JP7534749B2 - Healthcare Systems and Programs - Google Patents

Description

The present invention relates to a healthcare system and a healthcare program that support the healthcare of members of an organization.

In Patent Document 1, health status (health checkup data) is set as the objective variable, and data obtained by breaking down attendance data over time and aggregating it (monthly, etc.) is set as the explanatory variable to create a machine learning model for prediction using machine learning or other methods. When an employee's attendance data is input into this machine learning model, the influential explanatory variables and their values can be determined. The original attendance data for the influential explanatory variables is output as a summary explanation. Furthermore, when other attendance data is input, the health status for that attendance is predicted. In Patent Document 1, predictions are made using data in which employees are classified (grouped) by "attendance data," "health checkup data," and "items such as place of business and type of job."

Japanese Patent Laid-Open No. 2003-233693 describes classifying (grouping) people based on their health condition (health checkup results) and a questionnaire, extracting characteristics from the questionnaire of each group, and providing guidance.
Patent document 3 describes how stress check data, attendance data, and biometric data (saliva) are used as weighting parameters through multivariate analysis, and by applying the stress check data, attendance data, and biometric data (saliva) to this, stress prediction data over the course of years is created.

Patent No. 6105825 JP 2010-170534 A JP 2020-52757 A

In Patent Document 1, the explanatory variable is simply attendance data, so there is a possibility that similar health conditions will be predicted if working conditions are similar. For example, even if employees have similar working conditions, if their daily exercise habits, alcohol consumption, and other lifestyle habits are different, their health conditions will be dissimilar. Furthermore, Patent Document 2 does not predict an employee's future. Furthermore, Patent Document 3 uses weighting parameters and applies stress checks, attendance data, and biological data (saliva) to this to create stress prediction data over time, but does not take lifestyle habits into account.

In a healthcare system for solving the above problem, health checkup judgment aggregate data that is an aggregate of health checkup judgment data related to health checkup results of members of a plurality of organizations, stress check judgment aggregate data that is an aggregate of stress check judgment data related to stress check results of the members, attendance aggregate data that is an aggregate of attendance history related to attendance for each of the members, personnel history data related to personnel, and attendance history data related to attendance and lifestyle history data related to lifestyle for each member are each item data, and a collection of the item data is work style data. The healthcare system is a healthcare system that uses the work style data to predict the judgment of items related to health checkups and stress checks for a prediction target person. and a classification data generation unit which generates initial combined data that combines the health-related combined data with the working style data for a period going back from the implementation date of the previous judgment of the items related to the health checkup and stress check of the member, the initial combined data being generated by the working style data classification unit, and which divides the initial combined data for each classification, and a classification data generation unit which generates classification data by dividing the initial combined data for each classification, and a classification data generation unit which uses judgments related to the current health checkup judgment data and stress check judgment data of the member as objective variables for each of the classification data, The system includes a teacher data generation unit that generates teacher data by setting each item data of the working style data and the judgments related to the previous health check judgment data and stress check judgment data of the member as explanatory variables, a learning unit that generates a machine learning model for predicting the objective variable for each judgment of the explanatory variables, and a prediction unit that classifies the working style data for a period going back from the date of the current judgment of items related to the health checkup and stress check of the member who is the prediction target, and prediction target data that combines the current health check judgment data and stress check judgment data of the member who is the prediction target, using the machine learning model for classification, and predicts the judgment of the items related to future health checkups and stress checks, which are the objective variables, by setting the prediction target data as explanatory variables using the machine learning model for prediction according to the classification result.

According to the above configuration, in the learning phase, work style data is classified, and a predictive machine learning model is generated for each classification. Therefore, it is possible to predict the judgment of health-related items using a predictive machine learning model that matches the work style of the person to be predicted. This makes it possible to assist in predicting the judgment of items related to the health of the person to be predicted. Then, in the prediction phase, it is possible to predict the judgment of items related to the health of the person to be predicted using a predictive machine learning model that matches the work style of the person to be predicted, thereby improving prediction accuracy.

