JP6146813B2 - Numerical data documentation apparatus and program - Google Patents

Description

  The present invention relates to a numerical data documentation apparatus and program capable of generating a document to which a potential Dirichlet distribution method can be applied for continuous data such as health checkup data.

  As represented by Patent Document 1 and Patent Document 2, the health management system and the like are expanding greatly. In such a health management system, health advice is often given to users. However, as shown in Patent Document 3, advice is given after classifying users based on actual health data. Are more likely to lead to behavioral changes.

JP 2013-085626 JP 2010-264088 JP 2010-170534 A

  In order to effectively operate such a health management system, needs for classification of subjects from medical data are increasing in recent years. For example, patients are classified according to risk factors from biometric data, and measures for health promotion are performed for each classification. Appropriately implementing this measure makes it possible to clarify the disease risk of members of the health insurance association and reduce medical costs, for example.

  When applying a data classification method for general purposes that is not limited to medical data in classification of medical data, in general, simple classification is often performed based on a reference value by providing a reference value for the data. Therefore, if the general method is applied, medical data often has special statistical properties, and thus a large amount of such reference values must be prepared manually. Therefore, it is desirable to eliminate such trouble.

  Furthermore, on the premise that medical data is classified, it is considered that patient classification is further performed using the result. Generally, there is a clustering method such as the k-means method as a machine learning method. If this is applied to medical data classified as it is, specific patients / members can be grouped. .

  On the other hand, LDA (latent dirichlet allocation) is one of such clustering methods. LDA is a document clustering method that performs topic classification of multiple documents with high accuracy by estimating latent variables. Document clustering is known to have better performance than clustering methods such as the k-means method.

  Therefore, the present inventor has also proposed that the highly accurate clustering method is applied to medical data to classify patients with high accuracy and enable automatic extraction of dangerous patient groups. It was examined. However, since LDA is originally a technique used in document topic classification, it can only handle discretized data. Therefore, in order to apply LDA to continuous data such as medical data, first, the continuous value data must be appropriately discretized according to its characteristics, and second, the discretization The resulting results must be further documented to provide appropriate input to the LDA. However, there is no such method in the prior art.

  The present invention has been made in view of the above-described problems of the prior art, and an object thereof is to provide a numerical data documentation apparatus and program capable of generating a document applicable to LDA for continuous value data such as medical data. .

  Another object of the present invention is to classify continuous value data by applying LDA to the generated document.

  In order to achieve the above object, the present invention is a numerical data documenting apparatus, in which item data obtained by aggregating numerical values in each sample for a predetermined item is divided into sections according to their statistical properties, and each section is divided into sections. On the other hand, a numeric label assigning unit that assigns a numeric label that changes stepwise according to a deviation from the average as a category number, and for each sample, the category number assigned to the section to which the numeric value of the predetermined item belongs As a document for classifying the sample by the potential Dirichlet distribution method, a document including a predetermined number of predetermined words corresponding to the predetermined item and a predetermined number of predetermined words corresponding to the predetermined word And a documentation unit to be generated.

  Further, the numerical data documenting apparatus further applies a latent Dirichlet distribution method to a document set in which the generated documents are collected over all the samples for each sample, so that A classification unit that estimates a topic set and a topic ratio in each sample is provided.

  Furthermore, the present invention is a program that causes a computer to function as the numerical data documenting apparatus.

  According to the present invention, after numerical data is divided into sections, a numerical label that changes stepwise according to the deviation from the average is assigned as a category number, and each sample is assigned a predetermined item. Since a document is generated by generating a predetermined number of words corresponding to the assigned category number in a section to which a numerical value belongs, a document to which the potential Dirichlet distribution method can be applied can be generated. At this time, by generating not only the predetermined words but also the corresponding predetermined paired words, it is possible to generate a document to which the potential Dirichlet distribution method can be applied with high accuracy.

It is a functional block diagram of the numerical data documentation apparatus concerning one Embodiment. It is a functional block diagram of a numerical value label provision part. It is a figure for demonstrating notionally the process of the division by the distribution type | mold identification system discretization part. It is a figure for demonstrating notionally the process of the division by the numerical frequency system discretization part. It is a figure which shows the example of numerical value label provision. It is a figure which shows the example by which the division | segmentation was made | formed with respect to the item data of neutral fat, and the category number was provided. It is a figure which shows the example of 1st embodiment of a documentation part. It is a figure which shows the example of the word allocation number in 2nd embodiment of a documentation part. It is a figure which shows the example of the correlation coefficient of the item data in medical examination data. It is a figure which shows the example of three item data used as merge object, and the correlation coefficient between them.

