Method, apparatus, and program for creating a machine learning model to predict the presence or absence of disease from gut microbiota data.

Abstract

Creating a highly accurate machine learning model to predict the presence or absence of disease from gut microbiota data. [Solution] A preferred embodiment includes creating multiple composition data sets with the same sample ID by sampling multiple times for each intestinal microbiota data set, associating a probability distribution with the count value, and creating a machine learning model from the set of composition data using the Nested Cross Validation method. However, in both the Inner fold and Outer fold of Nested Cross Validation, the composition data are used individually for training during the model training stage, and the accuracy of the model is determined by averaging the multiple composition data sets with the same sample ID.

Description

Technical Field

[0001] The present invention relates to a method for creating a machine learning model for predicting the presence or absence of a disease from gut microbiota data. The present invention also relates to an apparatus and a computer program for creating a machine learning model for predicting the presence or absence of a disease from gut microbiota data.

Background Art

[0002] The relationship between gut microbiota and the diseases of the subjects carrying it has been reported for many diseases such as not only digestive diseases but also allergies, autoimmune diseases, lifestyle-related diseases such as obesity and diabetes, cancer, and neurological and mental diseases. Therefore, attempts have been made to develop detection and evaluation methods for diseases using gut bacteria as an index. The inventors of the present application have set a gut microbiota factor (latent variable) composed of a plurality of gut bacteria (observed variables) having the same physiological action using the method of structural equation modeling, and analyzed the relationship between this latent variable (a latent variable representing a specific physiological action) and a disease, thereby developing a method for calculating a disease risk as a probability value.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Patent Document 2

Non-Patent Documents

[0004]

Non-Patent Document 1

[0005] However, while existing estimation models have received some positive evaluations, there is a need to further improve the accuracy of disease risk estimation. Furthermore, current estimation models require significant manual effort to create, which limits the number of diseases analyzed. [Means for solving the problem]

[0006] To address at least some of the above issues, the present invention discloses a method for creating a machine learning model to predict the presence or absence of disease from a set of gut microbiota data. However, the gut microbiota data included in the set are all data created from samples of the same gender of the subjects, and each sample includes a sample ID, which is information that identifies a specific sample, and the subject of the sample. Specific diseases The method comprises at least health information that can indicate whether or not the person has the condition, and microbiome data obtained from the sample, wherein the microbiome data consists of a combination of specific information related to the identification of the bacteria and a count value related to the specific information. For each of the aforementioned intestinal microbiota data, a probability distribution is associated with the count value, and the probability distribution is sampled multiple times to create multiple composition data with the same sample ID; Using a cross-validation method, the machine learning model is created from the set of composition data; Including the use of the Cross Validation method, The set of composition data is divided into multiple folds, provided that the composition data having the same sample ID is divided so that it belongs to the same fold; In the Cross Validation described above, during the model training stage, the composition data is used individually for training. During the stage of determining the model's accuracy, multiple sets of composition data with the same sample ID are averaged to create average composition data, and the accuracy is determined using this average composition data. thing and ; Includes.

[0007] Current estimation models sample a probability distribution associated with count values ​​multiple times and use only the average of each sample value for training. However, since count values ​​are determined probabilistically, using only the average of the sample values ​​results in a loss of variance information, which can lead to a decrease in accuracy. On the other hand, according to the method of this application, each sample value is used individually for model training, so the essential fact that count values ​​are determined probabilistically can be reflected in the model training, thereby improving the accuracy of the created model. In this case, in cross-validation, the occurrence of leaks can be prevented by dividing composition data with the same sample ID so that they belong to the same fold. Furthermore, in the stage of determining the accuracy of the model, the above method uses the average of these sample values. This makes it possible to obtain accuracy that is appropriate for actual operation. That is, it is possible to obtain accuracy appropriate for determining the presence or absence of disease in unknown gut microbiota data using the created model.

[0008] In some embodiments, the specified information is a genus name, but in other embodiments, the specified information may be an ASV (Amplicon Sequence Variant, a taxonomic unit defined based on DNA sequences and independent of bacterial classification). Even when the specified information is a genus name, that genus name is identified by comparing the ASV with a bacterial genus database. However, the definition of bacterial classification may change as research progresses, and the contents of the bacterial genus database may change. For this reason, creating a model based on a genus name may cause problems when examining changes over time in the same individual.

[0009] In contrast, ASV is defined based on DNA sequences and is therefore independent of bacterial classification. For this reason, creating a model based on ASV avoids the problems that arise when creating a model based on bacterial genus names.

[0010] In this embodiment, the taxonomic genus name is used, but it is not limited to this. The name of the fungus may also be the species name or the family name. The identifying information may include fungal names at different hierarchical levels (for example, genus name and species name).

[0011] Depending on the embodiment, creating the machine learning model from the set of composition data may include creating it using a decision tree-based model generation algorithm.

[0012] Depending on the embodiment, each of the gut microbiota data included in the set may be associated with an enterotype. Alternatively, if the gut microbiota data is not associated with an enterotype, the method may include estimating an enterotype for each of the gut microbiota data based on comparison with reference data. Dividing into the plurality of folds is Based on the aforementioned health information, the set of compositional data will be divided into two groups; For each of the two aforementioned groups, the composition data belonging to that group is divided into multiple subgroups so that the distribution of enterotypes is even; • For one of the two groups, a fold number shall be assigned to the subgroup in order of the number of composition data included in that subgroup; • For the other of the two groups mentioned above, assign a fold number to the subgroup in reverse order of the number of composition data included in that subgroup; • Merging subgroups that have the same fold number; It may include.

[0013] By dividing the set of compositional data into multiple folds using the method described above, the ratio of healthy individuals to diseased individuals is balanced across different folds, as is the distribution of enterotypes. The inventors of this application have found that using folds in which both the distribution of health status and enterotypes are balanced improves the accuracy of the model created.

[0014] Depending on the embodiment, the Cross Validation method is a Nested Cross Validation method.

[0015] Depending on the embodiment, the method described above may be (c1) For each inner fold of the Nested Cross Validation, train the model using all of the composition data and calculate the feature importance for each explanatory variable; (c2) When the number of explanatory variables is n, the model is trained using the n explanatory variables for each of the Inner folds, in order of the feature importance of that Inner fold, and the accuracy of the obtained model is determined; (c3) From all the Inner folds belonging to a specific Outer fold, calculate the average value of the precision obtained in c2, when the explanatory variable is n; (c4) For the specific Outer fold, perform the above c2 and c3 while changing n, and store the n for which the average value obtained in c3 is the best as the number of explanatory variables for the Outer fold; It may include this. Here, the best criterion for the average value is determined by the accuracy index to be selected. If an accuracy index where a larger value is better is selected, the case where the average value is the largest is considered the best. Examples of accuracy indices where a larger value is better include, for example, the correct answer rate. Conversely, if an accuracy index where a smaller value is better is selected, the case where the average value is the smallest is considered the best. Examples of accuracy indices where a smaller value is better include, for example, the mean squared error.