In the above-mentioned healthcare system, the work style data classification unit is further configured to classify item combination data obtained by combining the classified item data in the work style data into multiple types by machine learning, and the classified data generation unit is further configured to classify item combination data obtained by combining the classified item data in the initial combined data into multiple types by the work style data classification unit.

According to the above configuration, item-combined data is generated by combining the classified item data in the work style data, and this data is then classified. The item-combined data is generated by combining the classified item data that constitutes the work style data, and allows classification from a holistic perspective across the entire work style data, rather than classification that focuses on each item. By using a predictive machine learning model that reflects such a holistic perspective, it is possible to improve the prediction accuracy of determining items related to the health of the person being predicted.

In the above-mentioned healthcare system, the work style data classification unit is configured to generate feature points for each classification. With the above-mentioned configuration, the person being predicted can know the features of the classification of his/her own work style.

In the above healthcare system, the working style data of the prediction target person is changeable, and the prediction unit is configured to set the prediction target person data including the changed working style data as an explanatory variable, and predict a judgment related to future health check judgment data and stress check judgment data, which are objective variables, using the predictive machine learning model different from the selected predictive machine learning model. According to the above configuration, by changing the working style data, it is possible to confirm the judgment of future health-related items when working in a different way than now.

In the above-mentioned healthcare system, the working style data is configured to include attendance history data for each of the members. According to the above-mentioned configuration, by including attendance history data such as attendance time data and departure time data for each individual employee in the working style data, it is possible to classify the working style data taking into account not only the length of working hours but also the working hours.

Furthermore, the present invention is a computer program that causes a computer to function as the above-mentioned healthcare system.

The present invention can assist in predicting the outcome of health-related items for members by taking into account not only the results related to the members' health, but also data related to their daily attendance and lifestyle habits.

1 is a diagram showing the configuration of a health care system to which the present invention is applied. FIG. 4 is a diagram for explaining health checkup judgment data. FIG. 4 is a diagram for explaining stress check judgment data. FIG. 4A is a diagram for explaining monthly attendance history data, and FIG. 4B is a diagram for explaining daily attendance history data. FIG. 1 is a diagram illustrating personnel history data. A diagram explaining salary history data. FIG. 4 is a diagram for explaining lifestyle history data. FIG. FIG. 1 is a diagram for explaining medication history data. FIG. 1 is a diagram illustrating company history data. 1 is a flowchart showing the steps of a learning phase. FIG. 13 is a diagram showing work style data, where (a) shows item A in the work style data, (b) shows category 1 of item A after classification, and (c) shows category 4 of item A after classification. 13A shows the G term to be combined with the work style data, FIG. 13B shows category 1 of the G term after classification, and FIG. 13C shows category 4 of the G term after classification. (a) shows the average of the explanatory variables for the A term of the working style data, (b) shows the explanatory variables that characterize category 1 of the A term, and (c) shows the explanatory variables that characterize category 4 of the A term. FIG. 4 is a diagram conceptually explaining the procedure of steps S2 to S5 in the learning phase. 13 is a diagram showing health-related combined data obtained by combining two sets of health checkup assessment data (current and previous) and stress check assessment data. A diagram showing work style data. A diagram showing initial combined data in which the work style data (Figure 17) and the health-related combined data (Figure 16) are combined. FIG. 13 is a diagram for explaining the time range of each item of data when generating initial combined data. (a) is a diagram showing item (G) which combines data from six items (A) to (F) x four types (classifications 1 to 4), and (b) and (c) are diagrams showing the classification data. A diagram conceptually explaining the procedure for generating a predictive machine learning model in the learning phase. 1 is a flowchart showing the steps of the prediction phase. FIG. 2 is a diagram conceptually explaining the procedure of steps S21 to S25 in the prediction phase. FIG. 13 is a diagram conceptually explaining the procedure of step S25 in the prediction phase. FIG. 13 is a diagram showing an example of an output of a future judgment of a person to be predicted. 11 is a flowchart showing the procedure for selecting another predictive machine learning model in the prediction phase. A diagram conceptually illustrating the procedure for selecting another machine learning model for prediction in the prediction phase.