  FIG. 1 is a functional block diagram of a numerical data documenting apparatus according to an embodiment. The numerical data documenting apparatus 1 includes a numerical label assigning unit 2, a documenting unit 3, and a classification unit 4. FIG. 2 is a functional block diagram of the numerical value label assigning unit 2. The numerical label attaching unit 2 includes a discretizing unit 21 and a label attaching unit 22. The discretization unit 21 includes a distributed identification method discretization unit 25 and a numerical frequency method discretization unit 26. The outline of each part will be described below.

  The numerical label assigning unit 2 receives numerical data as an input to the numerical data documenting apparatus 1, discretizes it, assigns a predetermined numerical label, and passes the result to the documenting unit 3. A typical example of the numerical data as the input is numerical data in each examination item such as a health check performed on a predetermined patient group (sample group). The contents of the numerical data indicate which numerical value is the numerical value in which examination item of which patient, and are given detailed information on all samples. (It is not data edited as a distribution graph.)

  Hereinafter, as an example for explanation, it is assumed that input numerical data is such health check data. In addition, as the test items, three types of triglycerides, BMI, and γ-GTP are used as appropriate.

  In the numerical label assigning unit 2, as a first process, the discretization unit 21 changes the statistical properties of a series of numerical data (referred to as “item data”) in each examination item with respect to the input health diagnosis data. Each item data is divided into sections in an appropriate manner.

  The section division has the following properties. That is, the samples (patients) belonging to the same section can be regarded as being homogeneous because the values for the examination item are close, and the samples separated in different sections have different values for the examination item. Therefore, it can be regarded as being heterogeneous. Here, the section is divided into a form suitable for each distribution characteristic of the item data. For example, the item data of triglyceride value is divided into sections according to the distribution characteristic of triglyceride value, the item data of BMI is divided into sections according to the distribution characteristic of BMI, and the item data of γ-GTP is γ -Sections are divided according to the distribution characteristics of GTP.

  Therefore, the discretization unit 21 needs to divide the sections for each item data by an individual method according to the distribution characteristics of the item data. For this reason, each of the distribution type identification method discretization unit 25 and the numerical frequency method discretization unit 26 implements an individual method for such section division.

  FIG. 3 is a diagram for conceptually explaining section division processing by the distribution type identification method discretization unit 25.

  First, as shown in (1), (2), and (3) of FIG. 3, the distribution type identification method discretization unit 25 displays the histogram of the target item data as “normal distribution type” and “log normal distribution”. By fitting in accordance with the corresponding model formula and determining its threshold value in this order, which of the three distribution types of “type” or “logarithmic normal distribution type and mixed distribution type of Pareto distribution” is best fitted. ,judge.

  The significance of the three types of distribution types is as follows and is suitable for classification of health check data. “Normal distribution type” means that the item data does not exhibit any behavior that deviates from normal. The “log normal distribution” is generally considered to be frequently used when the change rate of the target data draws a normal distribution. As a result, the behavior that deviates from normal seems to be mixed. In a case where “Pareto distribution” is mixed, there is a possibility that the target value is generated by a rule different from the rule that should be generally followed. For example, a case where an abnormal value has been recorded due to a metabolic abnormality or the like. As a result, it seems to be mixed with behavior that deviates significantly from normal.

  Next, the distribution type identification method discretization unit 25 performs section division in a form corresponding to the distribution type determined to fit the item data. If (1) and (2) are “normal distribution type” and “log normal distribution type”, the respective average m and standard deviation σ are obtained, and the width is a constant 2α times the width of σ around the average. What is necessary is just to divide into a series of sections (m ± ασ). (In the case of a lognormal distribution, it may be obtained on the logarithmic axis.) Thus, in (1) and (2), as a result, sections D11 to D15 as shown in (4) and (5), respectively. And division into sections D21 to D25 are obtained.

  In the case of (3) “logarithmic normal distribution type and Pareto distribution type mixed distribution type”, each of the two types of distributions is similarly divided into sections determined based on a predetermined width or the like according to the distribution. That's fine. Thus, in (3), as a result, as shown in (6), sections D31, D32, D33 inside the lognormal distribution type and sections D34, D35, D36 inside the Pareto distribution are divided. .

  FIG. 4 is a diagram for conceptually explaining section division processing by the numerical frequency method discretization unit 26.

  As shown in FIG. 4, the numerical frequency method discretization unit 26, in the cumulative frequency function of the item data, is a series of sections that can be regarded as performing a linear increase with a certain slope in each section. Division into D41 to D45. Since the slope is constant, the medical examination data can be regarded as being divided into sections according to the nature thereof (sections reflecting numerical bias and frequency bias as described below).

  Specifically, first, a fitting with a sigmoid function (the inverse function thereof) is performed on the cumulative frequency function, and the threshold value determination with respect to the cumulative relative frequency is a frequency abnormal value on the upper end side and / or the lower end side. And a central part that can be regarded as a normal frequency value.

