JP2017027510A - Selection device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a selection device that automatically selects data that is usable as a model for performing a medical expense prediction and other health state-related predictions with high accuracy using medical data that could include a partial loss or noise, etc., as input.SOLUTION: A clustering unit 2 clusters health state data in bag-of-words format. A feature amount vector generation unit 3 generates a feature amount vector of each individual data by a topic ratio of clustering results. A threshold determination unit 4 determines, for each element of the feature amount vector, a threshold as a value that would be determined as one that helps to obtain a correlation with a prescribed prediction object associated with a health state. A data division unit 5 determines, for each individual data, the one that contains an element of the feature amount vector whose value equals the threshold or greater as belonging to a prediction possible group and makes it as output of the selection device 10, and determines the other as belonging to a prediction impossible group.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a sorting apparatus, and in particular, receives medical data that may contain defects, noises, etc. as input, and does not require manual labor for extracting predictive factors, so that medical cost prediction and other predictions related to health conditions are highly accurate. The present invention relates to a sorting apparatus capable of sorting data that can be used as a model for performing the above.

  As shown in Patent Documents 1 and 2, etc., there are cases where it is desired to predict medical expenses in order to provide health guidance and alert to the person for the purpose of reducing medical expenses.

JP 2014-225176 A JP 2014-225175 A

  When trying to predict medical expenses in this way, receipt (medical specification) information is often used as input data for prediction, but the feature quantity dimension described in the medical specification is very large, There is a problem that it is difficult to apply a prediction method such as regression as it is.

  For this reason, in general, it is often the case that a disease name, etc. is artificially extracted and used as a feature amount, but the possibility that a potential medical cost predictor exists in the artificially deleted data is denied. Absent. In other words, artificial extraction causes a problem that a medical cost prediction factor is overlooked and the prediction accuracy is lowered. In the first place, even if there is no oversight, it is necessary to manually analyze the enormous receipts and determine whether or not to extract each enormous disease name. Will occur.

  The receipt is receipt information that is generated when a person visits a medical institution, and whether or not to go to the medical institution when a specific state is reached depends on the individual's free will. For this reason, there will always be a group whose future medical expenses cannot be predicted from the current reception. In other words, the receipt information itself for making predictions is in a state that is equivalent to including many deficits and noise, etc., so that the prediction accuracy is lowered regardless of the oversight of the medical cost prediction factor. There are challenges.

  In view of the problems of the prior art as described above, the present invention is related to medical cost prediction and other health conditions without requiring manual operation of extracting predictive factors by inputting medical data that may include defects, noise, and the like. An object of the present invention is to provide a sorting apparatus capable of sorting data that can be used as a model for performing prediction with high accuracy.

  In order to achieve the above object, the present invention is a sorting apparatus, which is a collection of individual data for each subject and each age group. A clustering unit that performs clustering by latent topic analysis to obtain a clustering result, a feature vector generation unit that generates a feature vector of each individual data from a topic ratio of each individual data in the clustering result, and the feature vector For each element, a threshold value determination unit that determines a threshold value as a value that is determined to be correlated with a predetermined prediction target related to the health state, and for each individual data, The element value that is equal to or greater than the threshold value is determined to belong to the predictable group, and the element value does not exist And a determining data distinguishing unit as belonging to the unpredictable group a, and outputs the individual data which are determined to belong to predictable group by the data distinguishing unit.

  According to the present invention, even if the input health status data includes missing data, noise, and the like, after performing clustering by latent topic analysis once, it is determined that each element of the topic ratio has a prediction capability. It is possible to automatically extract only individual data that is equal to or greater than the threshold value and output it as belonging to the predictable group.

It is a functional block diagram of the sorting device concerning one embodiment. It is a figure which shows the typical example of all the medical data input. It is a figure which shows the matrix decomposition result obtained in the clustering of a latent topic analysis. It is a figure which shows the cross tabulation table which a threshold value determination part produces and refers. It is a figure which shows distribution of an explanatory variable and an objective variable in the data of a 1st data group about each value at the time of sweeping a threshold value in a threshold value determination part. FIG. 6 is a diagram showing a schematic example in which the first data group and the second data group are separated by setting a threshold when the data shown in [1] of FIG. 5 is all data. The figure which shows the typical example of the single regression implemented with respect to the data of [2] of FIG. 5, and the typical example of the presence area | region of the target data which counts the hit / not-forecast of prediction by this It is. It is a figure which shows the cross tabulation table created and referred in the feature-value reduction part. It is a figure for demonstrating the content of the repetition process typically. 9 is a generalized table of the cross tabulation tables of FIGS. 11 is a table of probabilities in the dependent model corresponding to the cross tabulation table of FIG. 10. 11 is a table of probabilities in the independent model corresponding to the cross tabulation table of FIG. 10.

  FIG. 1 is a functional block diagram of a sorting apparatus according to an embodiment. The sorting device 10 includes a documenting unit 1, a clustering unit 2, a feature quantity vector generation unit 3, a threshold value determination unit 4, a data distinction unit 5, a model generation unit 6, a feature quantity reduction unit 7, and a prediction unit 8.

  In FIG. 1, each of the units 2 to 7 shown as the repeating unit 20 performs the repetitive processing by inputting / outputting each data in the process flow of the loop structure as illustrated. Details of the data generated by the repetition processing will be described later with reference to FIG. 9, but the processing contents of the units 2 to 7 in each iteration of the repetition processing are the same, and each unit 2 to 7 has each I of the loop processing. The input of the first time (I = 1, 2,..., N: where N is the number of times of loop processing) is received from the front side of each part, and the output of the I time is output to the subsequent stage part of each part.