Hereinafter, a health care system to which the present invention is applied will be described with reference to the drawings.
First Embodiment
[Overall structure]
The health care system to which the present invention is applied is a system provided for the health care of employees who are members belonging to a company, which is an organization. The system predicts the future health-related judgment data of each employee who is the prediction target, such as health-related judgment data of medical examinations and stress checks related to the employee's health, attendance history data, personnel history data, and lifestyle habit history data. Here, a health care system for employees who are members belonging to a company, which is an organization, is described. The prediction targets here are employees.

As shown in FIG. 1, the healthcare system includes a learning data input unit 10 and a data storage unit 20 that stores learning data. It also includes a learning prediction processing unit 30 that generates a predictive machine learning model for predicting the health state of a prediction subject and predicts the health state of the prediction subject using the predictive machine learning model, and a model recording unit 40 that records data for generating the predictive machine learning model. It also includes a prediction subject data input unit 50 for inputting data related to the health of the prediction subject, and a prediction subject data output unit 60 that outputs the prediction result.

The learning data input unit 10 provides the data storage unit 20 with various data related to employees belonging to the company. Specifically, the learning data input unit 10 includes a health check data input unit 11 for inputting health checkup judgment data (hereinafter referred to as "health checkup judgment data") of each employee belonging to the company, and a stress check data input unit 12 for inputting stress check result judgment data (hereinafter referred to as "stress check judgment data") of each employee. In addition, the learning data input unit 10 includes an attendance data input unit 13 for inputting attendance history data of each employee, a personnel data input unit 14 for inputting personnel history data of the employee, a salary data input unit 15 for inputting salary history data of the employee, and a lifestyle data input unit 16 for inputting lifestyle history data such as the employee's lifestyle. In addition, the learning data input unit 10 includes a medical data input unit 17 for inputting medical history data of the employee, and a company data input unit 18 for inputting history data of the company to which the employee belongs.

The health check data input unit 11, stress check data input unit 12, attendance data input unit 13, personnel data input unit 14, salary data input unit 15, and company data input unit 18 are, for example, computers equipped with databases managed by each of multiple companies that receive the services of this system. The lifestyle habit data input unit 16 is a wearable device terminal (such as a smart watch) that is worn by an employee or a tablet terminal such as a smartphone. The medical data input unit 17 is, for example, a personal terminal managed by an employee.

The health checkup data input unit 11 inputs to the data storage unit 20 the health checkup assessment data, such as the "overall assessment", "metabolic syndrome assessment", "diabetes assessment", "high blood pressure assessment", "waist circumference assessment", "GPT (ALT) assessment", "GOT (AST) assessment", "liver dysfunction assessment", "γ-GTP assessment", "systolic blood pressure assessment", "diastolic blood pressure assessment", and "dyslipidemia assessment" of each employee belonging to the company, are the result data measured and assessed in a general health checkup. The assessment may be based on the classification of the Japanese Society of Health Checkups or a classification determined by other medical institutions. Hereinafter, the Japanese Society of Health Checkups' classifications A (no abnormalities) to E (under treatment) will be used as an example of the assessment. In addition, test value data measured in the health checkup (weight, height, waist circumference, BMI, blood pressure (systolic blood pressure and diastolic blood pressure), uric acid, urinary protein, LDL cholesterol, fasting blood glucose, γ-GTP, GOT (AST), GPT (ALT), and other test value data) may be input. Specifically, the health checkup assessment data is configured as shown in FIG. 2. In FIG. 2 to FIG. 10, the company CD is company identification data that uniquely identifies a company, and the employee CD is individual identification data.

The stress check judgment data input unit 12 inputs to the data storage unit 20 includes, for example, the answer items and factor item evaluation scores of the simple occupational stress questionnaire (57 items) for each employee belonging to the company, as well as the result data of the judgments A (good) to E (interview required) for the stress check classifications "work burden," "work style and purpose," "health condition and satisfaction," and "work environment (department)" described in JP 2017-102673 A. Specifically, the stress check judgment data is configured as shown in FIG. 3.