  The center part is further fitted with a sigmoid function (the inverse function thereof) and divided into a part that can be regarded as a frequency abnormal value on the upper end side and / or the lower end side and a center part of a normal value. Perform recursively. In addition, for the part considered to be a frequency abnormal value on the upper end side and / or the lower end side, it is also recursively until it can be considered that the part shows a linear increase with a certain slope as a numerical behavior. Perform segmentation by straight line fitting (fitting is started from either end of the part).

  Then, when it is determined that the portion that can be regarded as a normal value and the portion that can be regarded as an abnormal value that are sequentially divided show a linear increase with a certain slope in the threshold determination, further division of the portion is stopped.

  In the example of FIG. 4, for example, division into sections is performed as follows. Fitting the inverse of the sigmoid function to the entire original cumulative frequency function, (1) D41 on the lower end side, (2) normal values D42 and D43, and (3) D44 and D45 on the upper end side are obtained. (1) The lower end D41 can be regarded as a straight line and is not divided any further. (2) The normal values D42 and D43 are further divided into (2-1) the lower end D42 and (2-2) the normal value D43 by fitting the inverse of the sigmoid function. Minami, do not divide any more. (3) D44 and D45 on the upper end side are divided into (3-1) D44 and (3-2) D45 by performing linear fitting from one end side (for example, D44 side) and whether or not the straight line is fitted. These can be regarded as straight lines and are not further divided. In this way, the division into the five sections D41 to D45 shown in FIG. 4 is completed.

  Details of the processing by the distribution type discretization unit 25 and the numerical frequency discretization unit 26 are described in Japanese Patent Application Nos. 2013-159323 and 2013-163207 by the present inventors, respectively.

  In addition, for each item data, which of the above distribution type identification method discretization unit 25 and numerical frequency method discretization unit 26 is applied, both are applied to the item data distribution. The degree of fitting may be compared and the better fitting may be applied. The degree of fitting may be quantified by a predetermined expression (such as difference absolute value integration in a histogram). Alternatively, it may be determined in advance as a user setting for each item data. Only one of the distribution type identification method discretization unit 25 and the numerical frequency method discretization unit 26 may be applied.

  The discretization unit 21 may be realized by a method other than the above-described method (independently or in combination with the method described above). For example, if the distribution type is known in the item data, the section may be divided in the known distribution type without examining the distribution type that fits best like the distribution type identification method discretization unit 25. Another type of fitting target distribution type may be added to the distribution type identification method discretization unit 25 in the same manner.

  As described above, as the first process by the numerical label assigning unit 2, the discretization unit 21 receives the result of dividing the item data into sections, and as the second process, the label attaching unit 22 assigns a numerical label to each section. FIG. 5 is a diagram showing an example of the numerical label assignment. When the item data is divided into five sections d1 to d5 as shown on the upper side, an example of assigning numerical labels (1) It is shown as (3).

  As shown in the example of (1), numerical labels such as “0, 1, 2, 3, 4” may be given to the sections d1 to d5 in ascending order, and conversely, as shown in the example of (2) Thus, “4, 3, 2, 1, 0” may be given in descending order. Or, as shown in the example of (3), assigning `` 0 '' to the middle section d3 and giving it sequentially to the left and right in ascending order, `` 3, 1, 0, 2, 4 '' It may be given.

  For such numerical label assignment by the label assigning unit 22, by providing predetermined rules in advance as in the above examples (1) to (3), numerical labels are automatically assigned to item data divided into sections. . In addition, although a non-negative integer is used for the numerical value of the numerical value label to be assigned and different numerical values are assigned to different sections in the same item data, there may be overlap in numerical values assigned to the sections of different item data.

  However, regarding the assigning rules, numerical values are given in a form that distinguishes between normal values and abnormal values in item data by providing stages. For example, “0” may be given to the most normal section, and the value to be given may be increased as the degree of abnormality increases. The assignment rule may be provided according to how the item data is divided into sections. It is preferable to assign values according to different rules to those having different abnormal modes (causes of abnormalities). For example, in the example of (6) in FIG. 3, “1, 2, 3” is assigned to the logarithmic normal distribution sections D31, D32, and D33, respectively, and “11,” is assigned to the Pareto distribution sections D34, D35, and D36, respectively. , 12, 13 "may be given.

  The documenting unit 3 generates a document representing the condition of each patient included in the medical checkup data, using the numerical labeling result for each item data of the medical checkup data by the numerical labeling unit 2 described above, This is the first output of the numerical data documentation device 1.

  The classification unit 4 reads each patient's document generated by the documenting unit 3 over all the patients who participated in the medical examination, applies the LDA, obtains the classification result of each patient, and creates a numerical data document. This is the second output of device 1.

