JP6270216B2 - Clustering apparatus, method and program - Google Patents

Description

  The present invention relates to a clustering apparatus, method, and program capable of performing clustering suitable for predicting the risk of a specific disease in medical data.

  There are cases where it is desired to cluster the target person based on medical 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. It is easier to change the behavior and effective advice is possible.

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

  Here, in recent years, a high-precision classification technique represented by latent topic analysis (particularly, latent dirichlet allocation (LDA)) has attracted attention. Using this LDA, it is also possible to classify people with similar health factors and extract high-risk people whose health conditions will deteriorate in the future.

  In general, unsupervised learning categorizes similar data from the obtained data. In topic classification represented by LDA, data is represented as a mixture of multiple topics by unsupervised learning. . When targeting medical data, a topic is expressed as a mixture of occurrence probabilities of words representing a health condition.

  In general, when categorization is performed using medical data such as receipt data (medical specification information) and medical examination data, and a specific disease risk such as diabetes is predicted, an unsupervised learning feature (the specific There is a problem that the accuracy is reduced because it deals with features not related to disease).

  FIG. 1 is a diagram showing an example for schematically explaining the problem, and shows two results R10 and R20 obtained by typifying the same medical data. As a result of typification when you want to predict the disease risk of diabetes in the medical data, the result R10 is an undesired example in which unnecessary features have appeared due to unsupervised learning, and the result R20 is a desirable example suitable for diabetes prediction Respectively.

  That is, in the result R10, a series of patients in the medical data are classified into a cluster C11 for healthy subjects, a cluster C12 for fractures, a cluster C13 for colds, and a cluster C14 for diabetics ( That is, it has been classified. In particular, since “fractures” and “colds” are basically independent of “diabetes”, such a classification of R10 is inappropriate as a risk prediction for “diabetes”. is there. Appearance of “unnecessary features” such as “fractures” and “cold” results in inability to accurately analyze “diabetes” itself.

  On the other hand, the result R20 appropriately categorizes a series of patients in the medical data with respect to the disease risk prediction of “diabetes”. That is, it is categorized into a cluster C21 for healthy subjects, a cluster C22 for diabetic reserve army A, a cluster C23 for diabetic reserve army B, and a cluster C24 for diabetic patients. The results R20 are all categorized from the viewpoint of diabetes, and since the “unnecessary features” such as “fracture” and “cold” do not appear like the results R10, it is possible to appropriately examine the disease risk of diabetes It has become a result.

  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 clustering apparatus, method, and program capable of performing clustering suitable for predicting the risk of a specific disease.

  In order to achieve the above object, the present invention is a clustering apparatus that accepts designation of a disease name to be analyzed and an onset period of the disease from a user, and evaluates the health of each subject over multiple ages in each era. The weight calculation unit that analyzes the medical data for weight calculation as a word list as a result and calculates the degree of relevance to the onset of the disease through the onset period as a weight for each word; Each of a series of subjects receives a feature vector related to the health status of the subject in the form of a bug of words with words related to health evaluation, and calculates a weight calculated for each word for each word frequency of the feature vector. And a clustering unit for clustering the series of subjects to be analyzed by latent topic analysis. And wherein the Rukoto.

  In addition, the present invention is a clustering device, which accepts designation of a disease name to be analyzed and an onset period of the disease from a user, and is a word list as a result of evaluating each subject's health over multiple ages in each era A weight calculation unit that analyzes medical data for weight calculation as a word, calculates a degree of relevance to the onset of the disease through the onset period as a weight, and a series of subjects to be analyzed Each of them receives feature vectors related to the health status of the subject in the form of a bug of words based on words related to health evaluation, and performs latent topic analysis on the matrix data listing the feature vectors of the series of subjects to be analyzed. The matrix data is decomposed into a product of the relationship matrix θ of the subject and the topic and the relationship matrix Φ of the topic and the word, The subject and topic-corrected relationship matrix θ ′, the topic, and the subject and topic obtained by performing the latent topic analysis on the matrix data W * D obtained by assigning the weight W calculated for each word to the matrix data D The corrected relation matrix Φ ′ of the word is obtained as an approximate value without performing latent topic analysis, and the corrected relation matrix θ ′ obtained as the approximate value is obtained as a clustering result of the series of subjects to be analyzed. A clustering unit, wherein the clustering unit decomposes the weight W into a product of a plurality of transformations determined to be an identity transformation plus a minute transformation when the approximate value is obtained. Are sequentially added as perturbations to the relationship matrix θ and the relationship matrix Φ, so that the modified relationship matrix θ ′ and the modified relationship matrix Φ ′ are respectively approximated values. It is characterized by obtaining it.

  In addition, the present invention is a clustering method, which accepts designation of a disease name to be analyzed and an onset period of the disease from a user, and a word list as a result of evaluating each subject's health over multiple ages in each era A weight calculation stage that analyzes medical data for weight calculation as a word, calculates a degree of relevance to the onset of the disease through the onset period as a weight, and a series of subjects to be analyzed Each of them receives a feature vector related to the health condition of the subject in the form of a bug of words related to the health evaluation, and assigns a weight calculated for each word to each word frequency of the feature vector. And a clustering step of clustering a series of subjects to be analyzed by latent topic analysis. To.

  In addition, the present invention is a clustering method, which accepts designation of a disease name to be analyzed and an onset period of the disease from a user, and a word list as a result of evaluating each subject's health over multiple ages in each era A weight calculation stage that analyzes medical data for weight calculation as a word, calculates a degree of relevance to the onset of the disease through the onset period as a weight, and a series of subjects to be analyzed Each of them receives feature vectors related to the health status of the subject in the form of a bug of words based on words related to health evaluation, and performs latent topic analysis on the matrix data listing the feature vectors of the series of subjects to be analyzed. Thus, the matrix data is decomposed into a product of the relationship matrix θ of the subject and the topic and the relationship matrix Φ of the topic and the word, The subject and topic-corrected relation matrix θ ′ and the topic obtained by performing latent topic analysis on the matrix data W * D to which the weight W calculated for each word is given to the matrix data D And the corrected relationship matrix Φ ′ of the word as an approximate value without performing the latent topic analysis, and the corrected relationship matrix θ ′ determined as the approximate value is clustered for the series of subjects to be analyzed. A clustering stage as a result, and in the clustering stage, when obtaining the approximate value, the weight W is decomposed into a product of a plurality of transformations determined to be a transformation of the identity transformation and the transformation By sequentially adding a plurality of transformations as perturbations to the relationship matrix θ and the relationship matrix Φ, the modified relationship matrix θ ′ and the modified relationship matrix Φ ′ are respectively It is obtained as an approximate value.

