JP2016200955A - Information processing device, program, and estimation method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a health management device capable of estimating a mental health state of a person with high accuracy at an earlier stage.SOLUTION: Information processing devices 20, 19 for estimating a metal health state of a person include read means 29 for reading action state information from a storage device of action state information recording the action state of a person, and reading the quality of the mental health state from the storage device storing at least the quality of the mental health state of a person, feature vector creation means 24 for creating a feature vector of the action state of a person by using the action state information, and estimation means generation means 23 for generating estimation means for estimating the quality of the mental health state from the feature vector with the quality of the mental health state as a teacher signal.SELECTED DRAWING: Figure 1

Description

  The present invention relates to an information processing apparatus, a program, and an estimation method.

  From the viewpoint of preventing occupational accidents, companies and various organizations (hereinafter referred to as companies, etc.) are required to grasp the health status of employees and perform health management. In addition, the number of employees who become unable to work normally due to poor mental health has increased, and it has become more important for companies to conduct mental health management. It was.

  For this reason, companies and the like sometimes endeavor to manage mental health of employees by placing industrial physicians and health staff (hereinafter referred to as industrial physicians, etc.) in the workplace. However, it is not easy for occupational physicians etc. to identify employees with mental problems from many employees, and employees with mental problems are not always proactively consulting occupational physicians. Absent. Also, depending on the number of employees, it may be impractical to require all employees to consult.

  In view of this, it has been considered to use an information processing apparatus to manage the mental health of employees (see, for example, Patent Document 1). Patent Document 1 prepares determination rule data that defines a work pattern peculiar to an employee who develops a mental illness, and determines the employee's work history (for example, work time and work time history). A mental health management method for determining the mental health status of an employee according to whether or not the rule data is met is disclosed.

  However, in the mental health management method described in Patent Document 1, since it is set in the rule data that the mental health has already shown signs of malfunction, it is sufficient for the mental health management in advance. There is a problem that it may not be. Further, the rule data described in Patent Document 1, such as resting on the day after a holiday, is merely a sign of malfunction set up by an industrial physician or the like, and it is unclear whether this rule data is appropriate. There is a problem.

  In view of the above problems, an object of the present invention is to provide a health management device that can estimate a person's mental health state with higher accuracy at an earlier stage.

  The present invention is an information processing apparatus that estimates a person's mental health state, the behavior state information is read from a behavior state information storage device in which the person's behavior state is recorded, and the person's mental health state Reading means for reading out the quality of the mental health state from a storage device in which at least the quality is stored, and feature vector creating means for creating a feature vector representing the characteristics of the behavior situation of the person using the behavior situation information; And an estimation means generating means for generating an estimation means for estimating the quality of the mental health condition from the feature vector using the quality of the mental health condition as a teacher signal.

  It is possible to provide a health management device capable of estimating a person's mental health state with higher accuracy at an earlier stage.

It is an example of the figure which illustrates typically the judgment method of the mental state by the health management apparatus of this embodiment. It is an example of the external view of a health care apparatus. It is a figure which shows an example of the hardware constitutions of a health management apparatus or a health management server. It is an example of the functional block diagram of a health care apparatus. It is an example of the flowchart figure which shows the operation | movement procedure of the health management apparatus in a learning process, and the operation | movement procedure of the health management apparatus in the determination process using a mental state determination model. It is an example of the figure which illustrates the generated feature vector typically. It is an example of the flowchart figure explaining the determination procedure of the weight of a weak discriminator. It is an example of the figure which shows a mental state determination model typically. It is an example of the figure explaining the determination result of a mental state. It is an example of a functional block diagram of a health management device in which the malfunctioning person information reflecting the interview information of the industrial physician is used.

  Hereinafter, embodiments for carrying out the present invention will be described with reference to the drawings.

<Outline of Health Management Device of this Embodiment>
FIG. 1 is an example of a diagram schematically illustrating a method for determining an employee's mental state by the health management apparatus of the present embodiment.
(1) The health management device creates the strong classifier 50 by performing machine learning in advance using the teacher signal 53. As a teacher signal 53, work information of an employee whose mental health level has failed in the past is stored in the database. When this work information (an example of behavioral situation information) is represented by a feature vector 51, a model for determining the mental state using the work information as explanatory variables and the mental health level (not good: +2, good: +1) as an objective variable ( A strong discriminator 50) can be generated. For example, an appropriate weak classifier 54 is prepared and the weight α of each weak classifier 54 is learned using the teacher signal 53. Each weak discriminator 54 is machine-learned by each learning algorithm such as AdaBoost, linear discriminant analysis, and Random Forest (learning algorithm for creating a plurality of decision trees by limiting a part of feature vectors to be learned). It is. In this case, the strong classifier 50 is a combination of the weighted weak classifier 54. Further, the strong discriminator 50 may be learned by a single learning algorithm.
(2) Next, when a feature vector 51 representing the health state of an employee whose mental state is unsatisfactory is input to the strong discriminator 50, a weak discriminator 54 (a weak discriminator if each is distinguished) 1 to T) calculate a score for judging the mental state.
(3) The strong classifier 50 determines the mental state from the sum of the scores of the weak classifiers 1 to T, and outputs a determination result (good or bad) 52.

  In this way, the health management device of this embodiment does not create rules based on the experience of human beings such as industrial physicians, but the working status (behavioral status of employees) whose mental health has been poor in the past. Example) pattern is mechanically learned. Thereby, the pattern of the working state in which mental health is likely to be unsatisfactory can be obtained. Therefore, if a company etc. wants to judge the mental condition, if the employee's working status is entered into the health management device, employees who are mentally unwell can be identified at an earlier stage than showing signs of malfunction such as actually taking a leave of absence. Can be estimated. In addition, it is possible to estimate employees who have a mental state malfunction with higher accuracy than relying solely on the experience of an industrial physician or the like.