The attendance history data that the attendance data input unit 13 inputs into the data storage unit 20 includes monthly and daily data for each employee belonging to the company. Monthly attendance history data includes data such as working hours, break times, overtime hours, night overtime hours, and working on holidays. Daily attendance history data includes data such as arrival time at work, departure time, start time of going outside, end time of going outside, start time of break, and end time of break. Specifically, monthly attendance history data is structured as shown in Figure 4(a), and daily attendance history data is structured as shown in Figure 4(b).

The personnel history data that the personnel data input unit 14 inputs to the data storage unit 20 is data such as the date of joining the company, the date of department transfer, and the date of promotion for each employee belonging to the company. Specifically, the personnel history data is structured as shown in FIG. 5.

The salary history data that the salary data input unit 15 inputs to the data storage unit is salary data such as monthly salary, bonus, annual salary, overtime pay, holiday work pay, etc. for each employee belonging to the company. Specifically, the salary history data is structured as shown in FIG. 6.

The lifestyle history data that the lifestyle data input unit 16 inputs to the data storage unit 20 is data related to the lifestyle habits of each employee belonging to the company, such as exercise habits such as the number of steps taken (daily average), smoking habits (number of cigarettes per day), number of drinks consumed (average number per week), drinking habits such as alcohol consumption (amount drunk at one time), eating habits (number of meals), calorie intake (daily average), and sleep time (daily average). Specifically, the lifestyle history data is configured as shown in FIG. 7.

The medical history data that the medical data input unit 17 inputs to the data storage unit 20 includes medical treatment history data and medication history data for each employee belonging to the company. The medical treatment history data is history data such as the date of consultation, name of medical institution, doctor's name, disease name, and treatment history, while the medication history data is history data such as the dispensing date, pharmacy name, drug name, dispensed quantity, and method of administration. Specifically, the medical treatment history data is structured as shown in Figure 8, and the medication history data is structured as shown in Figure 9.

The company history data that the company data input unit 18 inputs to the data accumulation unit 20 is data related to the company to which the employee belongs, such as the line of business/industry to which the company belongs, and the size of the company by year (number of employees, annual turnover, sales amount, sales quantity, order amount, order quantity, etc.). Specifically, the company history data is structured as shown in FIG. 10.

The data storage unit 20 includes a health check data storage unit 21, a stress check data storage unit 22, an attendance data storage unit 23, a personnel data storage unit 24, a salary data storage unit 25, a lifestyle habit data storage unit 26, a medical data storage unit 27, and a company data storage unit 28.

The health check data storage unit 21 receives health check judgment data from the health check data input unit 11 on a regular or irregular basis. The stress check data storage unit 22 receives stress check judgment data from the stress check data input unit 12 on a regular or irregular basis. The attendance data storage unit 23 receives attendance history data from the attendance data input unit 13 on a regular or irregular basis. The personnel data storage unit 24 receives personnel history data from the personnel data input unit 14 on a regular or irregular basis. The salary data storage unit 25 receives salary history data from the salary data input unit 15 on a regular or irregular basis. The lifestyle data storage unit 26 receives lifestyle history data from the lifestyle data input unit 16 on a regular or irregular basis. The medical data storage unit 27 receives medical history data from the medical data input unit 17 on a regular or irregular basis. The company data storage unit 28 receives company history data from the company data input unit 18 on a regular or irregular basis.

Each of the storage units 21-28 stores data on multiple companies and employees. Each data in each of the storage units 21-28 of the data storage unit 20 is provided with personal identification data that identifies an individual and company identification data for the company to which the individual belongs, and the personal identification data is associated with the company identification data to which the individual belongs. The stored data may be anonymized.

The learning prediction processing unit 30 includes a work style data generation unit 31, a work style data classification unit 32, a classification data generation unit 33, a teacher data generation unit 34, and a learning unit 35 as means for generating a predictive machine learning model. It also includes a prediction target data generation unit 36, a prediction unit 37, and a prediction result output unit 38 as means for predicting the future health condition of the prediction target.

The model recording unit 40 includes a working style data recording unit 41, a feature data recording unit 42, an initial binding data recording unit 43, a classification data recording unit 44, an initial classification teacher data recording unit 45, a classification machine learning model storage unit 47, and a prediction machine learning model storage unit 48.

