Information processing system, information processing method, and program

Abstract

To provide an information processing system capable of creating a learning model that can accurately impute missing data even if there are gaps in healthcare data, an information processing method, and a program.SOLUTION: The information processing system includes: a past health checkup unit 111 that acquires conditioning data, which is healthcare data with missing data, and candidates for data to be used when imputing missing data; a conditioning item selection unit 112 that selects conditioning items to be used when imputing missing data from conditioning data based on the correlation between the healthcare data and conditioning data; a missing data imputation model learning part 114 that generates a missing data imputation model, which is a learning model for filling in missing data, using data included in the conditioning items.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to an information processing system, an information processing method, and a program, and more particularly to an information processing system and the like that can complement missing healthcare data. [Background technology]

[0002] To achieve the optimization of medical costs and disease prevention, the government is promoting data health, which allows local governments to carry out their health programs more effectively and efficiently through the analysis of healthcare data such as health checkup data and medical receipt data. However, because many local governments have not undergone health checkups, there are gaps in the health checkup data, which can lead to problems such as reduced accuracy in disease risk prediction and analytical bias. For this reason, there are existing technologies that can impute missing data.

[0003] Patent Document 1 discloses a data interpolation device including a data storage unit for storing data, a data division unit, a data interpolation processing unit, and a control unit. The data division unit divides the data stored in the data storage unit into data containing missing data and data without missing data. The data interpolation processing unit performs interpolation processing on missing data in response to opt-out based on an interpolation accuracy different from the interpolation accuracy for general missing data. The control unit creates data to be provided by combining the interpolated data interpolated by the data interpolation processing means with the data without missing data. Non-Patent Document 1 discloses a method for imputing missing data based on a diffusion model, which is one of generative models. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Publication No. 2017-162271 [Non-patent literature]

[0005] [Non-Patent Document 1] Shuhan Zheng, Nontawat Charoenphakdee, “Diffusion models for missing value imputation in tabular data”, NeurIPS, 2022 Summary of the Invention [Problem to be solved by the invention]

[0006] To improve the accuracy of imputing missing data, it is necessary to impute missing data taking into account the distribution of data. In addition, data distribution often differs depending on background information such as age and gender. Therefore, in order to impute missing data with higher accuracy, it is desirable to impute missing data by capturing the characteristics of the data distribution taking into account background information. The present invention aims to provide an information processing system, an information processing method, and a program that can create a learning model that can accurately complement missing data even if missing data occurs in healthcare data. [Means for solving the problem]

[0007] In order to solve the above problems, the present invention provides an information processing system having a data acquisition unit that acquires healthcare data containing missing data and conditioning data that are candidates for data to be used when imputing the missing data, a conditioning item selection unit that selects conditioning items that are items to be used when imputing the missing data from the conditioning data based on the correlation between the healthcare data and the conditioning data, and a gap imputation model training unit that uses data included in the conditioning items to generate a gap imputation model that is a training model for imputing the missing data.In this case, it is possible to provide an information processing system that can create a training model that can accurately imput the missing data even when missing data occurs in the healthcare data.

[0008] Here, for example, the system further includes a missing data imputation model application unit that applies the missing data imputation model to healthcare data containing missing data to impute the missing data. In this case, even if missing data occurs in the healthcare data, the missing data can be imputed with high accuracy. In addition, for example, the system further includes a data analysis unit that analyzes healthcare data after missing data has been imputed. In this case, analysis can be performed with higher accuracy than when analyzing healthcare data before missing data has been imputed. Furthermore, for example, the data analysis execution unit generates display information for displaying a screen for comparing healthcare data before and after imputation for items including missing data. In this case, analysis results of healthcare data before and after imputation can be compared. Furthermore, for example, the display information includes display information of a screen showing the distribution of data included in items including missing values ​​for each of the healthcare data before and after interpolation. In this case, it is possible to view the change in the distribution of the healthcare data before and after interpolation. For example, the distribution is the distribution of data selected from the healthcare data before and after interpolation based on the conditions specified by the user. In this case, the user can view the change in the distribution of the healthcare data before and after interpolation for the conditions that the user wants to know. Furthermore, for example, the system further includes an imputation target setting unit that sets imputation target data, which is data to be used when generating a missing data imputation model, from observation data that is included in items with missing data and is not missing. In this case, the accuracy of the learning model is likely to be improved. Furthermore, for example, the interpolation target setting unit determines the proportion of data to be interpolated in accordance with the missing data rate, which is the proportion of missing data. In this case, the accuracy of the learning model is likely to be further improved. Furthermore, for example, the interpolation target setting unit determines the interpolation target data to be set in accordance with the tendency of missing data. In this case, the accuracy of the learning model is likely to be further improved. Then, for example, the missing data imputation model training unit generates a missing data imputation model using the data included in the conditioning items and the data to be imputed. In this case, a learning model that can impute missing data with higher accuracy can be generated. Furthermore, for example, the missing data imputation model training unit adds noise to the data to be imputed and performs training based on a diffusion model that learns the distribution characteristics of the data to be imputed through estimation of the added noise. In this case, it becomes easier to generate the missing data imputation model.

[0009] The present invention also provides an information processing method that acquires healthcare data containing missing data and conditioning data that are candidates for data to be used when imputing the missing data, selects conditioning items that are items to be used when imputing the missing data from the conditioning data based on the correlation between the healthcare data and the conditioning data, and generates a missing data imputation model that is a learning model for imputing the missing data using data included in the conditioning items.In this case, it is possible to provide an information processing method that can create a learning model that can accurately imput the missing data even when missing data occurs in the healthcare data.