<Terminology>
Mental health level: This is a mental and mental health condition, and is called a mental health condition in this embodiment. An unsatisfactory or good mental health level indicates a mental health condition. The mental state represents the poor or good mental health at the time of determination.

  Mental upset: No method has been established for quantifying mental health with an observation device. For example, if the number of mental symptom symptom matches a predetermined number or more, it is judged as mental upset. Is done. The mental health level may be observed by a blood test, an electroencephalogram, or the like, and it is preferable to make an objective determination as much as possible.

  Employees: Not limited to those who provide labor and receive remuneration, but may be any person who can obtain information corresponding to working conditions.

  Discriminator: A device for judging the mental state from information corresponding to the working state. If the terms weak classifier and strong classifier appear depending on how the classifier is made (machine learning method), the strong classifier is the classifier. In some cases, only a classifier is created without a strong classifier or a weak classifier.

<Configuration example>
FIG. 2A shows an example of an external view of the health management device 20. The health management device 20 is an information processing device that outputs a determination result 52 indicating that the mental state is healthy or unsatisfactory with the employee's working state as an input. Specifically, for example, PC (Personal Computer), smart phone, tablet device, mobile phone, car navigation terminal, wearable computer (head mounted display, wristwatch), camera, electronic blackboard, projector, game machine, or MFP (Multifunction Peripheral / Printer / Product).

  FIG. 2B shows an example of a system configuration diagram of the health management system 100. As shown in the figure, in the health management system 100, an information processing terminal 18 and a health management server 19 are connected to be communicable via a network N. That is, as shown in FIG. 2A, the health management apparatus 20 can output the mental state determination result 52 by a client server type system in addition to outputting the mental state determination result 52 alone.

  In such a form, the user operates the information processing terminal 18 to transmit the employee identification information or the employee work status to the health management server 19. The health management server 19 reads the work status of the employee specified by the identification information from a database or the like, determines the mental status, and transmits the determination result 52 to the information processing terminal 18.

  2A and 2B, the health management device 20 or the health management server 19 has the strong classifier 50 by learning the work status of the employee in advance. Although this embodiment may be operated in any manner, the following mainly describes the manner in which the health management device 20 determines the mental state.

<Hardware configuration diagram>
FIG. 3 is a diagram illustrating an example of a hardware configuration of the health management device 20 or the health management server 19. The health management device 20 of the present embodiment has a configuration as an information processing device. The health management device 20 includes an input device 201, an output device 202, a drive device 203, an auxiliary storage device 204, a memory device 205, an arithmetic processing device 206, and a communication interface device 207 that are mutually connected by a bus B.

  The input device 201 is a mouse or a keyboard, for example, and is used for inputting various kinds of information. The output device 202 is, for example, a display and is used for displaying (outputting) various signals. The drive device 203 writes data to and reads data from various recording media 208. The recording medium 208 electrically records information such as a CD-ROM, a flexible disk, a magneto-optical disk, etc., a recording medium that records information optically, electrically, or magnetically, a ROM, a flash memory, etc. Various types of recording media such as a semiconductor memory can be used.

  The auxiliary storage device 204 is a non-volatile storage device such as a hard disk drive or a flash memory. The auxiliary storage device 204 stores a health management program, an operating system, files, data, and the like.

  The memory device 205 is a main storage device such as a DRAM, and the health management program is read from the auxiliary storage device 204 and stored therein.

  The arithmetic processing unit 206 is, for example, a CPU and controls the entire health management device 20 according to a health management program stored in the memory device 205. Further, various functions or means are realized as will be described later.

  The communication interface device 207 is a modem, a LAN card, or the like, and performs processing for connecting to the network N.

  The health management program is distributed in a state stored in the recording medium 208 in an executable format or a distribution format, for example. Alternatively, the program may be distributed from a server device for program distribution via a network.

  The health management server 19 and the information processing terminal 18 used in the health management system 100 are also configured as general information processing devices, and the hardware configuration thereof is the same as that of the health management device 20. In addition, even if there is a difference, it is assumed that there is no problem in explaining the present embodiment.

<About the functions of the health management device>
The functional configuration of the health management device 20 will be described with reference to FIG. FIG. 4 is an example of a functional block diagram of the health management device 20.

  The health management device 20 includes an output unit 21, a mental state determination unit 22, a mental state determination model generation unit 23, a feature vector generation unit 24, and a storage / readout unit 29. Each part of the health management device 20 includes a command from the arithmetic processing unit 206 according to the health management program 2100 in which any of the components shown in FIG. 3 is expanded from the auxiliary storage device 204 to the memory device 205. A function realized by operating by means of, or a means to be functioned.

  In addition, the health management device 20 includes a storage unit 2000 constructed by any one or more of the auxiliary storage device 204, the memory device 205, and the drive device 203 illustrated in FIG. The storage unit 2000 stores a health management program 2100 and various holding units. First, various holding units constructed in the storage unit 2000 will be described.