[0016] Depending on the embodiment, the method may further, for each of the Outer folds, (d1) Use all of the above composition data to perform model learning and obtain the feature importance for each explanatory variable; (d2) Using the explanatory variables of the number obtained in c4 in descending order of the feature importance obtained in d1, perform model learning again; It may include this.

[0017] That is, in the above embodiment, in the inner Cross Validation of Nested Cross Validation, the number of (explanatory variables) that gives the best accuracy is obtained by the above steps, and in the outer Cross Validation, a model is created using the explanatory variables of that number in descending order of feature importance. The inventor of the present application has found that when creating a model in this way, the prediction accuracy is improved compared to creating a model using all explanatory variables.

[0018] Depending on the embodiment, the gut microbiota data may further have the age information of the subject of the sample, and the age information may be included in the learning target.

[0019] Each step of the method described above is performed by a computer or software application. Therefore, no human intervention is required to create the model, and the number of diseases to be analyzed can be easily increased.

[0020] The invention disclosed herein includes a computer program, which, when executed by a processing means of an apparatus, is configured to cause the apparatus to perform the method disclosed herein.

[0021] The invention disclosed herein includes an apparatus comprising processing means and storage means, wherein the storage means stores a computer program, and the computer program, when executed by the processing means, is configured to cause the apparatus to perform the method disclosed herein.

[0022] The claims include some of the embodiments of the present invention that are currently considered preferable. However, the configurations specified in these claims do not necessarily encompass all of the novel technical concepts disclosed in the specification and drawings. The applicant asserts that it has the right to obtain a patent for all of the novel technical concepts disclosed in the specification and drawings, regardless of whether they are described in the current claims. [Brief explanation of the drawing]

[0023] [Figure 1] This is a block diagram illustrating an example of an embodiment of the present invention. [Figure 2] This is a flowchart illustrating the first embodiment of the present invention. [Figure 3] Figures 3A to 3C show an example of gut microbiota data. [Figure 4] Figures 4A to 4C show an example of CLR data created by converting the gut microbiota data from Figure 3. [Figure 5] This is a flowchart illustrating a second embodiment of the present invention. [Figure 6A]This is a flowchart illustrating the process performed in step 542 in the second embodiment. [Figure 6B] Figure 4 shows a modified example of the CLR data exemplified above. [Figure 7] This shows an example of splitting a CLR dataset into folds. [Figure 8] This is a flowchart illustrating a third embodiment of the present invention. [Figure 9] This shows an example of the definitions of Outer fold and Inner fold in Nested Cross Validation. [Figure 10] This shows an example of a data format for storing the definition of an outer fold. [Figure 11] This is a flowchart illustrating the process performed in step 846 in the third embodiment. [Figure 12] This is a flowchart illustrating the process performed in step 848 in the third embodiment. [Figure 13] This is a flowchart illustrating an exemplary method for predicting the presence or absence of disease using the model obtained by the third embodiment. [Modes for carrying out the invention]

[0024] The present invention will be described below using the best possible embodiment as an example, with reference to the attached drawings.

[0025] Figure 1 is a block diagram illustrating an example of an embodiment of the present invention. The model creation means 100 depicted in the center is a central element in carrying out the present invention. The model creation means 100 is embodied by a physical or virtual computer device and a computer program. That is, a physical or virtual computer device can operate as the model creation means 100 by executing a computer program on its physical or virtual processor. Therefore, the model creation means 100 can be implemented as a computer device or a software application. Since the model creation means 100 can be a computer device or a software application, it may also receive user input from a user input means 102 such as a keyboard, mouse, or touch panel, or display output on a computer monitor device 104.

[0026] Figure 1 also depicts various libraries 112, 114, and 116 that provide specific functions. These libraries are used by the model creation means 100 and provide predetermined outputs to the model creation means 100. Although these are referred to as "libraries," in some embodiments, these libraries may also be separate physical or virtual computer devices that can be distinguished from the computer device of the model creation means 100. In some embodiments, these libraries may also be executable software applications that can be controlled by the model creation means 100. Therefore, in some embodiments, it may be more appropriate to refer to elements 112, 114, and 116 as physical or virtual computer devices or software applications rather than "libraries."

[0027] Figure 1 also depicts various storage means 132, 134, 136, and 138. Depending on the embodiment, each of these storage means may be a software-defined storage area on a general-purpose computer storage device such as a hard disk or SSD. However, depending on the embodiment, at least one of these storage means may be a physical or virtual computer device whose primary purpose is data storage, sharing, and management.

[0028] It should be noted that, depending on the embodiment, the elements depicted in Figure 1 may not be as clearly distinguishable as shown in Figure 1. For example, in some embodiments, the program code of at least one of libraries 112, 114, and 116 may be incorporated into the program code that constitutes the model creation means 100. In that case, it would be difficult to distinguish element 100 from elements 112, 114, and 116 in the manner shown in Figure 1. In some embodiments, at least one of the storage means 132, 134, 136, and 138 may be a software-defined storage area by the program code of the model creation means 100. It would be difficult to visually distinguish a software-defined storage area from the model creation means 100. In some embodiments, all elements of the figure, except for the user input means 102 and the monitor device 104, may be virtually implemented using a computer cloud. The block diagram in Figure 1 is merely an example, and it should be noted that since the present invention is embodied by a computer device, there may be various variations in its specific implementation. [Embodiment 1]

[0029] Figure 2 shows a flowchart 200 illustrating a first embodiment of the present invention. Flowchart 200 illustrates a method for creating a machine learning model that predicts the presence or absence of disease from a set of gut microbiota data. The method according to flowchart 200 can be implemented by an architecture as illustrated in Figure 1.

[0030] Step 210 indicates the start of processing. Step 220 indicates that the model creation means 100 reads a set of gut microbiota data for creating a machine learning model to predict the presence or absence of disease. The set of gut microbiota data may be stored in the gut microbiota data storage 132 shown in Figure 1.

[0031] Figure 3A shows an example of gut microbiota data 300, which can be read in step 220. The gut microbiota data that can be used in the present invention is: • Sample ID 302, which is information that identifies the sample (e.g., the subject's name, initials, or identification number); • Age information of the subjects in the sample (for example, the actual age of the subjects in the sample, or a binned version of the subjects' actual ages, such as in their 40s, 50s, etc.); • Disease ID 306, which can indicate whether or not the subject of the specimen has the disease whose presence or absence is being predicted; • Microbial flora data obtained from samples 320a, 320b, ..., 320n; It holds.