[0010] Here, for example, the missing data imputation model is applied to healthcare data containing missing data to impute the missing data. In this case, even if missing data occurs in the healthcare data, the missing data can be imputed with high accuracy.

[0011] Furthermore, the present invention is a program for causing a computer to implement the following functions: acquire healthcare data containing missing data and conditioning data that are candidates for data to be used when imputing the missing data; select conditioning items that are items to be used when imputing the missing data from the conditioning data based on the correlation between the healthcare data and the conditioning data; and generate a missing data imputation model that is a learning model for imputing the missing data using data included in the conditioning items. In this case, the computer can implement the function of creating a learning model that can accurately imput the missing data even when missing data occurs in the healthcare data.

[0012] Here, for example, the missing data imputation model may be applied to healthcare data containing missing data to further provide a function for imputing the missing data. In this case, even if missing data occurs in the healthcare data, the program may be provided with a function for accurately imputing the missing data. [Effects of the Invention]

[0013] According to the present invention, it is possible to provide an information processing system, an information processing method, and a program that can create a learning model that can accurately complement missing data even if missing data occurs in healthcare data. [Brief explanation of the drawings]

[0014] [Figure 1] 1 is a block diagram showing a configuration of a missing data completion system according to an embodiment of the present invention. [Figure 2] FIG. 4 is a diagram showing an example of medical checkup data stored in a medical checkup data storage unit. [Figure 3] 10 is a diagram showing an example of medical checkup data after a previous medical checkup has been acquired by a previous medical checkup acquisition unit. FIG. [Figure 4] FIG. 4 is a diagram illustrating an example of conditioning data stored in a conditioning data storage unit. [Figure 5] 10 is a diagram showing an example of shaped data acquired by a conditioning item selection unit; FIG. [Figure 6] 10A and 10B are diagrams illustrating examples of observation data acquired by a conditioning item selection unit. [Figure 7] 10A and 10B are diagrams illustrating examples of conditioning items selected by a conditioning item selection unit. [Figure 8] FIG. 10 is a diagram showing an example of a conditioning item selection screen output by a conditioning item selection unit to an output unit. [Figure 9] 10A and 10B are diagrams illustrating examples of observation data for which a complement target has been set, which is acquired by a complement target setting unit. [Figure 10] FIG. 10 is a diagram illustrating an example of data after missing data imputation, acquired by a missing data imputation model application unit. [Figure 11] FIG. 10 is a diagram showing an example of a data distribution display screen output by the data analysis execution unit to the output unit. [Figure 12] 10 is a flowchart illustrating a missing data complementation process. [Figure 13] 10 is a flowchart illustrating a learning process of a missing data imputation model. [Figure 14] FIG. 10 is a diagram illustrating an example of a learning process for a missing data imputation model. [Figure 15] 10 is a flowchart illustrating a selection process of a conditioning item (conditioning item selection process). [Figure 16] FIG. 10 is a diagram showing an example of selecting a conditioning item. DETAILED DESCRIPTION OF THE INVENTION

[0015] Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

[0016] <Overall explanation of the missing data imputation system 100> FIG. 1 is a block diagram showing the configuration of a missing data completion system 100 according to this embodiment. As shown in FIG. 1, the missing data completion system 100 according to this embodiment includes a missing data completion terminal 101 and a database 120.

[0017] The missing data complementation terminal 101 is an example of an information processing system, and is a computer (computer device) having an input unit 102, an output unit 103, a processor 104, a memory 105, and a storage medium 106. The input unit 102 is a human interface such as a mouse and a keyboard, and receives input to the missing data complement terminal 101 . The output unit 103 is a display or printer that outputs the calculation results from the missing data complementation terminal 101 .

[0018] The processor 104 is an arithmetic device that executes a program loaded into the memory 105, and is, for example, a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit). The processes and calculations described below are executed by the processor 104. Note that part of the processes that the processor 104 executes by executing a program may be executed by another arithmetic device (for example, hardware such as an FPGA (Field Programmable Gate Array) or an ASIC (Application Specific Integrated Circuit)).

[0019] The memory 105 includes a ROM (Read Only Memory), which is a nonvolatile storage element, and a RAM (Random Access Memory), which is a volatile storage element. The ROM stores unchanging programs (e.g., a BIOS (Basic Input Output System)). The RAM is a high-speed, volatile storage element such as a DRAM (Dynamic Random Access Memory), which temporarily stores programs executed by the processor 104 and data used when the programs are executed. That is, the programs stored in the storage medium 106 are loaded into the memory 105.

[0020] The storage medium 106 is a storage device that stores various programs that realize the missing data completion processing by the missing data completion terminal 101, as well as the execution results of the missing data completion processing, and is configured, for example, by a non-volatile storage medium (magnetic disk drive, non-volatile memory, etc.).

[0021] The programs executed by the processor 104 are provided to each server from removable media (such as CD-ROMs or flash memory) or via a network, and are stored in a non-volatile storage device, which is a non-transitory storage medium. For this reason, the computer system may be provided with an interface for reading removable media.

[0022] The missing data imputation system 100 according to this embodiment is a computer system configured on a single computer or on multiple logically or physically configured computers, and may operate in separate threads on the same computer, or may operate on a virtual computer constructed on multiple physical computer resources.

