Learning device, inference device, learning method, inference method, and program

The learning device classifies user vital data into groups to estimate the most strongly felt discomfort using machine learning models, overcoming conventional methods' limitations in detecting and differentiating discomfort types, especially for users without prior data, and achieving timely and accurate disorder identification.

WO2024247202A1PCT designated stage expired Publication Date: 2024-12-05NT T INC

Patent Information

Application Number
PCT/JP2023/020389
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Filing Date
2023-05-31
Publication Date
2024-12-05

AI Technical Summary

Technical Problem

Conventional methods struggle to accurately detect and differentiate between various types of discomfort such as stress, fatigue, and sleepiness from vital data, especially for users with no prior data collection, and fail to determine the predominant disorder in a timely manner for large user groups.

Method used

A learning device and method that classify user vital data into groups based on similarity, using machine learning models to estimate the type of discomfort most strongly felt by users, even without prior data, by acquiring and preprocessing vital data, and training user group and disorder estimation models for efficient classification.

Benefits of technology

Enables accurate estimation of the predominant discomfort type from vital data for users with no prior data, effectively addressing the limitations of conventional techniques by providing timely and precise disorder identification.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure JP2023020389_05122024_PF_FP_ABST
    Figure JP2023020389_05122024_PF_FP_ABST
Patent Text Reader

Abstract

A learning device according to the present invention is provided with: an acquisition unit that acquires training data including a plurality of combinations of vital data of a user and a type of poor condition most strongly felt by the user at the time when the vital data was measured; a user group inference model training unit that classifies the plurality of combinations included in the training data into user groups on the basis of the degrees of similarity of the vital data and that trains a user group inference model with training data for user group inference, including a plurality of combinations of the vital data included in the training data and the user group into which the vital data was classified so that, when vital data of a user is input thereto, the user group to which the vital data of the user belongs will be output as an inference result; and a poor condition inference model training unit that trains a poor condition inference model with the training data for each of the user groups so that, when the vital data is input, a type of poor condition most strongly felt by the user will be output as an inference result.
Need to check novelty before this filing date? Find Prior Art

Description

Learning device, estimation device, learning method, estimation method, and program

[0001] The disclosed technology relates to a learning device, an estimation device, a learning method, an estimation method, and a program.

[0002] One method for improving the productivity of organizations such as companies is health management, which aims to maintain and promote the health of employees, and typical measures include health checkups and health guidance to detect and prevent illness. On the other hand, it has been pointed out that a decline in physical or mental condition (hereinafter referred to as "illness"), even if not resulting in illness, can affect the productivity of an organization (see, for example, Non-Patent Document 1).

[0003] While it is desirable to accurately identify and improve health issues from the perspective of improving productivity, it is not easy to detect these issues as the individual may not be aware of them. In addition, in the case of teleworking, it is difficult for the individual's superiors or colleagues at work to discover that the individual is unwell.

[0004] In recent years, as biosensing devices have become smaller and cheaper, wearable devices that users can wear to easily measure and check vital data such as heart rate information anywhere have become more common, and it has been proposed that ailments such as anxiety, fatigue, and sleep disorders can be detected from the vital data measured by such devices (see, for example, Non-Patent Document 2).

[0005] As a literature on detecting illness from vital data, Non-Patent Document 3 reports that an index representing "anxiety," which is considered to be one type of illness, has a significant negative correlation with a parasympathetic index calculated from heart rate variability. Also, Non-Patent Document 4 reports that an autonomic nervous activity index calculated from heart rate variability has a negative correlation with the degree of the illness "fatigue."

[0006] Based on this knowledge, a method has been disclosed for detecting stress and the unhealthy state of "drowsiness" from heart rate information obtained using a wearable sensor or the like.

[0007] E.B. Faragher, M. Cass and C.L. Cooper, The relationship between job satisfaction and health: a meta-analysis, Occupational and Environmental Medicine, vol. 62, no. 2, pp. 105-112, 2005. Yuta Takahashi, Akihiro Mizumoto, Yutaka Arakawa and Keiichi Yasumoto, A study on a method for estimating physical condition using wearable devices to improve labor productivity, Information Processing Society of Japan Scientific Report, vol. 2016-ITS-67, no. 2, pp. 1-5, 2016. Junichi Igawa, Shiro Shiwa, Daisuke Nakanishi, Miho Kurumaji, Nobuyuki Kikumoto and Hisato Ideshita, Evaluation of autonomic nervous function related to anxiety using heart rate variability, Biofeedback Research, vol. 37, no. 2, pp. 97-103, 2010. Hirohiko Kuratsune and Yasuyoshi Watanabe Fatigue research as a national issue: Objective fatigue assessment methods, Clinical Laboratory Science Education, vol. 10, pp. 1-8, 2018.

[0008] By the way, illness is information that a person feels about their own physical condition and expresses it in words or numbers, and how it is perceived varies from person to person. In particular, illnesses that do not lead to illness vary widely. However, this point has not been taken into consideration in conventional technologies.

[0009] On the other hand, it is believed that the tendency for changes in biometric information varies from person to person and even for the same person, changes occur depending on time, environment, and psychological conditions. Therefore, there are methods that allow users to input their actual condition and correct the estimated mental and physical state based on vital data. However, if input of the state of illness from the user cannot be received, such methods cannot be used. When expecting a large number of users in a short period of time, it is necessary to be able to estimate the type of illness from vital data even at the first use, but the above-mentioned documents cannot address this issue.

[0010] Another possible service would be to estimate which of various disorders is dominant and suggest corresponding remedial measures. However, conventional technologies estimate the degree and presence of each disorder, such as stress, drowsiness, fatigue level, depression, and lethargy, but are unable to estimate which type of disorder is dominant.