  Furthermore, the present invention is a program for causing a computer to function as the clustering device.

  According to the present invention, for each word, the degree of relevance to the onset of the disease through the onset period is calculated as a weight, and when a series of subjects to be analyzed given in the form of bug of words are clustered, Since the weight is taken into consideration, a clustering result suitable for predicting the risk of a specific disease can be obtained.

FIG. 10 is a diagram illustrating an example schematically explaining a problem that accuracy is lowered because irrelevant features are handled in typification by unsupervised learning when prediction of a specific disease risk is desired. It is a functional block diagram of the clustering device concerning one embodiment. It is a figure which shows the typical example of the medical data which a weight calculation part reads. It is a figure which shows the example which calculated the weight as a co-occurrence rate about each word. It is a figure for demonstrating the age which investigates co-occurrence or non-co-occurrence in medical data. It is a figure for demonstrating the clustering by a latent topic analysis. It is a figure for demonstrating the correction calculation in 2nd embodiment.

  FIG. 2 is a functional block diagram of the clustering apparatus according to one embodiment. The clustering apparatus 10 includes a weight calculation unit 11, an analysis target documenting unit 12, and a clustering unit 13. The outline of each part is as follows.

  The weight calculation unit 11 reads medical data as all patient / age data for weight calculation, receives designation of the disease name and onset period of the analysis target from the user who performs the data analysis, and details processing will be described later To obtain the weight for each word and pass it to the clustering unit 13. By using the weight information for each word, there is a risk that the specified disease will develop after the specified onset period (for example, one year if the specified onset period is one year) It is possible to extract the patient with high accuracy in the clustering unit 13 and to analyze the risk factor in detail by referring to the clustering result.

  FIG. 3 is a diagram showing a schematic example of the medical data read by the weight calculation unit 11. In the example of FIG. 3, all the medical data relates to patients A to F, and medical data for each age of each patient is given. That is, medical data for patient A in each age group between 40 and 43, medical data for patient B in each age group between 42 and 45, and medical care for patient C in each age group between 41 and 43 years Data, medical data for patient D in each age group from 50 to 53 years old, patient E for medical data in each age group from 51 to 53 years old, and patient F in each age group from 52 to 55 years old Medical data.

  As specific contents of the medical data given for each age of each patient, receipt data (medical specification information) and / or medical examination data can be used. In the receipt data, the name and symptom diagnosed for each patient, the prescribed drug and its amount, and the like are described along with the date and time. In the medical examination data, each item of the medical examination such as blood pressure, weight, etc. and the evaluation value thereof are described together with the date and time of diagnosis for each patient.

  Therefore, the medical data is a list of specific words that reflect the health status of each patient for each age. In addition, depending on the type of word, a numerical value is associated. For example, the amount is linked to a word representing the name of a medicine prescribed in the receipt data. The evaluation value is associated with each item of the blood test in the medical examination data.

  In addition, by referring to the medical data, if there is an onset of a specific disease at a specific age for each patient, information to that effect can be obtained. In the example of FIG. 3, patient A developed hyperlipidemia at age 43, patient D developed fatty liver at age 51, and patient F developed diabetic nephropathy at age 52. doing. In order to obtain such information, it is preferable that the medical data is configured to include receipt data including information on the diagnosis of the disease. As long as the information can be obtained, the receipt data may not necessarily be included.

  In addition, as specific examples of the description items of the medical data, descriptions in a column C2 in FIG.

  The analysis object documenting unit 12 reads a series of medical data of patients to be analyzed for health, documents the medical data of each patient, and passes it to the clustering unit 13.

  The medical data read by the analysis object documenting unit 12 is the same type as that read by the weight calculation unit 11 and includes receipt data and / or medical examination data. Here, data about a patient different from the data read by the weight calculation unit 11 may be read, or a part or all of the data read by the weight calculation unit 11 may be read as an analysis target. It may be.

  However, as shown in FIG. 3, the data read by the weight calculation unit 11 is for a series of ages for each patient. However, the data read by the analysis target documenting unit 12 is one for each patient. It consists only of ages. In the example of FIG. 3, assuming that the medical data at the current age (for example, 2014) of each patient A to F is the last part shown in the figure, the analysis target documenting unit 12, for example, Medical data for the current age (A's 43 years old, B's 45 years old, ..., F's 55 years old medical data) can be read.

  The analysis object documenting unit 12 documents the medical data of each age of each patient as described above. This documentation corresponds to pre-processing that is converted in advance into a format that can be clustered by latent topic analysis in the clustering unit 13 in the subsequent stage.

Specifically, the analysis object documenting unit 12 performs documenting in a Bag of Words format that is a well-known format. That is, a plurality of m words i _1, i ₂ representing the health, ..., a i _m in the form of a m-dimensional vector word frequency of how used many times each and document. The plurality of words can be created from those appearing in the medical data.

  In other words, in the case of receipt data, each word representing the health status, health-related items, and health-related items described in the diagnosed disease name and symptom, prescribed medicine, etc. is used as it is, and the word frequency vector in Bug of Words is used. Use as a word for each element. Similarly, even in the medical examination data, each word representing the inspection item is used as it is as each word in Bug of Word. Moreover, if the description item is composed of phrases, only words may be extracted and used with a predetermined word extraction filter.

  Regarding the frequency, when a numerical value is associated with a word as in the case of the amount of medicine exemplified above or a blood test item, it is associated with a word according to a conversion rule set in advance for each word. Can be converted into a frequency value. Further, it may be documented in such a manner that the frequency is “1” if the word is in the medical data, and the frequency is “0” if the word is in the medical data, regardless of whether or not the numerical value is associated with each word.

  In addition, for the documentation in the analysis object documenting section 12 including frequency correspondence, Japanese Patent Application No. 2013-159323 (numerical data analysis apparatus and program) and Japanese Patent Application No. 2013-163207 (numerical data analysis) Apparatus and program), Japanese Patent Application No. 2013-217817 (numerical data documentation apparatus and program), and the like may be used. In addition, a method of associating the frequency described above with each word of each element of the word frequency vector in Bug of Words may be used properly.

  In the clustering unit 13, by performing a latent topic analysis such as LDA on the word frequency vector (corresponding to the feature vector of each patient to be analyzed) sent from the analysis target documenting unit 12, Obtain clustering results for each patient.

  In this embodiment, in particular, the feature vector of each patient to be analyzed is not clustered as it is, but the clustering unit 13 performs clustering after multiplying the weight for each word obtained by the weight calculation unit 11 in advance. A clustering result specialized in terms of whether or not a specific disease can develop within a specific period can be obtained.