記憶部２０００には、表１に示されているような社員情報テーブル１，２によって構成されている社員情報保持部２００１が構築されている。社員情報テーブル１，２に記載された情報を勤務情報という。表１（ａ）の社員情報テーブル１には、社員ＩＤに対応づけて、毎日の出勤時刻、退勤時刻、残業時間、休憩時間、遅刻、及び、早退が登録されている。社員ＩＤは社員を一意に識別するための識別情報である。出勤時刻は出勤した時刻であり、退勤時刻は退社した時刻である。休憩時間は出勤から退社までの間の休憩時間である。遅刻は、遅刻したか否か及び遅刻した場合に遅刻時間（勤務時間内に勤務しなかった時間）が登録される。早退は、早退したか否か及び早退した場合の早退時間（勤務時間内に勤務しなかった時間）が登録される。

In the storage unit 2000, an employee information holding unit 2001 configured by employee information tables 1 and 2 as shown in Table 1 is constructed. Information described in the employee information tables 1 and 2 is called work information. In the employee information table 1 of Table 1 (a), the daily working hours, working hours, overtime hours, rest hours, late arrivals, and early departures are registered in association with the employee IDs. The employee ID is identification information for uniquely identifying the employee. The attendance time is the time when the employee has worked, and the leaving time is the time when the employee has left the office. The break time is the break time from work to work. As for the lateness, whether or not it is late and when it is late, the late time (time that did not work within working hours) is registered. As for early departure, whether or not early departure and early departure time when leaving early (time not working within working hours) are registered.

  In addition, in the employee information table 2 of Table 1 (b), the number of absentee days per month, the number of leave days, the leave without prior application, the number of days off work, the number of late arrivals, and the number of early departures are registered in association with the employee ID. ing. Absence days are the number of working days of the month that were not attended. The number of leave days is the number of leave days applied in advance among the absence days, and the leave without prior application is the number of leave days without application in advance. The number of attendance on holidays is the number of days worked on days other than work days. The number of late arrivals is the number of late arrivals in the month, and the number of early departures is the number of early departures in the month. The working hours may be registered, but the working hours can be calculated from these.

  Since there are various types of vacations, for example, one day vacation, morning vacation, afternoon vacation, disappointing vacation, menstrual leave, substitute leave, transfer holiday (transfer work), time leave, etc. can be recorded. Is preferred. Also, leave without prior application may be recorded separately from leave with no application (that is, there is no application after taking a leave).

  In addition, it is essential for work information to include attendance information as described above. In addition to this, employee attribute information (department, office, occupation, age, etc.), health check results (weight) , Height, body fat percentage, blood pressure, blood glucose level, numerical values such as total protein and LDH obtained by blood tests), eating history (calories, nutrients, etc. included in menus eaten at employee cafeterias), questionnaire results, sensors Physiological information obtained from, etc. may be included.

The questionnaire result is, for example, a set of a question and an answer for estimating the employee's state. As an example,
“A. 6 hours” to “Q. How long is the sleep time?”
“A. Occasionally” to “Q. Do you feel depressed?”
Etc.

  Data obtained from sensors, etc. includes heartbeats, brain waves, image processing results of images taken by the camera (facial color, eye opening, etc. are extracted by image processing), PC work records, and mail transmission / reception history analysis. Results (received mail from related parties related to the business, received mail from the superior, and how many outgoing mails responded to them).

記憶部２０００には、表２に示されているようなメンタル健康度テーブルによって構成されているメンタル健康度情報保持部２００２が構築されている。メンタル健康度テーブルには、社員ＩＤに対応づけて、月ごとの過去のメンタル健康度が登録されている。メンタル健康度は良又は不調のどちらかである。後述するように、良は例えば＋１に、不調は＋２に変換される。メンタル健康度は用いられる機械学習が扱いやすい数値に変換されればよく表２と変換後の数値は一例に過ぎない。

In the storage unit 2000, a mental health information holding unit 2002 configured by a mental health table as shown in Table 2 is constructed. In the mental health table, past mental health levels for each month are registered in association with employee IDs. Mental health is either good or bad. As will be described later, good is converted to +1, for example, and malfunction is converted to +2. The mental health level only needs to be converted into a numerical value that can be easily handled by the machine learning used. Table 2 and the converted numerical value are merely examples.

  In Table 2, the mental health level is displayed in two levels, but may be registered in three or more levels.

  Further, the mental health table need not include the mental health levels of all employees. In this case, the mental health level of an employee whose mental health level is not registered is regarded as +1 (good).

  The mental state determination model holding unit 2003 stores the generated mental state determination model as described later. The determination result holding unit 2004 stores a determination result 52 indicating whether the mental state of the employee is good or bad.

  Next, each function of the health management device 20 will be described. The feature vector generation unit 24 of the health management device 20 is realized by an instruction from the arithmetic processing unit 206 shown in FIG. 3, and generates the feature vector 51 from the work information of each employee.

  The mental state determination model generation unit 23 is realized by an instruction from the arithmetic processing unit 206 shown in FIG. 3 and generates a mental state determination model using the feature vector 51 and the mental health degree table. The mental state determination model is stored in the mental state determination model holding unit 2003. Generating a mental state determination model corresponds to determining the weight of the weak classifier 54 or creating the strong classifier 50 in machine learning.

  The mental state determination unit 22 is realized by a command or the like from the arithmetic processing unit 206 shown in FIG. 3 and applies a feature vector 51 of work information representing an employee's work state to the mental state determination model. Judge whether the product is good or bad. The determination result 52 is stored in the determination result holding unit 2004.

  The output unit 21 is realized by an instruction from the arithmetic processing device 206 and the output device 202 shown in FIG. 3, and outputs a determination result 52 by the mental state determination unit 22.

  The storage / reading unit 29 is realized by an instruction from the arithmetic processing unit 206 and the auxiliary storage device 204 shown in FIG. 3, and stores various data in the storage unit 2000 or various data stored in the storage unit 2000. The process of reading out is performed.

<Overall operation process>
FIG. 5A is an example of a flowchart illustrating an operation procedure of the health management device 20 in the learning process. FIG. 5B is an example of a flowchart illustrating an operation procedure of the health management device 20 in the determination process using the mental state determination model. In the present embodiment, one of the features is that an appropriate determination criterion is automatically generated by applying machine learning to past work information by this learning process. Each step L1-L6 of Fig.5 (a) is demonstrated in order. Each step D1-D6 of FIG.5 (b) is mentioned later.

