JP2006227942A - Extraction system of combination set of clinical test data, determination system of neoplasm progress using the same and clinical diagnosis support system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a clinical diagnosis support system estimating progress of organization body pathological change of neoplasm or the like from clinical test data. <P>SOLUTION: In the system, an electronic computer mechanically extracts a combination set of clinical test data with deep connection with neoplasm progress from progress determination data of neoplasm determined by using a pathological view and clinical test data comprising a neoplasm marker by a method for generating a decision tree rule set. (1) A cluster of a sub-data set which can be described is obtained by using a common test item among all test items, and a sub-decision tree is generated. (2) The sub-decision tree comprising defect/unknown data is cut from the sub-decision trees stated in (1). The system extracts the combination set of clinical test data having an operation process for generating the sub-decision tree as the cluster without defect/unknown data. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a system for extracting a combination set of clinical laboratory data, a system for judging a degree of tumor progression using the system, and a clinical diagnosis support system. More specifically, the present invention relates to a lesion of a tissue such as a tumor from clinical laboratory data. The present invention relates to a new clinical diagnosis support system capable of estimating the degree of progress.

  Regarding the degree of tumor progression, Kobayashi et al. (Non-Patent Document 1) proposes a method for empirically diagnosing tumor progression from the relationship between clinical test data and tumor size. This is a method that targets 12 types of tumor markers, but at present, methods for experimentally deriving the degree of tumor progression have not been studied for more than 70 types of clinical laboratory items that are actually used. Absent.

  In such a situation, there is no doubt that if a method for inferring the degree of tumor progression is established from actual clinical test data and clinical diagnosis can be supported, it will be a very significant contribution.

  In addition, the progress of lesions in human tissues, not limited to tumors, can be estimated from a large number of clinical laboratory data, which is a breakthrough for clinical medicine as a whole.

As a method for enabling inference of the degree of progression of such tissue lesions from clinical test data, inference by a machine learning method is effective. However, which method is applied in the machine learning method depends on the nature of the data. Machine learning includes those that can be learned only from a data set and those that require background knowledge (reference rules) in advance. The reference rule is a rule framework given in a format such as “(A <a) ∧ (B> b)... → S = z”, for example. When inferring from clinical laboratory data, since such a reference rule has not been established, the existence of a reference rule cannot be assumed. A learning method is needed to acquire new rules for the relationship between clinical laboratory data. In addition, there is a need for a learning method that can handle both continuous and discrete values in clinical laboratory data, and can handle these simultaneously. As one of machine learning systems corresponding to such restrictions, there is C4.5 (Non-patent Document 2) that generates a set of rules called a decision tree.
Tsuneo Kobayashi, and Tomoko Kawakubo, Prospective Investigation of Tumor Markers and Risk Assessment in Early Cancer Screening, CANCER, Vol. 37, No. 7, 1994. Quinlan JR: C4.5: Programs for Machine Learning, Morgan Kaufman, 1993.

  From the background as described above, an object of the present invention is to provide a new clinical diagnosis support system that can estimate the progress of a tissue lesion such as a tumor from clinical test data.

  In order to solve the above-mentioned problems, the present inventor has studied using the C4.5 decision tree learning method by Quinlan as described above, and has the following basic object.

  <A> A method for constructing a clinical diagnosis system for estimating the degree of progression of a tissue lesion such as a tumor from clinical test data. In other words, using clinical progress data including a tumor marker in each disease case using clinical progress data including a tumor marker determined by using pathological findings, Extract combinations efficiently.

  <B> By applying to the clinical laboratory data of a patient whose tumor progression degree is unknown, on the condition that the combination of clinical laboratory data obtained by extraction in <A> above is a condition, the tumor progression degree of the patient is efficiently determined To do.

  <C> The relationship between the combination of the clinical test data obtained by the above extraction <A> and the degree of tumor progression is incorporated into the database as knowledge.

  More specifically, the invention of this application is characterized by the following.

First: A decision tree rule set is a combination set of clinical laboratory data closely related to tumor progression degree from tumor grade determination data determined using pathological findings and clinical laboratory data including tumor markers A system for mechanical extraction by an electronic computer by a method of creating
1) Create a sub-decision tree by finding a cluster of sub-data sets that can be explained using common inspection items among all inspection items. 2) Items containing missing / unknown data from the sub-decision tree of 1) above. A system for extracting a combination set of clinical laboratory data, characterized by having an arithmetic process for creating a sub-decision tree as a cluster free of cut-missing missing / unknown data.

