Feature amount estimation device, feature amount estimation method and classification system

Abstract

To provide a feature amount estimation device, a feature amount estimation method, and a classification system that generate feature amount estimation rules for extracting feature amounts for time-series data and support the processing of classifying objects to be classified.SOLUTION: A feature amount estimation device includes: a feature amount estimation rule generation unit 20 that generates feature amount estimation rules as rules for estimating feature amounts based on a relation between pieces of data; an estimated feature amount candidate creation unit that applies the feature amount estimation rules to time-series data to be classified and creates estimated feature amount candidates that are candidates for the feature amounts; a feature amount evaluation unit 21 that evaluates classification performance for the time-series data in relation to the created estimated feature amount candidates; and a feature amount estimation rule addition unit that adds a feature amount estimation rule with a changed evaluation value to the feature amount estimation rules corresponding to the estimated feature amount candidates that need to be redefined as a result of the evaluation.SELECTED DRAWING: Figure 3

Description

[Technical field]

[0001] The present invention relates to a feature estimation device, a feature estimation method, and a classification system. The present invention relates to a feature estimation device that estimates features used when classifying medical receipt data, for example. [Background technology]

[0002] As a conventional technique, there is a method for analyzing temporal fluctuations based on feature quantities for time-series data that changes over time.

[0003] Patent Document 1 discloses an explanatory information generating device for time series patterns. This explanatory information generating device includes a prediction model learning unit that configures a prediction model by learning time series data into a neural network, a sequence pattern candidate generating unit that uses the prediction model to extract sequence pattern candidates that indicate characteristic changes included in the time series data, a sequence pattern generating unit that classifies a plurality of sequence pattern candidates by calculating dissimilarity between them and outputs representative sequence pattern candidates included in each classification as sequence patterns of the time series data, and a time series data analyzing unit that calculates feature amounts from any sequence pattern extracted from the time series data, compares the feature amounts calculated in advance from sequence patterns under normal conditions with a change detection model that stores the feature amounts, and outputs an analysis result as to whether the monitored system is normal or not.

[0004] Patent Document 2 discloses a diagnostic system. An equipment information classification unit obtains combined equipment information by combining fixed first combined equipment information and variable second combined equipment information associated with a monitoring item including basic equipment information of a monitored equipment with basic equipment information. The equipment information classification unit classifies the combined equipment information. An equipment status diagnosis unit obtains a feature amount of an analysis result obtained by analyzing measurement data corresponding to the monitoring item. The equipment status diagnosis unit diagnoses the status of the monitored equipment using a diagnostic pattern of the feature amount of the analysis result corresponding to the classification result of the equipment information classification unit. The equipment information classification unit obtains the feature amount of the combined equipment information. The equipment information classification unit classifies the combined equipment information into one of a plurality of pre-stored equipment feature patterns based on the feature amount of the combined equipment information. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Patent Publication No. 2021-124819 [Patent Document 2] JP 2016-91378 A Summary of the Invention [Problem to be solved by the invention]

[0006] An example of time-series data is medical receipt data showing the progress of medical treatment given to a subject by a medical institution. In this case, the subject is an injured or sick person. However, medical receipt data only indicates whether or not a specific service has been received at a given time, and in order for medical professionals to understand the subject's condition, they need to extract characteristics from the relationships between missing data within the medical receipt data, which requires a significant amount of effort. An object of the present invention is to provide a feature estimation device, a feature estimation method, and a classification system that generate feature estimation rules for extracting features from time-series data and assist in the process of classifying objects to be classified. [Means for solving the problem]

[0007] In order to solve the above problems, the present invention provides a feature estimation device including a feature estimation rule generation unit that generates feature estimation rules as rules for estimating features based on relationships between data, an estimated feature candidate creation unit that applies the feature estimation rules to time-series data to be classified and creates estimated feature candidates that are candidates for features, a feature evaluation unit that evaluates the classification performance of the created estimated feature candidates for the time-series data, and a feature estimation rule addition unit that adds a feature estimation rule with a changed evaluation value to a feature estimation rule corresponding to an estimated feature candidate that needs to be redefined as a result of the evaluation.In this case, it is possible to provide a feature estimation device that generates feature estimation rules for extracting features from time-series data and supports processing for classifying objects to be classified.

[0008] Here, the feature estimation rule can include a combination of words that are elements of an estimated feature candidate and an evaluation value for the combination of words. In this case, the rule becomes more appropriate as a rule for estimating a feature. Furthermore, the feature estimation rule generation unit can generate feature estimation rules based on relevance information that indicates the relevance of each piece of data, feature combination element information that indicates a combination of words, and feature estimation rule element information that indicates candidates for evaluation values. In this case, feature estimation rules can be generated more easily. Furthermore, time-series data can be converted into medical receipt data, in which case patients can be classified. Furthermore, the candidate estimated features may be used to classify medical receipt data and classify patients into a plurality of groups, in which patients have similar health conditions. Furthermore, the feature quantity estimation rule generating unit can generate feature quantity estimation rules based on the relationships between data regarding guidelines for injuries and illnesses, making it possible to generate more appropriate feature quantity estimation rules. Furthermore, the estimated feature candidate creation unit can extract estimated feature candidates for time-series data that have a predetermined common element, thereby making it possible to extract more appropriate estimated feature candidates. Furthermore, the estimated feature candidate creation unit can repeat the process of extracting estimated feature candidates for multiple common elements to add estimated feature candidates, thereby making it possible to extract more appropriate estimated feature candidates for various elements. The feature evaluation unit can then evaluate whether the candidate estimated features function to classify the time-series data, thereby eliminating features that are not suitable as features. The system may further include a display processing unit that displays to a user at least one of the feature estimation rule, the estimated feature candidate, and the feature evaluation rule for evaluating the classification performance of the estimated feature candidate, and allows the user to select one. In this case, the system may allow the user to select a more suitable one of the feature estimation rule, the estimated feature candidate, and the feature evaluation rule. Furthermore, features for classifying time-series data can be determined based on the feature estimation rule, estimated feature candidates, and feature evaluation rule selected by the user. In this case, features with higher accuracy can be extracted. Furthermore, the feature amount estimation rule generation unit can add a feature amount estimation rule corresponding to an estimated feature amount candidate created in the past, whereby a more useful feature amount estimation rule can be added. The feature estimation rule generation unit can select a feature estimation rule to be added based on the estimated feature candidates created in the past and the usage rate at which the estimated feature candidates were used by users. In this case, a feature estimation rule can be added based on the track record of past use.

[0009] The present invention also provides a feature estimation method in which a processor executes a program recorded in a memory to generate feature estimation rules as rules for estimating features based on relationships between data, apply the feature estimation rules to time-series data to be classified, create estimated feature candidates that are candidates for features, evaluate the classification performance of the created estimated feature candidates for the time-series data, and add a feature estimation rule with a changed evaluation value to a feature estimation rule corresponding to an estimated feature candidate that needs to be redefined as a result of the evaluation.In this case, it is possible to provide a feature estimation method that generates feature estimation rules for extracting features from time-series data and supports processing for classifying objects to be classified.