[0032] Depending on the embodiment, the disease ID 306 may represent information indicating the name of a disease (such as cerebral infarction, diabetes, or colorectal cancer) or information indicating the absence of a disease. For example, a disease ID of 0 may indicate that the sample was obtained from a subject without a disease, a value of 1 may indicate that the sample was obtained from a subject with cerebral infarction, and a value of 2 may indicate that the sample was obtained from a subject with diabetes. However, if the gut microbiota data read in step 220 only has two possible cases—that it was obtained from a sample of a subject with a specific disease or that it was obtained from a sample of a subject without a disease—then the disease ID 306 only needs to be information that can identify the presence or absence of the disease.

[0033] Depending on the embodiment, the microbial community data 320a, 320b, ..., 320n may each include identification information 322 for ASV (Amplicon Sequence Variant) and its count value (number of reads) 324, as shown in Figure 3B, and may also include identification information 326 for the fungal genus and its count value (number of reads) 328, as shown in Figure 3C. Here, the symbol i in Figures 3B and 3C is used to indicate the i-th item.

[0034] As is well known, ASVs are obtained based on the nucleotide sequence of a stool sample, and the bacterial genus (identification information, e.g., genus name) is identified by comparing the ASV with a bacterial genus database.

[0035] Depending on the embodiment, the gut microbiota data 300 may include gender information 308 of the subject of the sample, and quantity information 310 indicating how many of the microbiota data 320a, 320b, ... 320n there are. It may also include a field 312 for later use. (The example field 312 shown here will be used in the example described later.)

[0036] It should be noted that the data formats shown are merely examples. The various data formats illustrated in the drawings of this application, including Figure 3, are actually usable and are excellent in terms of facilitating understanding of the embodiments. However, the illustrated data formats do not impose any limitations on the present invention. It goes without saying that the present invention can be implemented with various data formats.

[0037] In step 220, gut microbiota data for many samples is read from the gut microbiota data storage 132. In a preferred embodiment, it is preferable that the gender of the subjects in all samples is the same. Furthermore, it is preferable that the read gut microbiota data is either obtained from samples with a specific disease for which a machine learning model is being created to predict the presence or absence of that disease, or obtained from healthy samples without the disease. In some embodiments, the gut microbiota data stored in the gut microbiota data storage 132 may all be obtained from samples of a specific gender, and there may only be two types of data: data obtained from samples with a specific disease or data obtained from healthy samples. In that case, the model creation means 100 can simply read the gut microbiota data stored in the gut microbiota data storage 132 in step 220. If the gut microbiota data storage 132 contains data of different genders and different diseases, the model creation means 100 is configured to read the necessary data using gender information 308 and disease ID 306 as search keys. Depending on the embodiment, the model creation means 100 may be configured to load gut microbiota data related to a specific disease or a specific gender based on user instructions received through the user input means 102.

[0038] In step 230, the gut microbiota data read in step 220 is converted into compositional data. In a preferred embodiment, the count values ​​(number of reads) exemplified in Figures 3B and 3C are converted into CLR (Centered Log-Ratio Transformation) values.

[0039] Converting count values ​​(read counts) to CLR values ​​is a well-known process, and existing libraries capable of performing such conversions exist, such as ALDEx2. Therefore, the model creation means 100 may be configured to perform such conversions using an existing composition data creation library 112 such as ALDEx2.

[0040] When using ALDEx2, the conversion to CLR values ​​is performed by associating a probability distribution with the count values ​​and performing Monte Carlo (MC) sampling of that probability distribution multiple times. CLR values ​​are obtained from each sampled value. Therefore, in step 230, multiple CLR data sets with the same sample ID are created from a single gut microbiota data set.

[0041] Figure 4A shows an example of CLR data 400, which is created from the gut microbiota data 300 in step 230. However, the data format shown differs from the CLR data obtained when using ALDEx2. Nevertheless, for the sake of easier understanding of the embodiment, this specification will use the data format shown in Figure 4. Also, fields common to both the gut microbiota data 300 and the CLR data 400 are denoted by the same reference numerals and their explanation is omitted.

[0042] As shown in Figure 4A, the CLR data 400 has an additional field 402 that stores the Monte Carlo sampling (MCS) number. As mentioned above, multiple MC samplings are performed in the conversion to CLR values, and the number in field 402 indicates which sampling it was obtained from. In the CLR conversion by ALDEx2, 128 MC samplings are performed by default, so in step 230, 128 CLR data with the same sample ID may be created from one intestinal microbiota data. In some embodiments, the model creation means 100 may store the created CLR data in the composition data storage unit 134. In some embodiments, the model creation means 100 may be configured to store only a portion of the created CLR data, rather than all of it. For example, it may be configured to store one in several data points.

[0043] In addition, a field 404 may be added to the CLR data 400 for later use. In the embodiments described herein, this illustrated field 404 is used in step 242 of Figure 2 and step 542 of Figure 5.

[0044] In addition to the addition of fields 402 and 404, the CLR data 400 differs from the gut microbiota data 300 in that field 310, which stores the number of microbiota data, has been changed to field 410, which represents the number of CLR data, and that individual microbiota data 320a, 320b, ... 320n have been changed to their corresponding individual CLR data 420a, 420b, ... 420n. If the microbiota data 320 is ASV data as shown in Figure 3B, the individual CLR data 420 will be as shown in Figure 4B, with the count value 324 of the microbiota data 320 changed to a CLR value of 424. Similarly, if the microbiota data 320 is genus data as shown in Figure 3C, the individual CLR data 420 will be as shown in Figure 4C, with the count value 328 of the microbiota data 320 changed to a CLR value of 248.

[0045] In step 240, a machine learning model is created to predict the presence or absence of the specific disease using the set of CLR data 400 created in step 230. Embodiments of the present invention are configured to create the machine learning model from the set of CLR data 400 using the Nested Cross Validation method and a decision tree-based model generation algorithm. The model creation means 100 may be configured to create the machine learning model using an existing model generation library 116. Known existing model generation libraries that implement a decision tree-based model generation algorithm include LightGBM and Random Forest. In a preferred embodiment of the present invention, it is preferable to use LightGBM. That is, the model generation library 116 may be a library that implements LightGBM.

[0046] The model creation means 100 is configured to create a machine learning model in step 240 using the set of CLR data 400 created in step 230 and employing the Nested Cross Validation method. Nested Cross Validation has various translations, such as nested cross-validation, nested cross-validation, and double cross-validation, so in this application, the English name Nested Cross Validation is used. Nested Cross Validation itself is a well-known method and broadly consists of four steps, as shown in step 240 in Figure 2.

[0047] In step 242, the set of CLR data 400 is divided into multiple subsets. These subsets are referred to as "folds" in Cross Validation terminology. The present invention is characterized in that the CLR data 400 is divided such that CLR data 400 having the same sample ID belong to the same fold.