[0011] The technology disclosed herein has been made in consideration of the above points, and provides a learning device, estimation device, learning method, estimation method, and illness state estimation device that enables a program to estimate the type of illness a user feels most strongly from vital data, even for users for whom no data on illness or vital data has been collected in the past.

[0012] A first aspect of the present disclosure is a learning device comprising: a learning data acquisition unit that acquires learning data including a plurality of combinations of a user's vital data and the type of illness that the user feels most strongly at the time the vital data is measured; a user group estimation model learning unit that classifies the plurality of combinations included in the learning data into user groups based on the similarity of the vital data, and trains a user group estimation model using learning data for user group estimation that includes a plurality of combinations of the vital data included in the learning data and the user groups into which the vital data has been classified, so that when a user's vital data is input, the user group to which the user's vital data belongs is output as an estimation result; and an illness estimation model learning unit that trains an illness estimation model using the learning data so that when the vital data is input, for each user group, the type of illness that the user feels most strongly is output as an estimation result.

[0013] A second aspect of the present disclosure is an estimation device comprising: an estimation target acquisition unit that acquires vital data of a user who is to be estimated; a user group estimation unit that estimates a user group to which the vital data of the user who is to be estimated belongs, using a user group estimation model learned with learning data for user group estimation including multiple combinations of the user's vital data and user groups into which the vital data has been classified; and an illness estimation unit that estimates the type of illness that the user who is to be estimated feels the most strongly, using learning data including multiple combinations of the user's vital data and the type of illness that the user feels the most strongly at the time the vital data was measured, using an illness estimation model that corresponds to the user group estimated by the user group estimation unit, out of the illness estimation models learned for each user group.

[0014] Furthermore, a third aspect of the present disclosure is a learning method, which acquires learning data including multiple combinations of a user's vital data and the type of illness that the user feels most strongly at the time the vital data is measured, classifies the multiple combinations included in the learning data into user groups based on the similarity of the vital data, and trains a user group estimation model using learning data for user group estimation including multiple combinations of the vital data included in the learning data and the user groups into which the vital data has been classified, so that when a user's vital data is input, the user group to which the user's vital data belongs is output as an estimation result, and trains an illness estimation model using the learning data so that when the vital data is input, the type of illness that the user feels most strongly is output as an estimation result.

[0015] Furthermore, a fourth aspect of the present disclosure is an estimation method, which acquires vital data of a user to be estimated, estimates a user group to which the vital data of the user to be estimated belongs using a user group estimation model trained with learning data for user group estimation including a plurality of combinations of the user's vital data and user groups into which the vital data has been classified, and estimates the type of illness most strongly felt by the user to be estimated using an illness estimation model corresponding to the estimated user group out of the illness estimation models trained for each user group using learning data including a plurality of combinations of the user's vital data and the type of illness most strongly felt by the user at the time the vital data was measured.

[0016] A fifth aspect of the present disclosure is a program for causing a computer to function as each part of the learning device of the first aspect or each part of the estimation device of the second aspect.

[0017] According to the technology of the present disclosure, even for a user for whom no data relating to illness or vital data has been collected in the past, it is possible to estimate the type of illness that the user feels most strongly from the vital data.

[0018] 1 is a block diagram showing an example of the functional configuration of an illness estimation system of an embodiment. FIG. 1 is a diagram for explaining an example of learning data of an embodiment. FIG. 2 is a diagram for explaining an example of learning data 2 to which a user group classification result has been added. FIG. 3 is a diagram for explaining an example of a user group estimation model DB. FIG. 4 is a diagram for explaining an example of an illness estimation model DB. FIG. 5 is a diagram for explaining vital data of a user to be estimated. FIG. 6 is a diagram for explaining vital data of a user to be estimated to which an estimated user group has been added. A configuration diagram showing an example of the hardware configuration of an illness estimation system of an embodiment. A flowchart showing an example of learning processing in a learning device of an embodiment. A flowchart showing an example of user group estimation model learning processing in the learning processing of an embodiment. A flowchart showing an example of illness estimation model learning processing in the learning processing of an embodiment. A flowchart showing an example of estimation processing in an estimation device of an embodiment. A flowchart showing an example of user group estimation processing in the estimation processing of an embodiment. A flowchart showing an example of illness estimation processing in the estimation processing of an embodiment.

[0019] Hereinafter, an example of an embodiment of the disclosed technology will be described with reference to the drawings. Note that the same or equivalent components and parts in each drawing are given the same reference numerals.

[0020] First, an example of the configuration of a malfunction estimation system 10 according to the technology of this embodiment will be described. Fig. 1 shows a block diagram illustrating an example of the functional configuration of the malfunction estimation system according to this embodiment.

[0021] 1, the illness estimation system 10 of this embodiment includes a learning device 12 and an estimation device 14. The illness estimation system 10 also stores a user group estimation model DB 30 and an illness estimation model DB 32.

[0022] First, a description will be given of the learning device 12. As shown in Fig. 1, the learning device 12 includes a learning data acquisition unit 20, a user group estimation model learning unit 22, and an illness estimation model learning unit 24.

[0023] The learning data acquisition unit 20 has a function of acquiring learning data 2. As shown in Fig. 2 , the learning data 2 in this embodiment includes a plurality of combinations of a user ID (identifier) ​​for identifying a user, the user's vital data, and the type of ailment the user feels most strongly about at the time the vital data is measured. As shown in Fig. 2 , in this embodiment, a number ("No." in Fig. 2 ) is assigned to each combination.