  Note that LDA calculates the frequency of words in a document and estimates the topic frequency and the frequency of words in that topic as latent variables, thereby estimating which topic a specific document is likely to belong to. This is a document clustering method. In the present invention, by applying the LDA by the classification unit 4, the classification result is in the form of a topic set for representing the characteristics of all the patients who participated in the health examination and a topic ratio representing the specific characteristics of each patient. can get.

  For example, a patient a with “neutral fat level: 200 mg / dl, BMI: 30, γ-GTP: 40” automatically obtained a result that the ratio of topics corresponding to “metabo” was large. Patient b with values: 100 mg / dl, BMI: 23, γ-GTP: 100 automatically has the result that the ratio of topics corresponding to “Metabo Reserve” is large. In addition, specific topic contents such as “Metabo” and “Metabo Reserve” are grasped by manually interpreting the data, but according to LDA, a topic set having such a meaning and its ratio Can be obtained automatically.

  The document representing the condition of each patient generated by the documenting unit 3 is a document that is interpreted as representing the condition of each patient in the positioning of input in the application of the LDA. In the present invention, in particular, such a document is intentionally created by the numerical label assigning unit 2 and the documenting unit 3 intentionally from health checkup data as continuous data that has nothing to do with the document. This makes it possible to apply LDA. As a document format, the document generated by the documenting unit 3 only needs to be able to calculate the frequency of words in the document, and is not a normal document having human readability but a well-known Bag of Words format. It is a document.

  Details of the documenting unit 3 will be described below. In the description of the documenting unit 3, the numeric label assigned by the numeric label assigning unit 2 is referred to as “category number” as representing the category of each corresponding section.

  Using the category number assigned to each item data of each patient, the simplest method for the documentation unit 3 to generate each patient's document is the number of items equal to the category number of each item data. This is a technique for generating a document using a predetermined word corresponding to data.

  For example, it is assumed that the item data of neutral fat is divided into sections according to the range of values as shown in FIG. If the value of the triglyceride of the patient A is “33”, the category number is “2” because it belongs to the range “32-50”. Similarly, it is assumed that the patient A's BMI and γ-GTP category numbers are “5” and “3”, respectively.

  Thus, for patient A, all health check items “neutral fat”, “BMI”, and “γ-GTP” The category numbers are “2”, “5”, and “3”, respectively. Assume that the corresponding predetermined words are set to “fat”, “obesity”, and “sake” for the three items, respectively. In this case, according to the simplest method, the document generated for the patient A is “fat, fat, obesity, obesity, obesity, obesity, obesity, liquor, liquor, liquor”.

  However, in the simplest method described above, it may be inconvenient to apply LDA to medical examination data and classify patients with high accuracy. That is, generally, biometric data such as health checkup data is often correlated with each other. For example, people with high numerical values such as neutral fat tend to have very high values such as BMI and blood glucose level.

  Here, the numerical label assigning unit 2 assigns numerical labels, that is, category numbers, according to a predetermined rule that changes stepwise from normal values to abnormal values. For example, if a normal number interval is assigned a small number and an abnormal value interval is sequentially assigned a large number, the presence of the above correlation for a generally healthy patient and a generally unhealthy patient, The “vector direction” of the word frequencies in the generated document coincides, and although it is originally a completely different characteristic, it is classified into a similar characteristic by LDA.

  For example, the category numbers of all items “neutral fat”, “BMI”, and “γ-GTP” in the health examination are “1”, “1”, and “1” for patient B who is generally healthy, respectively. On the other hand, patient C who is generally unhealthy is “10”, “10” and “10”, respectively. Although the health characteristics of Mr. B and Mr. C are significantly different, the word frequency vectors are (1, 1, 1) and (10, 10, 10), and the directions of both coincide. When LDA is applied, Mr. B and Mr. C are classified in the same tendency. Thus, if the simplest method is adopted, there is a problem that an inappropriate document is generated as an input to the LDA.

  Therefore, in the following, various embodiments of the documenting unit 3 that can solve the problem in the simplest method will be described as a first embodiment to a third embodiment.

In the first embodiment, for each item data i (i = 1, 2, 3,..., I is an identifier of the type of item data), the contents related to the item i are expressed in the generated document. predetermining a predetermined total number N _i to be used for.

Then, in addition to the predetermined number w _i corresponding to the item data i, the number f (A _i ) corresponding to the number of category numbers A _i assigned to the interval corresponding to the numerical value of the item data i of the patient A Furthermore, the aforementioned predetermined total number N _i from the number f (a _i) a minus N _i -f (a _i) number of pairs of the predetermined word w _i to corresponding pairs word c _i (given word w _i Te A document is generated by some word c _i ) representing a concept.