That is, the feature vector obtained from the analysis object documenting unit 12 and giving the respective frequencies of a predetermined plurality of m words i ₁ , i ₂ ,..., I _m for each patient A is V [A] = ( f _{1 [A]} , f _{2 [A]} ,..., f _{m [A]} ), the clustering unit 13 does not cluster the feature vector V [A] as it is, but is obtained by the weight calculation unit 11. each word i _1, i _2, the, ..., each of the weights w ₁ of i _m, w _2, ..., obtained by multiplying _{w m, V '[a]} = (w 1 * f 1 [a], w ₂ * f _{2 [A]} ,…, w _m * f _{m [A]} ) is used as a clustering target, and latent topic analysis such as LDA is performed.

Here, as described later in the detailed description of the weight calculation unit 11, each word i _1, i ₂ to the multiplying, ..., each of the weights w _1, w ₂ of the i _m, ..., w _m is the specific disease It is calculated | required as a value which makes the word frequency large about the word relevant to onset within a specific period, and makes the word frequency small about the word which is not related. Therefore, by clustering the feature vector V ′ [A] multiplied by the weight for each patient A as a feature vector, a clustering result specialized in terms of whether or not a specific disease can develop within a specific period is obtained. be able to.

In the example of FIG. 1, by using the weights w ₁ , w ₂ ,..., W _m , the “fracture” and “cold” appearing in the result R10 as factors unrelated to “diabetes” to be analyzed. The frequency of related words can be reduced, and at the same time, the frequency of words related to “diabetes” can be increased. In this way, by converting the feature vector to be clustered into V ′ [A], which does not use the original data V [A] as it is, but largely reflects the features related to “diabetes”, as a result R20 As shown, a clustering result specialized for “diabetes” can be obtained.

Then, each word i _1, i ₂ as described above, ..., weights w _1, w ₂ of the i _m, ..., will be described in detail weight calculator 11 for determining the w _m.

  Here, a case where a user who performs data analysis inputs that he / she wants to predict a person who will develop diabetes after one year as a designation input to the weight calculation unit 11 will be described as an example. That is, the user designates “diabetes” as the disease name to be analyzed, and designates “after one year” as the onset period.

  In this case, among the words in the medical data for each age of each patient as illustrated in FIG. 2 as an input to the weight calculation unit 11, it is related to the case of developing “diabetes” in “one year later”. A weight that preferentially leaves a word is to be obtained.

  To calculate the weights specifically, the “co-occurrence” relationship between the medical data 1 year ago (at the age of n-1) and the medical data 1 year later (at the time of n years) is shown. Look at it. That is, the drug name “a” is described in the medical data of patient A one year ago (at the time of n-1 years), and the diagnosis result “ When “diabetes” is described, “co-occurrence” has occurred, it is considered to be related, and the vector v [a] = (relevance, non-relevance) for the drug name “a” In the vector element of = (number of co-occurrence, number of non-co-occurrence), “1” is added as a contribution to the number of co-occurrence cases from patient A. Similarly, the medical name “a” is written in the medical data of patient B 1 year ago (at the age of m-1), but the medical data 1 year later (at the time of m age) If there is no description of “diabetes”, “1” is added to the non-relevance degree of the vector v [a] as a contribution to the number of non-co-occurrence cases from patient B.

  Note that “adding” “1” to the relevance level or the non-relevance level as a contribution in the above means the next operation. That is, after setting (relevance, non-relevance) = (0, 0) as an initial value in the vector v [a] for the drug name “a”, the relevance (co-occurrence) or The final vector v [a] is calculated by examining whether it is unrelated (non-co-occurrence) The operation of “aggregation” when obtaining a value is an operation of “adding” “1”.

  In this way, all the words i constituting the element of the feature vector obtained by the analysis object documenting unit 12 (the drug name “a” is used as an example of the word i in the above description) are input to the weight calculating unit 11. All patients included in the medical data to be treated fall under the category of co-occurrence ("diabetes" occurs at age n and the word i is present in data at age n-1) or The total number of co-occurring patients and non-co-occurrence by investigating whether it falls under the origin ("diabetes" has not developed at age n and the word i is present in the n-1 year-old data) And the total number of patients. Here, whether or not “diabetes” has developed at the age of n may be determined by whether or not “diabetes” is included as a word in the n-year-old data. Similarly, when the user designates a specific disease other than “diabetes”, the determination may be made based on the presence or absence of a word representing the designated specific disease.

  It should be noted that patients who developed “diabetes” at n years old and the word i does not exist in the n-1 year old data, and those who did not develop “diabetes” at n years old, and n-1 year old data And the patient who does not have the word i can also be present in the medical data. Such patients are treated as not counting in the total number of co-occurring and non-co-occurring patients in word i.

Then, the vector v [i] = (relevance [i], non-relevance [i]) = (total number of co-occurring patients, total number of non-co-occurring patients) for the word i is formed by the obtained total number. . Based on the configured vector v [i], the weight w _i related to the word i can be obtained by the following equation (1).

As is clear from the formula (1), the weight w _i has the meaning of the co-occurrence rate related to the onset of the specific disease for the word i in the specific period (the onset of diabetes in one year).

FIG. 4 is a diagram showing an actual example in which the weight as the co-occurrence rate is calculated for each word in a table format. The column C1 indicates the number of lines to be referred to in the explanation, the column C2 indicates the description contents in the medical data corresponding to each word i, and the column C3 indicates the relevance [i] in the expression (1). That is, the number of patients who developed “diabetes” one year later and the word i was present in the medical data one year ago (the number of co-occurrence) is shown in column C4, the degree of unrelevance [i] in formula (1), ie one year The number of patients who do not develop “diabetes” later and the word i exists in the medical data one year ago (the number of non-co-occurrence) is the relevance [i], relevance [ The co-occurrence rate (weight w _i ) calculated from i] using equation (1) is described. In FIG. 4, only a part of the medical data is displayed by sorting in the order of data with the highest co-occurrence rate.

  The column C2 also shows an example of the description items of the medical data for each patient in each age. The word i for the disease name indicating that a specific disease has been diagnosed is “Type 2 diabetic nephropathy stage 1” on line 16, “esophageal hernia” on line 30, “type 2” on line 31 Diabetes and renal complications ”,“ severe infection ”on line 33. As the name of the specific disease designated by the analysis user for the weight calculation unit 11, in addition to “diabetes”, such a name can be designated, and a predetermined disease name to be analyzed may be designated.

