from the subject sample) acquired by the signal information acquisition unit 110 for each of the multiple subject observation objects (in this embodiment, cells C1) based on a trained model (in each embodiment, classification model L1, observation object classification model L2a and classification model L3a, or classification model L4b) that has learned the relationship between the training signal information (in this embodiment, signal information D1 acquired from the training sample), which is signal information for each of the multiple training observation objects for training, and the state of the training sample provider.

[0162] With this configuration, each condition prediction device (condition prediction device 10, condition prediction device 10a, or condition prediction device 10b) according to each embodiment can obtain morphological information in an unlabeled state from a subject observation object contained in bodily fluid obtained from the subject, and predict disease or the risk of morbidity, thereby enabling disease or the risk of morbidity to be predicted quickly and accurately using a simple configuration and a non-invasive method.

[0163] Here, the state prediction device according to each embodiment (state prediction device 10, state prediction device 10a, or state prediction device 10b) can acquire morphological information for each observation object in an unlabeled (label-free) state from the observation object, and therefore has a simpler configuration than conventional invasive testing. Furthermore, the signal information acquired by the structured illumination configuration or the structured detection configuration contains compressed morphological information of the observation object. This morphological information has such high resolution that an image of the observation object can be reconstructed. Therefore, the state prediction device 10 according to this embodiment can acquire morphological information of the observation object with higher accuracy than flow cytometry.

[0164] In each state prediction device (state prediction device 10, state prediction device 10a, or state prediction device 10b) according to each embodiment, morphological information, which would conventionally be obtained by observation under a microscope, is obtained as signal information (GMI waveform information) from a label-free observation object and used to predict disease or morbidity risk. Therefore, even if the state prediction device (state prediction device 10, state prediction device 10a, or state prediction device 10b) according to each embodiment requires labels (markers) when acquiring training data used to train a learning model, once a trained model is created, it can also be applied to predicting disease or morbidity risk for which markers are unknown.

[0165] In the above-described embodiments, an example in which the object of observation is label-free has been described, but the present invention is not limited to this. The object of observation does not have to be label-free, but if the object of observation is label-free, the condition prediction device according to each embodiment (condition prediction device 10, condition prediction device 10a, or condition prediction device 10b) is preferably able to predict the risk of developing a molecular disease for which the marker is unknown.

[0166] Although the embodiments of the present invention have been described above in detail with reference to the drawings, the specific configuration is not limited to this embodiment, and design modifications and the like are also included within the scope of the present invention. For example, a computer program for realizing the functions of each of the above-described devices may be recorded on a computer-readable recording medium, and the program recorded on this recording medium may be read and executed by a computer system. Note that the "computer system" referred to here may also include hardware such as an OS and peripheral devices.

[0167] Additionally, "computer-readable recording media" refers to writable non-volatile memory such as flexible disks, optical magnetic disks, ROMs, and flash memories, portable media such as DVDs (Digital Versatile Discs), and storage devices such as hard disks built into computer systems. Furthermore, "computer-readable recording media" also includes devices that retain a program for a certain period of time, such as volatile memory (e.g., DRAM (Dynamic Random Access Memory)) within a computer system that serves as a server or client when a program is transmitted via a network such as the Internet or a communication line such as a telephone line.

[0168] The program may also be transmitted from a computer system storing the program in a storage device or the like to another computer system via a transmission medium or by transmission waves in the transmission medium. Here, the "transmission medium" that transmits the program refers to a medium that has the function of transmitting information, such as a network (communication network) such as the Internet or a communication line (communication line) such as a telephone line. The program may also be a program that realizes part of the above-mentioned functions. Furthermore, the program may be a so-called differential file (differential program) that can realize the above-mentioned functions in combination with a program already recorded in the computer system.