Second: In the first system,
3) A set of clinical laboratory data characterized by having a calculation process for creating a sub-decision tree in which all data is incorporated by extracting and reviewing missing / unknown data not included in the sub-decision tree in 2) above Combined set extraction system.

  Third: In the first or second system described above, the sub-decision tree is created and the data including missing / unknown data is trimmed through a verification process with a knowledge database related to the strength of relationship between examination items. A system for extracting a combination set of clinical laboratory data.

  Fourth: A tree and decision tree having a tree-structured decision tree cluster extracted by any one of the first to third systems described above, and a combination and decision tree based on condition selection of clinical laboratory data of a patient whose tumor progression degree is unknown A tumor progression determination system characterized in that a tumor progression degree can be mechanically determined by calculating a match with a cluster.

  Fifth: The tumor progression degree determination system according to the fourth system, wherein the tumor progression degree is predicted from the tree structure path of the decision tree cluster from the calculation of matching by a combination of condition selections.

  Sixth: A clinical diagnosis support system comprising any one of the first to fifth systems as at least a part of its configuration.

  The present invention as described above is configured as a mechanical information system using an electronic computer, and has means such as a data input unit, a calculation unit, and an output / display unit. For example, in the first, second, and third inventions, the calculation unit performs creation of a decision tree cluster having a tree structure of a sub-decision tree and pruning for missing / unknown data. In the fourth and fifth inventions, the determination prediction is performed in the calculation unit, and the result is output and displayed in the output / display unit.

  According to the present invention, it is possible to estimate the degree of progression of a diseased tissue such as a tumor from clinical laboratory data, which has never been realized so far, and extremely great support is realized for clinical diagnosis and disposal based thereon. It will be.

  The present invention will be described in detail below.

<A> Clinical Diagnostic System for Estimating Progression of Tissue Lesions such as Tumors from Clinical Laboratory Data This clinical diagnostic system is a means for constructing a set of decision trees as the core of the diagnostic system. Unlike experimental data, patient clinical test data generally does not have all necessary test items, and a lot of test data is missing (for example, Table 1). The present invention proposes a method for finding out the relationship between diagnosis results and clinical laboratory data from such irregular data.

(1) First, clinical laboratory data (D ₀ ) having a configuration as shown in Table 1, for example, which is a source for constructing a set of decision trees has the following properties.
D ₀ is data in which the results of the clinicopathological determination of the degree of tumor progression for each patient are associated with the clinical laboratory data examined at the time of determining the degree of tumor progression.
-As shown in Table 1, for example, the degree of tumor progression is classified into 5 levels of 1 to 5 based on Kobayashi et. Al (CANCER, 1994).
Clinical examination data contained in D ₀ are sometimes the tumor progression important inspection item for judging the not always is always included, are deficient because it has not examined .
-Not only test items that are meaningful in determining tumor progression, but also many clinical laboratory data that are not closely related to tumor progression.
(2) With regard to clinical laboratory data (D ₀ ) having such properties, it is difficult to create a decision tree that explains all data by learning once with C4.5 by Quinlan.

  Quinlan C4.5 decision tree learning has the following properties.

  ・ C4.5 was obtained by examining the input data (clinical test data + tumor progression), calculating the condition to become the judgment value (tumor progression here), and expressed as a combination of the range of clinical laboratory data Output the decision tree.

If missing data is not allowed, there is only one leaf (decision tree end) corresponding to a combination of laboratory values of a patient in the decision tree. (One leaf may correspond to multiple patients.)
・ When missing data is allowed, multiple leaves corresponding to the same patient appear in one decision tree. This is because a condition that the missing portion data may take any value (don't care: does not consider the value), and possibly a plurality of conditions are arranged side by side. On the other hand, in the present invention, missing data is not allowed.
(3) Therefore, by using the inspection items common among all the inspection items, several clusters (D _{k in} FIG. 1) of the sub data set are obtained by learning, and the sub decision tree DT _k (kth decision tree) is obtained. ).