That is, particularly in the first embodiment, by introducing N _i −f (A _i ) number of paired words c _i , so-called dummy word frequencies are newly introduced. Therefore, the problem of coincidence of apparent word frequency vectors in the simplest method using only A _i predetermined words w _i corresponding to item data i is solved. At this time, previously determined total number of words N _i for each item data i, a predetermined word w _i is defined as using f (A _i) number within the N _i number of ranges. It should be noted that the use of both words w _i and c _i is generally more appropriate when using different words to represent both because the medical views differ between high and low numbers. There is also a meaning.

Since the category number A _i is given so as to change stepwise from the normal value to the abnormal value, the function f () that gives the number f (A _i ) so as to keep the tendency of the step change. Is given as a predetermined function of monotonic increase or monotonic decrease. Alternatively, the constraint may be relaxed and the function f () may be given as a non-decreasing or non-increasing predetermined function.

FIG. 7 is a diagram illustrating an example of the first embodiment as an example corresponding to the example of FIG. 6. In FIG. 7, the column C1 and the column C2 are the same as those in FIG. 6, and an example in which a category number is assigned to each section of the neutral fat item data is shown. For the same example as FIG. 6, in FIG. 7, the number f (A _i ) using the predetermined word w _i is shown in the column C3. That is, the column C2 is the domain of the function f () and the column C3 is the range of values. In FIG. 7, the number Ni-f (A _i ) of the paired words c _i when the predetermined total number N _i = 30 is also shown in the column C4.

In FIG. 7, the sum of the values in the columns C3 and C4 is the predetermined total number N _i = 30 in any category number. Thus, for example, the predetermined word w _i corresponding to the triglyceride item data i (here, i = neutral fat) is set to the same “neutral fat” as the item name, and the corresponding word c _i is “ Assuming that mathematics is defined, patients belonging to category number 8 (patients whose triglyceride values are within the interval 900 to 1400) use 30 words for triglyceride and 0 for mathematics. A document is generated. A patient belonging to category number 5 (a patient whose neutral fat value is in the interval 250 to 420) generates a document using 14 words of “neutral fat” and 16 words of “mathematics”.

  If there is item data other than “neutral fat”, the word, the paired word, the upper limit number, and the like are similarly determined for those item data, and a document for each patient is generated with the words in all item data.

Note that a predetermined word w _i and its counter word c _i are set in advance as a conceptual pair. At this time, it does not have to be strictly antonym, and for the above-mentioned “neutral fat”, it is paired in some sense to set “math” that is not directly related to the medical examination, What is necessary is just to set what distinguishes. The predetermined total number N _i, instead of setting the individual values for each item data i, may be set a common value for all the item data i.

Next, a second embodiment corresponding to an application of the first embodiment will be described. In the first embodiment, it was used as the predetermined word w _i corresponding to the item data i and the pair word c _i, in the second embodiment further introducing intermediate word m _i. The introduction generally has different medical views when the numbers are high and low, and also different medical views that represent normal states, so it is better to use different words to represent each. It also means that it is appropriate. Then, after providing the first embodiment similarly to the predetermined total number N _i, for the word of the three, the distribution of the number may be, for example, as follows.

That is, for example, in a section showing the most average tendency (referred to as “average section”) such as giving a mode value, all N _i are set as intermediate words m _i , a predetermined word w _i and a counter word c _i. Is 0 and is not used. In a section that deviates from the average section in a first tendency, the predetermined word w _i is increased while the intermediate word m _i is decreased. Further, in a section that deviates from the average section with a second tendency (the reverse tendency of the first tendency), the number of words c _i is increased while the intermediate word _mi is decreased. Accordingly, there is no section in which the predetermined word w _i and the counter word c _i are used simultaneously, and a series of sections in which the intermediate word _mi and the predetermined word w _i are used, with the average section as a boundary, and the intermediate words _mi and It is divided into a series of sections in which the word c _i is used.

FIG. 8 shows an example of the word allocation number of the second embodiment in a form corresponding to the examples of FIGS. 6 and 7. In Figure 8, the section of category number 4 is "average interval" above, to all the predetermined total number Ni = 30 amino intermediate word m _i. As the average interval of category number 4 deviates from the interval of category numbers 3, 2, 1, 0, the number of words c _i is increased while the intermediate word _mi is decreased. In the segment of category number 0 that is most distant, only the word c _i is used. Further, in the reverse direction, the predetermined word w _i is increased while the intermediate word m _i is reduced as the average section of the category number 4 deviates from the average section of the category number 5, 6, 7 and 8. Only the predetermined word w _i is used in the section of category number 8 that is most distant.

The distribution of the three types of words may be set in advance in accordance with the rules described above according to the setting of the total number of category numbers and the predetermined number of words. Moreover, you may be made by methods other than the above. For example, the number of the predetermined word w _i and the paired word c _i is provisionally determined by the method of the first embodiment, and then the predetermined number determined in the first embodiment within the predetermined total number N _i frames from the word w _i and a counter word c _i, only the number fraction of larger predetermined ratio closer to the average interval, may be replaced with an intermediate word m _i.