[0169] 1...flow cytometer, 10...state prediction device, 110...signal information acquisition unit, 112...state prediction unit, 3...light source, 6...photodetector, LE1...illumination light, SLE1...structured illumination light, D1...signal information, C1...cell

Claims

1. A state prediction device for use in a flow cytometry method, which predicts the state of a test specimen provider from a plurality of test observation objects contained in a test specimen obtained from the specimen provider, the state being an observation object contained in the test specimen obtained from the test specimen provider, comprising: a signal information acquisition unit which acquires, for each of the plurality of observation objects, signal information indicating time changes in intensity of the electromagnetic waves generated when modulated electromagnetic waves from the observation objects present in a light irradiation area irradiated with illumination light from a light source are received by a photodetector using one or more of a structured illumination configuration in which illumination light from a light source is converted into structured illumination light and irradiated onto the observation objects, or a structured detection configuration in which modulated electromagnetic waves from the observation objects are structured and detected; and a state prediction unit which predicts the state of the test specimen provider from the test signal information including the signal information acquired by the signal information acquisition unit for each of the plurality of test observation objects, based on a classification model which is a trained model in which the relationship between training signal information including the signal information for each of the plurality of training observation objects, which are the observation objects contained in the training specimen obtained from the training specimen provider, and the state of the training specimen provider, and A state prediction device comprising:

2. The state prediction device according to claim 1, wherein the classification model learns a relationship between the state of the training specimen provider and dimension-reduced data obtained by performing dimension reduction on the training signal information including the signal information for each of the plurality of training observation objects included in the training specimen, and the device further comprises a dimension reduction unit that performs dimension reduction on the testing signal information including the signal information for each of the plurality of testing observation objects acquired by the signal information acquisition unit, and the state prediction unit predicts the state of the testing specimen provider based on the classification model from dimension-reduced data generated by performing dimension reduction on the testing signal information including the signal information for each of the plurality of testing observation objects acquired by the signal information acquisition unit.

3. The state prediction device according to claim 1, wherein the classification model has learned a relationship between the state of a plurality of training specimen providers and a result of classification of the training signal information, including the signal information for each of the plurality of training observation objects, based on an observation object classification model that predicts classification for each individual cell, and further comprises an observation object classifying unit that classifies each of the plurality of test observation objects based on the observation object classification model from the test signal information, including the signal information for each of the plurality of test observation objects, and the state prediction unit predicts the state of the test specimen provider based on the trained model from the result of classification of the plurality of test observation objects by the observation object classifying unit from the test signal information, including the signal information for each of the plurality of test observation objects, acquired by the signal information acquisition unit.

4. The state prediction device according to claim 1, wherein the classification model has learned a relationship between training signal information including the signal information acquired from a sample obtained from the training sample provider at a predetermined first time period and the state of the training sample provider at a second time period after the first time period, and the state prediction unit determines the state of the test sample provider at a fourth time period after the third time period based on the trained model from the test signal information including the signal information acquired by the signal information acquisition unit from a sample obtained from the test sample provider at a third time period.

5. The condition prediction device according to claim 1, wherein the condition includes one or more of the type of disease, the severity of the disease, differences in responsiveness or resistance to drugs, and differences in productivity of the target substance.

6. The state prediction device according to claim 1, wherein the object of observation includes one or more of a cell and a microorganism.

7. The state prediction device according to claim 1, wherein the test specimen provider is a subject.

8. The state prediction device according to claim 1, wherein the test specimen provider is a unit of manufacture, preparation or storage of the object to be observed.

9. The state prediction device according to claim 1, wherein the object to be observed is unlabeled.

10. A classification model creation device for use in a flow cytometry method, which creates a classification model, which is a trained model that predicts the state of a test specimen provider, from a plurality of test observation objects contained in a test specimen obtained from the test specimen provider, the classification model creation device comprising: a signal information acquisition unit that acquires, for each of the plurality of observation objects, signal information indicating time changes in intensity of electromagnetic waves generated when modulated electromagnetic waves from the observation objects present in a light irradiation area irradiated with illumination light from a light source are received by a photodetector using one or more of a structured illumination configuration in which illumination light from a light source is converted into structured illumination light and irradiated onto the observation objects, or a structured detection configuration in which modulated electromagnetic waves from the observation objects are structured and detected; and a learning unit that creates the classification model based on machine learning from the relationship between training signal information, which includes the signal information for each of the plurality of training observation objects that are the observation objects contained in the training specimen obtained from the training specimen provider, and the state of the training specimen provider.