FIG. 2 shows a clustering calculation procedure for creating the sub-decision tree DT _k . D _k is a subset of laboratory data Ds _k used to create the sub-decision tree DT _k at the k-th step represents patient data covered by the sub-decision tree DT _k obtained from D _k . Since there is missing data in the decision tree DT _k obtained from D _k by the process <1> in FIG. 2, an unnecessary branch that cannot explain D _k is generated. In the process of <2> in FIG. 2, such unnecessary branches are removed and output as the _kth sub-decision tree DTk. As a result, the patient data of the original data set foo.data is learned so as to be covered by any sub-decision tree. By repeating this process repeatedly, a decision tree that can explain patient data is configured as a set of sub-decision trees {DT _k } (k = 1... K).

In this iterative process, if the patient data is missing values across multiple test items, the ambiguity is that the original decision tree is to determine based on the values and ranges of these test items. Therefore, each inspection item can take any value, and a sub-decision tree can no longer be generated. Therefore, in the sub-decision tree DT _k obtained in the process of FIG. 2 <1>, when there is patient data that can support DT _k in D _k is only any one of the leaves the patient data in the DT _k Ds _k is obtained under the condition that it does not exist. This means that the decision tree is generated under a condition that does not allow the case where the value of the inspection item is missing (don't care). On the other hand, it is possible in principle to create a decision tree under conditions that allow don't care, but because it becomes a sub-decision tree that can be taken under the condition that each inspection item can take any value, considered to be a decision tree that result in the determination of over-fitting (too much inference) for the unknown data, does not adopt the method of determining the Ds _k under the conditions that allow the do not care.

The procedure for obtaining the sub-decision tree cluster will be described in more detail with reference to FIG.
[1] C4.5 decision tree learning is performed using all data D ₀ as input data to obtain DT ₀ .

Here, a decision tree for the entire input data is created.
[2] The inside of DT ₀ is examined, and the clinical laboratory data of the patient that satisfies the conditions up to each leaf is extracted and set as Ds ₀ .

That is, a leaf that is not don't care is searched from the created decision tree, and it is examined which patient data is supported by the condition. Data for clinical test items included in the conditions up to the leaf are not missing.
[3] Since Ds ₀ is input and a sub-decision tree DT ₀ having a Ds ₀ element in the leaf is obtained, it is stored.
[4] Remove the patient data Ds ₀ which is obtained as a sub-decision tree already out of all the data D _0, the remaining ones, performs the evaluation by again C4.5 decision tree learning. From the D _0, remove the patient data _{Ds 0 (D 0 -Ds 0)} , and D _1.
[5] C4.5 decision tree learning is performed using D ₁ as input data (repeating [1], [2], [3], and [4]) to obtain a new sub-decision tree DT ₁ . This operation is repeated until D _k disappears or patient data that satisfies the condition in [2] is not obtained.

By subtracting the patient data evaluated as the sub-decision tree from all the data D ₀ every cycle, a cluster of sub-decision trees can be created with no missing data. In addition, if all the remaining patient data is don't care, it is not evaluated forcibly, but is checked data. The patient data determined as the check data required can be subjected to re-evaluation by adding clinical laboratory data.
By a series of operations on [6], the stored K sub decision tree _{DT 0, DT 1, DT 2} ,,, DT K to a decision tree clusters.

  By collecting the stored sub-decision trees, a sub-decision tree with missing data is obtained.

  3 and 4 are conceptual diagrams showing the relationship between all data and the sub-decision tree with respect to the above procedure.

  In order to efficiently extract data with clear relation between clinical laboratory data and tumor progression from all the data, don't care and don't care are determined from the decision tree obtained by using C4.5 decision tree learning. By cutting the branches that contain the leaves, a subset with a clear relationship between clinical laboratory data and tumor progression is extracted.