<< L1 Get work information L2 Generate feature vector 51 >>
First, the storage / reading unit 29 reads work information from the employee information holding unit 2001 (step L1). The same applies to step D1, but in step D1, only the work information of the employee whose mental state is to be determined needs to be read. All employees can be judged.

  Then, the feature vector generation unit 24 generates the feature vector 51 by the following processing. The feature vector generation unit 24 extracts or calculates a plurality of types of feature amounts from work information of each employee ID using work information for a certain period (for example, one month). A vector having these feature amounts as elements is a feature vector 51. The type of work information used for the vector is the number of dimensions.

  Specific examples of the feature amount include an average, a standard deviation, a maximum value, and a minimum value for a certain period of work hours, work hours, overtime hours, and break times. Also, the number of absentee days obtained, the number of leave days, the number of leave days without prior application, the number of days off work, the number of late arrivals, the number of early departures, and the like can be mentioned.

  In addition, since it is considered that the accumulation of past work conditions influences the mental health level, in addition to the period in which the user of the health management apparatus 20 wants to determine the mental condition, the feature amount of the past period is used together. It is also suitable. For example, when the mental state of June 20XX is determined, the feature vector 51 using the work information of June 20XX, the work information of May 20XX, and the work information of April 20XX is created. As a result, it is expected that a mental state judgment model that can judge the mental state of June 20XX more accurately can be created.

  FIG. 6 is an example of a diagram schematically illustrating the generated feature vector 51. In FIG. 6, feature values for any year and month are shown for each employee ID from 1 to N. The feature amount 1 is, for example, the average working hours on the first day of the year and month, and the feature amount N is, for example, the total number of holiday attendances for the year and month. The feature quantities 1 to N may be normalized so that the length of the feature vector 51 is 1.

<< Addition of L3 mental health information >>
Next, the mental state determination model generation unit 23 acquires the generated feature vector 51 from the feature vector generation unit 24. Then, the mental health level of the employee corresponding to each feature vector 51 is acquired from the mental health level information holding unit 2002. The mental health level to be acquired is the mental health level of the same month as the feature vector 51 serving as a teacher signal (or the last month when there are multiple feature vectors 51). Since the feature vector 51 and the mental health level are associated with the employee ID, both are associated with the employee ID.

  At the time of acquisition, the mental state determination model generation unit 23 converts good mental health level to +1 and malfunctions to +2. This is to make it easier for the health management device 20 to handle the mental health level (to learn the relationship between the feature vector 51 and the mental health level).

<< L4 learning of each weak classifier, L5 weighting determination >>
In the present embodiment, there is a case where the weight of the weak classifier 54 is learned using a single learning algorithm, or a case where the weak classifier 54 of a different learning algorithm is combined to learn the weight of each weak classifier 54. First, a weight learning method when learning with a single learning algorithm will be described.

(When learning with a single learning algorithm)
Hereinafter, as an example, a method of generating the weak classifier 54 included in the strong classifier 50 in the method of creating the strong classifier 50 by Boosting called AdaBoost will be described. The weak classifier 54 is a function that outputs +2 (not good) or +1 (good) for the feature vector 51 generated from the work information. Here, +2 (unsatisfactory) is output when the mental state is determined to be abnormal from the feature vector 51, and +1 (good) is output when the mental state is determined to be good from the feature vector 51. The Since the mental health level (+2, +1) is added to the feature vector 51, it can be determined whether or not the output of the weak classifier 54 is appropriate.

  The weak discriminator 54 only needs to output a result that is appropriate to some extent (for example, the mental health level and the output need only match about half of the number of feature vectors 51 prepared). Therefore, if the creation method is instructed by the creator of the health management device 20, the health management device 20 can be created automatically. For example, mental health problems are poor for employees who meet the average daily working hours, long overtime hours, many vacations without prior application, and many work hours on holidays. Can be estimated. Since these are replaced with feature amounts 1 to N, +2 (unsatisfactory) is output when some of feature amount 1> a, feature amount 2> b,..., Feature amount N> n is satisfied. What is necessary is just to produce | generate a function (an is a numerical value used as a threshold value). If this condition is not satisfied, +1 (good) is output. The range that the numerical values a to n can take is designated in advance by the creator of the health management device 20 or the like. The mental state determination model generation unit 23 changes a numerical value of “a” to “n” and a determination of magnitude, and generates a plurality of (in this embodiment, M) weak classifiers.

  Further, the weak classifier 54 may determine the mental state by an expression such as “feature quantity 1-a + feature quantity 2-b +... + Feature quantity N−n> threshold”. Further, the weak classifier 54 may determine the mental state by an expression such as “a × feature amount 1 + b × feature amount 2+... + N × feature amount N> threshold”. Further, the creator or the like may create all or part of the weak classifiers in advance. Such an expression will be referred to as a score calculation expression.

  By the way, in the process in which the weak discriminator 54 outputs +2 or +1, a numerical value for determining +2 or +1 is obtained. For example, when the determination of feature quantity 1> a, feature quantity 2> b,..., Feature quantity N> n is included in one weak classifier 54, the following numerical values are obtained. If feature amount 1> a, 1 is held otherwise, and if the determination based on feature amount N is performed, 1 or 0 is obtained for the number of feature amounts. Therefore, for example, a total value of 1 and 0 obtained by adding the number of feature amounts is obtained. In this case, the weak discriminator 54 compares the total value with a threshold value (for example, half of the number N of feature amounts), and finally outputs +2 (not good) or +1 (good).