[0024] The user ID is identification information for identifying a user. In this embodiment, an identification number is used, but any form of expression may be used as long as the information can identify an individual user. The same user ID indicates that the user is the same.

[0025] Vital data is data measured from a user that relates to vital signs and fluctuates depending on the user's condition. Vital data is an indicator of the user's vital activity and is obtained directly or indirectly from the user, for example, via a sensor device. In this embodiment, for example, values ​​obtained by aggregating time-series data of vital signs output from a sensor device over a fixed period of several seconds to several tens of seconds using statistical processing or values ​​calculated from the aggregated values ​​are used. Examples of time-series data of vital signs include statistical values ​​such as the average value and variance of vital signs, values ​​calculated from combinations of these, and values ​​further calculated from these calculated values. Vital data can include not only directly acquired indicators, but also information associated with the indicators, such as the measurement date and time and measurement environment when the sensor device measured the indicators, and indicators inferred from the measured indicators. In the example shown in FIG. 2 , the vital data used are "heart rate," "heart rate variability," "systolic blood pressure," "diastolic blood pressure," and the "measurement time period" during which these were measured. Examples of types of vital data other than those shown in FIG. 2 include respiratory rate, pulse rate, oxygen saturation, skin color, skin age, body temperature, workload, and stress level.

[0026] The type of ailment is the type of ailment the user experiences. "Illness" refers to a feeling of being physically or mentally unwell compared to normal, and is not limited to illness or injury. Examples of ailments other than those shown in FIG. 2 include fatigue, drowsiness, difficulty concentrating, irritability, stress, and lethargy. In this embodiment, since ailments classified as the same type may be expressed differently by individual users, words representing multiple types of ailments are prepared as ailment type candidates. The user then selects the type of ailment that the user feels most strongly from the multiple candidates, and the resulting ailment type is used as training data. Note that "Other" may be included in the options in case the type of ailment the user feels most strongly about is not included in the multiple candidates. Furthermore, if the user is not feeling ill, "health," which indicates no ailment, is used as the ailment type for convenience.

[0027] The learning data acquisition unit 20 outputs the acquired learning data 2 to the user group estimation model learning unit 22 .

[0028] The user group estimation model learning unit 22 has a function of learning the user group estimation model 30M. The user group estimation model learning unit 22 of this embodiment first performs preprocessing on the learning data 2. The learning data 2 includes multiple types of vital data, but some types of vital data may be missing depending on the user. For example, when measuring a user's vital signs with a wearable sensor device, measurement may become impossible. For combinations in which at least some types of vital data are missing, it is preferable to remove the combination from the learning data 2. In this case, the user group estimation model learning unit 22 determines whether all types of vital data are included for each combination included in the learning data 2, and if not, removes the combination from the learning data 2.

[0029] Furthermore, as a preprocessing step, the user group estimation model learning unit 22 may generate another type of vital data from the vital data included in the learning data 2. For example, the user group estimation model learning unit 22 may calculate the difference between the "systolic blood pressure" and the "diastolic blood pressure" to generate another type of vital data called "blood pressure difference."

[0030] Furthermore, as a preprocessing step, if the vital data contains non-numeric information, such as the "measurement time period" shown in FIG. 2, the user group estimation model learning unit 22 converts the information into a numerical value based on a predetermined criterion. For example, in the case of the "measurement time period" of the vital data, a criterion is set in advance that converts "morning" to "1," "afternoon" to "2," and "evening" to "3." Based on this criterion, the user group estimation model learning unit 22 converts the "measurement time period" of the vital data included in the learning data 2 into a numerical value.

[0031] Furthermore, bias may occur in the learning of the user group estimation model 30M, which has different trends (for example, mean values ​​and variances) for each type of vital data value, or in the learning of the subsequent illness estimation model 32M, affecting the accuracy of inference. In such cases, the user group estimation model learning unit 22 may perform preprocessing such as normalization so that the maximum and minimum values ​​between each type of vital data are the same, or standardization for each element so that the mean value is 0 and the variance value is 1.

[0032] The type of pre-processing that the user group estimation model learning unit 22 performs is not particularly limited, and may be determined in advance depending on the type of vital data included in the learning data 2, for example.

[0033] Next, the user group estimation model learning unit 22 classifies the multiple combinations included in the learning data 2 into user groups based on the similarity of the vital data. In other words, the user group estimation model learning unit 22 classifies the multiple combinations included in the learning data 2 so that combinations with similar vital data belong to the same user group.

[0034] The method by which the user group estimation model learning unit 22 classifies the multiple combinations included in the learning data 2 is not limited, but may be any automatic classification method in which one-dimensional vectors represented by each type of vital data that are similar belong to the same user group. As an example, the user group estimation model learning unit 22 of this embodiment performs classification using the k-means method (see J.A. Hartigan and M.A. Wong, Algorithm AS 136: A k-means clustering algorithm, Journal of the royal statistical society. Series C (applied statistics), vol. 28, no. 1, pp. 100-108, 1979.), which is one method of clustering data from unsupervised data. Note that, in the case of the k-means method, the number of classifications must be determined in advance, so the number of user groups is determined in advance before classification.

[0035] The user group estimation model learning unit 22 adds the classification results to the learning data 2. Fig. 3 shows an example in which the user group classification results are added to the learning data 2 shown in Fig. 2. Note that when the same user, i.e., multiple combinations with the same user ID, are included in the learning data 2 and each combination is classified into a different user group, the user group estimation model learning unit 22 uniquely determines a user group for that user (user ID) based on the tendency. For example, the user group may be determined to be the user group into which the user is most frequently classified, or, when the user is classified into multiple user groups with the same number of users, the user group may be determined randomly from among those.