As a result, decision trees (DT _{1 to} DT _K ) that do not include don't care are obtained from all the data. Data that is not included in DT _{1 to} DT ₅ is data for which a clear decision tree cannot be obtained because there is missing data (FIG. 5).
(4) A process of adding new clinical test data to data that is not included in the clustering process of a series of sub-decision trees and cannot be determined, and adding the data to the decision tree cluster will be described with reference to FIG.
[1] A decision tree cluster (DTC ₀ ) is created from original data (D ₀ ) in which clinical laboratory data and tumor progression degree are related for each patient.
[2] Data (DU ₀ ) that is not included in the decision tree cluster because it has become don't care is extracted from D ₀ .
[3] With respect to undecidable data (DU ₀ ), a missing value in clinical laboratory data is eliminated by reexamination or the like.
[4] A decision tree cluster (DUC ₀ ) is created using DU ₀ whose data has been reviewed.

As a result of the data review, data that was previously don't care is included in the decision tree.
To [5] decision tree that was created from the D ₀ (DTC _0), incorporated a new decision tree cluster (DUC0) that was created from DU _0, the new decision tree cluster (DTC _1).
[6] Undecidable data that has become don't care by the processing of [4] is extracted (DU ₁ ), and the processing of [3] is repeated to create a decision tree cluster.
[7] Even if the data is reviewed, if all the data becomes don't care or if all the data is incorporated into the decision tree cluster, the process is terminated.
<B> System for Determining Patient's Tumor Progression Next, using the L sub-decision trees (DTC _{0 to} DTC _L ) created as described above, the tumor progression is determined using unknown patient data. To do.
(1) First, the decision tree cluster created by the above <A> will be described. As illustrated in FIG. 7, the sub-decision tree constituting the decision tree cluster includes a set of conditions branched from one vertex. In the sub-decision tree, the path from the vertex to each leaf can be considered as a condition connected by AND. One of the tumor progression levels (1 to 5) is described in the leaf, and the same tumor progression level may be described in another leaf. Therefore, there are multiple routes (conditions) to reach the same degree of tumor progression. However, there is only one type of path from the vertex to the leaf in the entire decision tree cluster. In the leaf, the patient number corresponding to the data from which the sub-decision tree was created is added, but this is data for verifying the sub-decision tree. The number has no meaning.

  For example, in the illustrated sub-decision tree, three types of conditions for tumor progression = 4 are shown, and it is determined that tumor progression = 4 for any of these three conditions. The conditional expression that gives a tumor progression of −4 is expressed as follows. From this equation, the order of the elements included in the condition is irrelevant to the determination of tumor progression.

By applying data with unknown tumor progression to this decision tree cluster and examining which leaves in the decision tree cluster match from the combination of laboratory data described in the unknown data, the unknown data Predicts the degree of tumor progression.
(2) A flow for applying unknown data to a decision tree cluster will be described with reference to FIGS.

Tumor progression of unknown data with unknown tumor progression is determined using the generated decision tree clusters (DT _{1 to} DT _K ). Unknown clinical laboratory data is applied to the DT ₁ conditions, and if no matching condition is found in DT ₁ , the DT ₂ conditions are applied. In this way, the condition in the decision tree cluster is examined, and if a condition that matches somewhere is found, the degree of tumor progression can be determined.

  If a decision tree cluster that does not completely match a combination of unknown clinical test data (a combination of test data from the vertex to the leaf of the decision tree cluster matches) is not found, it is output as undecidable data. If a plurality of paths that completely contain unknown clinical test data combinations are found, this is also output as undecidable data.

The undecidable data that has been output can be subjected to determination by the decision tree cluster again by adding clinical laboratory data or the like.
(3) FIG. 10 illustrates an effect obtained by applying unknown data to a decision tree cluster.

[1] If a combination of clinical laboratory data extracted as a condition for tumor determination from the original data (D ₀ ) is present in unknown data, the degree of tumor progression can be determined.

  [2] Tumor progression degree determination can be performed by adding clinical laboratory data to unknown data determined to be indeterminate and determining the decision tree cluster again.

[3] Accumulation of tumor progression assessment of unknown data by accumulating tumor progression and clinical laboratory data determined using pathological findings, and updating the decision tree cluster as needed using the accumulated data Can be improved.
(4) In the above determination system, in practice, (a) a method for forcibly determining the degree of tumor progression regardless of the value of a missing test item that becomes don't care in unknown patient data; b) Two types of methods are possible: a method of strictly judging using only inspection items without missing items (don't care).

  If there is a difference in the tumor progression judgment results of the two methods by using these two judgment methods in combination, diagnosis such as reexamining the missing items that are don't care. There is a possibility of suggesting support to the doctor.