The data input unit 50 for the prediction subject includes a health check data input unit 51 for inputting the health check judgment data of the prediction subject, and a stress check data input unit 52 for inputting the stress check judgment data of the prediction subject. The health check data input unit 51 and the stress check data input unit 52 are, for example, personal terminals of the prediction subject.

The data output unit 60 for the person to be predicted also includes a prediction result display unit 61. The prediction result display unit 61 is a monitor of the personal terminal of the person to be predicted, and also a printer. The person to be predicted can view the future judgment via the monitor by opening a personal file on the personal terminal, and can also view it on paper media printed out by a printer.

<Learning Phase>
Here, FIG. 11 is a flowchart showing the procedure by which the learning prediction processing unit 30 generates a classification machine learning model and a prediction machine learning model.

In step S1, the work style data generation unit 31 reads various data from the data accumulation unit 20, and in step S2, generates work style data consisting of the following items (A) to (F). Items (A) to (C) are data from the company's perspective, and items (D) to (F) are data from the individual employee's perspective. The work style data for each employee includes company-perspective data consisting of items (A) to (C) and individual-perspective data for the employee consisting of items (D) to (F), making it possible to generate a predictive machine learning model using the individual employee's data and the data of the company to which the employee belongs, and to make judgment predictions for items related to the health of the person being predicted.

(A) The work style data generation unit 31 accesses the health checkup data accumulation unit 21 and reads out the judgments A (very good) to E (very bad) for "metabolic syndrome judgment," "diabetes judgment," etc. for each company and each year. The work style data generation unit 31 also accesses the stress check data accumulation unit 22 and reads out the judgments A (good) to E (requires interview) for "work burden," "work style and purpose," etc. for each company and each year. Then, the work style data generation unit 31 calculates the number of people for each judgment A to E and generates health-related judgment summary data by aggregating the health checkup judgment data and stress check judgment data for each year, i.e., health judgment summary data and stress check judgment summary data. The time range is two to three years prior to the time of aggregation.

(B) The work style data generation unit 31 accesses the attendance data storage unit 23, the personnel data storage unit 24, and the salary data storage unit 25 to calculate the average number of work days, average total working hours, average number of vacation days, frequency of personnel orders, and average salary (monthly) for each company each month, and generates attendance summary data, which is time-series data related to work conditions and personnel, and personnel history data. The time range is from six months to one year before the summary is made (see Figure 19).

(C) The work style data generation unit 31 accesses the salary data accumulation unit 25 and the company data accumulation unit 28, calculates the business type/industry and company size (number of employees, annual turnover, sales amount, sales quantity, order amount, order quantity, average salary (annual amount)) for each company by year, and generates company history summary data, which is time-series data related to the entire company. The time range is two to three years prior to the time of compilation (see Figure 19).

(D) The work style data generation unit 31 accesses the attendance data storage unit 23, reads out the daily attendance data and departure time data for each individual, and generates attendance history data, which is time series data of date and time attendance. The time range is three to six months prior to the date on which the health check or stress check was conducted (see FIG. 19).

(E) The work style data generation unit 31 accesses the attendance data storage unit 23, the personnel data storage unit 24, and the salary data storage unit 25, reads out data on the number of working days for each individual person each month, total working hours, number of vacation days, whether or not personnel orders have been issued, and average salary (monthly), and generates attendance history data, which is time-series data related to working conditions. The time range is six months to one year prior to the date on which the health check or stress check was conducted (see Figure 19).

(F) The work style data generation unit 31 accesses the lifestyle data storage unit 26 and the medical data storage unit 27, and generates lifestyle history data, which is time-series data such as exercise habits such as the number of steps taken each year and month for each individual, smoking habits, drinking habits, eating habits such as meal times, sleep hours, presence or absence of underlying diseases, number of days of hospitalization/visiting to the hospital, and medication history. The time range is two to three years prior to the date of the health check or stress check (see Figure 19).

The time range for collecting data in paragraphs (A) through (F) above is not a long period going back in time (e.g., 10 years), but a period that will have a significant impact on your current physical and mental health.