  In the following, the details of the processing of each part in FIG. 1 will be described, but by using the counter variable I as described above, each loop processing is described as I times (I = 1, 2,..., N). To do.

  The documenting unit 1 reads all medical data as input data for clustering (and prediction based on the result of the clustering) by the sorting device 10 (the clustering unit 2 thereof), and each subject X constituting the all data The documented medical data D (X, n) at each age n (age n) is generated and output to the clustering unit 2.

  Documenting to the medical data D (X, n) is converted into a well-known bug of words format, that is, a document vector format with the frequency (number of occurrences) of each predetermined word as an element. The data D (X, n) is a vector reflecting the health state of the subject X at the age of n. The documenting is performed as preprocessing for enabling clustering in the clustering unit 2 on the subsequent stage side. Specifically, it is as follows.

  First, all input medical data is an evaluation of the health status of a series of subjects at a series of times. Specifically, for example, the results of a health check conducted under a health association, etc. As a result of an inquiry, a receipt (medical remuneration statement), or a combination thereof can be used.

A predetermined plurality of m words i ₁ , i ₂ ,..., I _m that are described in the medical data or are known to be able to be described are prepared in advance, and the documenting unit in 1 by analyzing the text of the medical data in n age of the subject X, the word _{i 1, i 2, ...,} i m vector D (X, n) representing the health condition as the frequency vector of be generated it can.

For example, if the word i _b corresponding to the name of a specific disease in medical examination data or the like exists in the medical data of the subject X at the age of n, the value of the element of i _b of the vector D (X, n) is set to “1”. If it does not exist, the value of the same element can be set to “0”. Similarly, the word i _b such as the prescribed drug name in the receipt data or the like can be set to “1” or “0” depending on whether or not the word exists. The predetermined in number the word i _b is interview may be a value of number of times of the elements appear when appearing more than once in the data or the like, which appeared the as in the case of numerical evaluation items and the like to be described below The value to which the function is applied may be used as the element value.

  In addition, for items to be evaluated numerically, such as body weight and blood test results in health checkup data, a predetermined word corresponding to the item is prepared, and according to a predetermined rule (predetermined function, etc.) according to the evaluation numerical value The frequency of words can be calculated and used as the element value of the vector D (X, n). Regarding the conversion from the evaluation numerical value to the word frequency, Japanese Patent Application Laid-Open No. 2015-32013 (invention name: numerical data analysis device and program) and Japanese Patent Application No. 2013-163207 (numerical data analysis device and Program), Japanese Patent Application No. 2013-217817 (numerical data analysis apparatus and program) may be used.

  In addition to the case of numerical values (quantitative data) as described above, in the case of qualitative data (for example, the presence or absence of a smoking habit described in a questionnaire, etc.) To the value of the element of the vector D (X, n).

As described above, each of the words i ₁ , i ₂ ,..., I _m is a health condition evaluation item in the input medical data (not only a direct evaluation but also a drug name in the receipt data. Each of the predetermined rules (words i ₁ , i ₂ , ..., i _m) for the evaluation result of the subject X at the age of n The document unit 1 generates a document vector D (X, n).

  FIG. 2 shows a schematic example of all medical data input to the documenting unit 1. As shown in the example, it is assumed that all medical data as input has a defect. Here, in order to accurately construct a prediction model for medical expenses, etc., it is desired that each subject has long-term data such as several decades. In many cases, only data over a short period of several years can be used.

  Furthermore, even within the data that exists only for several years for each subject as described above, as described above, the health status of each subject is not necessarily fully evaluated, including defects, noise, etc. It is assumed that it is done. As will be described below, according to the present invention, it is possible to obtain a clustering result as a model that can realize medical cost prediction and the like with high accuracy from input data including such defects and noise.

  In the example of FIG. 2, for example, there is data for 40-year-old to 43-year-old for Mr. A, so in the documentation unit 1, D (A, 40), D (A, 41 ), D (A, 42), and D (D43) are output. For other subjects such as Mr. G and Mr. D, data corresponding to the age at which medical data exist will be output.

In the clustering unit 2, clustering is performed on a series of data D (X, n) of the age n in the series of subjects X output from the documenting unit 1, and the clustering result is sent to the feature vector generation unit 3. Output. As the clustering method, clustering based on latent topic analysis such as LDA (Latent Dirichlet allocation) based on a latent topic model disclosed in the following Non-Patent Document 1 or the like can be used.
[Non-Patent Document 1] D. Blei, A. Ng, and M. Jordan. Latent Dirichlet allocation. Journal of Machine Learning Research, 3: 993-1022, January 2003.

  In particular, data D (X, n) and D (X, m) whose age is different from that of n years old and m years old (m ≠ n) even for the common target person X is clustered as separate data. . For example, Mr. A's data D (A, 40), D (A, 41), D (A, 42), D (D43) in Mr. A in the example of FIG. It becomes. Note that the data D (X, n), D (Y, m), where the target is different from X and Y (where age n and age m may be the same or different) are naturally classified as clustering targets. Become.

  In addition, as indicated by a line L1 in FIG. 1, in the clustering unit 2, in the first iteration process (I = 1), a series of data of the generation n in the series of subjects X output by the documenting unit 1 Clustering is performed using all D (X, n) as input.

  On the other hand, as indicated by a line L2 in FIG. 1, in the clustering unit 2, the data D () output from the feature amount reduction unit 7 (to be described later) is output each time after the second iteration (I ≧ 2). Clustering is performed using all of X, n) as inputs.