Therefore, the case (b) above will be evaluated.
<Experiment details>
Using multiple clinical test data with overlapping test items such as tumor marker tests, inference learning by decision tree analysis is performed for the test item of interest, and the relevant information that is clinically meaningful between the test items based on the results The method of obtaining was evaluated. A method to evaluate possible methods of extracting relevant information as general trend data from the conditions of nodes included in the upper layers of decision trees obtained as a result of inference learning, and to narrow down possible diseases from clinical laboratory data Was evaluated.
<Evaluation method of experimental results>
We examined the results of inference learning and evaluated the method to obtain clinically relevant information from decision trees. Evaluate methods to obtain clinically meaningful general trends from the conditions of nodes in the upper layers of decision trees obtained as a result of inference learning, and evaluate methods to narrow down possible diseases from clinical laboratory data went.
<Experimental data>
Laboratory test data (51 types of test items, number of patients 432) in which doctors determined tumor progression by examining mainly tumor markers
<Experimental result>
(1) Evaluation of a method for extracting a general tendency from clinical test data by inference learning In general, items to be tested for a clinical test result are determined according to the patient's condition. For this reason, when clinical test data of many patients are collected, all of the clinical test items are not inspected without defects, and there are almost no items that have not been tested. From such a collection of clinical laboratory data, we examined a method to apply inference learning techniques to obtain general trends. In the experiment, in a decision tree DT _k by clinical examination data set D _k, present patient data to the leaf support of the decision tree, and, the condition that there is no loss of inspection items in conditions up to the leaf (loss The condition that does not allow is set. In the decision tree shown below, the degree of tumor progression is displayed on each leaf, and the patient data number (0 to 431) and the number of patient data are described on the right side.

The first [12] sub-decision tree DT ₁ obtained from the clinical laboratory data set D ₀ is shown (Table 2). This decision tree explains the degree of tumor progression for 40 people ([12] + [26] + [2]).
Sub-decision tree DT ₁ (explains 40 people)

Next, a sub-decision tree DT ₂ is obtained from the data set D _{2 for} 432-40 = 392 persons by an equivalent method (Table 3), whereby the tumor progression of 6 persons ([2] + [4]) The explanation is made.

Sub-decision tree DT ₂ (explains 6 people)

Subsequently, the following sub-decision trees (Tables 4 to 12) were created.

Sub-decision tree DT ₃ (explains 4 people)

Sub-decision tree DT ₄ (explains one person)

Sub-decision tree DT ₅ (explains 4 people)

Sub-decision tree DT ₆ (explains 9 people)

Sub-decision tree DT ₇ (explains 91 people)

Sub-decision tree DT ₈ (explains 28 people)

Sub-decision tree DT ₉ (explains 4 people)

Sub-decision tree DT ₁₀ (explains one person)

Sub-decision tree DT ₁₁ (explains one person)

Sub-decision trees DT _{1 to} DT ₁₁ were created, and a decision tree set explaining the degree of tumor progression for 189 out of 432 patient data sets was obtained. Since the decision tree obtained here can sort patients who have not yet been given the degree of tumor progression, it is possible to support a diagnosis by a doctor.

The 243 persons that could not be explained here are due to the influence of a lack of clinical test data, and can be explained by adding patient clinical test items.
<Evaluation of experimental results>
Clustering was performed in which sub-decision trees were sequentially generated using conditions that did not recognize a deficiency, and sub-decision trees that explained the degree of tumor progression were obtained from sample clinical laboratory data sets.

It was confirmed that there was at least one patient data that matched the condition of each leaf of the obtained sub-decision tree, and that the condition was generally clinically consistent.
<C> Knowledge database system for the relationship between clinical test data and tumor progression level By applying the machine inference learning method to clinical data, a decision tree that explains the combination of patient test data included in the clinical data is obtained. It is done. The decision tree information includes information that can be clinically affirmed and information that is not. For this reason, the decision tree obtained by opportunity reasoning learning cannot be used as a knowledge element in practice. Therefore, the decision tree obtained by machine inference learning is used to select necessary information from the decision tree information by examining the interface that mediates the information contained in the decision tree and existing knowledge elements. Guide the way.