[0010] Furthermore, the present invention provides a feature estimation device that creates estimated feature candidates, which are candidates for features, in order to classify time-series data to be classified, and a data classification device that classifies the time-series data by feature values ​​determined based on the estimated feature candidates, wherein the feature estimation device includes a feature estimation rule generation unit that generates feature estimation rules as rules for estimating feature values ​​based on a relationship between each piece of data; The classification system includes an estimated feature candidate creation unit that applies feature estimation rules to time-series data to create estimated feature candidates, a feature evaluation unit that evaluates the classification performance of the created estimated feature candidates for time-series data, and a feature estimation rule addition unit that adds a feature estimation rule with a changed evaluation value to a feature estimation rule corresponding to an estimated feature candidate that needs to be redefined as a result of the evaluation.In this case, it is possible to provide a feature estimation system that generates feature estimation rules for extracting features from time-series data and supports the process of classifying objects to be classified. Effect of the Invention

[0011] According to the present invention, it is possible to provide a feature estimation device, a feature estimation method, and a classification system that generate feature estimation rules for extracting features from time-series data and assist in the process of classifying objects to be classified. [Brief description of the drawings]

[0012] [Figure 1] 1 is a block diagram showing the overall configuration of a classification system according to an embodiment of the present invention. [Diagram 2] FIG. 1 is a block diagram showing a functional configuration of a data relationship extraction system. [Diagram 3] FIG. 2 is a block diagram showing a functional configuration of a feature amount estimation rule management system. [Figure 4] FIG. 2 is a block diagram showing a functional configuration of the data classification system. [Diagram 5] FIG. 2 is a block diagram showing a functional configuration of the data management system. [Figure 6] FIG. 13 is a diagram showing a relationship table between data created based on guidelines for injuries and illnesses. [Figure 7] FIG. 13 is a diagram illustrating feature amount estimation rules as a feature amount estimation rule table. [Figure 8] FIG. 13 illustrates a feature amount combination element table. [Figure 9] FIG. 13 is a diagram showing a feature amount estimation rule element table. [Figure 10] FIG. 13 is a diagram showing medical receipt data. [Figure 11] 13 is a diagram showing the result of the feature generating unit selecting prescription data having a specific ID. FIG. [Figure 12] 1A is a diagram showing an estimated feature candidate list of estimated feature candidates created by a feature generating unit, and FIG. 1B is a diagram showing feature evaluation rules used by a feature evaluating unit as a feature evaluation rule table. [Figure 13] FIG. 13 is a diagram showing a statistical processing result table. [Figure 14] 1A shows a feature estimation rule table to which a feature estimation rule with a changed evaluation value has been added by the feature generation unit, and FIG. 1B shows an estimated feature candidate list to which an estimated feature candidate has been added according to the added feature estimation rule. [Figure 15]1A is a diagram illustrating a generated feature utilization history in the form of a generated feature utilization history table, and FIG. 1B is a diagram illustrating a history utilization rule in the form of a history utilization rule table. [Figure 16] 1 is a flowchart illustrating a process performed by the classification system. [Figure 17] 17 is a flowchart illustrating in more detail the process of generating feature amount estimation rules in S602 of FIG. 16. [Figure 18] FIG. 13 is a diagram showing a main screen. [Figure 19] FIG. 13 is a diagram showing a registration screen for data relationship information and feature quantity estimation rule information. [Figure 20] FIG. 13 is a diagram showing a detailed rule setting screen. [Figure 21] FIG. 13 is a diagram showing a feature amount estimation rule presentation screen. [Figure 22] FIG. 13 is a diagram showing a screen for presenting estimated feature amounts. [Diagram 23] FIG. 13 is a diagram showing a screen for presenting a classification result. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0013] Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. <Overall explanation of classification system S> FIG. 1 is a block diagram showing the overall configuration of a classification system S of this embodiment. 1, the classification system S of this embodiment includes a data relationship extraction system 1, a feature amount estimation rule management system 2, a network 3, a data classification system 4, and a data management system 5. The data relationship extraction system 1, the feature amount estimation rule management system 2, the data classification system 4, and the data management system 5 are connected to each other via the network 3.

[0014] The data relationship extraction system 1, the feature estimation rule management system 2, the data classification system 4, and the data management system 5 are computer devices such as a server device, a PC (Personal Computer), a smartphone, and a tablet. The data relationship extraction system 1, the feature estimation rule management system 2, the data classification system, and the data management system 5 each include a processor such as a CPU (Central Processing Unit) as a computing means, and a storage such as a main memory and an HDD (Hard Disk Drive) as a storage means. Here, the processor executes various software such as an OS (operating system) and an application program (application software). The main memory is a storage area that stores various software and data used for executing the software. The storage is a storage area that stores input data for various software and output data from various software. In addition, the data relationship extraction system 1, the feature estimation rule management system 2, the data classification system, and the data management system 5 are each equipped with a communication interface for communicating with the outside, a display mechanism consisting of a video memory, a display, etc., and an input mechanism such as a keyboard, a mouse, and a touch panel.

[0015] The network 3 is a communication means used for information communication between the data relationship extraction system 1, the feature estimation rule management system 2, the data classification system 4, and the data management system 5, and is, for example, the Internet, a LAN (Local Area Network), or a WAN (Wide Area Network). The communication lines used for information communication may be wired or wireless, or a combination of these may be used. Furthermore, the data relationship extraction system 1, the feature estimation rule management system 2, the data classification system 4, and the data management system 5 may be connected via multiple networks or communication lines using relay devices such as gateway devices or routers.

[0016] FIG. 2 is a block diagram showing the functional configuration of the data relationship extraction system 1. As shown in FIG. The data relationship extraction system 1 extracts relationships between data included in a predetermined extraction target. For example, the data relationship extraction system 1 uses a natural language analysis method to divide a sentence into words (nouns, verbs, adjectives, etc.) and extracts relationships between data using text mining, which analyzes the frequency of occurrence and correlation between the words.

[0017] The illustrated data relationship extraction system 1 includes a data acquisition processing unit 10, a document reading processing unit 11, a data extraction processing unit 12, a UI (User Interface) display processing unit 13, and a data relationship storage unit 15.

[0018] The data acquisition processing unit 10 acquires data of a predetermined extraction target. In this embodiment, the extraction target is a guideline for an injury or illness. That is, it is a manual that serves as a standard for examining and treating a subject (injured or sick person). The guideline for an injury or illness describes, for example, the pathology of the injury or illness, diagnostic criteria, medical purpose, items to be asked during the examination, items to be examined, timing of examinations, treatment policy, flow of treatment, names of drugs used in the treatment, dosage, and timing of administration. The guideline for an injury or illness may be published by an academic society dealing with each injury or illness. In the case of cancer, the academic society may be the Japanese Society of Clinical Oncology, in the case of cardiovascular disease, the Japanese Circulation Society, and in the case of diabetes, the Japanese Society of Diabetes and Lifestyle-Related Disease Human Data. The actual name of the guideline for an injury or illness is not particularly limited, and it may be called a treatment guideline, clinical practice guideline, treatment manual, or the like.

[0019] When the data acquired by the data acquisition processing unit 10 is, for example, image data read by a scanner or the like, the document reading processing unit 11 recognizes the text portion of the image data and converts it into character data. The data extraction processing unit 12 extracts relationships between data included in the extraction target such as guidelines for injuries and illnesses. The specific contents of this will be described later. The UI display processing unit 13 displays the relationships between the data extracted by the data extraction processing unit 12 to the user. The data relationship storage unit 15 stores the data relationships extracted by the data extraction processing unit 12 .

[0020] FIG. 3 is a block diagram showing the functional configuration of the feature amount estimation rule management system 2. As shown in FIG. The feature estimation rule management system 2 is an example of a feature estimation device, and generates feature estimation rules based on the relationships between the data extracted by the relationship extraction system 1. Here, a "feature" is a phrase or a numerical value that serves as a clue for classification in the time series data to be classified. In this case, a "phrase" is, for example, a word. However, it is not limited to this, and may be a collocation, phrase, or sentence that is a combination of words. Examples of actual feature amounts will be described later. Also, a "feature estimation rule" is a rule for estimating a feature amount. In this embodiment, the feature estimation rule management system 2 is characterized in that it generates this rule automatically, so to speak. Examples of actual feature estimation rules will be described later. Then, the feature estimation rule management system 2 creates estimated feature candidates for classifying time-series data based on the feature estimation rules, evaluates the classification performance of the estimated feature candidates, and for estimated feature candidates with low evaluations, reviews the feature estimation rules used when creating these estimated feature candidates.