  In column C2, the word i relating to prescription and treatment is shown in the 20th and 34th lines, and the word i relating to instruments used in the treatment is shown in the 22nd and 25th to 27th lines. The other lines show the names and amounts of drugs used. For example, in the first line, “agent A 0.3 mg” is described, and “agent A” is used in an amount of “0.3 mg”. Only “agent A” may be used as the word i.

  In addition, by analyzing all the medical data input in the weight calculation unit 11, the relationship between each word i and a specific disease that develops at a specific deadline is co-occurrence or non-co-occurrence in each patient as described above. In determining whether or not, it is possible to examine the age of medical data of each patient (the age n for confirming the onset of diabetes and n-1 which is one year before). In this regard, various embodiments are possible. FIG. 5 is a diagram illustrating an example of a part of medical data input to the weight calculation unit 11 for explaining each embodiment related to the age to be examined. As in the above description, a case where an attempt is made to predict a person who will develop diabetes after one year will be described as an example. In FIG. 5, there are patients A to C, which are patients A to C independent of the example of FIG.

  In FIG. 5, the acquisition date (year) of medical data acquired at each age (age) of each patient is shown as the time axis on the uppermost side, and patient A is 41 years old or older in each year from 2010 to 2014. Medical data for 45 years old has been acquired, medical data for patients B from 47 years old to 50 years old has been acquired in each year from 2011 to 2014 for patient B, and each year from 2010 to 2014 for patient C Exemplifies the case where medical data of 45 years old to 49 years old is acquired. And as indicated by the broken mark, patient A developed diabetes at the age of 45 (2014), patient B developed diabetes at the age of 48 (2012), and patient C had diabetes. It is assumed that it does not develop.

  As one embodiment of setting the target age for determining co-occurrence or non-co-occurrence, as shown in the frame F1, age n to check whether or not diabetes has occurred, and one year before that The age n-1 for checking the co-occurrence / non-co-occurrence of each word i may be set to be common to all examinees according to the calendar.

  In the example of frame F1, for each patient, the medical data as of 2014 AD is first checked to see if it has developed diabetes, and then each word i in the medical data as of 2013 AD, one year before, is checked. , The word i included in the medical data of patients who have diabetes as of 2014 is counted as a co-occurrence (the contribution from the patient is counted as a co-occurrence for the word i), and diabetes as of 2014 The word i included in the medical data of the patient who does not develop is counted as non-co-occurrence (for the word i, the contribution from the patient is counted as non-co-occurrence).

  In the example of patients A to C in FIG. 5, word i in 2013 medical data of patients A and B is counted as co-occurrence, and word i in 2013 medical data of patient C is counted as non-co-occurrence. Will be. Regarding Patient B, diabetes developed for the first time in 2012, but the symptoms continued until 2014, and it was counted as a co-occurrence on the assumption that this can be determined from the 2014 medical data. ing.

  In addition, when investigating a common age n, n-1 in all patients according to the calendar, a predetermined setting such as the present age n may be used for setting the age n (and n-1). The user who performs the analysis may specify the age n when the disease name and the onset period are specified in the weight calculation unit 11.

  As another embodiment of setting the target age for determining co-occurrence or non-co-occurrence, the weight calculation unit 11 first calculates the medical data of all age ranges where the medical data exists for each patient. By examining whether or not there is diabetes onset, the age of first onset for patients with onset is n years old, and all words i included in the medical data of n-1 years are shared. It can be counted as a start. On the other hand, for a patient who has no onset, any age in which medical data exists can be defined as n years old, and all the words i included in the medical data of n-1 years old can be counted as non-cooccurrence.

  In the example of FIG. 5, for patient A, as shown as range R1, n = 45 years is set as the age of first onset of diabetes, and the word i in the medical data of n-1 = 44 years co-occurs. Count as. For patient B, as shown as range R2, n = 48 years is set as the age of first onset of diabetes, and word i in the medical data of n-1 = 47 years is counted as a co-occurrence. For patient C, n, n-1 years are arbitrarily set, such as ranges R3, R4, etc., and word i in the medical data of n-1 years is counted as non-co-occurrence.

  As described above, the weight calculation method in the weight calculation unit 11 for performing clustering specialized in the viewpoint of whether or not to develop diabetes as an example of the specific disease after one year has been described. It is possible to calculate weights for clustering specialized in terms of whether or not a specific disease will develop in 2 years, 3 years, and generally N years after receiving a specified input from the analysis user. is there.

  In other words, if the disease develops after one year, the patient is determined from the information on whether or not the specific disease has occurred at the age n and the word i in the medical data at the age n-1 in the medical data at the age n and n-1. Co-occurrence / non-co-occurrence was determined every time, and the weight of each word i was obtained according to equation (1).

  Exactly in the same way, if it occurs after N years, in the medical data of age n, nN, for each patient from the information of whether or not a specific disease has occurred at age n and the word i in the medical data of age nN Co-occurrence / non-co-occurrence can be determined, and the weight of each word i can be obtained according to equation (1). The setting for each patient of the ages n and n-N may be performed in the same manner as described with reference to FIG.

In addition, the weight calculation unit 11 may obtain a weight for performing clustering specialized in the viewpoint of whether or not a specific disease will occur within N years in response to a designation input from the analysis user in the clustering unit 13. Is possible. In this case after each k years after 1 year until after N years (k = 1, 2, ..., N) of whether developing a particular disease, each word _{i 1, i 2, ...,} i m After calculating the respective weights w _{1 (k)} , w _{2 (k)} ,…, w _{m (k)} , and taking the average of each year as shown in equation (2), specify within N years Weights w _{1 (within N years)} , w _{2 (within N years)} , ..., w _{m ( within N years )} for clustering specialized in terms of whether or not to develop a disease can be obtained.

  In addition, the weight calculation unit 11 receives a designation input from the analysis user, and performs clustering that specializes in whether or not (at least one) of a plurality of specific diseases will develop in N years. It is also possible to obtain weights to be performed by the unit 13. In this case, the weights related to the occurrence of individual specific illnesses in N years after are obtained, and the weights for the individual specific illnesses are averaged over the plurality of illnesses in the same manner as the above formula (2). do it.

  Further, the weight calculation unit 11 receives a designation input from the analysis user, and when the specific disease is not completely cured and the treatment is continued for a plurality of years, the specific disease is not already developed (not already developed). It is also possible to obtain weights for the clustering unit 13 to perform more specialized clustering for the first occurrence after a specific period.