[0048] As a result of the processing in step 242, each CLR data 400 will belong to one of the folds, but the identification information of the fold to which it belongs (for example, the fold number) may be stored in field 404.

[0049] In step 244, multiple Outer folds are defined by combining the multiple folds defined in step 242. As is well known, each Outer fold consists of a fold used as test data and a fold used as training data. The multiple Outer folds that are defined each have different folds used as test data. That is, the test data of one Outer fold is different from the test data of another Outer fold.

[0050] In step 244, multiple Inner folds are further defined for each of the defined Outer folds. Each Inner fold is created by combining multiple folds included in the training data, with some folds used as validation data and the remaining folds used as training data (training data in the Inner fold). Each of the defined Inner folds has different folds used as validation data. That is, the validation data of one Inner fold is different from the validation data of another Inner fold.

[0051] In step 246, cross validation is performed for each outer fold using multiple inner folds belonging to that outer fold. At this time, the present invention is, The 400 CLR data points included in the fold, which will serve as the training data, will be used individually for model training. • For the 400 CLR data points included in the fold used as validation data, the average CLR data will be created by averaging all 400 CLR data points with the same sample ID, and this average CLR data will be used to validate the model. It is characterized by the following:

[0052] As described above, in this embodiment, the model creation means 100 is configured to train the above-mentioned model using LightGBM.

[0053] As a result of step 230, each fold contains multiple CLR data 400 with the same sample ID. If these multiple CLR data 400 are training data, the model creation means 100 inputs each of these multiple CLR data 400 individually into the model creation library 116. On the other hand, for validation data, the model creation means 100 averages all CLR data 400 with the same sample ID to create average CLR data. Such average CLR data may have the same data format as the CLR data 400, but the individual CLR values ​​stored are averaged for each ASV or genus. That is, if the average CLR data has the data format exemplified in Figures 4A to 4C, the CLR values ​​424i or 428i are averaged for each ASV ID 322i or genus ID 326i, respectively. The model creation means 100 uses such average CLR data to perform the validation process in step 246.

[0054] In step 246, the explanatory variables that the model creation means 100 inputs into the model creation library 116 are at least the age information 304 and the individual CLR data 420i (i=1,2,...) from the information contained in the CLR data 400. Other information may also be included. As the target variable, information indicating the presence or absence of the disease to be predicted, based on the disease ID 306, is used. This information may be the disease ID 306 itself. Alternatively, it may be binary information, such as 1 if the disease is present and 0 if the disease is not present. In some embodiments, the model creation means 100 may be configured to process the CLR data 400 to remove information that is not necessary for model creation, and then input the processed data into the model creation library 116.

[0055] In step 248, cross-validation of the outer fold is performed. That is, for each outer fold, the model is trained using the training data, and the accuracy of the model is determined using the test data. As mentioned above, the training data at this stage includes both the data designated as training data and the data designated as validation data in step 246. In this embodiment, the model creation means 100 is configured to train the model using LightGBM in step 248 as well.

[0056] In step 248, as in step 246, the 400 CLR data points included in the fold used as training data are used individually to train the model, while the 400 CLR data points included in the fold used as test data are averaged by averaging the individual CLR values ​​within the 400 CLR data points with the same sample ID to create average CLR data, and the model's accuracy is determined using this average CLR data.

[0057] The model obtained in step 248 will be used to predict the disease. A model is created for each Outer fold. The prediction accuracy is the average of the prediction accuracy of each model obtained in step 248. The model obtained in step 248, along with its accuracy information, is stored in the model storage unit 138 (step 250).

[0058] Current estimation models sample a probability distribution associated with count values ​​(read counts) multiple times and use only the average of each sample value for training. However, since count values ​​are determined probabilistically, using only the average of the sample values ​​results in a loss of variance information, which can lead to a decrease in accuracy. On the other hand, according to the method of this application, each sample value is used individually for model training, so the essential fact that count values ​​are determined probabilistically can be reflected in the model training, improving the accuracy of the created model. In this case, when dividing the dataset into multiple folds for nested cross-validation, leakage is prevented by dividing CLR data with the same sample ID so that they belong to the same fold. Furthermore, in the stage of determining the accuracy of the model, the above method uses the average of these sample values. This makes it possible to obtain accuracy that is appropriate for actual operation. That is, it is possible to obtain accuracy appropriate for determining the presence or absence of disease in unknown gut microbiota data using the created model.

[0059] Furthermore, each step in the flowchart of Figure 2 is executed by the model creation means 100. The model creation means 100 executes each step in the flowchart of Figure 2 while utilizing the composition data creation library 112, the model creation library 116, and various storage means. Therefore, models can be created without requiring manual intervention. [Embodiment 2]

[0060] Figure 5 shows a flowchart 500 for illustrating a second embodiment of the present invention. Like flowchart 200, flowchart 500 also illustrates a method for creating a machine learning model that predicts the presence or absence of disease from a set of gut microbiota data. Blocks that perform the same processing as in flowchart 200 are denoted by the same reference numerals and their descriptions are omitted. The method according to flowchart 500, like the method according to flowchart 200, can be executed by an architecture as illustrated in Figure 1.

[0061] The method described in flowchart 500 differs from the method described in flowchart 200 in that the process of dividing the dataset used for model generation into folds (step 242 in flowchart 200) is improved. Due to this improvement, flowchart 500 includes step 522 compared to flowchart 200, and the sign of the step for dividing into folds has been changed from 242 to 542.

[0062] In step 522, enterotype estimation is performed for each of the gut microbiota data points read in step 220. Enterotypes are classified into types based on the similarity of the composition ratio of gut bacteria. For more detailed information on enterotypes, please refer to the following URL. https: / / www.symbiosis-solutions.co.jp / reaserch-results / reaserch-results-002 /

[0063] Enterotype estimation can be performed by examining the similarity between the reference data for each enterotype and the gut microbiota data. For example, the URL above introduces four enterotypes for both Japanese men and women, and shows the composition ratio of bacterial genus for each enterotype. Such data can serve as an example of reference data. If the gut microbiota data loaded in step 220 consists of bacterial genus and count values ​​(read count), as shown in the example in Figure 3C, the count values ​​can be converted to composition ratios, and then the similarity with the above reference data can be calculated. If the gut microbiota data loaded in step 220 consists of ASV and count values, as shown in the example in Figure 3B, the ASV can be mapped to a bacterial genus by comparing it with a bacterial genus database as described above, and then the count values ​​can be converted to composition ratios, and then the similarity with the above reference data can be calculated. Various existing similarity measures used in the field of machine learning, such as squared distance and cosine distance, can be used as the similarity measure. The enterotype determined to be the most similar will be the enterotype of the gut microbiota data.