  As a result of inference learning, the decision tree that is output is, for example, as shown in FIG. 11, all output information is a text file, has an 11-line header portion, and the decision tree body starts from the 12th line. ing.

  The total number of data supporting the conditions up to each node of the decision tree is obtained by a post-processing filter (FIG. 12).

  In this FIG. 12, the meaning of the part enclosed by the square is as follows.

As one of the conditions for tumor progression = 4,
Numbers of patients that meet this condition are 1, 3, 6, 12, 17, 20, 22, 23, 33,.

  When importing a clinical diagnosis support system, the class item that the decision tree was made based on is extracted from the header information, and the decision tree is recognized as a conditional expression assigned to each class item classification .

  Incorporation into the clinical diagnosis support system using the decision tree is the procedure shown in FIG.

  In view of this, an HGM-aid (Japanese Patent Application No. 2003-320449) developed and constructed by the present inventors as an existing system will be described below as an example.

  This HGM-aid can be called a comprehensive gene medical support system, and has the following structural features.

  First, the knowledge classification database for accumulating clinical data and basic research data in clinical medicine, and the meaning of terms included in clinical data and research data, as well as weather data, data related to living environment and preferences, A term definition database for accumulating range definitions (knowledge elements), an area mediating database for accumulating the strength of the relationships between knowledge elements that constitute data accumulated in the knowledge domain database, and a user management database A medical knowledge database support system provided with normalized discretized data processing function means for generating discretized data by collating clinical data with a reference value and compressing the amount of information Yes.

  Secondly, a knowledge classification database for accumulating clinical data and clinical research data in clinical medicine, and a term definition for accumulating definitions of the meaning and scope of terms included in clinical data and research data. A medical knowledge database support system comprising a database, a domain-mediated database for accumulating the strength of association between knowledge elements constituting the data accumulated in the knowledge field database, and a user management database. A medical knowledge database support system characterized by having a search function means using a plurality of related information defined between knowledge elements, and thirdly, a search order from a base element It has a search function means that limits "depth" and the "weight" of the relationship between elements as a threshold. It is characterized in that it is a medical knowledge database support system.

  Fourthly, a knowledge classification database for accumulating clinical data and clinical research data in clinical medicine, and a term definition for accumulating definitions of the meaning and scope of terms included in clinical data and research data. A medical knowledge database support system comprising a database, a domain-mediated database for accumulating the strength of ties between records constituting the data accumulated in the knowledge field database, and a user management database, wherein clinical laboratory data Are arranged on a concentric circle, and the secondary elements derived from the primary element by the association search are arranged linearly or radially from the primary element, and the elements related between the secondary elements are branched elements. Characterized by having search graph display means arranged radially from That a medical knowledge database support system, the fifth, is characterized by the clinical test data is a medical knowledge database support system that arranged concentrically.

  Based on data in the clinical medical field and basic research field, by searching for the relationship between the data in the database, it is possible to newly connect multiple clinical symptoms that have not been known so far with genes and proteins. It is possible to find the relationship, and by completing and operating HGM-aid, mutual knowledge exchange between the basic research field and the clinical medicine field is promoted, and the accumulation of knowledge to comprehensively grasp the disease is increased. It becomes possible.