For example, diabetes is not completely cured and treatment continues for multiple years. At this time, since a specific antidiabetic drug is continuously used, the word representing the drug has a strong tendency to co-occur in many patients who have already developed diabetes. Therefore, if the calculation is performed as it is using the method described above, the relevance [i] of the word i representing the specific diabetes therapeutic drug is a large value, and the weight w _i is also a large value.

  However, if you increase the weight of a word representing a drug that is used continuously after the onset of such a specific disease, you want to cluster patients who develop the specific disease for the first time with high accuracy and analyze the risk factors In addition, the accuracy of clustering is lowered.

Therefore, the weight may be calculated in such a way as to eliminate the influence of words that appear continuously after the onset of a specific disease, such as a therapeutic drug for diabetes. In order to eliminate the influence, the relevance [i] _{(same age)} in the _same age for each word i is calculated and subtracted from the weight calculated by the method described above. Specifically, it is as follows.

Here, as an illustrative example, a case is used in which the analysis designation from the user first develops the specific disease one year later and the second designation is the first onset of the specific disease one year later. In this case, by considering only the first designation in the above method, for each patient, confirm whether or not a specific disease has developed at n years old, and by examining each word i at n-1 years old, Co-occurrence / non-co-occurrence of each word i was determined. Furthermore, by considering the second designation, whether or not each patient has already developed a specific disease (specific disease that has been confirmed to have occurred at the age of n) at the age of n-1 for each patient. For each word i, the number of patients who have already developed a specific disease at the age of n-1 (number of pre-existing patients ₎ may be determined as relevance [i] _{(same age)} .

Then, by further reducing the obtained relationship [i] _{(same age)} in the above equation (1), the weight w _i may be calculated as in the following equation (3). In addition, when using the formula (3), the calculation of the degree of association [i] in the prerequisite formula (1) applies an embodiment that does not consider the occurrence of the specific disease at the time of n-1 years. Shall.

  Whether or not a specific disease has already developed at the age of n-1 is exactly the same as when the presence or absence of a word for a specific disease was examined with reference to the medical data at the time of n years old. It can be determined by the presence or absence of a specific disease word in the medical data at the age of 1 year. The determination of the onset of a specific disease at other age points described below is also possible.

In the above description, the case of developing a specific disease for the first time after one year has been described. However, when handling the case of developing a specific disease for the first time within N years by user designation, the following may be performed. That is, it is confirmed whether or not a specific disease has occurred at the age of n, and the degree of relevance [i] and the degree of unrelevance [i] are obtained by examining each word i at the age of nN. The total number of patients who have already developed is regarded as relevance [i] _{(same age)} , and the weight may be obtained from the above equation (3).

Similarly, when handling a specific disease that occurs for the first time in “after” N years instead of “within” N years by user designation, nN, n-N + 1, n-N + 2, What is necessary is just to obtain | require weight from said Formula (3), calculating | requiring the total number of patients who have already developed the specific disease at any time of n-1 years old as a relationship [i] _{(same age)} .

  Next, as another embodiment of the operation of the clustering apparatus 10, an embodiment (a second embodiment) in which the amount of calculation can be suppressed as compared with the embodiment described above (referred to as the first embodiment). ).

  In other words, in the first embodiment, the analysis user designates the specific disease and onset period and other contents that he / she wants to analyze the risk factor with high accuracy to the weight calculation unit 11, and the word i specialized for this is specified. Each weight was calculated. The clustering unit 13 performs latent topic analysis such as LDA after multiplying the weight calculated from the user-specified condition with the documented medical data of a series of patients to be analyzed, and specializes in the user-specified condition Clustering results were obtained. Therefore, in the first embodiment, the clustering unit 13 needs to perform latent topic analysis with a corresponding calculation amount such as LDA for each individual user-specified condition. If there were many types of user-specified conditions, it was necessary to carry out calculations with a corresponding amount of calculation such as LDA.

  On the other hand, in the second embodiment, the latent topic analysis with a corresponding calculation amount such as LDA in the clustering unit 13 is performed only once as a common calculation process that does not depend on user-specified conditions that can exist in many types. By doing so, the amount of calculation is suppressed, and the obtained clustering result is corrected by the weight calculated according to the user-specified condition, thereby obtaining a clustering result specialized for the user-specified condition. Since the correction process can be performed by simple calculation, unlike the case where the LDA is applied several times individually in the first embodiment, the calculation amount can be suppressed in the second embodiment. Specifically, it is as follows.

  First, in the second embodiment, clustering by latent topic analysis by LDA or the like performed in a form that does not depend on user-specified conditions in the clustering unit 13 is performed by directly using a series of patient feature vectors obtained from the analysis target documenting unit 12. Cluster by shape. That is, as described above, V ′ [A] obtained by multiplying the feature vector V [A] of the patient A by the weight in the first embodiment is a clustering target, whereas in the second embodiment, V ′ [A] is V [A] is clustered as it is.

  FIG. 6 is a diagram for explaining the result of clustering by the latent topic analysis in the clustering unit 13 and its meaning. First, FIG. 6 will be described as a premise for explaining the correction of the clustering result by the weight in the second embodiment to be described later.

In FIG. 6, all data to be analyzed obtained from the analysis target documenting unit 12 is D, and is data in a matrix format in which the feature vectors V [A] of each patient A are listed as row vectors. That is, each row of the matrix D is "document" as Baguobuwazu representing the health condition of each patient, the elements aligned in the column direction each word i _1, i _2, ..., a frequency count of i _m.

  In latent topic analysis such as LDA, a document (a document corresponding to each patient) is treated as a bug of words, that is, a word and its appearance frequency, and the topic is estimated in the document. For example, words such as “stock price”, “increased sales and profits” will appear from the “economy” topic, and words such as “baseball” and “soccer” will appear from the “sports” topic.

  This is because the observed bug-of-words expression, that is, the relationship matrix D between the word i (column component) and the document u (row component), as shown in FIG. 6, the word i (column component) and the topic k (row component). , And the product of the relationship matrix θ of the document u (row component) and the topic k (column component). This topic k is estimated by LDA, and the decomposition result “D = θ × Φ” as shown in FIG.

  Thus, the topic model represented by LDA assumes that each document has a unique topic ratio, and words are generated at a ratio specific to the topic after selecting the topic according to this topic ratio. .

  The clustering result of each patient is expressed as a relationship matrix θ between the document (patient) u and the topic k, but various modes are possible as to how to interpret them specifically.