[0064] Depending on the embodiment, the determined enterotype may be recorded in the gut microbiota data. For example, in the embodiment described below, information regarding the determined enterotype is stored in field 312 in Figure 3A.

[0065] Those skilled in the art would find it easy to embed the code for performing enterotype estimation as described above into the program code of the model creation means 100. Furthermore, to ensure versatility, the part that performs enterotype estimation may be configured as an enterotype estimation library 114, separate from the model creation program 110, as illustrated in Figure 1.

[0066] It goes without saying that the above reference data regarding enterotypes is merely illustrative. Those skilled in the art will be able to prepare and utilize various forms of reference data.

[0067] Next, step 542 will be described. In step 542, similar to step 242 in the first embodiment, the set of CLR data 400 created in step 230 is divided into multiple folds such that CLR data 400 with the same sample ID belong to the same fold. However, in step 542, this division is carried out in the steps shown in Figure 6A.

[0068] Figure 6A is a flowchart illustrating the process performed in step 542 in the second embodiment. The process shown in this flowchart may be a process performed by the model creation means 100.

[0069] Step 602 indicates the start of processing. In step 604, the set of CLR data 400 is divided into two groups: a data set for individuals with the disease and a data set for individuals without the disease. In some embodiments, the modeling means 100 may be configured to classify the CLR data into a data set for individuals with the disease and a data set for healthy individuals based on health information contained in the CLR data (in the case of the CLR data 400, disease ID 306).

[0070] In step 606, the data set of healthy individuals is divided into k subgroups such that the enterotype distribution is even, where k is the number of folds. Such division can be performed, for example, by assigning subgroup numbers sequentially in a loop to multiple CLR data 400 having a particular enterotype, and then grouping CLR data 400 having the same subgroup number into one subgroup. However, it is necessary to adjust the CLR data 400 having the same sample ID so that they belong to the same subgroup. Those skilled in the art will know that the model creation means 100 can be configured to perform subgroup division using various other methods. In some embodiments, the model creation means 100 may be configured to search for CLR data relating to a specific enterotype using enterotype information contained in the CLR data (in the example of CLR data 400, the enterotype information stored in field 312) as a search key.

[0071] In step 608, the data set of individuals with the disease is also divided into k subgroups so that the enterotype distribution is even. In step 610, for each subgroup of healthy individuals, a fold number is assigned to each subgroup in ascending order of the number of CLR data points (400) included in that subgroup. In this example, the fold numbers range from 0 to k-1. In step 612, for each subgroup of diseased individuals, a fold number is assigned to each subgroup in descending order of the number of CLR data points (400) included in that subgroup. In step 614, subgroups with the same fold number are joined. Step 616 indicates the end of the process.

[0072] Figure 7 shows an example of the results of the above processing. In this example, the set of CLR data created in step 230 was divided into 10 folds. The upper graph shows the number of CLR data from healthy individuals and the number of CLR data from diseased individuals in each fold. As shown, the balance between the number of CLR data from healthy individuals and the number of CLR data from diseased individuals is balanced across all folds. The lower graph shows the distribution of the number of CLR data with different enterotypes in each fold. As shown, the distribution of enterotypes is also balanced across all folds.

[0073] By the way, in any of steps 610, 612, or 614, it is determined which fold each CLR data 400 belongs to. The model creation means 100 may then store identification information (e.g., the fold number) of the fold to which each CLR data 400 belongs within the CLR data 400. For example, this information may be stored in the field 404 shown in Figure 4A.

[0074] Depending on the embodiment, the CLR data 400 may be transformed into a format (CLR data 650) as shown in Figure 6B. This is because the gender information 308 and enterotype information 312 in the CLR data 400 are no longer used. That is, in step 220, the model creation means 100 only reads sample data obtained from subjects of the same gender, so the gender information 308 is no longer needed. Similarly, the enterotype information 312 is no longer used once the fold to which each CLR data belongs has been determined, so this is also no longer needed. Furthermore, in step 220, only sample data from subjects with a specific disease and sample data from subjects without the disease are read, so the disease ID 306 in the CLR data 400 may also be changed to binary information indicating the presence or absence of the disease (in Figure 6B, these are represented as health IDs 654). The CLR ID 656 and CLR value 658, which are the contents of the individual CLR data 420i in the CLR data 650, are the same as those in Figure 4B or Figure 4C, respectively. In other words, if the CLR value 658 is a value obtained based on ASV, then CLR ID 656 and CLR value 658 are identical to ASV ID 322i and CLR value 424i in Figure 4B, respectively. Also, if the CLR value 658 is a value obtained based on the genus of the fungus, then CLR ID 656 and CLR value 658 are identical to genus ID 326i and CLR value 428i in Figure 4B.

[0075] It should be reiterated that the illustrated data formats are merely examples and do not impose any limitations on the present invention; the present invention can be implemented using various data formats. The reason why the data formats shown in CLR data 400 and CLR data 650 are introduced in this application is not only because they are actually implementable, but also because they are excellent from the perspective of facilitating the understanding of the embodiments. From this perspective, the following description will use CLR data 650 instead of CLR data 400. This is because gender information 308 and enterotype information 312 are no longer used.

[0076] Furthermore, it goes without saying that the order of each step in the flowchart in Figure 6A is illustrative. For example, it is clear that the order of steps 606 and 608 can be swapped, and the order of steps 610 and 612 can also be swapped. It is also clear that similar results can be obtained even if the fold numbers in step 610 are assigned in descending order and in step 612 in ascending order. In other words, it is clear that the data can be divided into folds in such a way that the balance between the number of CLR data from healthy individuals and the number of CLR data from diseased individuals is maintained across all folds, and the distribution of enterotypes is also balanced.

[0077] As described above, each processing step explained in relation to step 542 is a process performed by the model creation means 100. That is, the model creation means 100 may be configured to perform such processing. Therefore, as with the first embodiment, the second embodiment also allows for the creation of a model without manual intervention. Furthermore, the accuracy of the created model is improved by performing nested cross-validation using a fold in which the distributions of both health status and enterotype are balanced. [Embodiment 3]

[0078] Figure 8 shows a flowchart 800 for illustrating a third embodiment of the present invention. Like flowcharts 200 and 500, flowchart 800 also illustrates a method for creating a machine learning model that predicts the presence or absence of disease from a set of gut microbiota data. Blocks that perform the same processing as in flowcharts 200 and 500 are denoted by the same reference numerals and their descriptions are omitted. The method according to flowchart 800, like the methods according to flowcharts 200 and 500, can be executed by an architecture as illustrated in Figure 1.

[0079] In the third embodiment of the present invention, the processing content of the cross validation of the inner fold and the processing content of the cross validation of the outer fold differ from those of the first and second embodiments. Due to this difference, in flowchart 800, compared to flowcharts 200 and 500, the reference numeral for the step of performing cross validation of the inner fold is changed from 246 to 846, and the reference numeral for the step of performing cross validation of the outer fold is changed from 248 to 848.