[0021] The feature estimation rule management system 2 shown in the figure includes a feature estimation rule generation unit 20, a feature evaluation unit 21, a feature generation unit 22, a UI display processing unit 23, a data acquisition processing unit 24, a feature estimation rule storage unit 25, a feature candidate storage unit 26, and a feature evaluation rule storage unit 27.

[0022] The feature amount estimation rule generating unit 20 generates feature amount estimation rules as rules for estimating feature amounts based on the relationships between each data. The feature evaluation unit 21 evaluates the classification performance for time-series data of the estimated feature candidates generated by the feature generation unit 22. The feature evaluation unit 21 evaluates the classification performance of the estimated feature candidates according to feature evaluation rules set in advance by the user. The feature generation unit 22 is an example of an estimated feature candidate creation unit, and applies feature estimation rules to time-series data to be classified to create estimated feature candidates that are feature candidates. The feature generation unit 22 is also an example of a feature estimation rule addition unit, and adds feature estimation rules with modified evaluation values ​​to feature estimation rules corresponding to estimated feature candidates that need to be redefined as a result of the evaluation by the feature evaluation unit 21.

[0023] The UI display processing unit 23 is an example of a display processing unit, and displays the estimated feature candidates created by the feature generating unit 22 to the user. As will be described in detail later, the user can select the estimated feature candidates. In other words, the estimated feature candidates are presented to the user, and the user selects those that are determined to be usable as features for actually classifying the classification target. As will be described in detail later, the UI display processing unit 23 also displays the estimated feature candidates and feature evaluation rules for evaluating the classification performance of the estimated feature candidates to the user, and allows the user to select them. Then, based on the feature estimation rule, estimated feature candidates, and feature evaluation rule selected by the user, the feature for classifying the time-series data is determined. The data acquisition processing unit 24 acquires information on the relationships between data extracted by the data relationship extraction system 1 as data necessary for creating feature amount estimation rules. The feature amount estimation rule storage unit 25 stores the feature amount estimation rules generated by the feature amount estimation rule generation unit 20 . The feature candidate storage unit 26 stores the estimated feature candidates generated by the feature generation unit 22 . The feature amount evaluation rule storage unit 27 stores the feature amount evaluation rules used in the feature amount evaluation unit 21 .

[0024] FIG. 4 is a block diagram showing the functional configuration of data classification system 4. The data classification system 4 is an example of a data classification device that classifies time series data by features determined based on estimated feature candidates. The data classification system 4 actually classifies the time series data to be classified. Specifically, the data classification system 4 classifies medical receipt data as the time series data to be classified. As described above, medical receipt data indicates the progress of medical treatment given to a subject by a medical institution. Specifically, the medical receipt data includes hospital inpatient information when an injured or sick person is hospitalized, outpatient information of a hospital or clinic when an injured or sick person is outpatient, patient information of a dental clinic, and information of medicine when a medicine is received at a pharmacy. The medical receipt data allows the contents of medical examinations, tests, treatments, etc. for each injured or sick person and the profits of the medical institution to be understood. The illustrated data classification system 4 includes a data processing acquisition unit 40, a data classification processing unit 41, a data statistical processing unit 42, and a data processing storage unit 43.

[0025] The data processing and acquiring unit 40 acquires time-series data to be classified. Specifically, the data processing and acquiring unit 40 acquires medical receipt data. The data classification processing unit 41 classifies the time-series data to be classified. That is, in this embodiment, the medical receipt data is classified based on the feature amount. As a result, it is possible to classify a plurality of subjects (in this case, a plurality of sick or injured persons, a group of subjects) and divide the subjects (in this case, sick or injured persons) having similar health conditions into a plurality of groups. The data statistical processing unit 42 performs statistical processing for each group classified by the data classification processing unit 41 . The data processing storage unit 43 stores the results of the statistical processing performed by the data statistical processing unit .

[0026] FIG. 5 is a block diagram showing the functional configuration of the data management system 5. As shown in FIG. The data management system 5 manages feature estimation rules, estimated feature candidates, and feature amounts that have been estimated in the past. The data management system 5 shown in the figure manages the data obtained by the relationship extraction system 1, the feature estimation rule management system 2, and the data classification system 4. The illustrated data management system 5 includes a data processing acquisition unit 50, a utilization history processing unit 51, a UI display processing unit 52, a history utilization rule storage unit 53, and a processing history storage unit .

[0027] The data processing acquisition unit 50 acquires data obtained by the relationship extraction system 1, the feature estimation rule management system 2, and the data classification system 4. That is, the data processing acquisition unit 50 acquires information about the associations between data included in a predetermined extraction target (in this case, guidelines for injuries and illnesses) from the relationship extraction system 1. The data processing acquisition unit 50 also acquires feature estimation rules, estimated feature candidates, and features selected by the user obtained by the feature estimation rule management system 2. Furthermore, the data processing acquisition unit 50 acquires information on classified groups and information on statistical processing by the data classification system 4. The utilization history processing unit 51 obtains the usage ratio of the estimated feature candidate from the past history of actual usage. The UI display processing unit 52 displays the data acquired by the data processing acquisition unit 50 and the usage ratio of the estimated feature candidates calculated by the utilization history processing unit 51 to the user. The history utilization rule storage unit 53 stores rules that determine the usage ratio of estimated feature candidates, and that indicate whether or not the estimated feature candidates have actually been utilized. The processing history storage unit 54 stores the history of processing performed by the data management system 5.

[0028] <Detailed explanation of data relationship extraction system 1> Next, a detailed description will be given of the processing performed by the data relationship extraction system 1. The data relationship extraction system 1 creates a data relationship table as an example of relationship information that indicates the relationship between each data included in the predetermined extraction target. Here, a description will be given by taking an example of a guideline for injury or illness as the predetermined extraction target.

[0029] FIG. 6 shows a table of relationships between data created based on guidelines for injuries and illnesses. The illustrated data relationship table includes the following items: ID 100, illness name 101, related item name 102, condition 103, recovery 104, series dependency 105, dosage 106, and reissue period 107.

[0030] ID100 is an ID given to each relationship between data. The injury / disease name 101 is the name of the disease or injury. Here, the data relationship table in FIG. 6 shows that the data relationship regarding diabetes has been extracted. The related item name 102 indicates the name of an item related to the disease or injury of the injury / illness name 101. Here, it is shown that the drugs biguanoid drugs, DPP-4 inhibitors, and SGLT2 inhibitors have been extracted as related items for diabetes. In this case, biguanoid drugs, DPP-4 inhibitors, and SGLT2 inhibitors are drugs for treating diabetes. Similarly, it is shown that the re-examination premium has been extracted as a related item for diabetes. The condition 103 indicates a method of dealing with the disease or injury of the injury / illness name 101. Here, it indicates that medication is to be administered for diabetes. The complete recovery 104 indicates whether or not the disease or injury of the injury or illness name 101 will be completely cured. When the complete recovery 104 is 0, it will not be completely cured. On the other hand, when the complete recovery 104 is 1, it indicates that it may be completely cured. In this case, the complete recovery 104 is 0, which indicates that diabetes will not be completely cured. The sequence dependency 105 indicates the order of the related item names 102. As described above, biguanoid drugs, SGLT2 inhibitors, and DPP-4 inhibitors are drugs for diabetes. For example, the fact that two other drugs are listed in the "none" field of the sequence dependency 105 for biguanoid drugs indicates that there is no order in which these drugs are administered. If there is an order in which these drugs are administered, the drug that is administered before the biguanoid drug is entered in the "before" field. Also, the drug that is administered after the biguanoid drug is entered in the "after" field. The dosage 106 indicates the dosage of a drug used to treat the disease or injury indicated by the injury or illness name 101. For example, the dosage of a biguanoid drug used to treat diabetes is 500 mg / day. The recurrence period 107 is a period for evaluating whether or not the injury or illness has recurred. Here, 3 months is extracted as the complete recovery 104. The association table shown in the figure is extracted from guidelines for injuries and illnesses, and can be said to be information that shows the association between diseases or injuries and treatment methods and numerical values ​​such as dosage and duration corresponding to the treatment methods.