  For example, when the counter word “mathematics” is defined for the predetermined word “neutral fat” in the item data of neutral fat, for example, “desk” may be used as the intermediate word. In the same way that the predetermined word and the counter word are set so as to give some conceptual distinction in the first embodiment, the intermediate word in the second embodiment also has some conceptual for each of the predetermined word and the counter word. What is necessary is just to set as what gives a distinction. At this time, the intermediate word is preferably set to give an intermediate concept between the predetermined word and the counter word, but it is not always necessary to set an exactly intermediate concept.

  Of course, it is also desirable to increase the distance on the code between the paired word and the predetermined word and to set the word having the intermediate edit distance as the intermediate word. For example, it is assumed that the predetermined word is expressed by “0000” (binary number, 4 bits). If the counter word is expressed as “1111” at that time, the distance between the two codes is the largest, and thus the most preferable expression for the counter word. Further, it is also desirable that the intermediate word is a word having a code having the same code distance from both, such as “0011”.

  In 3rd embodiment, after calculating the correlation coefficient between item data, another word is produced | generated about the part with which the concept of item data with high correlation overlapped. Accordingly, item data with high correlation is determined as the first process, and another word is generated in the determined item data as the second process.

  The significance of the third embodiment is as follows. That is, there is a high possibility that the item data are correlated with each other due to the nature of the health diagnosis data. For example, if the BMI is high, the waist circumference is likely to be large. In this case, BMI and waist circumference are considered to represent the same information for the most part. On the other hand, there are items that are not highly correlated, even though there is a correlation, such as triglyceride and γ-GTP. On the other hand, in the application of LDA, the fact that highly correlated items are analyzed simultaneously, that is, the appearance of words that appear to be independent independently in spite of being essentially the same information, is the accuracy of word frequency. , Which is not preferable.

  Therefore, as a first process, the correlation coefficient between the item data is calculated from the target health checkup data, and the obtained correlation coefficient value is determined to be large on the basis of the second standard. As a process, assuming that the two item data have the concept of overlap, a word reflecting the concept of overlap is generated. The word generated for the overlap concept is used as an LDA input in the classification unit 4 instead of the word used in individual item data that does not consider the overlap of concepts.

  Specifically, in the first process, after calculating the correlation coefficient between the item data, a threshold value is set, and if there is a certain level of correlation, they are treated as the same index, otherwise they are different. Treat as an indicator. The item data is referred to as an “index” when referring to the meaning of the data as the health check data.

  For example, in a certain medical examination data, if the correlation coefficient as shown in FIG. 9 is obtained by the first process, the correlation coefficient between BMI and abdominal circumference is 0.9, so the correlation between these item data is very high, It can be seen that it is an index that can be almost identified. In the example of FIG. 9, when the threshold value is set to 0.8, for example, the abdominal circumference and BMI at which the correlation is 0.9 are determined as the same index, and are to be merged. Other indices are not merged. In the second process, specific word settings in the item data to be merged are as follows.

  That is, in the third embodiment, the word generation processing similar to that of the first embodiment is performed after newly setting the predetermined word and the counter word in the first embodiment for the item data i and j to be merged. Implemented on the mixed concept. On the other hand, for the item data k that has a low correlation coefficient and has not been merged, the first embodiment is applied as it is to generate a word. Hereinafter, the third embodiment based on the first embodiment will be described, but a third embodiment based on the second embodiment is also possible.

First, the assumptions for explanation are listed. Also in the third embodiment, setting the total number N _i of words to generate from each item data i in advance. It is assumed that the item data i and the item data j have a correlation coefficient value r and are determined to be merged. For any patient A, the category number of item data i is A _i and the category number of item data j is A _j . Further, when the first embodiment is applied to the item data i and j in the patient A, f (A _i ) predetermined words w _i and N _i −f (A _i ) pairs are respectively obtained. Assume that a word c _i , g (A _j ) predetermined words w _j and N _j −f (A _j ) counter words c _j are generated.

In the third embodiment, the setting of the number in the first embodiment is modified and used, and the number n1 of the predetermined word w _ij representing the mixed concept of the item data i and j and the predetermined counter word c _ij By determining the number n2 according to the following formulas (1) to (3), the words in the merged indices i and j are generated. That is, n1 predetermined words w _ij and n2 counter words c _ij of the mixed concept are words generated in the merged indexes i and j.