  As will be described in detail later with reference to FIG. 9 and the like, the data D (X, n) output from the feature amount reduction unit 7 is configured in the form of a word frequency vector in the same manner as that output from the documenting unit 1. However, all the data output by the documenting unit 1 at the beginning is configured in a form in which the elements to be reduced are excluded from the elements (that is, the form in which the number of vector dimensions is reduced). Those used in model generation by the model generation unit 6 are deleted (that is, the number of configuration data is reduced).

  Therefore, the clustering unit 2 performs clustering by the same method at each I-th iteration and outputs the result to the feature vector generation unit 3, but the input data at that time (and obtained from this) Output data) changes every I times. Similarly, for each of the subsequent units 3 to 7 constituting the repetitive processing unit 20, the input data to be handled (and output data obtained therefrom) will change although the processing contents of each I-time are common. .

  The feature vector generation unit 3 is a document corresponding to the topic ratio of each data D (X, n) in the clustering result output from the clustering unit 2 at each I-th iteration process. Is multiplied as the feature quantity vector of each data D (X, n) and output to the threshold value determination unit 4.

  Here, in order to explain the meaning of the above processing of the feature vector generation unit 3, the contents of the clustering result output from the clustering unit 2 as a premise will be described. That is, the clustering result by the clustering unit 2 is obtained as a matrix decomposition result as shown in FIG.

  As shown in FIG. 3, in latent topic analysis such as LDA, all data D to be classified (each I-th input data D of the clustering unit 2) is each document u given as a frequency vector of the word i (the present invention). Is equivalent to each data D (X, n) output from the documenting unit 1 or the feature amount reducing unit 7). The result of clustering all the data D is the product “D = θ × as a matrix of the θ matrix representing the relationship between the document u and the topic k and the Φ matrix representing the relationship between the topic k and the word i. Φ ”, and the clustering unit 2 outputs the matrix decomposition result.

  Assuming that each topic k corresponds to each cluster in the matrix decomposition result, each row of the θ matrix representing the topic ratio of the document u can be interpreted as the cluster membership probability of each document u. The cluster affiliation probability is the topic ratio of each topic k in each document u, and is a vector representing the health factor of the corresponding original data D (X, n). Therefore, for example, each document u (= each data D (X, n)) can interpret the clustering result as belonging to the cluster corresponding to the topic having the maximum topic ratio value.

  Then, in the feature vector generation unit 3, the original ratio of each topic k in each document u corresponding to each data D (X, n) (vector in each row of the θ matrix in FIG. 3) is the original. By multiplying the total number of words of the data D (X, n), the feature vector of each data D (X, n) is obtained. Hereinafter, the obtained feature vector corresponding to D (X, n) is expressed as θ (X, n). (Note that not the row vectors themselves of the θ matrix of FIG. 3 but the product of the total number of words corresponding to each row vector is represented as θ (X, n).)

Here, the total number of words to be multiplied to obtain the feature vector θ (X, n) is the total number of words of each data D (X, n) at the time immediately after the output of the documenting unit 1 (the number of repetitions I) Does not depend on the total number of words, which is a constant value for each D (X, n). That is, since each data D (X, n) is configured as a frequency vector of a predetermined plurality of m types of words i ₁ , i ₂ ,..., I _m representing a health state as described above, The total number of occurrences of each word i ₁ , i ₂ ,..., I _m as an element is the total number of words, and the feature vector θ (X, n) can be obtained by multiplying the total number of words.

  Note that by multiplying the total number of words, it is possible to generate a feature vector appropriately distinguishing the following cases. For example, Mr. A's receipt at a certain age n has two occurrences of "cold" and "diabetes", and Mr. B's receipt at a certain age m has 50 occurrences of "cold" and "diabetes". While Mr. B is considered to be more severe, the topic ratio (≠ word) is “topic ratio related to cold” and “topic ratio related to diabetes” as Mr. A (age n). There may be cases where there is no big difference with Mr. B (age m). Here, by multiplying the total number of words, it is possible to generate a feature vector that appropriately expresses that Mr. B is more severe.

  The threshold value determination unit 4 performs each element of the feature vector θ (X, n) corresponding to each data D (X, n) output from the feature vector generation unit 3 at each I-th iteration process. Then, a “meaning” threshold value is obtained as the feature amount, and the obtained threshold value is output to the data distinguishing unit 5.

  Here, the “meaning” threshold value as a feature quantity is each element of the feature quantity vector θ (X, n) (each element as a vector component, and corresponds to each topic as described in FIG. ) Is a threshold that is considered to contribute to achieving accuracy when the model generation unit 6 described below performs prediction when the value is equal to or larger than the threshold obtained for each element. On the contrary, when the value of each element is less than the threshold value, the value of the element is considered not to contribute to the achievement of accuracy when the model generation unit 6 performs prediction.

Here, the notation of various quantities is defined as follows. Feature vector θ (X, n) the topics that correspond to elements of k _1, k _2, ..., a k _M. That is, M is the number of topics, which is the number of elements (ie, size or dimension) of the vector θ (X, n), and is reduced by the processing of the feature amount reduction unit 7 described later for each I-th iteration. Although it is a value, the I dependency of M is not specified because the notation is complicated. In addition, the value of the element corresponding to the j-th (j = 1, 2, ..., M) topic k _j in the feature vector θ (X, n) (that is, the j-th feature vector θ (X, n) Is expressed as θ (X, n) _[j] . In addition, a threshold value that the threshold value determination unit 4 obtains for an element corresponding to the topic k _j is expressed as TH [k _j ].