[0036] Next, the user group estimation model training unit 22 uses training data including multiple combinations of vital data included in the training data 2 and the user groups added to the training data 2 as training data for user group estimation. The user group estimation model training unit 22 then trains the user group estimation model 30M using the training data for user group estimation so that, when a user's vital data is input, the user group to which the user's vital data belongs is output as an estimation result. The method by which the user group estimation model training unit 22 trains the user group estimation model 30M is not limited. As an example, in this embodiment, the vital data of the training data for user group estimation is used as an explanatory variable, the user group is used as a target variable, and discriminative training is performed to minimize a loss function (evaluation function), thereby training the user group estimation model 30M. Any discriminative training method may be used as long as it is capable of multi-class discriminative training. Examples of such discriminative training methods include neural networks, support vector machines, and Bayesian estimation.

[0037] The user group estimation model learning unit 22 stores the user group estimation model 30M obtained by learning in this manner in the user group estimation model DB 30, as shown in Fig. 4. Furthermore, the user group estimation model learning unit 22 of this embodiment also adds a list of user groups added to the learning data 2 (user group list 30U) to the user group estimation model DB 30, as shown in Fig. 4. The user group list 30U shown in Fig. 4 corresponds to the learning data 2 to which the user groups shown in Fig. 3 have been added, and lists user IDs corresponding to the combinations included for each user group.

[0038] The user group estimation model learning unit 22 outputs the learning data 2 (see FIG. 4 ) to which the user group has been added to the illness estimation model learning unit 24 .

[0039] The illness estimation model learning unit 24 has a function of learning an illness estimation model 32M so as to output the type of illness of the user as an estimation result when vital data for each user group is input using the learning data 2. In other words, the illness estimation model learning unit 24 obtains an illness estimation model 32M for each user group.

[0040] The method by which the illness estimation model learning unit 24 learns the illness estimation model 32M is not limited. As an example, in this embodiment, the illness estimation model 32M is learned by performing machine learning to minimize a loss function (evaluation function) using vital data from the learning data 2 as explanatory variables and the type of illness as the objective variable. Any machine learning method capable of predicting one of multiple classes from multiple explanatory variables may be used. Examples of such machine learning methods include Random Forest Classifier, neural networks, support vector machines, and Bayesian estimation. Note that in methods requiring numeric explanatory variables, non-numeric explanatory variables are deleted or quantified before machine learning. For example, the illness estimation model learning unit 24 quantifies the measurement time period of the vital data, similar to the preprocessing performed by the user group estimation model learning unit 22, and then performs machine learning.

[0041] In the case of learning data 2 to which the user groups shown in Figure 3 have been added, the illness estimation model learning unit 24 generates a user group estimation model 30M for user group 1 and a user group estimation model 30M for user group 2.

[0042] The illness estimation model learning unit 24 extracts the combinations of No. 1, No. 3, and No. 4 from the learning data 2 to which the user group shown in FIG. 3 has been added, and uses the combinations to create the illness estimation model 1 (illness estimation model 32M in FIG. 5 ). 1 5). In addition, the illness estimation model learning unit 24 extracts the combination No. 2 from the learning data 2 to which the user group shown in FIG. 3 has been added, and uses this combination to learn the illness estimation model 2 (illness estimation model 32M in FIG. 5). 2 (see reference).

[0043] The illness estimation model learning unit 24 learns the illness estimation model 32M (32M 1 , 32M 2 ) is stored in the illness estimation model DB 32 in association with the corresponding user group, as shown in FIG. 5 .

[0044] Next, the estimation device 14 will be described. As shown in FIG. 1, the estimation device 14 includes an estimation target acquisition unit 40, a user group estimation unit 42, and a malfunction estimation unit 44.

[0045] The inference target acquisition unit 40 has a function of acquiring vital data 3 of the user of the inference target. As shown in Fig. 6 , the vital data 3 of the user of the inference target in this embodiment is associated with a user ID for identifying the user of the inference target.

[0046] The estimation target acquisition unit 40 outputs the acquired vital data 3 of the estimation target user to the user group estimation unit 42 .

[0047] The user group estimation unit 42 has a function of estimating the user group to which the vital data 3 of the user to be estimated belongs. In this embodiment, the user group estimation unit 42 first determines whether the user to be estimated was also included in the learning data 2. Specifically, the user group estimation unit 42 references the user group list 30U (see FIG. 4 ) stored in the user group estimation model DB 30 and determines whether the user ID associated with the vital data 3 of the user to be estimated is included in the user group list 30U. In the case of the user group list 30U shown in FIG. 4 , if the user ID of the user to be estimated is "1001" or "1003," the user group estimation unit 42 determines that the user to be estimated is included in the user group list 30U. Furthermore, if the user ID of the user to be estimated is "1002," the user group estimation unit 42 determines that the user to be estimated is included in the user group list 30U. In this way, when the user ID of the user to be estimated is included in the user group list 30U, the user group estimation unit 42 has a function of estimating the user group to which the vital data 3 of the user to be estimated belongs, based on the user group list 30U. When the user ID of the user to be estimated is "1001" or "1003", the user group estimation unit 42 estimates that the user belongs to user group 1. When the user ID of the user to be estimated is "1002", the user group estimation unit 42 estimates that the user belongs to user group 2.

[0048] 6, the user ID is "2001," so the user group estimation unit 42 determines that the vital data 3 of the estimation target user is not included in the user group list 30U. In such a case, the user group estimation unit 42 estimates the user group to which the vital data 3 of the estimation target user belongs, using the user group estimation model 30M. Specifically, the user group estimation unit 42 inputs the vital data 3 of the estimation target user into the user group estimation model 30M, and acquires the user group output as an estimation result from the user group estimation model 30M.