As is clear from equation (2), N _ij = n1 + n2, where N _ij defines the total number of words generated in the merged index i, j. Then, as shown in Equation (3), the total number N _ij is _calculated by taking into account the total number N _i defined in the item data _i and the total number N _j defined in the item data j. The merge is determined by reflecting the correlation coefficient r. In Equation (3), the direct contribution of item data i is (1-r) N _i and the contribution considering the merged item data i is r / 2 * N _i It is. Here, dividing by 2 of the denominator is a reflection of the merge of both i and j. The same applies to the contribution from the item data j. Also, in Formula (1), similarly to Formula (3), n1 predetermined words w _ij of the mixed concept are determined by reflecting the merge of i and j.

Note that the predetermined word w _ij and the counter word c _ij of the mixed concept when the item data i and the item data j are merged are set in advance. The setting in advance is the same as the setting of the individual predetermined words w _i and c _{i in} the item data _i and the individual predetermined words w _j and the counter word c _j in the item data j. . Alternatively, a predetermined word and a paired word set in either item data i or j to be merged may be used as words to be used after merging. For example, w _ij = w _i and c _ij = c _i may be set.

  Here, a numerical example of the above word count calculation is introduced. In the health check data including the item data of each examination item as shown in FIG. 9, it is assumed that the BMI and the waist circumference are determined to be merged due to the correlation coefficient r = 0.9. In BMI and waist circumference, the total number of words is set as 30 in the first embodiment, and in a patient A, a category number of (predetermined number of words, number of words) = (30, 0) is given to BMI. It is assumed that the abdominal circumference is given a category number (predetermined number of words, number of words versus) = (25, 5).

In this case, in the patient A, the total number of words generated for the mixed concept merged by BMI and abdominal circumference is 33 words based on the following calculation.
(1-0.9) * 30 + 0.9 / 2 * 30 + (1-0.9) * 30 + 0.9 / 2 * 30 = 33
Further, out of the total number of 33 words, from the following calculation, the number of predetermined words in the mixed concept is 30.25 words, and as the remainder subtracted from the total number of 33 words, the number of counter words in the mixed concept is 2.75 words.
(1-0.9) * 30 + 0.9 / 2 * 30 + (1-0.9) * 25 + 0.9 / 2 * 25 = 30.25
33-30.25 = 2.75

  In addition, when the number of the predetermined word and the counter word in the mixed concept is actually calculated, it may be a decimal number as in the above numerical example, but may be an integer by rounding off, rounding down, or rounding up.

  Hereinafter, a case will be described in which all correlation coefficients between three or more item data satisfy the predetermined threshold determination, and three or more item data are to be merged. FIG. 10 is a diagram illustrating an example in which three item data are determined to be merged by determining to be merged when the correlation coefficient is 0.8 or more. In FIG. 3, the neutral fat, BMI, and abdominal circumference drawn with circles have the correlation coefficient written together with the line drawn between the circles, and thus are determined to be merged.

  In one example, when the three or more item data are determined to be merged at the same time, the two parties described by the above formulas (1) to (3) etc. Merge processing may be performed. At this time, the two merged parties are excluded from the subsequent merge targets. In the example of FIG. 10, “neutral fat and BMI” or “BMI and abdominal circumference” having a maximum correlation coefficient of 0.9 is set as a merge target. (Note that it is only necessary to select one at random because they are in the same order.) For example, if “neutral fat and BMI” are merged, only “abdominal girth” remains, but at this point Since there is no item data to be merged, the merge process is completed.

  Alternatively, in another example, when it is determined that the three or more item data are to be merged at the same time, the method of applying the two-party merge processing described by the above formulas (1) to (3), etc. You may make it merge with.

  In this case, the number of words is determined by the above-described distribution method between the two parties of each of the three or more item data (assuming n pieces), and the total number of words when all the n pieces are merged However, the number of words may be normalized so that it matches or matches the total number set in one item data. In the example of FIG. 10, assuming that all three of “neutral fat, BMI and abdominal circumference” are set as the total number of words 30, “neutral fat and BMI”, “BMI and abdominal circumference”, “abdominal circumference and neutral circumference” Each of the three integrations of “fat” may be performed by standardizing the total number of words to 30 ÷ 3 = 10, and then combining the three results to generate a final word.

Hereinafter, the determination of the number of words in the standardization method will be described in the general case where n item data are merged simultaneously. The preconditions are as follows. The total number of words set in one item data is L (a common value in all item data). Let r _i (i = 1, 2,..., N (n−1) / 2) be the correlation coefficient between each of the n item data. The total number of words N when the n pieces are integrated is given by the following equation (4).

Further, out of the total number of words N when the n pieces are integrated, the number M of predetermined words in the mixed concept may be determined as in the following formula (5). (Thus, the number of words in the mixed concept is NM.) Here, w _a and w _b are assigned to the target patient in the item data a, b giving the correlation coefficient r _i . This is the number of predetermined words set in the category number. As is clear from the form, the expression (5) is the expression (1) corresponding to the n cases.