Specifically, the threshold value determination unit 4 can calculate the threshold value TH [k _j ] for the elements of each topic k _j as follows. That is, for each topic k _j element, the threshold TH was actually increased from 0 to 1 (as a variable) little by little (as each value of the variable), and the “goodness” of the threshold TH was calculated and determined to be optimal. The threshold TH [k _j ] is obtained as a result of obtaining the time TH. Here, when TH is increased little by little from 0 to 1 (sweep), the value of a predetermined step may be used. For example, a 5-step sweep may be performed with an increment of 0.2, such as TH = 0, 0.2, 0.4, 0.6, 0.8.

Note that sweeping the sweep variable TH from “0 to 1” as described above means that the length of the range that the value θ (X, n) _[j] of a series of elements in the topic k _j can take is “ It means “from 0 (minimum) to 1 (maximum)” after standardization as “1”, and the same applies to the following description (including descriptions of FIGS. 5, 6, etc.). In other words, the actual sweep range is the range between the minimum value min {θ (X, n) _[j] } and the maximum value max {θ (X, n) _[j] } (or includes this) (Predetermined range), but in order to simplify the description, it is assumed that the range is normalized to a range from 0 to 1, and the sweep is performed from 0 to 1. Therefore, the obtained threshold value TH [k _j ] as the optimum value is output to the data discriminating unit 5 as an actual value “cancelled” so to speak. For example, when the threshold value TH [k _j ] as an optimum value is obtained as x (standardized value in the range of 0 ≦ x ≦ 1), the standardized interval is from the above minimum value to the maximum value. When it corresponds to the range, the following values are output to the data distinction unit 5 as actual values.
(1-x) * min {θ (X, n) _[j] } + x * max {θ (X, n) _[j] }

  Here, the above “goodness” of TH was selected by TH as an estimation of “goodness” when actually predicting a target to be predicted by the prediction unit 8 described later, that is, to predict prediction accuracy. What is necessary is just to calculate by quantifying the ease of prediction of data. Specifically, the “goodness” of the threshold TH may be calculated as in the following [Procedure 1] to [Procedure 3] for each value of the threshold TH to be swept as a variable. Here, a value to be predicted by the prediction unit 8 using the feature vector θ (X, n) as an input (that is, output as a prediction result) is denoted as P (X, n).

  In addition, according to the following [Procedure 1] to [Procedure 3], the “goodness” of the threshold TH is easy to predict the data sorted by TH, but the data not sorted by the same TH other than the data is It is calculated in the form of the strength of the tendency that it is difficult to predict.

[Procedure 1] According to the threshold value TH for the element k _j of the topic, all the data θ (X, n) (all data θ ( X, n)), and the first data group in which the value θ (X, n) _[j] for the element k _j is greater than or equal to the threshold TH and the first data group that is less than the threshold TH Divide into two data groups.

  [Procedure 2] In each of the first data θ (X, n) group and the second data θ (X, n) group classified as not less than / less than the threshold TH in [Procedure 1], the prediction unit 8 A single regression is performed with the target value P (X, n) to be predicted as an objective variable.

Here, in the simple regression, the value θ (X, n) _[j] for the element k _j of the target topic is used as an explanatory variable. That is, the slope a and the intercept b in simple regression given by the following linear expression (1) are determined for each of the classified first data group and second data group. This determination can be made by a least square method or the like as is well known in simple regression.
P (X, n) = aθ (X, n) _[j] + b (1)

  Note that the value P (X, n) to be predicted is known in advance (as so-called teacher data) for each data θ (X, n), and the data D ( It shall be given in the form linked to X, n).

[Procedure 3], single regression model “Y = aX + b” obtained in [Procedure 2] (X is an explanatory variable, Y is an objective variable). By counting the data in each of the first data group and the second data group, a 2 × 2 cross tabulation table as shown in FIG. 4 is created, and the AIC value (of Akaike information criterion) Value) is calculated as AIC (k _j , TH) obtained by quantifying the “goodness” of the threshold TH for the topic element k _j .

Here, whether or not the obtained single regression model “Y = aX + b” is correct depends on the predicted value “Equation (1) applied to each data θ (X, n) _[j] ”. aθ (X, n) _[j] + b ”is compared with the actual value“ P (X, n) ”and the absolute value of the difference is“ | P (X, n)-(aθ (X, n) _{If [j]} + b) | "is within the predetermined threshold for judgment, it is judged that the prediction is correct. If the absolute value of the difference exceeds the threshold, the prediction is not correct. It can be judged.

Note that the “goodness” AIC (k _j , TH) of TH calculated as an AIC value has a meaning that the threshold TH is “good” as the value is smaller. Therefore, as shown in the following equation (2), TH that minimizes the value AIC (k _j , TH) calculated for each sweep variable TH can be determined as the optimum threshold TH [k _j ]. it can.

A specific method for calculating AIC (k _j , TH) as an AIC value from the cross tabulation table of FIG. 4 will be described later. Using the method described later, AIC (k _j , TH) is obtained by subtracting the “AIC value of the independent model” from the “AIC value of the dependent model” and the obtained value is minimized (or (TH in any of the obtained values being “−2” or less) can be determined as the optimum value.

FIG. 5 shows that the target data to be subjected to simple regression is selected by the threshold TH for each value of the threshold TH as the sweep variable by the above [Procedure 1] to [Procedure 3], and the explanatory variable θ (X, n) It is a figure which shows the example which plotted the objective variable P (X, n) on the horizontal axis _| shaft with _[j] on the vertical axis _| shaft regarding the 1st data group. Here, the case where the objective variable P (X, n) is the medical expenses for the next year is taken as an example. That is, when the data θ (X, n) is the data at the time of n years of the subject X, the medical cost required at the time of n + 1 years of the subject X is set as P (X, n) Take an example.