[0023] The storage medium 106 stores programs for realizing a past medical checkup acquisition unit 111, a conditioning item selection unit 112, a completion target setting unit 113, a missing data completion model learning unit 114, a missing data completion model application unit 115, and a data analysis execution unit 116.

[0024] The past medical checkup acquisition unit 111 acquires past medical checkup data from the medical checkup data input to the input unit 102 (FIG. 2, described later) using medical checkup result items from before the time point to be subjected to missing value completion, and acquires medical checkup data after the acquisition of the past medical checkup (FIG. 3, described later). In addition, as will be described in detail later, it acquires conditioning data (FIG. 4, described later) which is a candidate for data to be used when completing missing values. The past medical checkup acquisition unit 111 functions as a data acquisition unit that acquires healthcare data and conditioning data including missing values. Here, "healthcare data" is not particularly limited as long as it is data used to understand health conditions and diseases, maintain health, and promote health. Medical checkup data is an example of healthcare data. Hereinafter, the case where the healthcare data is medical checkup data will be described. Furthermore, "missing value" refers to a state in which a value is not stored in an observation.

[0025] The conditioning item selection unit 112 selects conditioning items, which are items used to complement missing data, from the conditioning data based on the correlation between the medical examination data and the conditioning data. Specifically, the conditioning item selection unit 112 links the medical examination data after the past medical examination acquisition acquired by the past medical examination acquisition unit 111 (FIG. 3, which will be described later) with the conditioning data input to the input unit 102 (FIG. 4, which will be described later) using the personal ID to acquire shaped data (FIG. 5, which will be described later). Then, from the shaped data, data that has no missing data in the medical examination result items at the time point to be complemented (hereinafter referred to as complement target items) is acquired as observed data (FIG. 6, which will be described later). Then, using the observed data, the conditioning items are selected based on the absolute value of the correlation coefficient. That is, here, the absolute value of the correlation coefficient is used as the correlation. The selection method will be described later with reference to FIGS. 16 and 17. The selected conditioning items (FIG. 7, which will be described later) are displayed on the output unit 103 (FIG. 8, which will be described later).

[0026] The complement target setting unit 113 sets complement target data, which is data used when generating a missing data completion model, which is a learning model for complementing missing data, from observation data that is data included in items where missing data exists and is not missing. Specifically, the complement target setting unit 113 sets some of the observation data as complement target items, and acquires the observation data for which the complement target has been set (completion target data) (FIG. 9, which will be described later). As will be described in detail later, the method of setting the complement target may be a random setting method or a setting method that takes into consideration the missing data rate and the method of missing data.

[0027] The missing data imputation model learning unit 114 uses data included in the conditioning items (in this case, imputation target data) to generate a missing data imputation model, which is a learning model for imputing missing data. Specifically, the missing data imputation model learning unit 114 uses the observation data (imputation target data) for which imputation targets have been set, acquired by the imputation target setting unit 113, and the data included in the conditioning items selected by the conditioning item selecting unit 112, to generate a missing data imputation model for estimating the value of the health check data set as the imputation target. At this time, the missing data imputation model learning unit 114 learns the missing data imputation model based on a diffusion model. A learning method for the missing data imputation model will be described later with reference to FIGS. 14 and 15.

[0028] The missing data imputation model application unit 115 applies the missing data imputation model to the medical examination data including missing data to imputate the missing data. Specifically, the missing data imputation model application unit 115 inputs the shaped data acquired by the conditioning item selection unit 112 to the missing data imputation model learned by the missing data imputation model learning unit 114, and acquires data after missing data imputation (FIG. 10 described later).

[0029] The data analysis execution unit 116 analyzes the screening data after missing data has been imputed. This allows for more accurate analysis compared to when analyzing based on screening data before missing data is imputed. Specifically, the data analysis execution unit 116 performs data analysis using the data after missing data imputation acquired by the missing data imputation model application unit 115. As one of the data analysis results, the distribution of the data before missing data imputation and the data after missing data imputation is displayed on the output unit 103 (see FIG. 11, which will be described later).

[0030] As shown in FIG. 1, the database 120 is composed of a health checkup data storage unit 121, a conditioning data storage unit 122, a past health checkup data storage unit 123, a conditioning item storage unit 124, a data storage unit 125 for which completion targets have been set, a missing data completion model storage unit 126, a missing data completion result storage unit 127, and a data analysis result storage unit 128.

[0031] The medical checkup data storage unit 121 stores the medical checkup data input to the input unit 102 . The conditioning data storage unit 122 stores the conditioning data input to the input unit 102 . The past medical checkup data storage unit 123 stores the medical checkup data obtained by the past medical checkup obtaining unit 111 after the past medical checkup has been obtained. The conditioning item storage unit 124 stores the conditioning items selected by the conditioning item selection unit 112 .

[0032] The interpolation target set data storage unit 125 stores the observation data (interpolation target data) for which an interpolation target has been set, which is acquired by the interpolation target setting unit 113. The missing data imputation model storage unit 126 stores the missing data imputation model learned by the missing data imputation model learning unit 114 . The missing data imputation result storage unit 127 stores the missing data imputed by the missing data imputation model application unit 115 . The data analysis result storage unit 128 stores the data analysis results obtained by the data analysis execution unit 116 .