  Note that predetermined words and counterwords for the mixed concept are also set in advance in the n simultaneous merging. In the above formula (4), the total number of words set in each item data is set to a constant L for simplification of explanation, but it can be calculated in the same manner even if a different total number is set for each item data. .

  Hereinafter, supplementary matters in the present invention will be described.

  (Supplement 1) In the first embodiment, it has been described that for each item data, a counter word corresponding to a predetermined word is prepared. However, as a simplified method, all the item data constituting the health check data On the other hand, a predetermined number of words of only one type may be provided and used. In this case, after generating the predetermined word for the category number or the predetermined number of words corresponding to the category number for each item data by the above-mentioned “simplest method”, the one type of counter word is further added. What is necessary is just to make it the document finally obtained by adding a predetermined number.

  (Supplement 2) Although the present invention has been described by taking the health diagnosis data as an example, it can also be applied to other medical data in general. As long as they have similar properties, they can be applied to any kind of continuous value data.

  When general data that is not limited to health checkup data is targeted, the properties described as “normal value” or “abnormal value” in the health checkup data are generalized as “average” or “deviation from average”. Is done. That is, whether it is “normal” or “abnormal” is an example property, and a value judgment is added in the case of health checkup data. For example, when dealing with sports performance, “normal” may correspond to “normal” and “abnormal” may correspond to “excellent”.

  Then, when handling general continuous data, the rule for assigning by the label assigning unit 22 may be a rule for assigning a numerical label so as to change stepwise according to the deviation from the average. For example, “0” may be assigned to the most average interval, and larger values such as “1”, “2”,... May be assigned as the distance from the average interval increases. Note that the “average” is an “average property” and is not limited to an arithmetic mean, but can be obtained as an arbitrary statistical index representing the group, such as the mode, median, etc. It's okay. The required statistical index may be different for each item data.

  (Supplement 3) The present invention is read by a computer and causes the computer to function as the numerical data documentation apparatus 1 of FIG. 1 (and FIG. 2), or corresponds to each part of FIG. 1 (and FIG. 2). It may be provided as a numerical data documentation program that causes the procedure to be executed.

  1 ... Numerical data documentation device, 2 ... Numeric labeling unit, 3 ... Documenting unit, 4 ... Classification unit

Claims (7)

The item data obtained by aggregating numerical values for each sample for a given item is divided into sections according to their statistical properties, and category labels that change in stages according to the deviation from the average are assigned to each section. A numerical label attaching unit to be given as
For each sample, a predetermined number of predetermined words corresponding to the predetermined item and a predetermined number of predetermined words corresponding to the predetermined word, corresponding to the category number assigned to the section to which the numerical value of the predetermined item belongs , And a documenting unit that generates a document as a document for classifying the sample by the latent Dirichlet distribution method.   The predetermined number is obtained by subtracting the number corresponding to the category number assigned to the section to which the numerical value of the predetermined item belongs from the predetermined upper limit number set in advance for the predetermined item. The numerical data documenting apparatus according to claim 1. The documenting unit, for each sample, a predetermined number of intermediate words corresponding to the predetermined word and a predetermined pair word corresponding to the category number assigned to the section to which the numerical value of the predetermined item belongs, Generating a document including a word and the predetermined paired word;
When the assigned category number corresponds to a section to which a mode of numerical values of a predetermined item belongs, a document including a predetermined upper limit number of predetermined intermediate words is generated for the sample,
When the assigned category number corresponds to the first section that deviates from the mode value, a document composed of a predetermined intermediate word and a predetermined word whose total number is the predetermined upper limit number for the sample. Produces
When the assigned category number corresponds to a second side section that is out of the mode value, the total number of the sample is composed of a predetermined intermediate word and a predetermined paired word whose total number is the predetermined upper limit number. 2. The numerical data documenting apparatus according to claim 1, wherein the numerical data documenting apparatus generates a document. The documentation section
Item data with a high correlation coefficient is determined according to a predetermined standard,
For each sample, when the document is generated for the first predetermined item and the second predetermined item whose correlation coefficient is high, the first predetermined item and the second predetermined word are used instead of the predetermined word and the predetermined pair word. Generating a document composed of a predetermined mixed concept word and a predetermined mixed concept pair word corresponding to a predetermined item;
The number of the predetermined mixed concept word and the predetermined mixed concept pair word generated in the document is determined based on the predetermined upper limit number, the number corresponding to the category number, and the correlation coefficient. The numerical data documenting apparatus according to claim 2.   Applying the latent Dirichlet distribution method to a document set in which the generated document is collected over all samples for each sample, thereby obtaining a topic set in the document set, a topic ratio in each sample, The numerical data documenting apparatus according to claim 1, further comprising a classification unit for estimating   6. The numerical data documenting apparatus according to claim 1, wherein the item data is related to medical care.   A program for causing a computer to function as the numerical data documenting device according to any one of claims 1 to 6.