  In FIG. 5, plots of data selected when a 5-step sweep is performed at an increment of 0.2, such as TH = 0, 0.2, 0.4, 0.6, and 0.8, are shown as [1] to [5], respectively. It can be seen how the single regression fits better as the threshold is raised.

  However, with respect to the calculated AIC value, it shows the behavior that the monotonic decrease is changed to the monotonic increase with the increase of TH, and the minimum value is taken while the fit of the single regression is improved. For example, FIG. Then, the minimum value is taken at the threshold TH = 0.4 shown in [3]. (Note that depending on the step width of the threshold TH and the specific data distribution, increasing the threshold TH may cause the AIC value to monotonously increase or monotonously decrease. In addition, the behavior of the monotonous increase / decrease is noise. This is the behavior when excluding typical fluctuations.)

  Although only the first data group is shown in FIG. 5, the second data group, which is the remaining data divided by the threshold TH, is also subject to single regression for each value of the sweep variable TH. . For example, in the case of [1] in FIG. 5, the second data group is an empty set. In this case, simple regression is not possible, but if the two element values in the cross tabulation table in FIG. Good.

  FIG. 6 is a schematic example of separating the first data group and the second data group by setting the threshold value TH when the data shown in [1] of FIG. Is described).

FIG. 7 shows a schematic example of simple regression performed on the data [2] in FIG. 5 (first data group), and whether or not the prediction is correct in the above [Procedure 3]. A schematic example of the existence area of target data to be counted is shown. In FIG. 7, the straight line L0 is a regression line obtained from the data, and the lower limit straight line L0 _{[lower] which} is lowered by a predetermined value in the vertical direction on the graph, and the upper limit straight line L0 _{[upper] which is} increased by the same predetermined value on the graph _. The data in the area between and is determined as the data that has been predicted correctly, and the data outside the area is determined as the data that has not been predicted.

As described above, the threshold value determination unit 4 obtains the threshold value TH [k _j ] for which it is determined that the prediction performance can be ensured if the element value corresponding to each topic kj of the feature vector is equal to or higher than the value, and the results are classified into data. Output to part 5.

In addition, the threshold value determination unit 4 further confirms whether the threshold value TH [k _j ] is appropriate for the threshold value TH [k _j ] obtained by the procedure as described above. The data may be output to the data distinction unit 5. Specifically, the first data group in [Procedure 1] of the above procedure determined by the threshold value TH [k _j ] obtained in the above procedure (the data group that is determined to be more than the threshold value and the predicted performance is ensured. ), A correlation test that is well-known in the statistical field is performed, and it is confirmed that the test is passed, and then the passed result is output to the data distinction unit 5.

As a result of performing the above uncorrelated test, if the test is not passed, the first data group in a threshold value smaller than the threshold TH [k _j ] that did not pass is sequentially determined. Then, an uncorrelated test may be performed, and the threshold value TH having the maximum value after passing the test may be output to the data distinguishing unit 5 as a result. The small threshold value may be selected from the same value as when the variable is swept by a predetermined step width as the variable to be swept in the above procedure.

In the data distinction unit 5, the data (feature quantity vector generation unit 3 is the target of the I-th processing) according to the threshold value TH [k _j ] obtained from the threshold value determination unit 4 at each I time of the iterative processing. All of the output feature quantity vectors θ (X, n)) are classified into a predictable group and an unpredictable group, and the discrimination results are output to the model generation unit 6 and the feature quantity reduction unit 7 To do.

Specifically, for each data θ (X, n), the value θ (X, n) _[j] of the element corresponding to each topic k _j is compared with the threshold TH [k _j ], and the value of the element Data θ (X, n) that has a threshold value or more, that is, at least one element that satisfies “θ (X, n) _[j] ≧ TH [k _j ]” is predicted. Judge as belonging to the possible group. On the other hand, if the determination is impossible, that is, the values θ (X, n) _{[j] of} the elements corresponding to all the topics k _j are less than the threshold TH [k _j ], that is, “θ (X , n) _[j] <TH [k _j ] ”is determined as belonging to the unpredictable group.

For example, if there are three types of topics (k ₁ , k ₂ , k ₃ ) (ie, data θ (X, n) is three-dimensional) and the corresponding threshold is (10, 20, 30), Data (11,9,5) is judged to belong to the predictable group because the value 11 in at least the first element of the three elements is greater than or equal to the threshold 10, and data (1, 2, 3) is all Since the value of the element is less than the corresponding threshold value, it is determined that it belongs to the unpredictable group.

  In the model generation unit 6, at each I time of the iterative process, the predictable group data output from the data distinction unit 5 is input, and a model for predicting the health condition and the like is generated, and the prediction unit 8 Output.

  Here, as the simplest embodiment, the entire data θ (X, n) of the predictable group output by the data distinguishing unit 5 is output as it is to the predicting unit 8 (or the user who performs data analysis). be able to. Here, the data θ (X, n) of the predictable group is configured as data in which the constituent elements are reduced every I times of the iterative processing by the feature amount reduction unit 7 described later. (Note that, in the first iteration process, since the process of the feature amount reduction unit 7 is not performed, the data is composed of all elements when the documenting unit 1 outputs.)

  Since the data of the predictable group is configured such that defects, noise, and the like as described in the above problem are automatically excluded, it is used as original data when performing various predictions. be able to.