11. A flow cytometry method for predicting the state of a test specimen provider from a plurality of test observation objects contained in a test specimen obtained from the test specimen provider, the method comprising: a signal information acquisition step for acquiring, for each of a plurality of observation objects, signal information indicating a time change in intensity of the electromagnetic waves generated when modulated electromagnetic waves from the observation objects present in a light irradiation area irradiated with illumination light from a light source are received by a photodetector using one or more of a structured illumination configuration in which illumination light from a light source is converted into structured illumination light and irradiated onto the observation objects, or a structured detection configuration in which modulated electromagnetic waves from the observation objects are structured and detected; a state prediction step of predicting the state of the test specimen provider from test signal information including the signal information acquired in the signal information acquisition step for each of the plurality of test observation objects, based on a classification model which is a trained model that has learned the relationship between training signal information including the signal information for each of a plurality of training observation objects, which are the observation objects contained in the training specimen acquired from the training specimen provider, and the state of the training specimen provider.

12. In a flow cytometry method, a computer of a state prediction device that predicts the state of a test specimen provider from a plurality of test specimens contained in a specimen obtained from the specimen provider, the test specimen being an object of observation contained in the specimen obtained from the specimen provider, includes: a signal information acquisition step of acquiring, for each of a plurality of objects of observation, signal information indicating time variations in the intensity of the electromagnetic waves generated when modulated electromagnetic waves from the objects of observation present in a light irradiation area irradiated with illumination light from a light source are received by a photodetector using one or more of a structured illumination configuration that converts illumination light from a light source into structured illumination light and irradiates the objects of observation, or a structured detection configuration that structures and detects modulated electromagnetic waves from the objects of observation; a state prediction step of predicting the state of the test specimen provider from test signal information including the signal information acquired in the signal information acquisition step for each of the plurality of test observation objects, based on a classification model which is a trained model that has learned the relationship between training signal information including the signal information for each of a plurality of training observation objects, which are the observation objects contained in the training specimen acquired from the training specimen provider, and the state of the training specimen provider.

13. A classification model creation method for creating a classification model, which is a trained model that predicts the state of a test specimen provider, from a plurality of test observation objects that are observation objects contained in a specimen obtained from the specimen provider and that are observation objects contained in a test specimen provided from the test specimen provider in a flow cytometry method, the classification model creation method comprising: a signal information acquisition step of acquiring, for each of the plurality of observation objects, signal information indicating time changes in intensity of electromagnetic waves that are generated when modulated electromagnetic waves from the observation objects present in a light irradiation area irradiated with illumination light from a light source are received by a photodetector using one or more of a structured illumination configuration that converts illumination light from a light source into structured illumination light and irradiates the observation objects, or a structured detection configuration that structures and detects modulated electromagnetic waves from the observation objects; and a learning step of creating the classification model based on machine learning from the relationship between training signal information, which includes the signal information for each of the plurality of training observation objects that are observation objects contained in a training specimen obtained from the training specimen provider, and the state of the training specimen provider.

14. A program for causing a computer of a classification model creation device that creates a classification model, which is a trained model that predicts the state of a test specimen provider, from a plurality of test observation objects that are observation objects contained in a specimen obtained from a test specimen provider in a flow cytometry method, to execute the following steps: a signal information acquisition step of acquiring, for each of a plurality of observation objects, signal information indicating time changes in the intensity of electromagnetic waves that are generated when modulated electromagnetic waves from the observation objects present in a light irradiation area irradiated with illumination light from a light source are received by a photodetector using one or more of a structured illumination configuration that converts illumination light from a light source into structured illumination light and irradiates the observation objects, or a structured detection configuration that structures and detects modulated electromagnetic waves from the observation objects; and a learning step of creating the classification model based on machine learning from the relationship between training signal information, which includes the signal information for each of a plurality of training observation objects that are observation objects contained in a training specimen obtained from the training specimen provider, and the state of the training specimen provider.

Images