[0031] <Details of the Feature Estimation Rule Management System 2> Next, the process performed by the feature estimation rule management system 2 will be described in detail. Here, the processes performed by the feature estimation rule management system 2 according to the first and second embodiments will be described.

[0032] [First embodiment] In the first embodiment, a case will be described in which the feature quantity estimation rule management system 2 generates feature quantity estimation rules and creates estimated feature quantity candidates in a state in which there is no past history of estimated feature quantity candidates.

[0033] First, a method in which the feature quantity estimation rule generating unit 20 of the feature quantity estimation rule management system 2 generates feature quantity estimation rules will be described. The feature quantity estimation rule generation unit 20 creates a feature quantity combination element table and a feature quantity estimation rule element table based on the data relationship table extracted by the data relationship extraction system 1. Then, the feature quantity estimation rule generation unit 20 generates feature quantity estimation rules based on the feature quantity combination element table and the feature quantity estimation rule element table. This will be described in detail below.

[0034] FIG. 7 is a diagram showing feature amount estimation rules as a feature amount estimation rule table. As described above, the "feature estimation rule" is a rule for estimating a feature. The illustrated feature estimation rule table includes the following items: ID 110, feature 111, evaluation value 112, evaluation granularity 113, target period 114, feature combination element ID 115, and feature estimation rule element ID 116.

[0035] The ID 110 is an ID given to each feature estimation rule. Feature 111 indicates the content of the feature. As described above, a "feature" is a word or a numerical value that serves as a clue for classification in the time series data to be classified. Therefore, feature 111 is expressed as one or more combinations of words or numerical values ​​that serve as clues for classification. For example, when ID 110 is 1, this means that the feature is for one word, injury or illness. Also, when ID 110 is 2, injury or illness_relapse period_continuation of treatment section, this means that the feature is for a combination of three words, injury or illness, relapse period, and continuation of treatment section. The evaluation value 112 indicates the content when the feature amount 111 is evaluated as a numerical value. Moreover, the evaluation granularity 113 is a numerical unit or a numerical range of the evaluation value 112. For example, in the case of an injury or illness whose ID 110 is 1, the evaluation value 112 is the presence or absence of an injury or illness, and the evaluation granularity 113 is 0 / 1 (0 is no occurrence of an injury or illness, and 1 is occurrence of an injury or illness). In the case of an injury or illness_relapse period_continuation of treatment section whose ID 110 is 2, the evaluation value 112 is the continuation of treatment within the relapse period for the injury or illness, and the evaluation granularity 113 is in months. The target period 114 indicates the period to be classified. In this case, the entire period means the entire period for which the medical receipt data exists. The feature amount combination element ID 115 is an ID of a feature amount combination element table used to generate a feature amount estimation rule table, and means ID 120 described in FIG. The feature amount estimation rule element ID 116 is an ID of a feature amount estimation rule element table used to generate a feature amount estimation rule table, and means ID 130 described in FIG.

[0036] FIG. 8 is a diagram showing a feature combination element table. As described above, a "feature" is a word or a numerical value that serves as a clue for classification in the time-series data to be classified. The feature combination element table is a table that extracts combinations of words that serve as clues for classification. The feature combination element table is an example of feature combination element information that represents combinations of words. The illustrated feature combination element table includes the following items: ID 120, main axis feature 121, and combination candidate 122.

[0037] The ID 120 is an ID given to each combination of feature amounts. The main axis feature 121 and the combination candidates 122 are combinations of features, and are created by dividing the main axis feature 121, which is a feature that serves as the main axis, into the corresponding combination candidates 122. For example, when the main axis feature 121 with ID 1 is an injury or illness, the combination candidates 122 to be combined with this as a feature are, for example, related items, condition, complete recovery, dosage, and the like.

[0038] FIG. 9 is a diagram showing a feature estimation rule element table. The feature amount estimation rule element table is a table in which numerical values ​​and numerical ranges serving as clues for classification are extracted. The feature amount estimation rule element table is an example of feature amount estimation rule element information representing candidates for evaluation values. The illustrated feature estimation rule element table includes the following items: ID 130 , evaluation granularity 131 , granularity extension 132 , target period 133 , period extension 134 , extension source 135 , evaluation method 136 , and details 137 .

[0039] ID 130 is an ID given to each numerical value or numerical range that serves as a clue for classification. The evaluation granularity 131 indicates the size and fineness of a numerical value or a numerical range. For example, when the evaluation granularity 131 is 0 / 1, it indicates that the numerical value is 0 or 1. When the evaluation granularity 131 is in months, it indicates that the numerical value is in months. The granularity extension 132 is a criterion for extending the evaluation granularity 131. For example, when the granularity extension 132 is [multiple of 10], it means that the evaluation granularity 131 can be extended by a multiple of 10. The target period 133 indicates the period that is the target of the classification. The period extension 134 is a criterion for extending the target period 133. For example, if the period extension 134 is [all recent years, specific event period], it means that the target period 133 can be extended until the end of the most recent year or until within a specific event period. The extension source 135 is an ID 130 in which the evaluation granularity and the target period that are the extension source are described when granularity extension and period extension are performed. Evaluation method 136 is an evaluation method for the evaluation granularity, and details 137 are details of the evaluation method. For example, when evaluation granularity 131 is 0 / 1, the occurrence of a flag once or more within a period becomes evaluation method 136. Specifically, as shown in details 137, a flag is determined based on the presence or absence of an injury or illness, and when an injury or illness occurs once or more within a target period, the flag becomes 1, and when an injury or illness does not occur within a target period, the flag becomes 0. Also, when evaluation granularity 131 is in months, the number of months that the flag continues within a period becomes evaluation method 136. Specifically, as shown in details 137, it is the duration of the occurrence of the flag within a period, and is determined based on how many months the injury or illness has continued to occur within the target period.

[0040] The feature estimation rule generation unit 20 creates a feature combination element table shown in FIG. 8 and a feature estimation rule element table shown in FIG. 9 from the data relationship table. Furthermore, the feature estimation rule generation unit 20 generates a feature estimation rule table shown in FIG. 7 from the feature combination element table and the feature estimation rule element table. The feature estimation rules are phrases and values ​​that serve as clues for classification, and among them, phrases that serve as clues for classification are obtained from the feature combination element table in FIG. 8. Specifically, the feature 111 in FIG. 7 is a phrase combining the main feature 121 and the combination candidate 122 in FIG. 8. Furthermore, the numerical values ​​and numerical ranges that serve as clues for classification are obtained from the feature estimation rule element table in FIG. 9. Specifically, the evaluation value 112 in FIG. 7 is the details 137 in FIG. 9. Furthermore, the evaluation granularity 113 in FIG. 7 is the evaluation granularity 131 in FIG. 9. Furthermore, the target period 114 in FIG. 7 is the target period 133 in FIG. 9. As a result, the feature estimation rule includes word combinations and evaluation values ​​for the word combinations, which are elements of estimated feature candidates.

[0041] Next, a method in which the feature quantity generating unit 22 of the feature quantity estimation rule management system 2 generates estimated feature quantity candidates, which are feature quantity candidates, will be described. The feature quantity generation unit 22 applies the feature quantity estimation rules generated by the feature quantity estimation rule generation unit 20 to the medical receipt data to be classified, thereby generating estimated feature quantity candidates. Specifically, the feature generator 22 generates estimated feature candidates as follows.