  Further, when the weak classifier 54 outputs “+2 (unsatisfactory)” when “feature quantity 1−a + feature quantity 2-b +... + Feature quantity N−n> threshold”, the value of the left side is obtained in the classification process. Will be. Further, when the weak discriminator outputs “2” (malfunction) when “a × feature amount 1 + b × feature amount 2+... + N × feature amount N> threshold”, the value on the left side is obtained.

  A numerical value obtained in the process in which the weak classifier 54 outputs +2 or +1 can be an index representing the degree of employee malfunction. The numerical value of only one weak classifier 54 is poor in reliability, but the total of the numerical values of each weak classifier 54 is considered as an index indicating the degree of employee malfunction (after learning and weighting is determined). Good. In the present embodiment, a numerical value calculated by each weak classifier 54 is referred to as a score, and a numerical value calculated by the strong classifier 50 (total of each score) is referred to as a mental score (an example of an estimated index). Therefore, the health management apparatus 20 can not only determine the malfunction or good, but also estimate the degree of malfunction.

  Next, learning using the weak classifier 54, that is, a weighting determination method for each weak classifier 54 will be described with reference to FIG. FIG. 7 is an example of a flowchart for explaining the weighting determination procedure of the weak classifier 54.

  First, the mental state determination model generation unit 23 initializes the weight D of each feature vector 51 created in step L2 (S10). If the number of feature vectors 51 is m, it can be expressed as “weight D = 1 / m”. That is, at the beginning of learning, the weights of the feature vectors 51 are equal, and the sum thereof is 1. The reason why the weight is given to the feature vector 51 for learning is to give a large weight to the feature vector 51 that the weak discriminator 54 cannot correctly discriminate when evaluating the weak discriminator 54.

  Next, learning is repeated T times (S20). The number of times T is estimated to have been sufficiently learned, and about several hundred times is sufficient. Alternatively, it may be terminated when the ratio at which the strong discriminator 50 outputs the correct determination result 52 becomes a predetermined value or more.

  Next, the mental state determination model generation unit 23 inputs m feature vectors 51 to all M weak discriminators 54 (30). As a result, m +2 or +1 are obtained for each weak classifier 54.

  The mental state determination model generation unit 23 calculates the error rate e of each weak classifier 54 (S40). As described above, since the mental health level (+2, +1) is added to each feature vector 51, the output of the weak classifier 54 is +2 and the mental health level is +2, or the output of the weak classifier 54 is When +1 and mental health are +1, the output of the weak classifier 54 is correct. On the other hand, in other cases, the output of the weak classifier 54 is wrong (an error is output). The mental state determination model generation unit 23 calculates an error rate e at which an error is output for the number m of the feature vectors 51 for each weak classifier 54.

  Next, the mental state determination model generation unit 23 selects the weak classifier 54 that minimizes the error rate e (S50).

  Next, the mental state determination model generation unit 23 calculates the weight α of the weak classifier 54 selected in step S40 (S60). For example, the weight α is calculated as follows.

これで重み付きの弱識別器５４が１つ選択されたことになる。重みαの決定については式（１）に限られず、他の手法で決定してもよい。

As a result, one weighted weak classifier 54 is selected. The determination of the weight α is not limited to the expression (1), and may be determined by another method.

  Next, each weight D of the feature vector 51 is updated (S70). The updating formula of the weight D can be expressed as shown in Expression (2). The reason for updating the weight D is to make it easier to select the weak classifier 54 that correctly classifies the feature vector 51 that could not be correctly classified by the weak classifier 54 selected so far.

Ｄ_t+1(i)…ｉ番目の特徴ベクトル５１の更新後の重み。ｔは学習回数。
Ｄ_t(i)…ｉ番目の特徴ベクトル５１の更新前の重み。
α_t…選択された弱識別器の式（１）で計算された重み。
ｙ_i…ｉ番目の特徴ベクトル５１に付加されたメンタル健康度。
ｈ_t（ｘ_i）…選択された弱識別器５４に対しｉ番目の特徴ベクトル５１が適用された場合の弱識別器５４の出力。
分母…各特徴ベクトル５１の重みの合計を表しており、各特徴ベクトル５１の重みの合計を１にするための正規化係数。

D _{t + 1} (i)... Weight after updating the i-th feature vector 51. t is the number of learnings.
D _t (i): Weight before updating the i-th feature vector 51.
α _t is a weight calculated by the formula (1) of the selected weak classifier.
y _i ... Mental health level added to the i-th feature vector 51.
h _t (x _i )... output of the weak classifier 54 when the i-th feature vector 51 is applied to the selected weak classifier 54.
Denominator: A normalization coefficient that represents the sum of the weights of the feature vectors 51 and sets the sum of the weights of the feature vectors 51 to 1.

In this equation (2), if the output of the selected weak classifier 54 matches the mental health level added to the feature vector 51, “−α _t y _{i ht} (xi)” becomes a negative value. Therefore, exp is less than 1 and the weight D is reduced. When the added mental health in the output feature vector 51 of the selected weak classifier 54 does not match, _{"- α t y i h t (} xi) " is a positive value. Therefore, exp is greater than 1 and the weight D is increased.

  Thereafter, the process is repeated from step S20. For example, by repeating T times, the weight α of T weak classifiers is determined, and the T weak classifiers 54 are combined to obtain the strong classifier 50. The strong classifier 50 can be expressed as follows.

式（３）の出力が、任意の従業員のメンタル状態の判定結果５２を表す。すなわち、値が大きいほど不調の可能性が高い。また、各弱識別器５４が＋２、＋１を判定する際の数値（スコア）を取り出し合計すればメンタルスコアが得られる。メンタルスコアが大きいほど不調の度合いが高いことを意味する。

The output of the expression (3) represents the determination result 52 of an arbitrary employee's mental state. That is, the larger the value, the higher the possibility of malfunction. Also, a mental score can be obtained by extracting and summing up numerical values (scores) when each weak classifier 54 determines +2, +1. The higher the mental score, the higher the degree of malfunction.