For example, each of K topics k _i corresponds to a cluster, and each patient belongs to the cluster corresponding to the topic having the maximum topic ratio (k ₁ , k ₂ ,…, k _K ). A clustering result can be obtained by giving an interpretation. Alternatively, each of K topics k _i corresponds to a cluster, and each patient has the number of contributions of the topic ratio (k ₁ , k ₂ ,…, k _K ) (in integers such as 1 or 0) (The number of people expressed in decimal), and the clustering result can be obtained by interpreting that each cluster belongs to each cluster.

  The use of weights in the clustering unit 13 in the second embodiment will be described above on the premise of the details of clustering in the latent topic analysis as shown in FIG. Note that the calculation of the weight for each word i in the weight calculation unit 11 upon receiving a designation input such as a specific disease from the user is exactly the same as in the first embodiment.

In the clustering unit 13, in the decomposition result “D = θ × Φ” of all the data D composed of the intact feature vector V [A] of each patient A to which no weight is applied, the weight calculation unit 11 according to the designation input from the user Express the clustering result by performing simple correction calculation (described later) using the information on the weights w ₁ , w ₂ ,…, w _m of the words i ₁ , i ₂ ,…, i _m Assuming that the matrix “θ” is modified to “θ ′” and the matrix “Φ” is modified to “Φ ′”, the final clustering result is obtained.

  In order to explain the correction calculation, first, the decomposition result “D = θ × Φ” of all data D is displayed as an element as in the following equation (4). That is, the matrix D has n rows and m columns, the number of patients is n, and the number of words is m. The matrix θ is n rows and K columns, the number of patients is n, and the number of topics is K. The matrix Φ has K rows and m columns, similarly, the number of topics is K, and the number of words is m.

  FIG. 7 is a diagram listing the main stages of the calculation in order to explain the correction calculation. First, as shown in the column [10], the equation (10) is the same as the equation (4) displaying the above elements, and the result obtained by decomposing all the data D without weighting into “D = θ × Φ” is as follows. Represents. The policy of the correction calculation is as shown by the arrow, and the decomposition result “W * D = θ ′ × Φ of the weighted data W * D as shown in the equation (20) starting from the equation (10). '"Is reached by repeated successive approximations. Here, the calculation of each step in the successive approximation is sequentially shown in the columns [1], [2], ..., [i], ..., [M], and finally the equation (10 ) And equation (20) are shown in column [20].

In addition, in Expression (20) and other places, “*” is a Hadamard product, and the matrix W is the weight w ₁ , w _{2 of} each word i ₁ , i ₂ ,..., I _m obtained by the weight calculator 11. ,..., W _m is a weighting matrix W obtained by arranging a row vector having a common weight for each patient by arranging the total number n of patients. That is, when the right side of the equation (20) is displayed as a component following the component display of the equation (4), the following equation (5) is obtained.

  Each step of the sequential calculation is as follows, and the calculations shown in the columns [1], [2], ..., [i], ..., [M] are executed in this order. First, in the column [1], equations (1-1) to (1-4) are calculated in this order.

First, in equation (1-1), equation (10), which is the starting point, is approximated. That is, the weight matrix W in Equations (5) and (20) and the common size (n rows and m columns), and a matrix W ₁ whose elements are sufficiently close to 1 is defined as a Hadamard product with respect to Equation (10) An approximate value of the result obtained by decomposing the multiplied "D ₁ = W ₁ * D" by latent topic analysis such as LDA (hereinafter referred to as "LDA etc.") has already been obtained as a result of decomposition by LDA "D As a minute change with respect to = θ × Φ, “D ₁ ≈θ ₁ × Φ” is assumed assuming that only the portion of “θ” is changed to “θ ₁ ” without changing the portion of “Φ”. The symbol “≈” means “approximately equal”. In FIG. 7, the symbol “≈” is drawn as “˜” (tilde) added to the normal “=”, but it has the same meaning.

By approximation of the formula (1-1), the approximate value “θ ₁ ” is obtained by multiplying “D ₁ ” by the inverse matrix Φ ⁻¹ of Φ from the right as shown in the formula (1-2). It can be obtained as “θ ₁ = D ₁ × Φ ⁻¹ ”. Here, the inverse matrix Φ ⁻¹ may be obtained as a known pseudo inverse matrix (Moore Penrose pseudo inverse matrix) for the matrix Φ.

Then, the following approximate expressions (1-3) and (1-4) are carried out by performing the same approximate calculation on “Φ” by taking over the results of the above expressions (1-1) and (1-2). That is, in the expression (1-3), for the matrix “D ₁ ” defined in the expression (1-1), the weight matrix W in the expressions (5) and (20) is a common size (n rows and m columns). LDA and other potential topic analysis (hereinafter referred to as “D ₂ = W ₂ * D ₁ ”) obtained by multiplying the expression (1-1) by a Hadamard product with a matrix W ₂ whose elements are all sufficiently close to 1 The approximate value of the result of decomposition by LDA etc. is already obtained as the approximate value by the formulas (1-1) and (1-2) as “D ₁ = θ ₁ * As a minute change with respect to Φ, it is assumed that “θ ₁ ” is not changed, and only “Φ” is changed to “Φ ₁ ” as “D ₂ ≈θ ₁ × Φ ₁ ”.

The approximation of the equation (1-3), the approximation value "[Phi _1", it multiplies the inverse matrix theta ₁ ^-1 of theta ₁ with respect to "D _2" as shown in equation (1-4) from left Thus, it can be obtained as “Φ ₁ = θ ₁ ⁻¹ × D ₂ ”. Here, the inverse matrix θ ₁ ⁻¹ may be obtained as a well-known pseudo inverse matrix (Moor Penrose pseudo inverse matrix) for the matrix θ ₁ .

As described above, in the first stage calculation shown in the column [1], by applying an approximation to the starting point equation (10) “D = θ × Φ”, the data matrix “ For the data matrix “D ₂ ” in which “D” is slightly changed, an approximation result “D ₂ ≈θ ₁ × Φ ₁ ” of matrix decomposition by LDA or the like was obtained. From the second stage onward, it is sufficient to continuously perform the approximate calculation exactly the same as the first stage. Therefore, the calculation of an arbitrary i-th stage (i = 2, 3,..., M) shown in the column [i] will be described.