[0080] Since steps 522 and 542 are present in flowchart 800, the third embodiment is shown as a modification of the second embodiment. However, steps 846 and 848 are also applicable to modifications of the first embodiment.

[0081] Before explaining steps 846 and 848, let's provide some supplementary information about the process in step 244, which defines the outer fold and inner fold. Figure 9 shows an example of defining the outer fold and inner fold, using the case where there are 5 folds, 5 outer folds, and 4 inner folds. The upper row shows an example of the outer fold, and the lower row shows an example of the inner fold.

[0082] Nested Cross Validation performs cross validation in a nested manner, with the outer fold defined for the outer cross validation. In this example, there are 5 folds and 5 outer folds, so 5 outer folds are defined, each containing a different fold that serves as the test data. In Figure 9, the label 902 indicates the outer fold whose test data is fold number 4. Since the test data for outer fold 902 is fold #4, the 4 folds that will be used as training data are fold numbers 0, 1, 2, and 3.

[0083] An inner fold is defined for inner cross-validation and is defined individually for the training data of each outer fold. In this example, each outer fold has four training data sets, and there are also four inner folds. Therefore, for each outer fold, four inner folds are defined, each containing different validation data. For outer fold 902, four inner folds are defined as illustrated in the lower part of Figure 9. Of these, code 904 represents the inner fold whose validation data is fold number 0.

[0084] In some embodiments, the model creation means 100 may be configured to store the definition of the outer fold in data in a format illustrated in Figure 10. In some embodiments, the model creation means 100 may be configured to store such data in the fold storage unit 136 (see Figure 1). Again, it should be noted that such data formats are merely illustrative and do not limit the scope of the present invention. The data format 1000 depicted in Figure 10 includes identification information 1002 for the outer fold, identification information 1004 for the test fold of the outer fold, identification information 1006 for the train fold of the outer fold, information on the number of inner folds 1010, and definition information 1020#i (where i is equal to the number of inner folds). Field 1008 will be described later.

[0085] The definition information 1020#i of the inner fold includes, respectively, the identification information 1022 of the inner fold, the identification information 1024 of the validation fold of the inner fold, and the identification information 1026 of the train fold of the inner fold.

[0086] If we store the information of Outer fold 902 in Figure 9 in data format 1000, Outer fold 902 is the 5th Outer fold in Figure 9, so the identification information of Outer fold 1002 might be 4 (because in the example in Figure 7, the fold number of the 1st Outer fold is 0). Also, the identification information of test fold 1004 might be 4, the identification information of train fold 1006 might be 0, 1, 2, 3, and the number information of Inner fold 1010 might be 4. Furthermore, if we store the information about the first Inner fold 904 of Outer fold 902 in the first column of the Inner fold definition information 1020, then since Inner fold 904 is the first Inner fold of Outer fold 902, the Inner fold identification information 1022 would be 0, the validation fold identification information 1024 would be 0, and the train fold identification information 1026 would be 1, 2, and 3. It should be reiterated that such a data format is merely an example and does not limit the scope of the present invention. Also, while Figures 9 and 10 show a division into folds of 5, this is also merely illustrative. If the number of samples exceeds 10, it is preferable that the number of divisions into folds be at least 10.

[0087] Next, referring to the flowchart in Figure 11, we will explain step 846, which performs cross-validation of the inner fold. The cross-validation in step 846 is characterized by its purpose of determining the optimal number of explanatory variables for each outer fold. This number is used for model training in step 848, which is shown as the next step after step 846 in Figure 8.

[0088] Step 1104 indicates the start of processing. The loops shown in steps 1108 and 1152 indicate that processing is performed for each of the k outer folds. Therefore, the symbol i used in steps 1108 and 1152 in Figure 11 indicates the outer fold number.

[0089] The loops shown in steps 1112 and 1144 indicate that the subsequent processing is carried out by varying the number of explanatory variables. Therefore, the symbol n used in these steps in Figure 11 represents the number of explanatory variables. For example, if the subsequent processing is carried out by changing the number of explanatory variables by 10 each time, n1 in step 1112 could be 10 and n2 could be 20. N represents the maximum possible number. Since N may not be a multiple of 10, the intervals of n are not necessarily equal.

[0090] These explanatory variables are data used for training or validating the model in steps 1120, 1128, and 1132. These are at least the age information 304 and the individual CLR data 420x (x=1,2,...) from the information contained in the CLR data 650. However, in step 1132, the individual CLR data 420x is the mean of the individual CLR data 420x of all CLR data 650 with the same sample ID. The dependent variable is the health ID (654,306). The age information 304 may be the actual age of the subject in the sample, but it is preferable to use the actual age of the subject in the sample binned, such as 40s, 50s, etc. These are the same as those described in the description of the first embodiment, particularly in relation to step 246.

[0091] The loops represented by steps 1116 and 1136 indicate that processing is performed for each of the k-1 inner folds. Therefore, the symbol j used in these steps in Figure 11 indicates the inner fold number.

[0092] Step 1120 involves training the model to determine feature importance. In this step, the model is trained using all explanatory variables. Specifically, in this step, the model creation means 100 inputs all individual CLR data 420x from all CLR data 650 designated as training data and age information 304 into the model generation library 116 to obtain the model. Step 1124 involves calculating feature importance. If the model generation library 116 utilizes LightGBM, it is possible to calculate feature importance based on the number of splits (Split Count Importance), information gain (Gain Importance), and coverage (Cover Importance). Steps 1120 and 1124 may not be clearly separate depending on the embodiment. In any case, model training and feature importance calculation may be performed using an existing model creation library such as LightGBM. Furthermore, various other types of feature importance are known besides the three mentioned above, and other types may also be calculated.

[0093] In step 1128, the model is trained again using the n explanatory variables that were determined to be of high importance in step 1124. That is, in this step, the model creation means 100 inputs only the n pieces of information that were determined to be of high importance from all the individual CLR data 420x and age information 304 for each CLR data 650 used as training data into the model generation library 116.

[0094] In step 1132, the accuracy of the model is calculated and recorded using the 650 CLR data points designated as validation data. Again, the accuracy is calculated using only the n explanatory variables deemed to be of high importance. As mentioned earlier, in this step, the average CLR data is created by averaging the individual CLR data points 420x of all 650 CLR data points with the same sample ID, and the accuracy is calculated using this average CLR data.

[0095] In step 1128, the model is trained using individual CLR data corresponding to individual sample values, and in step 1132, validation is performed using average CLR data obtained by averaging multiple individual CLR data. This allows the inherent fact that count values ​​are determined probabilistically to be reflected in the model training, improving the accuracy of the created model, while also making it possible to determine that accuracy as a value that is more relevant to actual operation.