[0033] FIG. 2 is a diagram showing an example of medical checkup data 200 stored in the medical checkup data storage unit 121. As shown in FIG. The medical checkup data 200 includes medical checkup result items at multiple time points, including the time points to be imputed. In the example shown, the medical checkup data 200 includes an ID, a BMI (Body Mass Index) for year x to be imputed, and BMIs for years x-1 and x-2 as BMIs before year x.

[0034] FIG. 3 is a diagram showing an example of medical checkup data 300 obtained by the previous medical checkup obtaining unit 111 after the previous medical checkup has been obtained. The health checkup data 300 after the acquisition of a past health checkup includes health checkup result items at the time point to be subject to missing value imputation (items to be imputed) and past health checkup data items obtained from health checkup result items prior to the time point to be subject to missing value imputation. Here, the past health checkup data items are the most recent health checkup results prior to the time point to be subject to missing value imputation. In the example shown, the health checkup data 300 after the acquisition of a past health checkup includes, in addition to the ID, the BMI for year x to be subject to missing value imputation and the past BMI. In this case, as the past BMI, subjects with IDs 1 and 2 will obtain the health checkup results for year x-1, which is the most recent health checkup result, and subjects with ID 3 will obtain the health checkup results for year x-2, which is the most recent health checkup result.

[0035] FIG. 4 is a diagram showing an example of conditioning data 400 stored in the conditioning data storage unit 122. As shown in FIG. The conditioning data 400 includes data items that serve as auxiliary information for supplementing missing data in the health check data 200. In the illustrated example, the conditioning data 400 includes a total of 12 items, including ID, sex, age, and the presence or absence of 10 diseases, Disease 1 to Disease 10.

[0036] FIG. 5 is a diagram showing an example of shaped data 500 acquired by the conditioning item selection unit 112. As shown in FIG. The shaping data 500 is data obtained by linking, based on ID, the medical checkup data 300 obtained after a previous medical checkup and the conditioning data 400. In the example shown, the shaping data 500 includes an ID, BMI as an item to be complemented, past BMI, sex, age, and an item indicating the presence or absence of disease 1 to disease 10.

[0037] FIG. 6 is a diagram showing an example of observation data 600 acquired by the conditioning item selection unit 112. As shown in FIG. The observed data 600 includes data extracted from the shaped data 500 that has no missing values ​​for the items to be complemented. In the example shown, the observed data 600 includes data from the shaped data 500 for subjects with IDs 1 and 4 who have no missing values ​​for BMI, which is an item to be complemented.

[0038] FIG. 7 is a diagram showing an example of a conditioning item 700 selected by the conditioning item selection unit 112. As shown in FIG. In the illustrated example, the conditioning items 700 include five items: sex, age, presence or absence of disease 1, presence or absence of disease 3, and past BMI.

[0039] FIG. 8 is a diagram showing an example of a conditioning item selection screen 800 that the conditioning item selection section 112 outputs to the output section 103. As shown in FIG. The conditioning item selection screen 800 includes a conditioning item number setting field 810 , a correlation coefficient display field 820 , a conditioning item display field 830 , and a conditioning item edit field 840 .

[0040] The number of conditioning items setting field 810 is a field where the user sets the number of conditioning items to be selected. In the example shown, "5" is set as the number of conditioning items to be selected. The correlation coefficient display field 820 is a field that displays absolute values ​​821 to 825 of the correlation coefficient between the items selected by the conditioning item selection unit 112 and the items to be complemented. In the example shown, the number of selected items is set to "5", so the absolute values ​​of the correlation coefficient between the five selected items and the items to be complemented are displayed. The conditioning item display field 830 displays the item names of the items selected by the conditioning item selection unit 112. In the example shown, past BMI, gender, presence / absence of disease 3, age, and presence / absence of disease 1 are displayed in descending order of the absolute value of the correlation coefficient with the item to be complemented. The conditioning item edit field 840 is a field for editing the items selected by the conditioning item selection unit 112. When the user presses the conditioning item selection start button 850, for example, the user can manually add "presence / absence of disease 5" as a conditioning item or delete the selected "presence / absence of disease 1." Then, when the user presses the conditioning item selection completion button 860, the conditioning items 700 are finalized.

[0041] FIG. 9 is a diagram showing an example of observation data 900 for which a complement target has been set, which is acquired by the complement target setting unit 113. As shown in FIG. The observation data for which the complement target has been set (completion target data) 900 includes data for which some of the complement target items have been set as the complement target out of the observation data 600. In the example shown in the figure, the BMI of a subject with ID 4 is set as the complement target.

[0042] FIG. 10 is a diagram showing an example of data 1000 after missing data imputation, acquired by the missing data imputation model application unit 115. The data 1000 after missing value imputation includes data in which missing values ​​in the shaped data 500 have been imputed using the missing value imputation model acquired by the missing value imputation model learning unit 114. In the example shown in the figure, the missing values ​​of BMI for subjects with IDs 2 and 3 in the shaped data 500 have been imputed, and values ​​of 20.3 and 22.2, respectively, have been stored.

[0043] FIG. 11 is a diagram showing an example of a data distribution display screen 1100 that the data analysis execution unit 116 outputs to the output unit 103. As shown in FIG. As shown in FIG. 11, a data distribution display screen 1100 includes a distribution display setting field 1110, a distribution display field 1120, and a basic statistics display field 1130.