[0042] FIG. 10 shows the medical receipt data. The illustrated receipt data consists of the following items: ID 200, diabetes mellitus 201, re-examination surcharge 202, biguanoid drug 203, DPP-4 inhibitor 204, calcium antagonist 205, and Date 206. The ID 200 is an ID given to each piece of medical receipt data. Here, there are three types of ID 200, A to C. Diabetes 201 indicates whether or not the patient has diabetes. When Diabetes 201 is 1, it means that the patient has diabetes, and when it is 0, it means that the patient does not have diabetes. The re-examination surcharge 202 indicates whether or not the re-examination surcharge is applicable. When the re-examination surcharge 202 is 1, it means that the re-examination surcharge is applicable, and when it is 0, it means that the re-examination surcharge is not applicable. Biguanoid drug 203 indicates whether or not a biguanoid drug has been administered. When biguanoid drug 203 is 1, it means that biguanoid drug 203 has been administered, and when it is 0, it means that biguanoid drug 203 has not been administered. The DPP-4 inhibitor 204 indicates whether or not a DPP-4 inhibitor has been administered. When the DPP-4 inhibitor 204 is 1, it means that the DPP-4 inhibitor 204 has been administered, and when it is 0, it means that the DPP-4 inhibitor 204 has not been administered. Calcium antagonist 205 indicates whether or not a calcium antagonist has been administered. When calcium antagonist 205 is 1, it means that a calcium antagonist 205 has been administered, and when it is 0, it means that a calcium antagonist 205 has not been administered. Calcium antagonist 205 is, for example, a drug prescribed for a hypertensive patient complicated with diabetes. Date 206 is the date on which the medical receipt data was created.

[0043] In the illustrated receipt data, when ID200 is A or C, it is diabetes, and when ID200 is B, it is not diabetes. And when ID200 is B, none of the re-examination surcharge 202, biguanoid drug 203, DPP-4 inhibitor 204, and calcium antagonist 205 is administered. When ID200 is A, biguanoid drug 203 and DPP-4 inhibitor 204 are not administered, but calcium antagonist 205 is administered. On the other hand, when ID200 is C, biguanoid drug 203 and DPP-4 inhibitor 204 are administered, but calcium antagonist 205 is not administered. Therefore, ID200 is assigned to the receipt data that has a common predetermined element.

[0044] The feature generating unit 22 selects medical receipt data having a specific ID. FIG. 11 is a diagram showing the result of the feature generating unit 22 selecting medical receipt data having a specific ID. Here, the case where ID 200 is selected as A is shown. Next, the feature generation unit 22 applies the feature estimation rule to the selected medical receipt data to generate estimated feature candidates. Then, estimated feature candidates are generated for other IDs 200. That is, the IDs 200 B and C are selected in sequence, and estimated feature candidates are added. In this case, the feature generation unit 22 extracts estimated feature candidates for prescription data that have a predetermined common element (in this case, the ID), and further repeats the process of extracting estimated feature candidates for multiple common elements, thereby adding estimated feature candidates.

[0045] FIG. 12A is a diagram showing the estimated feature quantity candidates generated by the feature quantity generating unit 22 as an estimated feature quantity candidate list. The illustrated estimated feature candidate list includes the following items: ID 300, feature 301, and estimation rule ID 302. ID 300 is an ID given to each of the estimated feature candidates. The feature amount 301 is a candidate for an estimated feature amount. The estimation rule ID 302 is the ID 110 (see FIG. 7) of the feature estimation rule used to generate the estimated feature candidate.

[0046] For example, the estimated feature candidate with ID 300 of 1 is "diabetes_1_all time", which indicates that it was created using the feature estimation rule with ID 110 of 1 in FIG. 7. In FIG. 7, the feature estimation rule with ID 110 of 1 has feature 111 which is injury or illness, so "diabetes" is extracted from "diabetes_1_all time" as an element of the estimated feature candidate to which this rule is applied. Similarly, evaluation value 112 is the presence or absence of occurrence, and evaluation granularity 113 is 0 / 1, so "1" is extracted from "diabetes_1_all time" as an element of the estimated feature candidate to which this rule is applied. Furthermore, target period 114 is the entire period, so "all time" is extracted from "diabetes_1_all time" as an element of the estimated feature candidate to which this rule is applied.

[0047] For example, the estimated feature candidate with ID 300 of 2 is "diabetes_within 3 months_duration of treatment_0 months_entire period", which indicates that it was created using the feature estimation rule with ID 110 of 2 in FIG. 7. In FIG. 7, the feature estimation rule with ID 110 of 2 has feature 111 of "injury / disease_recurrence period_duration of treatment", so "diabetes_within 3 months_duration of treatment" is extracted from "diabetes_within 3 months_duration of treatment_0 months_entire period" as an element of the estimated feature candidate to which this rule is applied. Similarly, the evaluation value 112 is the duration of recurrence within the recurrence period, and the evaluation granularity 113 is in months, so "0 months" is extracted from "diabetes_within 3 months_duration of treatment_0 months_entire period" as an element of the estimated feature candidate to which this rule is applied. Furthermore, since the target period 114 is the entire period, "entire period" is extracted from "diabetes_within 3 months_duration of treatment_0 months_entire period" as an element of an estimated feature candidate to which this is applied.

[0048] Next, a method in which the feature evaluation unit 21 of the feature estimation rule management system 2 evaluates the classification performance of the estimated feature candidates generated by the feature generation unit 22 using the feature evaluation rules will be described.

[0049] FIG. 12B is a diagram showing feature amount evaluation rules used by the feature amount evaluation unit 21 as a feature amount evaluation rule table. The illustrated feature amount evaluation rule table includes the following items: ID 400 , evaluation target condition 401 , evaluation exclusion 402 , threshold 403 , and evaluation target feature amount ID 404 . The ID 400 is an ID given to each feature amount evaluation rule. The evaluation target condition 401 is a condition to be evaluated. Evaluation exclusion 402 is a case where a condition is excluded from being evaluated. The threshold value 403 is a specific numerical value of a threshold value in the condition to be evaluated. The evaluation target feature amount ID 404 is the ID 110 (see FIG. 7) of the feature amount estimation rule to be evaluated.

[0050] For example, the feature evaluation rule with ID 400 being 1 is a "feature with an evaluation value of 0", which is applied to the estimated feature candidate created based on the feature estimation rule with ID 110 being 2 in FIG. 7. In FIG. 7, the feature estimation rule with ID 110 being 2 has feature 111 as "injury / illness_recurrence period_duration of treatment", and in this case, for example, if the duration of treatment is 0, this means that no treatment has been given and is meaningless. Therefore, such estimated feature candidates are excluded because they are not suitable as features. Note that in evaluation exclusion 402, evaluation granularity = 0 / 1, but when the evaluation granularity is 0 / 1, even 0 is meaningful, so this is excluded from this evaluation.

[0051] For example, the feature evaluation rule for ID 400=2 is "the occurrence ratio in the input data is equal to or greater than a threshold value," and estimated feature candidates with an occurrence ratio of 70% or greater as a threshold value are excluded. This means that it occurs regardless of the type of injury or illness, and for example, body temperature measurement falls under this category. In this case, it is excluded because it is not suitable as a feature. Furthermore, for example, the feature evaluation rule for ID 400=3 is "the occurrence ratio in the input data is less than a threshold value," and estimated feature candidates with an occurrence ratio less than the threshold value of 10 are excluded. This means that the occurrence is too low, and it is excluded because it is not suitable as a feature. In other words, the feature evaluation unit 21 evaluates whether or not the estimated feature candidate functions to classify the medical receipt data.