  On the other hand, in one embodiment, the model generation unit 6 generates a specific prediction model from the predictable group in a form consistent with the specific prediction method performed by the prediction unit 8, and then outputs it to the prediction unit 8. You can also That is, with data θ (X, n) as an input, a predetermined type of predicted value P (X, n) such as medical expenses for the next year (at the time of n + 1) of the subject X who is n years old is output. A model formula for this may be obtained and output to the prediction unit 8.

  Here, for the construction of the model formula, input data θ (X, n) (that is, data determined as a predictable group in the I-th iteration of the data input to the documenting unit 1) On the other hand, the output data P (X, n) can be given as teacher data in advance, and can be constructed as a regression calculation expression such as SVR (support vector regression).

  In the feature amount reduction unit 7, the original word frequency vector D (X, n corresponding to the data θ (X, n) (belonging to the unpredictable group output from the data discrimination unit 5 at each I time of the iterative processing, respectively. )) Is output to the clustering unit 2 as the I + 1th input, which is the next iteration. At this time, the feature amount reduction unit 7 identifies a word i that is considered to be noise and does not contribute to the prediction performance from the word frequency vector D (X, n), and deletes the identified word i. Then, each vector D (X, n) of the I-th unpredictable group is output to the clustering unit 2 as an I + 1-th processing target in a form in which the number of vector dimensions is reduced.

  Here, for the word i that does not contribute to the prediction performance, a cross tabulation table as shown in FIG. 8 is created (as in the case where the threshold determination unit 4 calculates using the cross tabulation table of FIG. ) By calculating the AIC value, it is determined that a word i for which a dependent model as described later is determined to be good is determined to contribute to the prediction performance, and the word i that has not been determined is left. Is to be deleted.

  That is, for a word i for which it is determined that the dependent model is good, it is determined that there is a significant difference (difference) between the distribution of the word i in the predictable group and the distribution of the word i in the unpredictable group. Therefore, it has the ability to predict the unpredictable group to be analyzed (the unpredictable group distinguished at the current I time) in the subsequent I + 1 and subsequent iterations. It is assumed that there is a possibility of being left. Conversely, the word i that is not determined is a word that is determined to have no difference between the distribution in the predictable group and the distribution in the unpredictable group. It is considered as a deletion target because it has no capability (thus, there is a possibility that it may become noise when analyzing by clustering or the like). (Note that FIG. 9 will be described later as an example of leaving / deleting words.) When counting the cross tabulation table of FIG. 8, whether or not each word i exists is determined based on the original word frequency vector D ( In X, n), if the frequency of the word i is greater than or equal to the threshold, it exists, and if it is less than the threshold, it may be determined that it does not exist. The threshold value defined in advance for each word i may be used.

  The prediction unit 8 holds the prediction model formula output by the model generation unit 6 at each I-th iteration process, receives a user instruction, and outputs a prediction result to which the prediction model is applied to the user. By giving an instruction as to which of the I prediction models is to be used and prediction target data as the user instruction, the prediction unit 8 can output a prediction result for the prediction target data.

  Note that the prediction target data is prepared on the user side as being composed of the same elements as the feature quantity vector θ (X, n) when the I-th prediction model is constructed by the model generation unit 6. Give above. As a result, the prediction unit 8 can apply the prediction expression (referred to as function f) output from the model generation unit 6 and output the predicted value P (X, n) as f (θ (X, n)). it can.

  In the above, each part 1-8 of FIG. 1 was demonstrated, respectively. Here, with reference to FIG. 9, data obtained by the iterative processing by the iterative processing unit 20 in FIG. 1 will be described.

  In FIG. 9, [0] shows all data D output from the documenting unit 1, and data D1 determined to be a predictable group in the first iteration of all the data D according to the Venn diagram format. (Shown as a white area), data D2 determined as a predictable group in the second iteration (represented as a vertical stripe area), and data D3 determined as a predictable group in the third iteration (as a horizontal stripe area) Notation). As shown in the figure, in this example, by repeating the process three times, all data D is determined as a predictable group at each time, and there is a relationship of D = D1∪D2∪D3.

The results obtained in each of the first to third iterations are shown in columns [1] to [3], respectively. As shown in [1], it is assumed that the initial all data D is configured as a word frequency vector of 6 word elements, and in the first iteration, all the 6 words (i ₁ , i ₂ , i ₃ , i ₄ , i ₅ , i ₆ ) constitute a word frequency vector, and all the data D is input to perform processing from the clustering unit 2 onward. As a result, the data D1 is determined to be a predictable group, and prediction based on D1 A model is generated. Further, the feature amount reduction unit 7 deletes two words (i ₁ , i ₂ ) from the element.

As shown in [2], in the second iteration, the word frequency vector is composed of four words (i ₃ , i ₄ , i ₅ , i ₆ ) that remain without being deleted, and predicted in the first process Data D \ D1 determined as an impossible group is input to perform processing after the clustering unit 2, and as a result, the data D2 is determined as a predictable group, and a prediction model based on D2 is generated. Further, the feature amount reduction unit 7 deletes two words (i ₃ , i ₄ ) from the element.

  As is well known as a mathematical symbol, “\” means “exclude” in the set operation, and in the above, data D \ D1 means the remaining data obtained by removing data D1 from data D.

Finally, as shown in [3], in the third process, the word frequency vector is composed of two words (i ₅ , i ₆ ) that remain without being deleted, and the second process determines that the group is unpredictable The data D \ (D1∪D2) = D3 is input, and the processing after the clustering unit 2 is performed. As a result, the entire data D3 is determined as a predictable group, and a prediction model based on D3 is generated.