[0044] The distribution display setting field 1110 is a field for setting conditions 1111 to 1116 for the distribution displayed in the distribution display field 1120. The conditions include data items 1111 and 1114, data item values ​​1112 and 1115, and conditional expressions 1113 and 1116. Subjects whose set data item values ​​satisfy the conditional expressions are extracted. The data items to be set can be any data item in the formatted data 500 other than the items to be complemented. Conditional expressions can include "match," "greater than or equal to," and "less than or equal to." In the illustrated example, subjects whose gender value matches 0 and whose age value is 60 or greater are extracted. Additional conditions can be added using the condition addition button 1117. After setting the data item values ​​and conditional expressions, the user can press the distribution display button 1118 to change the display in the distribution display field 1120 and basic statistics display field 1130 according to the settings.

[0045] The distribution display field 1120 displays the distribution of the items to be complemented for the subjects extracted in the distribution display setting field 1110. At this time, a data distribution 1121 before missing data imputation for the items to be complemented and a data distribution 1122 after missing data imputation for the items to be complemented are displayed. In the example shown in the figure, the distribution of BMI before imputation and the distribution of BMI after imputation for the extracted subjects are displayed. In this case, it can be said that the data analysis execution unit 116 generates display information for displaying the data distribution display screen 1100 as a screen for comparing the pre-interpolation and post-interpolation medical examination data for items including missing data. This makes it possible to compare the analysis results of the pre-interpolation and post-interpolation medical examination data. The display information can also be said to include screen display information showing the distribution of data included in items including missing values ​​for each of the pre- and post-interpolation screening data, which allows users to view the change in the distribution of screening data before and after interpolation. Furthermore, the distribution can be said to be the distribution of data selected from the pre-interpolation and post-interpolation medical examination data based on the conditions specified by the user in the distribution display setting field 1110. This allows the user to view the change in the distribution of medical examination data before and after the interpolation for the conditions that the user wants to know.

[0046] The basic statistics display field 1130 displays basic statistics of the items to be complemented for the subjects extracted in the distribution display setting field 1110. At this time, basic statistics 1131 before complementation for the items to be complemented and basic statistics 1132 after complementation for the items to be complemented are displayed. In the example shown, the average BMI values ​​before complementation and the average BMI values ​​after complementation for the extracted subjects are displayed. Note that, in addition to the average value, basic statistical quantities such as the median and mode may also be displayed as basic statistics.

[0047] Next, the missing data imputation process performed by the missing data imputation system 100 will be described. FIG. 12 is a flowchart illustrating the missing data completion process. 12 starts, first, the past medical checkup acquisition unit 111 executes a data input step S1201. In the data input step S1201, the input unit 102 of the missing measurement complement terminal 101 receives input of the medical checkup data 200 (see FIG. 2) and the conditioning data 400 (see FIG. 4). The past medical checkup acquisition unit 111 acquires the input medical checkup data 200 and stores it in the medical checkup data storage unit 121, and also acquires the input conditioning data 400 and stores it in the conditioning data storage unit 122.

[0048] Next, in a past medical checkup acquisition step S1202, the past medical checkup acquisition unit 111 acquires past medical checkup data from the medical checkup result items before the time point to be subjected to missing value completion from the medical checkup data 200 (see FIG. 2) input in the data input step S1201, and acquires medical checkup data 300 after the past medical checkup has been acquired (see FIG. 3). The acquired medical checkup data 300 after the past medical checkup has been acquired is stored in the past medical checkup data storage unit 123.

[0049] Next, in a conditioning item selection step S1203, the conditioning item selection unit 112 links the health checkup data 300 (see FIG. 3) obtained after the previous health checkup and acquired in the data input step S1201 with the conditioning data 400 (see FIG. 4) input in the data input step S1201 using the personal ID to acquire shaped data 500 (see FIG. 5). Then, the conditioning item selection unit 112 acquires data from the shaped data 500 (see FIG. 5) that has no missing data in the items to be complemented as observed data 600 (see FIG. 6). Thereafter, the conditioning item selection unit 112 uses the observed data 600 to select conditioning items 700 based on correlation coefficients. The selection method will be described later with reference to FIGS. 16 and 17. The selected conditioning items 700 (see FIG. 7) are displayed on the output unit 103 (see FIG. 8).

[0050] Next, in the complement target setting step S1204, the complement target setting unit 113 sets some of the complement target items from the observation data 600 (see FIG. 6) acquired in the conditioning item selection step S1203 as complement targets, and acquires the observation data (complement target data) 900 (see FIG. 9) for which the complement targets have been set. Here, the method of setting the data to be interpolated may be a random setting method or a setting method that takes into consideration the data missingness rate and the manner in which the data is missing. In the case of a method based on the data missingness rate, the interpolation target setting unit 113 determines the proportion of data to be interpolated based on the missingness rate, which is the proportion of data that is missing. In other words, the interpolation target setting unit 113 sets the missingness rate and the proportion of data to be interpolated to be approximately the same. In addition, in the case of a method that takes into consideration the manner in which data is missing, the interpolation target setting unit 113 determines the data to be interpolated based on the tendency of missing data. For example, if there is a tendency for missing data to occur more frequently among men, the interpolation target data will also include more data about men. In the next missing data imputation model learning step S1205, a missing data imputation model is generated using observation data (interpolation target data) 900 for which an imputation target has been set, and in this case, the accuracy of the missing data imputation model is likely to be improved. However, the missing data imputation model may be generated using all of the observation data 600 without setting observation data (interpolation target data) for which an imputation target has been set.