[0096] To reiterate, in steps 1120, 1128, and 1132, the dependent variable is the health ID, which is binary information indicating whether or not a person has a disease.

[0097] Once the loops represented by steps 1116 and 1136 have completed and the model accuracy for the k-1 inner folds with n explanatory variables has been calculated, step 1140 calculates the average accuracy for the case with n explanatory variables by averaging all the accuracies.

[0098] When the loop represented by steps 1112 and 1144 is completed, the average accuracy for each n is obtained. Therefore, the n that gives the best average accuracy for the Outer fold can be determined (step 1148).

[0099] When the loop represented by steps 1108 and 1152 has completed, the value of n that gives the best accuracy average for each outer fold is determined. Step 1156 indicates the completion of the process.

[0100] In this embodiment, the processing of the flowchart in Figure 11 is performed by the model creation means 100 using the model generation library 116. Depending on the embodiment, the model creation means 100 may use the data format 1000 shown in Figure 10 or a similar data format to perform the processing of the flowchart in Figure 11. That is, the data format 1000 shown in Figure 10 may be used to select appropriate outer folds and inner folds in the loops starting in steps 1108, 1112, and 1116, or to input appropriate training data and validation data into the model generation library 116 in steps 1120, 1128, and 1132. The model creation means 100 may also be configured to store the value n, which gives the best accuracy average determined in step 1148, in the field 1008 of the data format 1000.

[0101] Note that the processing flow shown in Figure 11 is merely illustrative in the implementation of step 846. For example, steps 1120 and 1124 are the same regardless of j and n, so they can be completed before step 1112. Then, in the loop starting from step 1116, the processing flow may be such that the already calculated feature importance information is read out and step 1128 is performed.

[0102] Next, referring to the flowchart in Figure 12, we will explain step 848, which performs cross-validation of the outer fold. In step 848, a model is created for each outer fold using the number of explanatory variables defined in step 846, and its accuracy is calculated.

[0103] Step 1204 indicates the start of processing. The loops shown in steps 1208 and 1228 indicate that processing is performed for each of the k outer folds. Therefore, the symbol i used in steps 1208 and 1228 in Figure 12 indicates the outer fold number.

[0104] In step 1212, the model is trained for the purpose of determining feature importance. In this step, the model is trained using all explanatory variables. Therefore, the processing in this step is virtually identical to the processing in step 1120. However, the train data in step 1212 is different from the train data in step 1120, and all CLR data other than the CLR data designated as test data in the Outer fold is used as train data. The model creation means 100 may be configured to use the data format 1000 shown in Figure 10 to extract the test data and train data in the Outer fold, or to obtain the number of explanatory variables for the Outer fold.

[0105] In step 1212, the model creation means 100 inputs all individual CLR data 420x and age information 304 from all CLR data 650 used as training data, as explanatory variables into the model generation library 116, similar to step 1120. The target variable is the health ID, i.e., information on whether or not a person has a disease. In step 1216, feature importance is calculated. The processing content of this step is also virtually identical to that of step 1124.

[0106] In step 1220, the model is trained again using the n explanatory variables determined for the Outer fold as a result of step 848. That is, the model creation means 100 inputs only the n pieces of information that are determined to be of high importance from all the individual CLR data 420x and age information 304 for each CLR data 650 which constitute the training data, into the model generation library 116. The model obtained in step 1220 forms part of the machine learning model that predicts the presence or absence of disease, which is obtained in the third embodiment.

[0107] In step 1224, the model's accuracy is calculated and recorded using the test data, CLR data 650. Again, the accuracy is calculated using only the n explanatory variables deemed to be of high importance. As mentioned earlier, in this step, the average CLR data is created by averaging the individual CLR data 420x of all CLR data 650 with the same sample ID, and the accuracy is calculated using this average CLR data.

[0108] Similar to steps 1128 and 1132, in step 1220 the model is trained using individual CLR data corresponding to individual sample values, and in step 1224 the model is validated using average CLR data obtained by averaging multiple individual CLR data. This allows the inherent fact that count values ​​are determined probabilistically to be reflected in the model training, improving the accuracy of the created model, while also making it possible to determine that accuracy as a value that is more relevant to actual operation.

[0109] When the loop represented by steps 1208 and 1228 has completed, step 1232 calculates the average accuracy by averaging the individual accuracies obtained in step 1224 for each outer fold. Step 1236 indicates the end of the process.

[0110] Step 848 completes the machine learning model obtained in the third embodiment for predicting the presence or absence of disease. That is, one model is created for each outer fold, and its predicted value is the average of the values ​​estimated by each model. The prediction accuracy is the average accuracy calculated in step 1232. In some embodiments, the model creation means 100 may be configured to store these models and their accuracy in the model storage unit 138 (see Figure 1). Preferably, these models are stored in the model storage unit 138 along with information representing the disease to be predicted and the gender of the subject of the sample for which the model was created.

[0111] In the third embodiment, the number of explanatory variables that yield the best accuracy is determined in the inner Cross Validation, and in the outer Cross Validation, a model is created using that number of explanatory variables in order of feature importance. This model provides predictions with better accuracy than a model created using all explanatory variables.

[0112] Furthermore, each step of the method described in relation to the third embodiment is also performed by the model creation means 100. Therefore, as with the first and second embodiments, the model can be created without human intervention.

[0113] Figure 13 is a flowchart of a method 1300 for predicting the presence or absence of disease in gut microbiota data of a specific specimen using a model obtained by the third embodiment. This method is performed, for example, by a disease prediction means 140 depicted in Figure 1. The disease prediction means 140, like the model creation means 100, is embodied by a physical or virtual computer device and a computer program. That is, a physical or virtual computer device can operate as the disease prediction means 140 by executing a computer program with its physical or virtual processor. Since the disease prediction means 140 can be a computer device, it may be configured to receive user input from a user input means 102 such as a keyboard, mouse, or touch panel, output to a computer monitor device 104, or provide output to other computer devices not shown.

[0114] Step 1302 indicates the start of processing. Step 1304 reads intestinal microbiota data that is the subject of disease presence / absence prediction. In this step, the disease prediction means 140 may read specific intestinal microbiota data from the intestinal microbiota data storage 132 based on user input (e.g., sample ID) received, for example, through the user input means 102. This intestinal microbiota data may be the same as the intestinal microbiota data 300 introduced earlier. In step 1306, the disease prediction means 140 converts the read intestinal microbiota data into CLR data. This conversion may be the same as the conversion described in relation to step 230 in Figure 2. This conversion creates multiple CLR data with the same sample ID. However, in step 1306, the CLR values ​​contained in these CLR data are averaged for each CLR ID, so that one average CLR value is associated with each CLR ID. That is, the data is the same as the data used when determining the accuracy of the model when the model was created.