[0052] FIG. 13 is a diagram showing a statistical processing result table. When creating estimated feature candidates from feature estimation rules, the feature generation unit 22 performs statistical processing on the words and numbers that serve as clues for classification. The statistical processing result table shows the results of this statistical processing. This makes it possible to create more appropriate estimated feature candidates. The statistical processing result table includes the following items: ID 410 , evaluation value 411 , calculation target 412 , calculation value 413 , and calculated evaluation value 414 . The ID 410 is an ID given to each statistical processing result. Evaluation value 411 is the same as evaluation value 112 shown in FIG. 7, and indicates here that, among the evaluation values, statistical processing has been performed on the recurrence duration within the recurrence period. A calculation target 412 is a calculation target when performing statistical processing. Here, it is shown that the recurrence duration is the calculation target. The calculated value 413 is a numerical standard for the calculation target 412. Here, the evaluation granularity 113 of the recurrence duration is in months, so this indicates that statistical processing was performed every 3 to 6 months. If a matching patient is found as a result of statistical processing, that value is entered into the calculated evaluation value 414. In this case, it means that for a matching patient, when the ID is 1 and the playback period is 3 months or less, the recurrence duration within the playback period is 0 months, but when the ID is 2 and the playback period is 4 months or less, the recurrence duration within the playback period is 0, 4, 6, and 10 months.

[0053] Furthermore, the feature generation unit 22 adds a feature estimation rule with a changed evaluation value to the feature estimation rule corresponding to the estimated feature candidate that needs to be redefined as a result of the evaluation by the feature evaluation unit 21.

[0054] FIG. 14A shows a feature amount estimation rule table to which the feature amount generating unit 22 has added a feature amount estimation rule with a changed evaluation value. The illustrated feature amount estimation rule table includes items of ID 420, feature amount 421, evaluation value 422, evaluation granularity 423, and target period 424. These are the same as the ID 110, feature amount 111, evaluation value 112, evaluation granularity 113, and target period 114 in the feature amount estimation rule table shown in FIG. Comparing Fig. 14(a) with Fig. 7, it can be seen that a feature estimation rule with ID 420 of 5 has been added. In this case, the feature estimation rule with ID 420 (ID 110) of 2 may be inappropriate for creating estimated feature candidates as explained in Fig. 12(b), and this shows a case where a feature estimation rule with a redefined evaluation value has been added. Specifically, a feature estimation rule has been added in which the evaluation value has been changed from "reoccurrence duration within reoccurrence period" to "reoccurrence duration within (reoccurrence period + 1)".

[0055] FIG. 14B is a diagram showing an estimated feature quantity candidate list in which estimated feature quantity candidates have been added according to the added feature quantity estimation rule. The illustrated estimated feature candidate list is made up of items of ID 430, feature 431, and estimation rule ID 432. These are the same as the ID 300, feature 301, and estimation rule ID 302 in the feature estimation rule table shown in FIG. Comparing FIG. 14(b) with FIG. 12(a), it can be seen that an estimated feature candidate with ID 430 of 5 has been added. In this case, the estimated feature candidate with ID 430 (ID300) of 2 is redefined as "Diabetes_within 3 months_duration of treatment_0 months_entire period," and "Diabetes_within 4 months_duration of treatment_8 months_entire period" is added as an estimated feature candidate with ID 430 of 5.

[0056] [Second embodiment] In the second embodiment, a case will be described in which estimated feature quantity candidates have a past history, and the feature quantity estimation rule management system 2 generates feature quantity estimation rules and creates estimated feature quantity candidates.

[0057] Here, first, the utilization history processing unit 51 of the data management system 5 obtains the usage ratio of the estimated feature candidates from the past history of actual use. Then, the data acquisition processing unit 24 of the feature estimation rule management system 2 acquires the usage ratio of the estimated feature candidates. The usage ratio of the estimated feature candidates is the ratio selected by the user as a feature that can actually be used to classify the classification target. The feature estimation rule generating unit 20 obtains a generated feature utilization history based on the usage ratio of the feature candidates.

[0058] FIG. 15A is a diagram illustrating the generated feature value utilization history as a generated feature value utilization history table. The generated feature utilization history table shown in the figure includes the following items: ID 500, feature 501, corresponding estimation rule ID 502, and usage ratio during estimation 503. ID500 is an ID given to each of the estimated feature candidates. The feature amount 501 is a candidate for an estimated feature amount. The correspondence estimation rule ID 502 is the ID 110 (see FIG. 7) of the feature estimation rule used to generate the estimated feature candidate. The usage ratio 503 during estimation is the usage ratio of the feature candidate. Therefore, the generated feature utilization history table in FIG. 15(a) is obtained by adding the usage ratio 503 during estimation to the estimated feature candidate list in FIG. 12(a).

[0059] Next, the feature estimation rule generation unit 20 evaluates the estimated feature candidates using the history utilization rule, which is a rule for extracting estimated feature candidates with a high usage rate as those that are utilized. FIG. 15B is a diagram showing the history utilization rules as a history utilization rule table. The history utilization rule table shown in the figure includes the following items: ID 510, utilization rule 511, and threshold value 512. The ID 510 is an ID given to each history utilization rule. The utilization rule 511 is a condition that is the subject of evaluation as to whether or not it is utilized. Threshold 512 is the threshold of the condition to be evaluated. In this case, the history utilization rule with ID 510 set to 1 indicates that the usage rate at the time of estimation is equal to or greater than a threshold value as utilization rule 511, and this threshold value is 70% as determined by threshold value 512. In other words, when the usage rate of an estimated feature candidate is 70% or greater, feature estimation rule generation unit 20 determines that the corresponding estimated feature candidate is utilized.

[0060] Next, the feature amount estimation rule generating unit 20 enters the feature amount estimation rule corresponding to the estimated feature amount candidate that is determined to be utilized into a feature amount estimation rule table. The rest is the same as in the first embodiment. That is, similarly to the first embodiment, feature amount estimation rules are generated for the medical receipt data, and these feature amount estimation rules are further added.

[0061] FIG. 16 is a flowchart illustrating the processing performed by the classification system S. First, the data acquisition processing unit 24 acquires information on relationships between data from the data relationship extraction system 1 and registers it (S600). Next, the data acquisition processing unit 24 acquires the generated feature utilization history table and the history utilization rule table described with reference to FIG. 15 as past feature utilization history information (S601). Next, the feature quantity estimation rule generating unit 20 generates feature quantity estimation rules based on information about the relationships between data (S602). Furthermore, the data acquisition processing unit 24 acquires medical receipt data as input data (S603). Then, the feature generating unit 22 estimates candidates for estimated feature quantities from the relationships between data included in the medical receipt data, which is the input data (S604). Next, the feature generating unit 22 calculates the number of people to be classified for each estimated feature candidate (S605). Furthermore, the feature generating unit 22 judges whether there is any item in which the number of classification subjects is a value outside the threshold (S606). That is, the feature generating unit 22 judges whether each estimated feature candidate is suitable for classification based on the number of classification subjects of each estimated feature candidate. If the number of classification subjects is too small and falls outside the threshold, the estimated feature candidate is not suitable as a feature for classifying medical receipt data.

[0062] As a result, if there is no item for which the number of classification targets is outside the threshold (YES in S606), the UI display processing unit 23 presents to the user a list of relationships between the data used for the estimation and corresponding estimated feature candidates (S607). Then, the data acquisition processing unit 24 receives a selection of features to be used for classification from the user (S608). Furthermore, the feature generating unit 22 generates the selected feature based on the input data, and the data classification system 4 classifies the input data based on the feature (S609). Then, the data classification system 4 presents the classification result of the object group based on the selected feature amount to the user (S610). The object group corresponds to injured or sick people. In addition, in S606, if there is an item in which the number of classification target persons is a value outside the threshold (NO in S606), the feature amount estimation rule is additionally defined by referring to the corresponding feature amount (S611). After that, the process returns to S602.

[0063] FIG. 17 is a flowchart illustrating in more detail the process of generating feature quantity estimation rules in S602 of FIG. First, the feature amount estimation rule generation unit 20 acquires a feature amount combination element table and a feature amount estimation rule element table (S620). Next, the feature amount estimation rule generation unit 20 determines whether or not an additional definition of a feature amount estimation rule is required (S621). As a result, if additional definition of feature quantity estimation rules is not required (NO in S621), the feature quantity estimation rule generation unit 20 acquires a generated feature quantity utilization history table and a history utilization rule table (S622). Next, the feature amount estimation rule generation unit 20 determines whether or not there is an item that corresponds to the history utilization rule in the generated feature amount utilization history (S623). As a result, if there is no item that corresponds to the history utilization rule (YES in S623), the process proceeds to S625. This corresponds to the case described in the first embodiment. On the other hand, if there is an item that corresponds to the history utilization rule (NO in S623), the feature estimation rule generation unit 20 generates a feature estimation rule and adds it to the feature estimation rule table (S624). This corresponds to the case described in the second embodiment.