  FIG. 3 shows an example in which all data is determined as a predictable group and each prediction model is generated by performing the iteration process three times. However, the number of iteration processes is given by user designation. In addition, the repetition process may be aborted when the number of times is reached. In addition, the repetition process may be terminated when a predetermined termination condition is satisfied without giving the number of times for termination in advance. For example, all of the data distinguished at the time of the I-th time in the data distinction unit 5 may be terminated when it becomes a predictable group or an unpredictable group, or the prediction at the time of the I-th time You may make it censor when the number of data determined as an impossible group or the number of dimensions becomes below a threshold value.

  As described above, according to the present invention, as illustrated in FIG. 9, a plurality of feature amounts and a group that can be predicted from the feature amounts are finally specified. Use of to improve prediction accuracy. At this time, feature quantities other than these feature quantities may be used as auxiliary information.

  That is, in the present invention, a topic topic, in particular, latent topic analysis (Latent dirichlet allocation: LDA) is used to clarify potential topics of data, and medical costs are predicted using the topics as feature quantities. At that time, topics generated from medical data such as general medical details always include noise. For example, certain drugs may be used for multiple illnesses, and both diabetics and healthy people can get a cold. Therefore, a specific cut-off criterion is set for the obtained topic value, and smoothing is performed to identify a target person having only a feature amount having a certain level of predictive power. Next, a target person who can only have a feature quantity that cannot be used for prediction is separated, a feature quantity is selected, and topic classification is performed again. By repeating this procedure, it is possible to predict the medical cost prediction of a group that is difficult to predict using only important feature amounts, and the prediction accuracy can be improved.

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

  (1) About calculating the AIC value from the cross tabulation table of FIG. 4 or FIG.

FIG. 10 is a generalized table of the cross tabulation tables of FIG. 4 and FIG. That is, the total number n _ij in FIG. 10 shows the common case as in FIG. 8, FIG. 9, etc. as a general case. The result is configured with the horizontal axis (column element). In FIG. 10, the size is generalized to 2 × m, but by setting m = 2, the case of FIGS. 4 and 8 can be applied.

  That is, “number of people who predicted” and “number of people who did not hit” in FIG. 4 correspond to “applicable” and “not applicable” in FIG. 10, respectively, and “first data group” in FIG. "And" second data group "correspond to" cluster 1 "and" cluster 2 "in FIG. 10, respectively. Further, “the number of people in which the word exists” and “the number of people in which the word does not exist” in FIG. 8 correspond to “applicable” and “not applicable” in FIG. 10, respectively, and “predictable group” and “ “Unpredictable group” corresponds to “cluster 1” and “cluster 2” in FIG.

As shown in FIG. 10, the peripheral frequency ki (i = 1, 2,..., M), h, Nh, etc. can be immediately calculated from the total number n _ij in the cross tabulation table, and these values are used. The AIC value can be calculated as follows.

  The AIC value is obtained as one of the following methods. In the first method, by applying a dependent model to the cross tabulation table, the AIC value AIC (DM) of the dependent model as shown in [Formula 1] below, where DM is an abbreviation of Dependent Model Asking. In the second method, an independent model is applied to the cross tabulation table, and the AIC value AIC (IM) of the independent model as shown in [Equation 2] below, where IM stands for Independent Model. Then, as [Equation 3], the difference is obtained by subtracting the AIC value of the independent model from the AIC value of the dependent model.

  In [Expression 1] and the like, MLL (DM) is the maximum log likelihood in the dependent model, and can be obtained as a value like [Expression 1-2]. In [Expression 2] and the like, MLL (IM) is the maximum log likelihood in the independent model, and can be obtained as a value as in [Expression 2-2]. The characters in the above equations are as described in the cross tabulation table of FIG. 12, and the same applies to the equations described below.

  Hereinafter, the maximum log likelihood MLL (DM) in the dependent model and the maximum log likelihood MLL (IM) in the independent model are calculated as [Equation 1-2] and [Equation 2-2] above, respectively. And using each calculated maximum log likelihood, the AIC value in the dependent model is calculated as [Equation 1], and the AIC value in the independent model is as [Equation 2]. The calculation will be described.

  FIG. 11 is a table of probabilities in the dependent model corresponding to the cross tabulation table of FIG. 10 for explaining the calculation in the dependent model shown as [Expression 1] and [Expression 1-2]. Calculation is performed as follows according to the probability shown in the table.

  First, the random variables of the dependent model are as follows.

  On the other hand, not all of the 2m pieces shown in FIG. 13 can be moved freely, and there are the following restrictions.

  Therefore, the degree of freedom of the dependent model is 2m-1, and the 2 * (2m-1) term of [Equation 1] is obtained from the definition of AIC (AIC = -2 × MLL + 2 × degree of freedom). Further, when the log likelihood LL is calculated from the above random variable, it is as follows.

  The conditions for maximizing the log likelihood LL are as follows.

  The following is obtained from the maximum condition.

  As above, further

Etc. are obtained. there,

Then,

And so on,

Therefore, the following case is the maximum likelihood estimation.

  Therefore, by substituting the above value into LL, the above-described [Equation 1-2] is obtained as the maximum value.

  FIG. 12 is a table of probabilities in the dependent model corresponding to the cross tabulation table in FIG. 10 for explaining the calculation in the independent model shown as [Expression 2] and [Expression 2-2]. Calculation is performed as follows according to the probability shown in the table.