[0049] The user group estimation unit 42 outputs the vital data 3 (see FIG. 7 ) to which the estimated user group has been added to the illness estimation unit 44. FIG. 7 illustrates a case in which user group 1 has been estimated for the vital data 3 of the estimation target user with user ID "2001."

[0050] The illness estimation unit 44 has a function of estimating the type of illness of the user to be estimated using an illness estimation model 32M corresponding to the user group estimated by the user group estimation unit 42, out of the multiple illness estimation models 32M.

[0051] Specifically, the malfunction estimation unit 44 uses a plurality of malfunction estimation models 32M (32M in FIG. 5 ) stored in the malfunction estimation model DB 32. 1 , 32M 2 ), the illness estimation unit 44 selects an illness estimation model 32M corresponding to the user group estimated by the user group estimation unit 42. The illness estimation unit 44 then inputs the vital data 3 of the user to be estimated into the identified illness estimation model 32M, and acquires the type of illness output as an estimation result from the illness estimation model 32M.

[0052] In the example shown in FIG. 7, the estimated user group is user group 1, so the illness estimation unit 44 selects the illness estimation model 32M from among the plurality of illness estimation models 32M stored in the illness estimation model DB 32. 1 Then, the malfunction estimation unit 44 selects the malfunction estimation model 32M 1 Vital data 3 is input to the malfunction estimation model 32M 1 The type of malfunction output as an estimation result is obtained from

[0053] The illness estimation unit 44 outputs to the outside the illness type 4 of the user to be estimated, which is estimated in this way by the illness estimation model 32M.

[0054] The malfunction estimation system 10, which includes the learning device 12 and estimation device 14 described above, is configured with the hardware shown in Fig. 8. Fig. 8 shows a configuration diagram illustrating an example of the hardware configuration of the malfunction estimation system 10 of this embodiment. As shown in Fig. 8, the malfunction estimation system 10 includes a CPU (Central Processing Unit) 51, a ROM (Read Only Memory) 52, a RAM (Random Access Memory) 53, storage 54, a display unit 57, and a communication I / F (Interface) 58. Each component is connected to each other so as to be able to communicate with each other via a bus 59 such as a system bus or a control bus.

[0055] The CPU 51 is a central processing unit that executes various programs, such as a learning program 55 and an estimation program 56, stored in the storage 54, and controls each part.

[0056] The ROM 52 stores various programs and data executed by the CPU 51. The RAM 53 temporarily stores programs or data as a work area when the CPU 51 executes various programs. That is, the CPU 51 reads a program from the storage 54 and executes the program using the RAM 53 as a work area.

[0057] In this embodiment, the storage 54 stores a learning program 55 and an estimation program 56. Each of the learning program 55 and the estimation program 56 may be a single program or a group of programs configured by multiple programs or modules. The storage 54 is configured by an HDD (Hard Disk Drive) or an SSD (Solid State Drive). The storage 54 also stores various programs, including an operating system, and various data (neither of which are shown). The storage 54 also stores a user group estimation model DB 30 and an illness estimation model DB 32 trained by the learning device 12.

[0058] The display unit 57 displays various information such as information related to the learning of the user group estimation model 30M, the learning of the illness estimation model 32M, and the estimation of the type of illness felt by the estimation target user. The display unit 57 is not particularly limited, and various displays can be used.

[0059] The communication I / F 58 is an interface for communicating with external devices, and uses standards such as Ethernet (registered trademark), FDDI, and Wi-Fi (registered trademark).

[0060] Next, the operation of the malfunction estimation system 10 of this embodiment will be described.

[0061] First, the operation of the learning device 12 included in the illness estimation system 10 will be described. Fig. 9 shows a flowchart of an example of a learning process executed by the learning device 12 of this embodiment. In the learning device 12, the CPU 51 executes a learning program 55 stored in storage 54, causing the CPU 51 to function as a learning data acquisition unit 20, a user group estimation model learning unit 22, and an illness estimation model learning unit 24, and executes the learning process shown in Fig. 9. Note that the learning process shown in Fig. 9 is executed at a predetermined timing, such as when an execution instruction is received from a user who is to estimate the type of illness.

[0062] 9, the learning data acquiring unit 20 acquires the learning data 2 (see FIG. 2) as described above. The learning data acquiring unit 20 outputs the acquired learning data 2 to the user group estimation model learning unit 22.

[0063] In the next step S12, the user group estimation model learning unit 22 executes a user group estimation model learning process shown in FIG. 10, the details of which will be described later.

[0064] In the next step S14, the malfunction estimation model learning unit 24 executes a malfunction estimation model learning process shown in Fig. 11, the details of which will be described later. When the process of step S14 ends, the learning process shown in Fig. 9 ends.

[0065] The user group estimation model learning process will be described with reference to Fig. 10. In step S120 of Fig. 10, the user group estimation model learning unit 22 performs predetermined preprocessing on the learning data 2, such as excluding from the learning data 2 combinations for which not all types of vital data are present, and converting non-numeric information contained in the vital data into numerical values ​​based on predetermined criteria, as described above.

[0066] In the next step S122, the user group estimation model learning unit 22 classifies the multiple combinations included in the learning data 2 into user groups based on the similarity of the vital data using the k-means method or the like, as described above.

[0067] In the next step S124, the user group estimation model learning unit 22 learns the user group estimation model 30M using the learning data for user group estimation. As described above, the user group estimation model learning unit 22 sets the learning data for user group estimation as learning data including multiple combinations of vital data included in the learning data 2 and the user groups added to the learning data 2. The user group estimation model learning unit 22 then learns the user group estimation model 30M by performing discriminative learning using the vital data of the learning data for user group estimation as an explanatory variable and the user group as a target variable.