[0064] Furthermore, the feature amount estimation rule generation unit 20 repeats the processes of S626 to S629 below for the number of unique IDs included in the feature amount combination element table (S625). That is, the feature amount estimation rule generation unit 20 repeats for the number of combinations of the main axis feature 121 and the combination candidate 122 (see FIG. 8) (S626). Also, the feature amount estimation rule generation unit 20 repeats for the number of unique IDs in the feature amount estimation rule element table (S627). At this time, the feature amount estimation rule generation unit 20 generates a feature amount estimation rule for each combination of the combination of the main axis feature and 121 and the combination candidate 122 and the combination of the feature amount estimation rule elements (S628). Then, the feature amount estimation rule generation unit 20 adds the generated feature amount estimation rule to the feature amount estimation rule table (S629).

[0065] Note that in S621, when additional definition of the feature amount estimation rule is necessary (YES in S621), the feature amount estimation rule generation unit 20 acquires each data of the generation source of the corresponding feature amount estimation rule (S630). Next, the feature amount estimation rule generation unit 20 calculates the evaluation value of the input data for the extension candidate of the corresponding feature amount estimation rule (S631). Furthermore, the feature amount estimation rule generation unit 20 selects an extension candidate that satisfies the violated feature amount evaluation rule (S632). Then, the feature amount estimation rule generation unit 20 adds the selected extension candidate to the feature amount estimation rule table (S633).

[0066] <UI Explanation> Next, an explanation of the UI presented to the user is given. FIG. 18 is a diagram showing the main screen. In the illustrated main screen, a registration button 700 for relationship information and feature amount estimation rule information between data, a detailed rule setting button 701, an input file field 702, and a registration button 703 for the feature amount estimation rule are displayed. The registration button 700 for data relationship information / feature amount estimation rule information is pressed when registering a data relationship table (see FIG. 6), a feature amount combination table (see FIG. 8), and a feature amount estimation rule element table (see FIG. 9). A detailed rule setting button 701 is pressed when setting the feature amount evaluation rule table (see FIG. 12(b)) or the history utilization rule table (see FIG. 15(b)). When inputting medical receipt data as input data, the file name is input in the input file field 702. In this example, the file name Test.csv is input. A feature amount estimation rule registration button 703 is pressed when registering a feature amount estimation rule table (see FIG. 7).

[0067] FIG. 19 is a diagram showing a registration screen for data relationship information and feature quantity estimation rule information. The screen shown in the figure is displayed when a registration button 700 for registering data relationship information / feature amount estimation rule information in FIG. 18 is pressed. On this screen, an input file field 710, a data relationship information registration button 711, and a diabetes treatment guideline projection screen button 712 are displayed in the upper part. The file name of the data used to create the data relationship table (see FIG. 6) is input in the input file column 710. Here, it is shown that the file name of the diabetes treatment guideline has been input. A data relationship information registration button 711 is pressed when registering a data relationship table (see FIG. 6). A diabetes treatment guideline projection screen button 712 is pressed when it is desired to display the diabetes treatment guideline.

[0068] This screen also has display columns for data relationship information, feature combination information, and feature estimation rule element information in the middle and lower sections. Here, data relationship table 713, feature combination table 717, and feature estimation rule element table 721 described in Figs. 7 to 9 are displayed, respectively. Also, corresponding to these tables, add buttons 714, 718, 722, modify buttons 715, 719, 723, and delete buttons 716, 720, 724 are displayed. These are pressed when adding, modifying, or deleting from each table. When modifying or deleting, the radio button Rb provided adjacent to the ID column of each table can be selected to modify or delete each element of the corresponding ID. At the bottom of this screen is provided a registration completion button 725. When the registration completion button 725 is pressed, each displayed table is registered.

[0069] FIG. 20 is a diagram showing a detailed rule setting screen. The screen shown in the figure is displayed when a detailed rule setting button 701 in FIG. 18 is pressed. This screen has display columns for feature evaluation rule information and history utilization rule information. Here, the feature evaluation rule table 730 described in Fig. 12(b) and the history utilization rule table 735 described in Fig. 15(b) are displayed. Also displayed corresponding to these tables are add buttons 731 and 736, modify buttons 732 and 737, and delete buttons 733 and 738. The functions of these buttons and the radio buttons Rb provided adjacent to the ID columns of the respective tables are the same as those in FIG. At the bottom of this screen, a setting completion button 739 is provided. When the setting completion button 739 is pressed, each of the displayed tables is registered.

[0070] FIG. 21 is a diagram showing a feature amount estimation rule presentation screen. The screen shown in the figure is displayed when the feature estimation rule registration button 703 in FIG. 18 is pressed. This screen has a display field for feature amount estimation rules, in which the feature amount estimation rule table 740 described with reference to FIG. Also displayed corresponding to this table are an Add button 741, a Modify button 742, and a Delete button 743. The functions of these buttons and the radio buttons Rb provided adjacent to the ID columns of the respective tables are the same as in FIGS. In addition, on this screen, a feature quantity candidate generation button 745 is provided at the bottom. When the feature quantity candidate generation button 745 is pressed, estimated feature quantity candidates are generated according to the displayed feature quantity estimation rule.

[0071] FIG. 22 is a diagram showing a screen for presenting estimated features. The screen shown in the figure is displayed when the feature candidate generation button 745 in FIG. 21 is pressed. This screen has display columns for feature estimation rules and estimated feature candidates. Here, the feature estimation rule table 750 described in FIG. 7 is displayed in the upper part. Then, below the feature estimation rule table 750, a rule generation basis presentation button 751 is displayed, and when pressed, the basis for generating the feature estimation rule is presented. Here, the estimated feature candidate list 752 described in FIG. 14(b) is displayed in the lower part. Here, when the user selects the radio button Rb, this is selected as the feature. Then, below the estimated feature candidate list 752, a classification implementation button 753 is displayed, and when pressed, classification is performed on the prescription data specified in the input file column 702 in FIG. 18.

[0072] FIG. 23 is a diagram showing a screen for presenting the classification results. The screen shown in the figure is displayed when the classification implementation button 753 in FIG. 22 is pressed. This screen displays a classification result table 760 based on the selected feature. The classification result table 760 is a classification result when the features have IDs 1 and 4 in the estimated feature candidate list 752 in Fig. 22. That is, this is a classification result when "diabetes_1_entire period" and "diabetes_within 3 months_treatment duration_3 months_entire period" are used as features. A back button 761 is displayed below the classification result table 760, and when this button is pressed, the screen returns to, for example, the screen in FIG. 18 or FIG.

[0073] Receipt data only indicates whether or not a specific service was received at that time. In other words, while the receipt data includes the type and content of the specific service received by the injured or ill person (date and time, number of times, amount of medicine, etc.), the data does not necessarily include information showing the relationship between each piece of data. In order for medical staff to understand the condition of the injured or ill person, it is necessary to extract the relationship between the missing data in the receipt data as a feature, which requires a large amount of work. On the other hand, even if the guidelines for injuries and illnesses state that treatment should be continued for no more than three months, this can change depending on the characteristics of the injured or sick person and the doctor's judgment, such as every four months. When it comes to prescriptions, the first drug to be applied and the amount to be administered are basically in accordance with the relationships in the guidelines for injuries and illnesses, but the information showing these relationships is not absolute. Therefore, the relationships in the guidelines for injuries and illnesses cannot necessarily be applied directly to all data.