(In the case of a combination of weak classifiers 54 of different learning algorithms)
The feature vector generation unit 24 creates a plurality of weak classifiers 54 using at least a plurality of learning algorithms. As described with reference to FIG. 7, the weak classifier 54 may be created by a method suitable for each machine learning method. As a result, in addition to the above AdaBoost, for example, linear discriminant analysis and the Random Forest weak classifier 54 are obtained. Note that the machine learning learning algorithm to be combined is not limited to these.

  When weak discriminators 54 of different learning algorithms are combined, the weight α of each weak discriminator 54 is determined by the following three methods or other methods. Hereinafter, a method of determining the weight α by the three methods will be described.

1. (i) Each weak classifier 54 calculates a score.
(ii) For each weak classifier 54, for example, the top 100 (or top XX%) of the scores are determined.
(iii) For each weak discriminator 54, calculate the percentage of mental health level +2 (strange) among the top 100 people.
(iv) “100 × the ratio of each weak classifier / the total ratio of all weak classifiers” is determined as the weight of each weak classifier 54.

2. (i) Each weak classifier 54 calculates a score.
(ii) A score threshold is set so that the percentage of employees whose mental health level is +2 (not good) is XX% (for example, 50%). That is, the threshold value is determined so that the mental health degree of A / 2 of the employees A from the top of the score is +2 (not good).
(iii) With the determined threshold value, each weak classifier 54 again determines +2 (malfunction) or good (+1) for the feature vector 51.
(iv) The recall of all weak classifiers 54 is calculated. The recall rate is the ratio of employees (feature vectors) whose weak classifier 54 outputs +2 (malfunction) to the number of employees whose mental health level is +2 (malfunction).
(v) “100 × recall rate of each weak classifier / total recall rate of all weak classifiers” is determined as the weight α of each weak classifier 54.

3. (i) The weight α of each weak classifier 54 is arbitrarily determined.
(ii) The F value of the strong discriminator 50 is defined using the weight α and the threshold value for outputting the mental state as parameters. The F value is the harmonic mean of recall and precision, defined by
F value = 2 / (1 / Fit rate + 1 / Recall rate)
The relevance rate is the number of employees whose mental health level is +2 (not good) among the number of employees determined by the weak classifier 54 to be +2 (not good). Therefore, the F value can be obtained by changing some of the weight α and the threshold value for determining the mental state.
(iii) Parameters (α, threshold) are determined using various optimization algorithms (such as the steepest descent method and Newton method) so that the F value is maximized.

<< Save L6 Mental State Judgment Model >>
The mental state determination model generation unit 23 stores the mental state determination model expressed by Equation (3) in the mental state determination model holding unit 2003. This completes the learning process of FIG.

  In the above, as an example, a mental state determination model was generated using machine learning called AdaBoost. As machine learning to be used, linear discriminant analysis, decision tree, support vector machine, logistic regression, Hidden Markov model, perceptron Any network such as a neural network, deep learning, or Bayesian network may be used. A plurality of machine learning may be combined.

  FIG. 8 is an example of a diagram schematically illustrating a mental state determination model. As described so far, the mental state determination model has one or more weak classifiers 54. Each weak classifier 54 calculates a different score for the same feature vector 51 in order to identify the mental state by different analysis methods or different judgment rules. Further, when the weak classifier 54 performs machine learning using different learning algorithms, not only the analysis method or the determination rule but also the feature vector suitable for learning is different. For example, linear discriminant analysis uses work information for June 20XX (when judging the mental state of June 20XX). On the other hand, it may be preferable to use work information from April to June 20XX in Random Forest. If the feature vectors used for learning are different, each weak classifier 54 calculates a different score. Therefore, by adding the scores output by the weak classifiers 54 machine-learned with different learning algorithms, it is possible to output the mental score obtained by multifaceted judgment. Note that the scores of the weak classifiers 54 may be output individually without being summed.

The strong discriminator 50 adds the outputs (scores) of the weak discriminators 54 to calculate a mental score. In the figure, it is calculated as follows based on the scores (3.0, 1.8, 0.8, 2.4) of the four weak classifiers 54.
Mental score = 3.0 + 1.8 + 0.8 + 2.4 = 8.0
If the mental score is large, it can be determined that the degree of malfunction is high. Therefore, if the feature vector 51 of the employee to be determined for the mental state is input to the strong discriminator 50, the mental state can be determined based on the mental score.

  When the weak classifiers 54 of a plurality of learning algorithms are combined, the score of each weak classifier 54 can be calculated as follows, for example.

For example, when linear discriminant analysis is used as the learning algorithm of the weak classifier 54, the score can be calculated by the following score calculation formula. For example, suppose that the following is a plane that divides an N-dimensional feature vector dotted in an N-dimensional space into two groups of bad and good.
a ₁ × k ₁ + a ₂ × k ₂ +... + a _n × k _n = 0 (4)
Therefore, it is possible to determine which group belongs to the positive or negative value of the value on the left side (not good if positive, good if negative). In other words, the larger the value on the left side, the stronger the degree of malfunction, and this value can be treated as a score.

Further, when a neural network is used as a learning algorithm, it is as follows. The neural network using one or more intermediate layers and one neuroserpin the output layer, is multiplied by the weighting w ₁ to w _n to the input value x ₁ ~x _n to neuroserpin, a value obtained by subtracting the threshold value v The output is u.
u = Σx _i w _i −v (5)
The value of u is input to a function such as a sigmoid function to obtain a neurocell output z. If the output z of all the intermediate layers is input to the output layer, z output from the output layer can be used as a score.