[0051] Next, in missing data imputation model learning step S1205, the missing data imputation model learning unit 114 generates a missing data imputation model that estimates the value of the health check data 200 set as the imputation target, using the observation data (imputation target data) 900 for which the imputation target has been set and acquired in imputation target setting step S1204, and the data included in the conditioning item 700 selected in conditioning item selection step S1203. This makes it possible to generate a learning model that can imputate missing data with higher accuracy. At this time, the missing data imputation model learning unit 114 learns the missing data imputation model based on a diffusion model. This makes it easier to generate the missing data imputation model. The method for learning the missing data imputation model will be described later with reference to Figures 14 and 15.

[0052] Next, in the missing data imputation model application step S1206, the missing data imputation model application unit 115 inputs the shaped data 500 acquired in the conditioning item selection step S1203 into the missing data imputation model learned in the missing data imputation model learning step S1205, and acquires the data 1000 after missing data imputation (see Figure 10).

[0053] Finally, in data analysis step S1207, the data analysis execution unit 116 performs data analysis using the data 1000 after missing data imputation (see FIG. 10) acquired in missing data imputation model application step S1206. As one of the data analysis results, the distribution of the data before missing data imputation and the data 1000 after missing data imputation is displayed on the output unit 103 (see FIG. 11). This completes the missing data complementation process.

[0054] Next, the learning process of the missing data imputation model performed by the missing data imputation model learning unit 114 will be described. FIG. 13 is a flowchart illustrating the learning process of the missing data imputation model. The missing data imputation model is trained based on a diffusion model in which the missing data imputation model training unit 114 adds noise to input data and learns the distribution characteristics of the input data through estimation of the added noise. This method is characterized by training taking into account the distribution of the conditioning items 700 by inputting selected conditioning items 700 into a neural network.

[0055] 13 starts, first, the missing data imputation model learning unit 114 executes a data input step S1301. In the data input step S1301, the observation data 900 for which the imputation target has been set and acquired in the imputation target setting step S1204 and the conditioning item 700 selected in the conditioning item selection step S1203 are input. The conditioning item 700 is set to c.

[0056] Next, in initialization step S1302, the missing data imputation model learning unit 114 sets values ​​for the number of learning times N and the number of steps T. Values ​​of 1 or greater are set for N and T. Also, a variable i that counts the number of learning times is initialized to 1.

[0057] Next, in step number acquisition step S1303, the missing data imputation model learning unit 114 randomly acquires a value between 1 and T, and assigns the acquired value to a variable t representing step information.

[0058] Next, in a noise addition step S1304, the missing data imputation model learning unit 114 adds a Gaussian noise ε to the data set as the imputation target among the observation data (imputation target data) 900 for which the imputation target has been set and which has been input in the data input step S1301. t Gaussian noise ε t The data added is x t Let's say.

[0059] Next, in noise estimation step S1305, the missing data imputation model learning unit 114 calculates the Gaussian noise ε t Data x with the sum of t The step information t acquired in the step number acquisition step S1303 and the conditioning item 700(c) input in the data input step S1301 are input to the neural network, and the Gaussian noise ε added to the observation data 600 in the noise addition step S1304 is t and estimate the estimated Gaussian noise ε t ' is output.

[0060] Next, in the parameter update step S1306, the missing data imputation model learning unit 114 calculates the Gaussian noise ε estimated in the noise estimation step S1305. t ' and the Gaussian noise ε added to the observation data 600 in the noise addition step S1304. t The neural network parameters are updated to minimize the squared error of

[0061] Next, in learning count determination step S1307, the missing data imputation model learning unit 114 compares the value of N set in initialization step S1302 with the value of i. If the value of N is equal to or less than the value of i, the process proceeds to missing data imputation model output step S1309. Otherwise, the process proceeds to variable i increment step S1308.

[0062] In variable i increment step S1308, the missing data imputation model learning unit 114 performs processing to increment the value of the variable i by one.

[0063] In the missing data imputation model output step S1309, the missing data imputation model learning unit 114 outputs the noise estimation function ε θ (x t ,t,c) is output.

[0064] This completes the learning process for the missing data imputation model. In the missing data imputation model application step S1206, Gaussian noise ε t The learned noise estimation function ε θ (x t , t, c) is applied to obtain data 1000 after missing data imputation (FIG. 10).

[0065] Next, an example of the learning process of the missing data imputation model will be described. FIG. 14 is a diagram showing an example of the learning process of the missing data imputation model. 14 shows an example 1400 of noise estimation function learning in the noise estimation step S1305 and the parameter update step S1306. As shown in FIG. 14, in the noise estimation step S1305, first, a Gaussian noise ε t Data x to which t The step information t and conditioning data c including five items, namely, gender, age, presence / absence of disease 1, presence / absence of disease 3, and past BMI, are input to the neural network, and the estimated Gaussian noise ε t Then, in parameter update step S1306, the estimated Gaussian noise ε t ' and Gaussian noise ε added to the BMI t The neural network parameters are updated so that the squared error of

[0066] Next, the selection process of the conditioning item 700 performed by the conditioning item selection unit 112 will be described. FIG. 15 is a flowchart illustrating the selection process of the conditioning item 700 (conditioning item selection process). The conditioning item selection process is characterized by selecting the conditioning item 700 based on the absolute value of the correlation coefficient with the item to be complemented.