[0068] In the next step S126, the user group estimation model learning unit 22, as described above, stores the user group estimation model 30M learned in step S124 in the user group estimation model DB 30. Furthermore, the user group estimation model learning unit 22 of this embodiment also stores the user group list 30U in the user group estimation model DB 30, as described above.

[0069] In the next step S128, the user group estimation model learning unit 22 outputs, as described above, the learning data 2 to which the estimated user group has been added to the illness estimation model learning unit 24. When the processing of step S128 ends, step S12 of the learning process shown in FIG. 9 ends, and the process proceeds to step S14 of the learning process described above.

[0070] Next, the malfunction estimation model learning process will be described with reference to Fig. 11. In step S140 of Fig. 11, the malfunction estimation model learning unit 24 selects a user group for which a malfunction estimation model 32M is to be generated from among a plurality of user groups.

[0071] In the next step S142, the illness estimation model learning unit 24 extracts combinations to which the user group selected in step S140 above has been added from the learning data 2 input from the user group estimation model learning unit 22, as described above.

[0072] In the next step S144, the illness estimation model learning unit 24 learns the illness estimation model 32M using the combinations extracted in step S142 from the learning data 2. As described above, the illness estimation model learning unit 24 learns the illness estimation model 32M by performing machine learning using the vital data included in the extracted learning data 2 as an explanatory variable and the type of illness as a target variable.

[0073] In the next step S146, the illness estimation model learning unit 24 stores the illness estimation model 32M learned in step S144 in the illness estimation model DB 32 in association with the user group selected in step S140, as described above.

[0074] In the next step S148, the illness estimation model learning unit 24 determines whether or not the selection in step S140 has been performed for all user groups. That is, it determines whether or not learning of the illness estimation model 32M has been performed for all user groups. If there is a user group that has not yet been selected, the determination in step S148 is negative, and the process returns to step S140, and the processing in steps S140 to S146 is repeated. On the other hand, if all user groups have been selected, the determination in step S148 is positive, and step S14 of the learning processing shown in FIG. 9 is completed, thereby terminating the learning processing.

[0075] Next, the operation of the estimation device 14 included in the malfunction estimation system 10 will be described. Fig. 12 shows a flowchart of an example of estimation processing executed by the estimation device 14 of this embodiment. In the estimation device 14, the CPU 51 executes an estimation program 56 stored in the storage 54, causing the CPU 51 to function as the estimation target acquisition unit 40, the user group estimation unit 42, and the malfunction estimation unit 44, and executes the estimation processing shown in Fig. 12. Note that the estimation processing shown in Fig. 12 is executed at a predetermined timing, such as when an execution instruction is received from a user who is to estimate the type of malfunction.

[0076] 12, the inference target acquisition unit 40 acquires the vital data 3 (see FIG. 6) of the user who is the inference target, as described above. The inference target acquisition unit 40 outputs the acquired vital data 3 to the illness inference unit 44.

[0077] In the next step S22, the user group estimation unit 42 executes a user group estimation process shown in FIG. 13, the details of which will be described later.

[0078] In the next step S24, the malfunction estimation unit 44 executes a malfunction estimation process shown in Fig. 14, the details of which will be described later. When the process of step S24 ends, the estimation process shown in Fig. 12 ends.

[0079] The user group estimation process will be described with reference to Fig. 13. In step S220 of Fig. 13, the user group estimation unit 42 refers to the user group list 30U in the user group estimation model DB 30, as described above.

[0080] In the next step S222, the user group estimation unit 42 determines whether or not the user ID of the user to be estimated is included in the user group list 30U, as described above. If the user ID of the user to be estimated is included in the user group list 30U, the determination in step S222 is positive, and the process proceeds to step S224.

[0081] In step S224, the user group estimation unit 42 estimates the user group to which the user ID of the user to be estimated belongs from the user group list 30U, as described above, and then proceeds to step S228.

[0082] On the other hand, if the user ID of the user to be estimated is not included in the user group list 30U, the determination in step S222 is negative, and the process proceeds to step S226.

[0083] In step S226, the user group estimation unit 42 estimates a user group using the user group estimation model 30M, and then proceeds to step S228. As described above, the user group estimation unit 42 inputs the vital data 3 of the user to be estimated into the user group estimation model 30M, and acquires the user group output as the estimation result from the user group estimation model 30M as the estimation result.

[0084] In step S228, the user group estimation unit 42 outputs the vital data 3 to which the user group estimated in step S224 or step S226 has been added, as described above, to the illness estimation unit 44. When the processing of step S228 ends, step S22 of the estimation processing shown in FIG. 12 ends, and the process proceeds to step S24.

[0085] Next, the malfunction estimation process will be described with reference to Fig. 14. In step S240 of Fig. 14, the malfunction estimation unit 44 selects a malfunction estimation model 32M corresponding to the estimated user group from among the plurality of malfunction estimation models 32M as described above.

[0086] In the next step S242, the illness estimation unit 44 estimates the type of illness of the estimation target user using the illness estimation model 32M selected in the above step S240. As described above, the illness estimation unit 44 inputs the vital data 3 of the estimation target user into the selected illness estimation model 32M, and obtains the type of illness 4 of the estimation target user output as the estimation result from the illness estimation model 32M.

[0087] In the next step S244, the illness estimation unit 44 outputs the illness type 4 of the estimation target user estimated in step S242 to the outside, as described above. When the processing of step S244 ends, step S24 of the estimation processing shown in FIG. 12 ends, and the estimation processing ends.