  When Random Forest is used as an algorithm, the score can be calculated by, for example, summing the discrimination results (+1, +2) by each decision tree.

  Therefore, the mental score is obtained by summing up the scores of these weak classifiers 54.

<Judgment process>
Next, the determination process in FIG. 5B will be described.

<< D1 Get work information D2 Generate feature vector 51 >>
Steps D1 and D2 are the same as steps L1 and L2. However, in step L1, work information of an employee whose mental state is determined is read out.

<< Load D3 model >>
The mental state determination unit 22 reads the mental state determination model from the mental state determination model holding unit 2003.

<< D4 model feature vector 51 input, D5 mental score calculation >>
The mental state determination unit 22 inputs the feature vector 51 generated in step D2 to the mental state determination model. Thereby, the mental score mentioned above is calculated. The determination result 52 of the mental state is stored in the determination result holding unit 2004.

<< D6 Result output >>
The output unit 21 reads the mental state determination result 52 from the determination result holding unit 2004 and outputs the result to the output device 202 or the like.

  FIG. 9A is an example for explaining the mental state determination result 52. The determination result 52 in FIG. 9 includes an employee ID, a mental score, and a determination reason that contributes to the mental score. As explained in the above calculation method, the mental score represents the degree of malfunction of the targeted employee (the higher the numerical value, the higher the degree of malfunction). For example, in FIG. 9A, the mental scores of four employees are displayed, and it can be seen that the degree of malfunction is high in the order of K98765> K01234> K77777> K99999.

  As a result, occupational physicians and occupational physician health staff (an example of a specialist) can prioritize interviews from employees with high mental scores. Alternatively, an employee to be interviewed may be determined using a rule that an interview is conducted if the mental score is equal to or greater than a threshold value. Moreover, you may output a malfunction with respect to a person whose mental score is more than a threshold value.

  Next, the reason for determination will be described. The item of work information that contributes the most to the mental score is described in the judgment reason. When the mental state determination model has a plurality of weak classifiers 54 learned by, for example, Ada Boost, the determination reason in the weak classifier 54 having the largest weight α is adopted. For example, when the weak classifier 54 having the largest weight α compares the feature amounts 1 to N and the threshold values a to n to determine the mental state, the feature amount having the largest difference between the feature amount and the threshold values a to n is determined. For reasons. When the score calculation formula of the weak classifier 54 calculates the score by linear combination of the coefficients an to n and the feature vector 51, the feature amount having the largest coefficient or the feature amount having the largest product of the coefficient and the feature amount is set as the determination reason. . A plurality of weak classifiers 54 may be specified in descending order of the weight α, or a plurality of feature quantities may be specified from the top of the coefficient.

  Even when the mental state determination model has a plurality of weak classifiers 54 learned by different learning algorithms, first, the weak classifier 54 having the largest weight α is selected. Then, the feature amount is determined according to the calculation result of the score calculation formula of the weak classifier 54. For example, when the score is calculated by a linear combination of the score calculation formulas, the feature amount having the largest coefficient or the feature amount having the largest product of the coefficient and the feature amount is set as the determination reason. Alternatively, the weak discriminator 54 using a linear discriminant analysis or a decision tree as a learning algorithm determines the feature quantity as described below.

  When linear discriminant analysis is used as a learning algorithm, “a1 to an”, which is a coefficient (weight) of equation (4) obtained by learning, and the product of each feature amount, “a1 × k1, a2 × k2,. , An × kn ”, the term having the largest value (or a value greater than a certain threshold value) is specified. The item of work information corresponding to this item is the reason for determination.

  In addition, when a decision tree (a learning algorithm that extracts items of work information that becomes unsuccessful by tracing the items of work information that cause trouble) is used, enumerates the items that have arrived unsatisfactory . For example, it is assumed that a decision tree is obtained that is out of order when feature quantity 1> a and feature quantity 2> b, or when feature quantity 1> a and feature quantity 3> c. If the feature vector of the employee to be judged satisfies feature quantity 1> a and feature quantity 3> c for this decision tree, feature quantity 1 and feature quantity 3 are the reason for judgment.

(Problem probability)
Further, as shown in FIG. 9B, a malfunction probability may be output as the determination result 52 instead of the mental score. For example, the malfunction probability is calculated as follows.
(i) First, the mental scores of all employees are divided into several groups (for example, 5 to 20 groups) that may be regarded as the same.
(ii) Next, count how many employees in each group have a mental health level of +2 (unsatisfactory) (employees who are actually unsuccessful).
(iii) Find the ratio of employees whose mental health is +2 (stuck) to the number of people in each group. (iv) When the group to which the mental score calculated by the mental state determination unit 22 belongs is determined, the ratio obtained in (iii) of the group is a malfunction probability.

  For example, if there are 10 employees with a mental score of 20 (or grouped into 20 groups because they are close to 20), and 8 of them have a mental health level of +2, then the mental state is judged. The malfunction probability of the employee whose part 22 calculates a mental score of 20 or close to 20 is 80%.

  According to such a malfunction probability, since the approximate probability of malfunction is known, the user of the health management apparatus 20 can easily understand the meaning of the numerical value output by the health management apparatus 20.

  As described above, the health management apparatus 20 according to the present embodiment does not create rules based on the experience of human beings such as industrial physicians, but instead uses the pattern of the working status of those who have experienced mental problems in the past. To learn. Therefore, it is possible to estimate employees who are mentally malfunctioning with higher accuracy at an earlier stage than actually showing signs of malfunction such as actually taking a leave of work.

<Other application examples>
The best mode for carrying out the present invention has been described above with reference to the embodiments. However, the present invention is not limited to these embodiments, and various modifications can be made without departing from the scope of the present invention. And substitutions can be added.