[0067] 15 starts, first, the conditioning item selection unit 112 performs a data input step S1501. In the data input step S1501, the health checkup data 300 after the past health checkup acquisition acquired in the past health checkup acquisition step S1202 and the conditioning data 400 input in the data input step S1201 are input.

[0068] Next, in data linking step S1502, the conditioning item selecting unit 112 links the medical checkup data 300 after the previous medical checkup, which was input in data input step S1501, with the conditioning data 400 using the personal ID, and acquires the shaped data 500.

[0069] Next, in observed data acquisition step S1503, the conditioning item selection unit 112 acquires, as observed data 600, data with no missing data in the items to be complemented from the shaped data 500 acquired in data linking step S1502.

[0070] Next, in correlation coefficient calculation step S1504, the conditioning item selection unit 112 uses the observation data 600 acquired in observation data acquisition step S1503 to calculate the absolute value of the correlation coefficient between the item to be complemented in the observation data 600 and other items.

[0071] Next, in conditioning item selection step S1505, the conditioning item selection unit 112 compares the absolute values ​​of the correlation coefficients of the items calculated in correlation coefficient calculation step S1504, and selects the top n items with the largest values ​​as conditioning items 700. This completes the conditioning item selection process.

[0072] Next, an example of selecting the conditioning item 700 will be described. FIG. 16 is a diagram showing an example of selecting a conditioning item 700. In FIG. 16 shows an example 1600 of conditioning item selection in correlation coefficient calculation step S1504 and conditioning item selection step S1505. As shown in FIG. 16, in correlation coefficient calculation step S1504, first, the absolute values ​​of correlation coefficients with BMI, which is the item to be complemented, are calculated for all items other than the item to be complemented. Then, in conditioning item selection step S1505, the absolute values ​​of the calculated correlation coefficients are compared, and the top five conditioning items 700 with the largest values ​​are selected. In the example shown, past BMI, gender, presence / absence of disease 3, age, and presence / absence of disease 1 are selected in descending order of the absolute value of the correlation coefficient. Here, the number n of items to be selected is set in advance by the user. In addition, in the conditioning item selection step S1505, the number of items n to be selected may not be set by the user, but may be determined automatically, for example, by setting the number of items whose absolute value of the correlation coefficient is 0.4 or more.

[0073] As described above, the missing data imputation system 100 of this embodiment can create a missing data imputation model as a learning model that learns the characteristics of the conditional data distribution based on the diffusion model. By using this missing data imputation model, missing data in the health check data 200 can be imputed with high accuracy. In this way, this embodiment uses a conditional data distribution as a data distribution that takes background information into account. Therefore, more accurate missing data imputation is possible than with conventional methods that do not take the conditional data distribution into account. The past medical checkup acquisition unit 111 of the missing data imputation system 100 according to this embodiment acquires past medical checkup data from the medical checkup data 200 input to the input unit 102 using medical checkup result items prior to the time point targeted for missing data imputation, and acquires medical checkup data 300 after the acquisition of the past medical checkup (see FIGS. 2 and 3). This enables the effective use of past medical checkup data that is useful as auxiliary information for missing data imputation. Furthermore, the conditioning item selection unit 112 selects conditioning items 700 based on the correlation coefficient using the medical checkup data 300 after the past medical checkup acquired by the past medical checkup acquisition unit 111 and the conditioning data 400 input to the input unit 102. This makes it possible to select items that are more suitable for use in conditioning. Furthermore, the missing data imputation model learning unit 114 generates a missing data imputation model based on the diffusion model as a learning model for estimating values ​​of the health checkup data 200 set as the imputation target, using the observation data 900 for which the imputation target has been set acquired by the imputation target setting unit 113 and the conditioning items 700 selected by the conditioning item selecting unit 112. This makes it possible to generate a missing data imputation model that imputs missing data while taking into account the characteristics of the conditioning distribution.

[0074] It should be noted that the present invention is not limited to the above-described embodiments, and includes various modifications and equivalent configurations within the spirit of the appended claims. For example, the above-described embodiments have been described in detail to clearly explain the present invention, and the present invention is not necessarily limited to configurations including all of the described configurations. Furthermore, part of the configuration of one embodiment may be replaced with the configuration of another embodiment. Furthermore, the configuration of one embodiment may be added to the configuration of another embodiment. Furthermore, the aforementioned configurations, functions, processing units, processing means, etc. may be realized in part or in whole in hardware, for example by designing them as integrated circuits, or may be realized in software by having a processor interpret and execute a program that realizes each function. Information such as programs, tables, and files that realize each function can be stored in a storage device such as a memory, a hard disk, or an SSD (Solid State Drive), or in a recording medium such as an IC card, an SD card, or a DVD. In addition, the control lines and information lines shown are those that are considered necessary for explanation, and do not necessarily represent all the control lines and information lines that are necessary for implementation. In reality, it can be assumed that almost all components are interconnected.