[0088] The illness estimation system 10 of this embodiment can be modified in various ways. For example, when estimating the type of illness, if a probability value is calculated for each type of illness, the probability value may be output along with the type of illness. For example, in an implementation example of the Random Forest learning device (reference: sklearn.ensemble.RandomForestClassifier, https: / / scikit-learn.org / stable / modules / generated / sklearn.ensemble.RandomForestClassifier.html), the probability of each class in each learned decision tree is averaged, and the class with the highest averaged probability value is output as the estimation result. When using this method, the averaged probability value may be output along with the estimated type of illness.

[0089] As described above, the above-described illness inference system 10 trains an illness inference model 32M for each user group classified based on the similarity of vital data. Furthermore, the above-described illness inference system 10 uses the illness inference model 32M for each user group obtained by training to infer the type of illness that is most strongly felt from the vital data of the inference target.

[0090] Therefore, with the above-described illness inference system 10, it is possible to infer the type of illness that the user feels most strongly from the vital data even for a user for whom no data related to illness or vital data has been collected in the past. In particular, even if the estimation target is a user different from the users included in the learning data 2, the above-described illness inference system 10 can infer the type of illness that the user feels most strongly from the vital data of the user who is the estimation target.

[0091] Furthermore, according to the above-described illness estimation system 10, it is possible to estimate the type of illness that the user feels most strongly, and therefore, even if the user feels multiple types of illness, it is possible to determine which type of illness is predominant.

[0092] Although the above-described malfunction estimation system 10 is configured to include the learning device 12 and the estimation device 14, the learning device 12 and the estimation device 14 may be configured as separate devices. Also, the user group estimation model DB 30 and the malfunction estimation model DB 32 may be stored in a storage device or the like outside the malfunction estimation system 10.

[0093] Furthermore, the various processes executed by the CPU after reading the software (program) in each of the above embodiments may be executed by various processors other than the CPU. Examples of processors in this case include dedicated electrical circuits, such as programmable logic devices (PLDs) whose circuit configuration can be changed after manufacture, such as field-programmable gate arrays (FPGAs), and application-specific integrated circuits (ASICs), which are processors having a circuit configuration designed specifically for executing specific processes. Each of the learning process and the estimation process may be executed by one of these various processors, or by a combination of two or more processors of the same or different types (e.g., multiple FPGAs, a combination of a CPU and an FPGA, etc.). The hardware structure of these various processors is, more specifically, an electrical circuit that combines circuit elements such as semiconductor elements.

[0094] In addition, in each of the above embodiments, the learning program 55 and the estimation program 56 are each pre-stored (installed) in the storage 54, but this is not limiting. Each of the learning program 55 and the estimation program 56 may be provided in a form stored in a non-transitory storage medium such as a CD-ROM (Compact Disk Read Only Memory), a DVD-ROM (Digital Versatile Disk Read Only Memory), or a USB (Universal Serial Bus) memory. Also, each of the learning program 55 and the estimation program 56 may be downloaded from an external device via a network.

[0095] The following additional clauses are disclosed in relation to the above-described embodiment.

[0096] (Addendum 1) A learning device comprising: a learning data acquisition unit that acquires learning data including multiple combinations of a user's vital data and the type of illness that the user feels most strongly at the time the vital data is measured; a user group estimation model learning unit that classifies the multiple combinations included in the learning data into user groups based on the similarity of the vital data, and trains a user group estimation model using learning data for user group estimation that includes multiple combinations of the vital data included in the learning data and the user groups into which the vital data has been classified, so that when a user's vital data is input, a user group to which the user's vital data belongs is output as an estimation result; and an illness estimation model learning unit that trains an illness estimation model using the learning data, so that when the vital data is input, for each user group, a type of illness that the user feels most strongly is output as an estimation result.

[0097] (Supplementary Item 2) An estimation device comprising: an estimation target acquisition unit that acquires vital data of a user to be estimated; a user group estimation unit that estimates a user group to which the vital data of the user to be estimated belongs, using a user group estimation model trained with learning data for user group estimation including multiple combinations of the user's vital data and user groups into which the vital data has been classified; and an illness estimation unit that estimates the type of illness most strongly felt by the user to be estimated, using an illness estimation model corresponding to the user group estimated by the user group estimation unit, out of illness estimation models trained for each user group, using learning data including multiple combinations of the user's vital data and the type of illness most strongly felt by the user at the time the vital data was measured.

[0098] (Supplementary Item 3) A learning method comprising: acquiring learning data including multiple combinations of a user's vital data and the type of illness that the user feels most strongly at the time the vital data is measured; classifying the multiple combinations included in the learning data into user groups based on the similarity of the vital data; training a user group estimation model using learning data for user group estimation including multiple combinations of the vital data included in the learning data and the user groups into which the vital data has been classified, so that when a user's vital data is input, the user group to which the user's vital data belongs is output as an estimation result; and training an illness estimation model using the learning data for each user group, so that when the vital data is input, the type of illness that the user feels most strongly is output as an estimation result.

[0099] (Supplementary Item 4) An estimation method comprising: acquiring vital data of a user to be estimated; estimating a user group to which the vital data of the user to be estimated belongs using a user group estimation model trained with learning data for user group estimation including multiple combinations of the user's vital data and user groups into which the vital data has been classified; and estimating the type of illness most strongly felt by the user to be estimated using an illness estimation model corresponding to the estimated user group from among the illness estimation models trained for each user group using learning data including multiple combinations of the user's vital data and the type of illness most strongly felt by the user at the time the vital data was measured.

[0100] (Supplementary Item 5) A program for causing a computer to function as each part of the learning device according to Supplementary Item 1 or each part of the estimation device according to Supplementary Item 2.