First, the peripheral power k _m of FIG. 10, and the marginal probability q _m of the corresponding FIG. 12, in, the following restrictions.

Therefore, since q _{1 to} q _m-1 and p can be moved freely, the degree of freedom of the parameter is (m-1) + 1 = m. From the definition of AIC calculation, A 2 × m term is obtained. The random variable of the independent model is as follows.

  Therefore, the log likelihood LL is as follows.

In order to obtain the condition that gives the maximum value of the logarithmic likelihood, this is partially differentiated by p, q ₁ ...

  Therefore,

And again

Then,

And so on,

And

Therefore, the maximum likelihood is

And so on. Therefore, by substituting the above value into LL, [Equation 2-2] as the maximum value is obtained.

  (2) The documentation unit 1 converts the input medical examination data and other medical data into data D (X, n) as a bug of word corresponding to the health state of each subject X in each age n As described above, the documenting unit 1 may be omitted when the input data is converted into the bug of word format in advance.

  (3) The documenting unit 1 generates documented medical data D (X, n) for each subject X at each age n (age n), and the age n is given every year. However, the data D (X, n) is generated not only for one year but by dividing it by the age n for a given period of arbitrary length (two years or six months). Also good.

  (4) The present invention can also be provided as a program that causes a computer to function as all of the units 1 to 8 of the sorting device 10 or any part thereof. The computer can adopt a known hardware configuration such as a CPU (Central Processing Unit), a memory, and various I / Fs, and the CPU executes instructions corresponding to the functions of each part of the sorting device 10. It becomes.

  DESCRIPTION OF SYMBOLS 10 ... Sorting device, 1 ... Documenting part, 2 ... Clustering part, 3 ... Feature quantity vector generation part, 4 ... Threshold determination part, 5 ... Data distinction part, 6 ... Model generation part, 7 ... Feature quantity reduction part, 8 ... Prediction unit

Claims (8)

A clustering unit that obtains a clustering result by performing clustering by latent topic analysis on the health status data of a series of subjects given in the form of a bug of word as a collection of individual data for each subject and age;
From the topic ratio of each individual data in the clustering result, a feature vector generation unit that generates a feature vector of each individual data;
For each element of the feature vector, a threshold determination unit that determines a threshold as a value that is determined to be correlated with a predetermined prediction target related to a health state;
For each individual data, it is determined that the element value of the feature vector that is equal to or greater than the threshold value belongs to the predictable group, and the non-existing data value is determined to belong to the unpredictable group A data distinction unit that
A sorting apparatus that outputs individual data determined to belong to a predictable group by the data distinction unit. Furthermore, only words that are determined to have a significant difference in the existence distribution of the words of the bug of words in the individual data between the predictable group and the unpredictable group determined by the data distinguishing unit are distinguished. A feature amount reduction unit that outputs individual data of the unpredictable group to the clustering unit after leaving as a bug of word element,
The clustering unit, the feature quantity vector generation unit, the threshold value determination unit, the data distinction unit, and the feature quantity reduction unit perform I iteration processing each time, and the feature quantity reduction unit outputs the I-th unpredictable result. 2. The sorting apparatus according to claim 1, wherein the individual data of the group is subjected to clustering in the clustering unit at the next I + 1 time.   Further, a model generation for receiving a feature vector of individual data distinguished as belonging to the predictable group at each I time of the iterative process and constructing a prediction model for a predetermined prediction target related to the health state The sorting apparatus according to claim 2, further comprising a section.   Further, the prediction model constructed by the model generation unit at each I time of the iterative process is received from the user an instruction as to which one to use for prediction, and an instruction of the prediction target data, 4. The apparatus according to claim 3, further comprising: a prediction unit that outputs a prediction result of a predetermined prediction object related to the health state by applying the instructed prediction object data to the instructed prediction model. The sorting device described.   The feature amount reduction unit includes, for each word of the bug of word in individual data, an element of a bug of word corresponding to the word in each of the predictable group and the unpredictable group determined by the data distinguishing unit. Whether there is a significant difference based on the cross tabulation table created by counting the number of individual data determined to be determined by the threshold determination and the number of individual data determined not to be determined by the threshold determination The screening apparatus according to claim 2, wherein the determination is performed. The threshold determination unit sweeps a temporary threshold as a variable for each element of the feature vector,
A value of a predetermined prediction target related to the health state in the individual data corresponding to the feature quantity vector is used as an explanatory variable, with the element value in a series of feature quantity vectors such that the element value is equal to or greater than the temporary threshold value. Perform regression prediction with objective variables, count the number of hits and hits as the first count,
A value of a predetermined prediction target related to the health state in the individual data corresponding to the feature quantity vector is used as an explanatory variable, and the element value in a series of feature quantity vectors such that the element value is less than the temporary threshold value. Regression prediction is performed with the objective variable, and the number of hits and hits in the forecast is counted as the second count.
Based on the value of the Akaike information amount standard calculated by the cross tabulation table created by the first count and the second count, a predetermined prediction target related to the health condition from the temporary threshold values to be swept 6. The sorting apparatus according to claim 1, wherein a threshold value is determined as a value that is determined to be capable of obtaining a correlation with.   7. The sorting apparatus according to claim 1, wherein the predetermined prediction target related to the health condition is a medical cost in the future ahead of a predetermined age for each individual data.   The feature vector generation unit calculates the total number of words when the individual data is given in the form of the bug of words with respect to a vector configured with the topic ratio of each individual data in the clustering result as an element. 8. The sorting apparatus according to claim 1, wherein a feature quantity vector of each individual data is generated by multiplication.