That is, when the calculation at the i-1 stage is completed (when the calculation at the i stage is started), an approximate decomposition result “D _{2 (i-1)} ≈θ _i-1 × Φ _i-1 ” is obtained. ing. Thus, by introducing a "theta _i-1" was minimal change in the formula of the column [i] (i-1) "theta _i" be a n row m column size "D _{2 (i-1)"} Approximate decomposition result “D _2i-1 ” of “D _2i-1 = W _2i-1 * D _{2 (i-1)} ” obtained by multiplying the matrix “W _2i-1 ” whose elements are sufficiently close to ₁ as a Hadamard product = θ _i × Φ _i-1 ”. Then, the specific value of the slightly changed “θ _i ” is obtained by multiplying the pseudo inverse matrix “Φ _i-1 ⁻¹ ” of “Φ _i-1 ” as shown in the equation (i-2). _i = D _2i-1 × Φ _i-1 ^-1 ".

Subsequently, in the same manner as the formula (i-3), formula (i-1), (i -2) is "[Phi _i-1" of introducing "[Phi _i" was slightly changed, "D _2i-1" Approximate decomposition result “D _2i = θ _i ” of “D _2i = W _2i * D _2i-1 ” obtained by multiplying the matrix “W _2i ” with n rows and m columns size and all elements sufficiently close to 1 as Hadamard product XΦ _i ”. The specific values for the minimal change "[Phi _i" is by multiplying a pseudo-inverse "theta _i ^-1" of the formula (i-4) as shown in "theta _i" "Φ _i = _{θ i} ^-1 x D _2i ".

The general i-th case has been described above. The above description also applies to the i = 1 stage if characters “D = D ₀ ”, “θ = θ ₀ ”, and “Φ = Φ ₀ ” are set.

Above-described approximations of the i = 1, 2, ..., a result of repeated until M stage, as shown as formula (M-3) in the column of FIG. 7 [M], the approximate decomposition result "D _2M = theta _M × Φ _M ”is obtained. The decomposition result “W * D = θ ′ × Φ ′” of the equation (20) to be finally obtained is “θ ′ ==” as shown in the equation [13] in the column [20] based on the calculation result of the M stage. θ _M ”and“ Φ ′ = Φ _M ”. The basis for this is shown in the column [20] as equations (11) and (12). In particular, the expression (11) is an expression of a condition that makes it possible to obtain the expression, and the expression (12) is an expression that is a basis for obtaining the decomposition result under the condition of the expression (11).

That is, as shown in the equation (11), matrixes W ₁ , W ₂ , W ₃ , W in which all elements used at each stage of the approximate calculation repeated up to the M = 1 stage are close to 1 as shown in Equation (11). ₄ ,…, W _2i-1 , W _2i ,…, W _2M-1 , W _2M are prepared so that all their Hadamard products are equal to the weight matrix W in equation (20). . Therefore, for Equation (20), Equation (11) and Equations (1-3), (2-3), [M] shown in columns [1], [2],…, [i],…, [M] By substituting…, (i-3),…, (M-3), Equation (12) is obtained. As an approximate decomposition result, θ ′ = θ _M ”and“ Φ ′ = Φ _M ”can be adopted.

Note that all elements used at each stage enabling approximate calculation are matrices W ₁ , W ₂ , W ₃ , W ₄ ,…, W _2i−1 , W _2i ,…, W _2M−1 , Each specific value of W _{2M and} the number of times 2M to repeat the calculation are predetermined values that satisfy Equation (11) and that each matrix is determined to be close to 1 on a predetermined basis. What is necessary is just to set by a method.

For example, as shown in the following equation (6), the matrices W ₁ , W ₂ , W ₃ , W ₄ ,…, W _2i-1 , W _2i ,…, W _2M-1 , W _2M are all mutually valued. It can be set to be equal to the “2M root” of the weight matrix W.

As shown in Equation (6), the “2M power root” of the weight matrix W means that each element w _i (i = 1, 2,..., M) is a 2M power root in the sense of Hadamard product. It may be defined as a 2M root w _i ^{1 / 2M} . The number of repetitions 2M may be set as a number at which each 2M power root element w _i ^{1 / 2M} (i = 1, 2,..., ^M ) is determined to be sufficiently close to 1 on a predetermined basis.

As described above, in the second embodiment, the weight matrix W is a plurality of matrices W ₁ , W ₂ , W that can be determined as a transformation obtained by adding a small transformation to the identity transformation obtained by setting all elements to 1. ₃ , W ₄ ,…, W _2i-1 , W _2i ,…, W _2M-1 , W _2M is decomposed into a product (Hadamard product), and the decomposition result “ For “θ” and “Φ” of “D = θ × Φ”, the decomposed matrices W ₁ , W ₂ , W ₃ , W ₄ ,..., W _2i−1 , W _2i ,…, W _2M-1 , W _2M is added as a perturbation sequentially, so that the decomposition results “W * D = θ ′ × Φ ′” when weighting is applied are actually “θ ′” and “Φ ′”. An approximate value can be obtained without performing a heavy load calculation such as LDA. Here, each stage of the sequential perturbation is as shown as columns [1] to [M] in FIG.

It should be noted that “small conversion” added to the identity conversion is to cause a slight change as a difference value with respect to the identity conversion. For example, when the number of words is m = 3, the identity transformation is defined by the weight (1, 1, 1) of all elements being 1. By adding a small transformation (+0.01, -0.02, +0.03) to this, Weights (1.01, 0.98, 1.03) are obtained. Each of the matrices W ₁ , W ₂ , W ₃ , W ₄ , ..., W _2i-1 , W _2i , ..., W _2M-1 , W _2M decomposed by such weights (similar to the weight matrix W) In addition, it is possible to perform sequential perturbation calculation by arranging common row vectors such as (1.01, 0.98, 1.03) by the total number of patients n. Here, the determination as to whether or not it is a minute conversion may be made according to whether or not each element is close to 0 on a predetermined basis. For example, if it is determined that the conversion is minute when the value is within ± 0.1, the exemplified conversion (+0.01, −0.02, +0.03) is a minute conversion.

  Hereinafter, supplementary items (1) to (4) in the present invention will be described.

  (1) In the medical data of a series of patients to be analyzed, the data of each patient is converted into a feature vector related to the health state of the patient in the form of a bug of words using words related to health evaluation in the analysis target documenting unit 12. Described as (documented). If the conversion is performed in advance on the medical data of a series of patients to be analyzed, the analysis target documenting unit 12 may be omitted.

  (2) The medical data acquisition target has been described as “patient”. However, as long as similar data can be acquired, the present invention is not limited to the “patient” for doctors but the general “target person”. Applicable.

  (3) The present invention can also be provided as a program that causes a computer to function as the clustering apparatus 10. The computer can be configured by known hardware such as a CPU (Central Processing Unit), a memory, and various I / Fs, and the CPU that reads and executes the program functions as each unit of the clustering device 10.