[0115] In step 1308, the disease prediction means 140 may read a machine learning model from the model storage unit 138 that is suited to the disease to be predicted and the gender of the subject of the sample read in step 1304. The machine learning model read here may be a model obtained in the third embodiment, and is created for each disease for which the presence or absence of the disease is to be predicted, and for each gender. Also, since there are as many models as there are Outer folds defined in step 244 of Figure 8, it might be more appropriate to call the models read in step 1308 a "set of models". This "set of models" also includes, or is associated with, information on the number of models included in the set of models, information on what should be input as explanatory variables, and accuracy information calculated in step 1232 of Figure 12.

[0116] In the loop represented by steps 1310 and 1316, the presence or absence of disease is predicted using the individual models included in the "set of models" loaded in step 1308. Therefore, the symbol i used in step 1310 represents the model number, and k represents the number of models. In step 1312, for example, if i=0, the average CLR data calculated in step 1306 is input to the first model, and the predicted value is obtained in step 1314. After exiting the loop, the k predicted values ​​obtained from the k models are averaged (step 1318), and this average predicted value is taken as the predicted value from the "set of models" mentioned above. Its accuracy is the accuracy calculated in step 1232 in Figure 12, which is the value loaded in step 1308.

[0117] In step 1320, the user input means 102 may display the obtained predicted value and accuracy on the display device 104 or output it to other computer devices or computer software (not shown). Step 1322 indicates the end of processing.

[0118] The present invention has been described above using what is considered to be the best possible embodiment at present, but the present invention can be implemented in various forms other than those described herein.

[0119] For example, while gut microbiota data may include all gut bacteria of a subject, it may also include other intestinal microorganisms such as fungi and yeasts, in addition to gut bacteria (bacteria, archaea, etc.) that inhabit the intestines.

[0120] As mentioned above, the model creation means 100 is often realized by the execution of a computer program by a general computer device, whether physical or virtual, and is often realized as a computer device or software application. The same applies to the disease prediction means 140. In the embodiments described, the model creation means 100 called and used various libraries, but in some embodiments, one or more program codes of these libraries may be integrated into the program code that constitutes the model creation means 100. Alternatively, in some embodiments, these libraries may also be separate physical or virtual computer devices or software applications that can be distinguished from the computer device or software application of the model creation means 100. In such embodiments, the expression "library" may not always be appropriate. As mentioned above, the various storage means 132, 134, 136, and 138 can be software-defined storage areas on storage devices used in general computers, such as hard disks and SSDs. However, in some embodiments, at least one of these storage means may be a physical or virtual computer device whose primary purpose is data storage, sharing, and management. Embodiments may exist in which the illustrated elements are divided into multiple sub-elements, or in which multiple elements are integrated into a single, indivisible element. Various steps have been introduced in the various flowcharts, but there are embodiments in which the order of these steps differs from the order in which they were introduced. Furthermore, embodiments may exist in which a certain step is broken down into multiple sub-steps, or in which multiple steps are integrated into a single step. Various variations are also possible for the target variable used for training and validating the model. For example, embodiments in which age information 304 is not used as the target variable in steps 1128 and 1132, and steps 1220 and 1224, etc. All of these variations are within the scope of the invention disclosed in this application.

Claims

1. A method for creating a machine learning model to predict the presence or absence of disease from a set of gut microbiota data, the method being performed by a computer, The gut microbiota data included in the set is all data created from samples of subjects of the same sex, and each set comprises at least a sample ID which is information that identifies a specific sample, health information which can at least indicate whether or not the subject of the sample has a specific disease, and microbiota data obtained from the sample, and the microbiota data consists of a combination of specific information related to the identification of bacteria and a count value related to the specific information. The aforementioned method, For each of the aforementioned intestinal microbiota data, a probability distribution is associated with the count value, and the probability distribution is sampled multiple times to create multiple composition data with the same sample ID; Using a cross-validation method, create the machine learning model from the set of composition data; Including the use of the Cross Validation method, The set of composition data is divided into multiple folds, provided that the composition data having the same sample ID is divided so that it belongs to the same fold; In the Cross Validation described above, during the model training stage, the composition data is used individually for training, and during the stage of determining the accuracy of the model, the average composition data is created by averaging multiple composition data having the same sample ID, and the accuracy is determined using the average composition data; Methods that include...

2. If each of the gut microbiota data included in the set is associated with an enterotype, or if no enterotype is associated with the gut microbiota data, the method includes estimating the enterotype for each of the gut microbiota data based on comparison with reference data. Dividing into the aforementioned multiple folds is Based on the aforementioned health information, the set of composition data is divided into two groups; - For each of the two groups mentioned above, the composition data belonging to that group is divided into multiple subgroups so that the distribution of enterotypes is even; - For one of the two groups, a fold number is assigned to the subgroup in order of the number of composition data included in that subgroup; - For the other of the two groups, assign a fold number to the subgroup in reverse order of the number of composition data included in the subgroup; - Combining subgroups that have the same fold number; The method according to claim 1, including the method described in claim 1.

3. The method according to claim 2, wherein the Cross Validation method is a Nested Cross Validation method.

4. (c1) For each inner fold of the Nested Cross Validation, train the model using all of the composition data and determine the feature importance for each explanatory variable; (c2) When the number of explanatory variables is n, the model is trained using the n explanatory variables for each of the Inner folds in descending order of feature importance for that Inner fold, and the accuracy of the obtained model is determined; (c3) From all the Inner folds belonging to a specific Outer fold, calculate the average value of the precision obtained in c2, when the explanatory variable is n; (c4) For the specific Outer fold, perform c2 and c3 by changing n, and store the n for which the average value obtained in c3 is best as the number of explanatory variables for the Outer fold; The method according to claim 3, including the method described in claim 3.

5. The above method applies to each of the Outer folds, (d1) Train the model using all of the aforementioned composition data and determine the feature importance for each explanatory variable; (d2) The model is trained again using the number of explanatory variables obtained in c4, in descending order of feature importance obtained in d1; The method according to claim 4, including the method described in claim 4.

6. The method according to claim 1, wherein the specified information is the name of the fungus genus.

7. The method according to claim 1, wherein the specified information is ASV.

8. The method according to claim 1, wherein the gut microbiota data further includes age information of the subject of the sample, and the learning target includes the age information.

9. The method according to claim 1, wherein creating the machine learning model from the set of composition data is done using a decision tree-based model generation algorithm.

10. A computer program, when executed by the processing means of the device, is configured to cause the device to perform the method described in any one of claims 1 to 9.

11. An apparatus comprising processing means and storage means, wherein the storage means stores a computer program, and the computer program, when executed by the processing means, is configured to cause the apparatus to perform the method described in any one of claims 1 to 9.

Images