[0074] In the first embodiment, the feature estimation rule management system 2 automatically generates feature estimation rules for extracting the relationship between data missing in the medical receipt data as features, taking into account the characteristics of the medical receipt data. The feature estimation rule management system 2 also evaluates the classification performance of the estimated feature candidates created by the feature estimation rules, and generates an alternative feature estimation rule if there is a problem. The feature estimation rule can be selected by the user. Therefore, the feature estimation rule management system 2 can be said to be a device that supports the generation of feature estimation rules. Furthermore, the medical receipt data is classified based on the features selected by the user from the estimated feature candidates, and the injured or sick person in a similar condition is divided into a plurality of groups. In this embodiment, the estimated feature candidates are created to classify the medical receipt data and divide the injured or sick person into a plurality of groups, and the one selected by the user from these is actually used to classify the medical receipt data. Therefore, the feature estimation rule management system 2 can be said to be a device that supports the creation of features. Then, by dividing the injured or sick person into a plurality of groups, it can be used to understand the condition of each injured or sick person. It can also be used to estimate the type and dosage of medication appropriate for each patient.

[0075] In the second embodiment, the feature amount estimation rule generation unit 20 adds a feature amount estimation rule corresponding to an estimated feature amount candidate created in the past. At this time, the feature amount estimation rule generation unit 20 selects a feature amount estimation rule to be added based on the estimated feature amount candidates created in the past and the usage rate at which the estimated feature amount candidates were used by the user. This makes it possible to further add feature amount estimation rules that were generated and utilized in the past. This improves the accuracy of the feature amount estimation rules.

[0076] In the above detailed embodiment, the prescription data is used as the time series data, but the present invention is not limited thereto. The time series data may be, for example, signal data from a plurality of sensors provided in an industrial plant. In this case, the feature estimation rule management system 2 generates a feature estimation rule based on information on the association between a plurality of pieces of signal data. Then, the feature estimation rule management system 2 applies the feature estimation rule to the feature estimation rule to create an estimated feature candidate. Furthermore, the feature estimation rule management system 2 evaluates the classification performance for the signal data for the created estimated feature candidate. Then, the feature estimation rule management system 2 adds a feature estimation rule with a changed evaluation value to the feature estimation rule corresponding to the estimated feature candidate that needs to be redefined as a result of the evaluation. In this case, the sensor is not particularly limited, and may be, for example, a displacement sensor, a strain sensor, a gas sensor, a temperature sensor, a humidity sensor, a pressure sensor, a flowmeter, a voltmeter, an ammeter, or the like.

[0077] <Explanation of feature estimation method> The above-described processing performed by the feature quantity estimation rule management system 2 is realized by cooperation between software and hardware resources. That is, a processor within a computer provided in the feature quantity estimation rule management system 2 loads software for realizing each of the above-described functions into memory and executes it to realize each of these functions. Therefore, the processing performed by the feature estimation rule management system 2 can be considered as a feature estimation method in which a processor executes a program recorded in memory to generate feature estimation rules as rules for estimating features based on the relationships between each piece of data, applies the feature estimation rules to the time-series data to be classified, creates estimated feature candidates that are candidates for features, evaluates the classification performance of the created estimated feature candidates for the time-series data, and adds feature estimation rules with modified evaluation values ​​to feature estimation rules corresponding to estimated feature candidates that need to be redefined as a result of the evaluation.

[0078] Although the present embodiment has been described above, the technical scope of the present invention is not limited to the scope described in the above embodiment. It is clear from the claims that various modifications and improvements to the above embodiment are also included in the technical scope of the present invention. [Explanation of symbols]

[0079] 1...data relationship extraction system, 2...feature estimation rule management system, 3...network, 4...data classification system, 5...data management system, 20...feature estimation rule generation unit, 21...feature evaluation unit, 22...feature generation unit, 23...UI display processing unit, 24...data acquisition processing unit, 25...feature estimation rule storage unit, 26...feature candidate storage unit, 27...feature evaluation rule storage unit, 112...evaluation value, S...classification system

Claims

1. a feature estimation rule generation unit that generates a feature estimation rule as a rule for estimating a feature based on a relationship between each data; an estimated feature candidate creation unit that applies the feature estimation rule to time-series data to be classified and creates estimated feature candidate candidates that are feature candidates; a feature evaluation unit that evaluates classification performance of the generated estimated feature candidates with respect to the time-series data; a feature estimation rule addition unit that adds a feature estimation rule with a changed evaluation value to the feature estimation rules corresponding to the estimated feature candidates that need to be redefined as a result of the evaluation; A feature estimation device comprising:

2. The feature estimation device according to claim 1 , wherein the feature estimation rule includes a combination of words that is an element of the estimated feature candidate, and the evaluation value for the combination of words.

3. 3. The feature estimation device according to claim 2, wherein the feature estimation rule generation unit generates the feature estimation rules based on relevance information indicating a relevance of each piece of data, feature combination element information indicating a combination of the wordings, and feature estimation rule element information indicating candidates for the evaluation value.

4. The feature estimation device according to claim 1 , wherein the time-series data is medical receipt data.

5. The feature estimation device according to claim 4 , wherein the estimated feature candidates are used to classify the medical receipt data and divide patients into a plurality of groups.

6. The feature amount estimation device according to claim 4 , wherein the feature amount estimation rule generation unit generates the feature amount estimation rule based on a relationship between each piece of data regarding a guideline for an injury or disease.

7. The feature estimation device according to claim 1 , wherein the estimated feature candidate creation unit extracts the estimated feature candidates from the time series data that have a predetermined element in common.

8. The feature estimation device according to claim 7 , wherein the estimated feature candidate creation unit repeats a process of extracting the estimated feature candidates for a plurality of common elements, and adds the estimated feature candidates.

9. The feature estimation device according to claim 1 , wherein the feature evaluation unit evaluates whether the estimated feature candidates function to classify the time-series data.

10. 2. The feature estimation device according to claim 1, further comprising a display processing unit that displays at least one of the feature estimation rule, the estimated feature candidate, and a feature evaluation rule for evaluating classification performance of the estimated feature candidate to a user and allows the user to select one of the features.

11. The feature estimation device according to claim 10 , further comprising: determining features for classifying the time-series data based on the feature estimation rule, the estimated feature candidate, and the feature evaluation rule selected by a user.

12. The feature estimation device according to claim 1 , wherein the feature estimation rule generation unit adds a feature estimation rule that corresponds to the estimated feature candidate that was previously created.

13. The feature estimation device according to claim 12 , wherein the feature estimation rule generation unit selects the feature estimation rule to be added based on the estimated feature candidates created in the past and a usage rate at which the estimated feature candidates were used by a user.

14. The processor executes the program stored in the memory, Based on the relationships between each piece of data, feature estimation rules are generated to estimate the features. applying the feature estimation rule to the time series data to be classified, and generating estimated feature candidates which are feature candidates; evaluating classification performance of the generated candidate estimated feature quantities for the time-series data; adding a feature estimation rule with a changed evaluation value to the feature estimation rule corresponding to the estimated feature candidate that needs to be redefined as a result of the evaluation; Feature estimation methods.

15. A feature estimation device that generates estimated feature candidates, which are feature candidates, in order to classify time-series data to be classified; a data classification device that classifies the time series data by a feature determined based on the estimated feature candidate; having The feature estimation device comprises: a feature estimation rule generation unit that generates a feature estimation rule as a rule for estimating a feature based on a relationship between each data; an estimated feature candidate creation unit that applies the feature estimation rule to the time-series data to create the estimated feature candidate; a feature evaluation unit that evaluates classification performance of the generated estimated feature candidates with respect to the time-series data; a feature estimation rule addition unit that adds a feature estimation rule with a changed evaluation value to the feature estimation rules corresponding to the estimated feature candidates that need to be redefined as a result of the evaluation; A classification system comprising:

Images