(4) In the description of FIG. 7 regarding the second embodiment, the number of repetitions is 2M, and it is assumed that the number of repetitions is an even number. However, as is clear from the gist of the description, the odd number of times is 2M + 1 or 2M-1. May be. In the case of the former (odd number 2M + 1), matrices W ₁ , W ₂ , W ₃ , W ₄ ,…, W _2i-1 , W _2i ,…, W _2M-1 , W are all elements close to 1 _{Using 2M} and W _{2M + 1} , “θ ′ = θ _{M + 1} ” and “Φ ′ = Φ _M ” may be adopted as approximate decomposition results. In the latter case (odd number 2M-1), matrices W ₁ , W ₂ , W ₃ , W ₄ ,…, W _2i-1 , W _2i ,…, W _2M-2 , W are all elements close to ₁ . _{Using “2M−1} ”, “θ ′ = θ _M ” and “Φ ′ = Φ _M-1 ” may be adopted as approximate decomposition results. Similarly, in the description of FIG. 7, approximate values are obtained in the order of “θ” → “Φ” as “θ ₁ ” → “Φ ₁ ” → “θ ₂ ” → “Φ ₂ ” →. Conversely, it may be obtained in the order of “Φ” → “θ”. The order (“θ” → “Φ” or “Φ” → “θ”) may be determined for each stage i.

  10 ... clustering device, 11 ... weight calculation unit, 12 ... analysis target documenting unit, 13 ... clustering unit

Claims (7)

Accepts designation of the disease name to be analyzed and the onset period of the disease from the user, analyzes the medical data for weight calculation as a word list as a result of evaluating the health of each subject over multiple ages in each era A weight calculation unit that calculates, as a weight, a degree of relevance to the onset of the disease through the onset period for each word;
For each of a series of subjects to be analyzed, a feature vector relating to the health status of the subject is received in the form of a bug of words with words related to health assessment, and is calculated for each word for each word frequency of the feature vector. And a clustering unit for clustering the series of subjects to be analyzed by latent topic analysis after adding a weight. Accepts designation of the disease name to be analyzed and the onset period of the disease from the user, analyzes the medical data for weight calculation as a word list as a result of evaluating the health of each subject over multiple ages in each era A weight calculation unit that calculates, as a weight, a degree of relevance to the onset of the disease through the onset period for each word;
Receive a feature vector for the subject's health status in the form of a bug of words with words related to health assessment, for each of a series of subjects to be analyzed,
By performing a latent topic analysis on the matrix data D enumerating the feature vectors of the series of subjects to be analyzed, the matrix data is subjected to a relationship matrix θ between the subjects and topics and a relationship matrix Φ between topics and words. Break into products,
Subject and topic-corrected relationship matrix θ ′ and topic obtained by performing latent topic analysis on matrix data W * D with weight W calculated for each word on matrix data D And a modified relation matrix Φ ′ of words as an approximate value without performing latent topic analysis,
A clustering unit that sets the modified relation matrix θ ′ obtained as the approximate value as a clustering result of the series of subjects to be analyzed, and
When obtaining the approximate value, the clustering unit decomposes the weight W into a product of a plurality of transformations determined to be an identity transformation plus a minute transformation, and sequentially converts the plurality of transformations into the relation matrix. A clustering apparatus characterized in that the corrected relation matrix θ ′ and the corrected relation matrix Φ ′ are respectively obtained as approximate values by adding as a perturbation to θ and the relation matrix Φ. The weight calculator analyzes the medical data for weight calculation,
Regarding the subject in which the disease name exists in a specific age, for each word existing in the age returned from the specific age by the onset period, the subject is counted as contributing to co-occurrence,
Regarding the subject who does not have the disease name in a specific age, for each word that exists in the age returned from the specific age by the onset period, the subject is counted as contributing to non-co-occurrence,
The clustering apparatus according to claim 1, wherein a weight for each word is calculated as a ratio of the total number of co-occurrence counted to the total number of co-occurrence and non-co-occurrence counted. The weight calculation unit further analyzes the medical data for weight calculation,
Regarding the subject who has the disease name in the specific age, for each word existing in the age that has returned by the onset period from the specific age, the subject is counted as contributing to co-occurrence, and the subject's If the disease name already exists in the age that has returned only to the onset age, the subject is counted as an onset,
Calculating the weight for each word as a percentage of the total number of co-occurrence subtracted from the total number of co-occurrence counted to the total number of co-occurrence and non-co-occurrence counted. The clustering apparatus according to claim 3, wherein the clustering apparatus is characterized in that: A clustering method executed by a computer,
Accepts designation of the disease name to be analyzed and the onset period of the disease from the user, analyzes the medical data for weight calculation as a word list as a result of evaluating the health of each subject over multiple ages in each era A weight calculation stage for calculating, as a weight, a degree of relevance to the onset of the disease through the onset period for each word;
For each of a series of subjects to be analyzed, a feature vector relating to the health status of the subject is received in the form of a bug of words with words related to health assessment, and is calculated for each word for each word frequency of the feature vector. And a clustering step of clustering the series of subjects to be analyzed by latent topic analysis after adding a weight. A clustering method executed by a computer,
Accepts designation of the disease name to be analyzed and the onset period of the disease from the user, analyzes the medical data for weight calculation as a word list as a result of evaluating the health of each subject over multiple ages in each era A weight calculation stage for calculating, as a weight, a degree of relevance to the onset of the disease through the onset period for each word;
Receive a feature vector for the subject's health status in the form of a bug of words with words related to health assessment, for each of a series of subjects to be analyzed,
By performing latent topic analysis on matrix data listing feature vectors of a series of subjects to be analyzed, the matrix data is multiplied by the relationship matrix θ of subjects and topics and the relationship matrix Φ of topics and words. Disassembled into
Subject and topic-corrected relationship matrix θ ′ and topic obtained by performing latent topic analysis on matrix data W * D with weight W calculated for each word on matrix data D And a modified relation matrix Φ ′ of words as an approximate value without performing latent topic analysis,
A clustering stage that uses the corrected relation matrix θ ′ obtained as the approximate value as a clustering result of the series of subjects to be analyzed, and
In the clustering step, when the approximate value is obtained, the weight W is decomposed into a product of a plurality of transformations determined to be an identity transformation plus a minute transformation, and the plurality of transformations are sequentially converted into the relation matrix. A clustering method, wherein the corrected relation matrix θ ′ and the corrected relation matrix Φ ′ are respectively obtained as approximate values by adding as a perturbation to θ and the relation matrix Φ.   A program for causing a computer to function as the clustering apparatus according to any one of claims 1 to 4.