  For example, instead of the mental health information holding unit 2002, unsuccessful person information reflecting interview information of an industrial physician may be used.

  FIG. 10 is an example of a functional block diagram of the health management device 20 in which the malfunctioning person information reflecting the interview information of the industrial physician is used. In the storage unit 2000 of the health management device 20, an occupational physician interview information holding unit 2005 and a malfunctioning person information holding unit 2006 are constructed. For example, the storage / reading unit 29 generates the malfunctioning person information based on the occupational physician interview information and records it in the malfunctioning person information holding unit 2006. Therefore, the same information as the mental health information holding unit 2002 is stored in the poor person information holding unit 2006.

  The occupational physician determines the mental malfunction at the interview with each employee, and holds the result determined from the psychosomatic medicine standpoint in the occupational physician interview information holding unit 2005 as a part of the interview record. With such an aspect, the mental state of the employee can be determined more accurately.

  4 and 10 shown in the above embodiments are divided according to main functions in order to facilitate understanding of the processing of the health management device 20. For example, a plurality of health management devices 20 may exist, and the plurality of health management devices 20 may perform the processing of the present embodiment in cooperation. The same applies to the health management server 19.

  In addition, the storage unit 2000 of the health management device 20 may not be included in the health management device 20, and each holding unit of the storage unit 2000 may be in a location where the health management device 20 can read and write data.

  In FIG. 6, the function of the health management device 20 is divided into several processing units. However, the present invention is not limited by the way of dividing or the name of each processing unit. The processing of the health management device 20 can be divided into more processing units according to the processing contents. Moreover, it can also divide | segment so that one process unit may contain many processes.

DESCRIPTION OF SYMBOLS 18 Information processing terminal 19 Health management server 20 Health management apparatus 21 Output part 22 Mental state determination part 23 Mental state determination model generation part 24 Feature vector generation part 50 Strong classifier 54 Weak classifier 100 Health management system

Japanese Patent No. 4913647

Claims (10)

An information processing apparatus for estimating a person's mental health state,
Reading out the behavior status information from a storage device of behavior status information in which the behavior status of the person is recorded, and reading out the quality of the mental health status from a storage device in which at least the quality of the mental health status of the person is stored Means,
Feature vector creating means for creating a feature vector representing the feature of the behavior situation of the person using the behavior situation information;
An estimation means generating means for generating an estimation means for estimating the quality of the mental health condition from the feature vector using the quality of the mental health condition as a teacher signal;
An information processing apparatus comprising: Index generating means for inputting the feature vector created from the behavior status information of the person by the feature vector creating means to the estimating means, and generating an estimated index of the mental health state;
Output means for outputting the estimated index generated by the estimation means by the index generation means,
The index generation means calculates the ratio of the number of persons stored in the storage device that the mental health condition is not good, which is obtained for the number of persons for which the estimated index that can be regarded as the same is generated. The information processing apparatus according to claim 1, wherein a probability that the mental health state of the person is not good is output based on the information. The index generation means identifies an element of the feature vector that is the reason for determining that the mental health state is not good when the estimation means generates the estimation index,
The information processing apparatus according to claim 2, wherein the output unit outputs the item of the behavior situation information corresponding to the element based on the correspondence between the item of the behavior situation information and the element.   As the elements of the feature vector, the average attendance time, the average overtime hours, the average break time, the standard deviation of the attendance time, the standard deviation of the overtime hours, the standard deviation of the break hours, the number of absentee days of the person were acquired as the elements of the feature vector The feature vector creation means uses at least one of the number of days of vacation, the number of days of acquired physiological leave, the number of days of substitute leave acquired, the number of days of leave without prior application, and the number of late arrivals. Item 4. The information processing device according to any one of Items 1 to 3.   The feature vector creation means includes the behavior status information of the person in a period prior to the period in addition to the behavior status information of the period that is the determination target of the mental health state of the person. The information processing apparatus according to claim 4, wherein the feature vector is created by including the feature vector.   The information according to any one of claims 1 to 5, wherein the quality of the mental health state stored in the storage device is created based on interview records of the person and an expert. Processing equipment.   The estimation means has a plurality of weak classifiers that have been machine-learned with different learning algorithms, and the sum of the products of the outputs of the weak classifiers and the weights set for each weak classifier is the index generation The information processing apparatus according to claim 2, wherein the means is generated as the estimated index. The feature vector creation means includes the behavior status information,
The person's department, business location, occupation, age, health check result, eating history, questionnaire result with a set of questions for estimating the person's mental health and answers to the question, heart rate, brain wave, and camera 8. One or more of an image processing result of a photographed image, a work record of a PC (Personal Computer), and an analysis result of a mail transmission / reception history are elements of the feature vector. The information processing apparatus according to item 1. An information processing device that estimates human mental health
Reading out the behavior status information from a storage device of behavior status information in which the behavior status of the person is recorded, and reading out the quality of the mental health status from a storage device in which at least the quality of the mental health status of the person is stored Means,
Feature vector creating means for creating a feature vector representing the feature of the behavior situation of the person using the behavior situation information;
An estimation means generating means for generating an estimation means for estimating the quality of the mental health condition from the feature vector using the quality of the mental health condition as a teacher signal;
Program to function as. An estimation method for estimating a person's mental health state performed by an information processing device,
Reading out the behavior status information from a storage device of behavior status information in which the behavior status of the person is recorded, and reading out the quality of the mental health status from a storage device in which at least the quality of the mental health status of the person is stored Steps,
A feature vector creating step for creating a feature vector that represents the feature of the behavior situation of the person using the behavior situation information;
An estimation unit generating step for generating an estimation unit for estimating the quality of the mental health state from the feature vector using the quality of the mental health state as a teacher signal;
An estimation method characterized by comprising:

Images