In the present invention, an interface for fetching a decision tree cluster by the HGM-aid database as described above is performed according to the following procedure along FIG. By this procedure, it is possible to import the information into the system as a part of the association information between knowledge elements accumulated in the HGM-aid database. In HGM-aid, information that supports patient status, diagnosis (including tumor progression), and the like by searching for knowledge information associated with clinical laboratory data of patients using association information between knowledge elements as clues Can be presented to the user.
[1] The decision tree cluster data created in <A> is read into the system and displayed. Decision tree cluster data is passed as information on a file or memory.
[2] The decision tree cluster having a tree structure is interpreted, and all paths from the vertex of the decision tree to the leaf are extracted. One route is described in one record. A node included in the route is configured by (examination item + range), and a determination value (in this case, a tumor progression degree) is described in the leaf (for example, FIGS. 15 and 16).
[3] If the same inspection item is included in one record, the inclusion relation of the range is examined and the range data is rewritten, and the same inspection item is included in only one place in the same record. So that the data is aggregated.
[4] All the collected records are examined, and all the inspection items contained therein are taken out and arranged in a table (for example, FIG. 17).
[5] By sorting the sorted table according to the degree of tumor progression (leaf), it becomes a table that describes the conditions leading to it by the degree of tumor progression.
[6] Since the range for each inspection item constituting each node of one record is a range determined when the decision tree is created, it does not necessarily match the range of the abnormal value of the inspection item. . In HGM-aid, each inspection item is classified into grades in a range of abnormally low value, low value tendency, normal range, high value tendency, abnormally high value, and the like. The inclusive relation between the grade classification range and the range of the inspection value by the decision tree is examined (for example, FIGS. 18 and 19).
[7] As a condition for diagnosing the degree of tumor progression, a combination of HGM-aid grade classifications including a test value range combined by a decision tree is selected.
[8] In HGM-aid, information in which a plurality of related knowledge elements are linked is stored as a database. The grade classification of the examination items constituting the combination that reaches the degree of tumor progression selected in [7] is normalized as a knowledge element and registered in the database as related information (for example, FIG. 20).
[9] It is conceivable to analyze the range of the inspection value by the decision tree and use it as information for updating the grade range.

It is a conceptual diagram which shows the relationship between all the data Do and the sub decision tree DTk. It is a flowchart of clustering for creating sub-decision tree DTk. It is the first half of the clinical test data and the sub-decision tree related conceptual diagram for creating a decision tree cluster. It is the figure of the second half part following FIG. It is the conceptual diagram which showed the relationship between a sub decision tree and the set of don't care data. It is the conceptual diagram explaining the post-processing method of the data which became undecidable by appearance of don't care. It is the conceptual diagram explaining the element which comprises a sub decision tree. It is a flowchart which determines the tumor progression degree of unknown clinical test data by a decision tree cluster. It is the figure which showed the flow of data when determining unknown clinical data with a decision tree cluster. It is a figure explaining the effect by application of a decision tree cluster. It is the figure which showed the decision tree output example (part) by C4.5. It is the figure which showed the example of data which added the patient number by post-processing the decision tree by C4.5. It is the figure which showed the taking-in flow of the inference learning effect. It is a figure explaining the procedure which takes in a decision tree cluster in a HGM-aid database. It is the figure which showed reading of the decision tree cluster and the example of display. It is the figure which showed the example different from FIG. It is the figure which showed the example of a display which aggregated-reorganized the read decision tree cluster. It is the figure which showed the example of a target display of the condition range by a decision tree cluster, and the great classification of HGM-aid. It is the figure which showed the example different from FIG. It is the figure which showed the example of the registration and display of the conditions by the combination of a great classification.

Claims (6)

A decision tree rule set is created from a combination set of clinical laboratory data closely related to the degree of tumor progression from data on the degree of progression of tumors determined using pathological findings and clinical laboratory data including tumor markers A system for mechanical extraction by a computer by a method,
1) Take out the parts that are not missing from all the inspection items and find the cluster of the subdata set to create a subdecision tree. 2) Cut out the data that contains missing / unknown data from the subdecision tree of 1) above. A system for extracting a combination set of clinical laboratory data characterized by having an arithmetic process for creating a sub-decision tree as a cluster having no unknown data. The system of claim 1, further comprising:
3) A set of clinical laboratory data characterized by having a calculation process for creating a sub-decision tree in which all data is incorporated by extracting and reviewing missing / unknown data not included in the sub-decision tree in 2) above Combined set extraction system.   3. The laboratory test data according to claim 1 or 2, wherein a sub-decision tree is created and data including missing / unknown data is trimmed through a verification process using a knowledge database related to the strength of relationship between examination items. Extraction system for combination set.   A database of decision tree clusters having a tree structure extracted by the system according to any one of claims 1 to 3, wherein a combination of condition selection of clinical laboratory data of a patient whose tumor progression degree is unknown matches a decision tree cluster A tumor progression determination system characterized by mechanically determining tumor progression by sex calculation.   5. The tumor progression degree determination system according to claim 4, wherein the tumor progression degree is predicted from the tree structure path of the decision tree cluster from the calculation of matching by a combination of condition selections.   A clinical diagnosis support system comprising the system according to any one of claims 1 to 5 as at least a part of its configuration.