[0075] <Explanation of information processing method> The processing performed by the above-described missing data imputation system 100 can be considered to be an information processing method that acquires healthcare data containing missing data and conditioning data that are candidates for data to be used when imputing the missing data, selects conditioning items that are items to be used when imputing the missing data from the conditioning data based on the correlation between the healthcare data and the conditioning data, and generates a missing data imputation model that is a learning model for imputing the missing data using the data included in the conditioning items. This makes it possible to provide an information processing method that can create a learning model that can accurately imput the missing data even when missing data occurs in the healthcare data. Furthermore, the processing performed by the missing data imputation system 100 can also be considered as an information processing method that applies a missing data imputation model to healthcare data containing missing data and imputes the missing data.

[0076] <Program Description> The processing performed by the above-described missing data imputation system 100 is realized, for example, by the cooperation of software and hardware resources. That is, the processor 104 provided in the missing data imputation system 100 loads programs for realizing each function of the missing data imputation system 100, namely, the past medical checkup acquisition unit 111, the conditioning item selection unit 112, the imputation target setting unit 113, the missing data imputation model learning unit 114, the missing data imputation model application unit 115, and the data analysis implementation unit 116, from the storage medium 106 to the memory 105, and executes the programs to realize each of these functions. Therefore, the program running on the missing data imputation system 100 can be considered to be a program that causes a computer to realize the following functions: a function to acquire healthcare data containing missing data and conditioning data that are candidates for data to be used when imputing the missing data; a function to select conditioning items that are items to be used when imputing the missing data from the conditioning data based on the correlation between the healthcare data and the conditioning data; and a function to generate a missing data imputation model that is a learning model for imputing the missing data using data included in the conditioning items. This allows the computer to realize the function of creating a learning model that can accurately imputate missing data even when missing data occurs in the healthcare data. Furthermore, the program that operates in the missing data imputation system 100 can be considered to be a program that further has the function of applying the missing data imputation model to healthcare data that includes missing data and imputing the missing data.

[0077] The program for realizing this embodiment can be provided not only by communication means but also by being stored on a recording medium such as a CD-ROM. [Explanation of symbols]

[0078] 100...missing data imputation system, 101...missing data imputation terminal, 102...input unit, 103...output unit, 104...processor, 105...memory, 106...storage medium, 111...past health check acquisition unit, 112...conditioning item selection unit, 113...imputation target setting unit, 114...missing data imputation model learning unit, 115...missing data imputation model application unit, 116...data analysis implementation unit, 120...database, 200...health check data, 300...health check data after past health check acquisition, 400...conditioning data, 500...shaped data, 600...observation data, 700...conditioning item, 800...conditioning item selection screen, 900...observation data with imputation target set, 1000...data after missing data imputation, 1100...data distribution display screen

Claims

1. a data acquisition unit that acquires healthcare data including missing data and conditioning data that is a candidate for data to be used when completing the missing data; a conditioning item selection unit that selects, from the conditioning data, conditioning items that are items used when complementing missing data based on a correlation between the healthcare data and the conditioning data; a missing data imputation model learning unit that uses data included in the conditioning items to generate a missing data imputation model, which is a learning model for imputing missing data; An information processing system having the above.

2. The information processing system according to claim 1 , further comprising a missing data imputation model application unit that applies the missing data imputation model to healthcare data containing missing data to impute the missing data.

3. The information processing system according to claim 2 , further comprising a data analysis execution unit that analyzes the healthcare data after missing data has been imputed.

4. The information processing system according to claim 3 , wherein the data analysis unit generates display information for displaying a screen for comparing healthcare data before and after imputation for items including missing data.

5. The information processing system according to claim 4 , wherein the display information includes display information for a screen showing a distribution of data included in items including missing data for each of the healthcare data before and after interpolation.

6. The information processing system according to claim 5 , wherein the distribution is a distribution of data selected from pre-interpolation and post-interpolation healthcare data based on conditions specified by a user.

7. 2. The information processing system according to claim 1, further comprising an imputation target setting unit that sets imputation target data, which is data to be used when generating the missing data imputation model, from observation data that is included in items where missing data exists and is not missing.

8. The information processing system according to claim 7 , wherein the interpolation target setting unit determines the proportion of the data to be interpolated in accordance with a missing data rate, which is the proportion of data that is missing.

9. The information processing system according to claim 7 , wherein the complement target setting unit determines the complement target data to be set in accordance with a tendency of missing data.

10. The information processing system according to claim 7 , wherein the missing data imputation model learning unit generates the missing data imputation model using data included in the conditioning items and the imputation target data.

11. The information processing system according to claim 10 , wherein the missing data imputation model learning unit adds noise to the data to be imputed and learns distribution characteristics of the data to be imputed through estimation of the added noise based on a diffusion model.

12. Obtain healthcare data containing missing data and conditioning data that are candidates for data to be used when imputing the missing data. selecting, from the conditioning data, conditioning items to be used when imputing missing data based on a correlation between the healthcare data and the conditioning data; generating a missing data imputation model, which is a learning model for imputing missing data, using the data included in the conditioning items; Information processing methods.

13. The information processing method according to claim 12 , wherein the missing data imputation model is applied to healthcare data containing missing data, and the missing data are imputed.

14. On the computer, A function to acquire healthcare data containing missing data and conditioning data, which are candidates for data to be used when completing missing data, and a function of selecting, from the conditioning data, conditioning items to be used when complementing missing data, based on a correlation between the healthcare data and the conditioning data; a function of generating a missing data imputation model, which is a learning model for imputing missing data, using data included in the conditioning items; A program to achieve this.

15. The program according to claim 14 , further comprising a function of applying the missing data imputation model to healthcare data containing missing data to impute the missing data.

Images