[0101] (Supplementary Item 6) A learning device comprising at least one processor, wherein the processor: acquires learning data including multiple combinations of a user's vital data and the type of illness that the user feels most strongly at the time the vital data is measured; classifies the multiple combinations included in the learning data into user groups based on the similarity of the vital data; trains a user group estimation model using learning data for user group estimation including multiple combinations of the vital data included in the learning data and the user groups into which the vital data has been classified, so that when a user's vital data is input, a user group to which the user's vital data belongs is output as an estimation result; and trains an illness estimation model using the learning data for each user group, so that when the vital data is input, a type of illness that the user feels most strongly is output as an estimation result.

[0102] (Supplementary Item 7) An estimation device comprising at least one processor, wherein the processor acquires vital data of a user to be estimated, estimates a user group to which the vital data of the user to be estimated belongs using a user group estimation model trained with learning data for user group estimation including multiple combinations of the user's vital data and user groups into which the vital data has been classified, and estimates the type of illness most strongly felt by the user to be estimated using an illness estimation model corresponding to the estimated user group out of the illness estimation models trained for each user group using learning data including multiple combinations of the user's vital data and the type of illness most strongly felt by the user at the time the vital data was measured.

[0103] (Addendum 8) A storage medium having stored therein a learning program for causing a computer to execute processes including: acquiring learning data including multiple combinations of a user's vital data and the type of illness that the user feels most strongly at the time the vital data is measured; classifying the multiple combinations included in the learning data into user groups based on the similarity of the vital data; learning a user group estimation model using learning data for user group estimation including multiple combinations of the vital data included in the learning data and the user groups into which the vital data has been classified, so that when a user's vital data is input, a user group to which the user's vital data belongs is output as an estimation result; and learning an illness estimation model using the learning data, so that when the vital data is input for each user group, the type of illness that the user feels most strongly is output as an estimation result.

[0104] (Supplementary Item 9) A storage medium having stored therein an estimation program for causing a computer to execute the following processes: acquiring vital data of a user to be estimated; estimating a user group to which the vital data of the user to be estimated belongs using a user group estimation model trained with learning data for user group estimation including multiple combinations of the user's vital data and user groups into which the vital data has been classified; and estimating the type of illness most strongly felt by the user to be estimated using an illness estimation model corresponding to the estimated user group out of the illness estimation models trained for each user group using learning data including multiple combinations of the user's vital data and the type of illness most strongly felt by the user at the time the vital data was measured.

[0105] 2 Learning data 10 Illness estimation system 12 Learning device 14 Estimation device 20 Learning data acquisition unit 22 User group estimation model learning unit 24 Illness estimation model learning unit 30 User group estimation model DB, 30M User group estimation model 32 Illness estimation model DB, 32M Illness estimation model 40 Estimation target acquisition unit 42 User group estimation unit 44 Illness estimation unit

Claims

1. A learning device comprising: a learning data acquisition unit that acquires learning data including a plurality of combinations of a user's vital data and the type of illness that the user feels most strongly at the time the vital data is measured; a user group estimation model learning unit that classifies the plurality of combinations included in the learning data into user groups based on the similarity of the vital data, and trains a user group estimation model using learning data for user group estimation including a plurality of combinations of the vital data included in the learning data and the user groups into which the vital data has been classified, so that when a user's vital data is input, a user group to which the user's vital data belongs is output as an estimation result; and an illness estimation model learning unit that trains an illness estimation model using the learning data for each user group, so that when the vital data is input, the type of illness that the user feels most strongly is output as an estimation result.

2. An estimation device comprising: an estimation target acquisition unit that acquires vital data of a user to be estimated; a user group estimation unit that estimates a user group to which the vital data of the user to be estimated belongs, using a user group estimation model learned from learning data for user group estimation including multiple combinations of the user's vital data and user groups into which the vital data is classified; and an illness estimation unit that estimates the type of illness felt most strongly by the user to be estimated, using an illness estimation model corresponding to the user group estimated by the user group estimation unit, out of the illness estimation models learned for each user group, using learning data including multiple combinations of the user's vital data and the type of illness felt most strongly by the user at the time the vital data was measured.

3. A learning method comprising: acquiring learning data including multiple combinations of a user's vital data and the type of illness that the user feels most strongly at the time the vital data was measured; classifying the multiple combinations included in the learning data into user groups based on the similarity of the vital data; training a user group estimation model using learning data for user group estimation including multiple combinations of the vital data included in the learning data and the user groups into which the vital data has been classified, so that when a user's vital data is input, a user group to which the user's vital data belongs is output as an estimation result; and training an illness estimation model using the learning data for each user group, so that when the vital data is input, the type of illness that the user feels most strongly is output as an estimation result.

4. An estimation method comprising: acquiring vital data of a user to be estimated; estimating a user group to which the vital data of the user to be estimated belongs using a user group estimation model trained with learning data for user group estimation including multiple combinations of the user's vital data and user groups into which the vital data is classified; and estimating the type of illness felt most severely by the user to be estimated using an illness estimation model corresponding to the estimated user group among the illness estimation models trained for each user group using learning data including multiple combinations of the user's vital data and the type of illness felt most severely by the user at the time the vital data was measured.

5. A program for causing a computer to function as each part of the learning device according to claim 1 or each part of the estimation device according to claim 2.

Citation Information

Patent Citations

  • Medical information processor, medical information processing method, and medical information processing program

    JP2009205464A

  • Pre-disease diagnosis device, pre-disease diagnosis method, and learning model generation device

    JP6949277B1

Cited By

  • Training device, estimation device, and training method

    WO2026186036A1