Here, generally, health checkups and stress checks are often conducted once a year. For example, if the current year is 2002, the previous year was 2001. Also, if the current year is 2001, the previous year was 2000. The month may be off by 2 to 3 months. When generating work style data, attendance history data, personnel history data, and lifestyle history data are acquired while allowing overlap in time range. That is, the attendance history data, personnel history data, and lifestyle history data each have a time range to be acquired as shown in FIG. 19, and there are three-month, six-month, one-year, three-year, and so on. Depending on the item, the time range of the year may overlap with the time range of the previous year or the following year due to the difference in the implementation date of the health checkup or stress check. For example, if the time range is one year or three years, time overlap is likely to occur or occurs. Even in such cases, overlapping time ranges are permitted, and work style data is generated from attendance history data, personnel history data, lifestyle history data, etc., for the required time range.

In step S3, the work style data classification unit 32 performs cluster classification (grouping) on each item data of the above items (A) to (F) using machine learning such as K-Means. The number of classifications is not particularly limited, but as an example, a number that allows employees to recognize the differences is a guideline, and here it is set to four types. As a result, four types of classifications (24 types in total) are created for the six items of the above items (A) to (F). Figure 12 shows the classification of the data of the above item (A). As shown in Figure 12 (a), the data generated by the work style data generation unit 31 is arranged in chronological order for each company, with the health checkup judgments A to E in chronological order. Then, the work style data classification unit 32 classifies the data of Figure 12 (a) into "1" to "4". Figure 12 (b) shows classification "1" after classification, and Figure 12 (c) shows classification "4" after classification.

In step S4, the work style data classification unit 32 generates item combination data of item combination data (G) combining six items (A) to (F) x four types (classifications 1 to 4) of the work style data, and further performs cluster classification using machine learning such as K-means. Figure 13 shows the item combination data of the (G) item, and (a) shows data combining data classified by items (A) to (F). Figure 13 (b) shows classification "1" after classification, and Figure 13 (c) shows classification "4" after classification. This results in a total of seven items ((A) to (G)) x four types ("1" to "4"). The work style data classification unit 32 stores the classification machine learning model learned by machine learning in the classification machine learning model storage unit 47. The classification of items (A) through (F) is a classification from an individual perspective according to the definition of each item, and the classification of item (G) is a classification from a holistic perspective that combines items (A) through (F) that do not appear individually in items (A) through (F).

In step S5, the work style data classification unit 32 extracts feature points in the seven items x four classifications (A) to (G). Specifically, the work style data classification unit 32 performs four classifications for each of the items (A) to (G) and averages the parameters of the explanatory variables of the four classifications (see FIG. 14(a)). Then, the difference between the value of the explanatory variable of each classification and the average value is calculated, and the explanatory variable with the largest difference is set as the feature point of each classification (see FIGS. 14(b) and (c)). The work style data classification unit 32 then stores data on the feature points of each classification in the feature data recording unit 42.

In other words, to summarize steps S4 and S5, as shown in FIG. 15, in the learning phase, the work style data classification unit 32 classifies each of the above items (A) to (F) into, for example, four categories, and further classifies the item combination data of item (G) which combines items (A) to (F) into, for example, four categories. The classification machine learning model is then stored in the classification machine learning model storage unit 47. Also, in step S5, the feature points of each of the categories "1" to "4" for each of (A) to (G) are stored in the feature data recording unit 42.

The above is the process for generating a machine learning model for classification, and the following is the process for generating a machine learning model for prediction.
In step S6, the classified data generating unit 33 generates health-related combined data by combining the current and previous health check assessment data and the stress check assessment data based on the implementation dates of the health check and stress check for each individual (see FIG. 16). Then, in order to grasp the trend of the health condition over the past year, for example, the results of the current health check are combined with those of the previous health check.

In step S7, the classification data generation unit 33 generates working style data consisting of (A) to (F) starting from the implementation date of the previous health check or stress check for each individual (see FIG. 17), and combines this working style data with the health-related combined data generated in step S6. Alternatively, the health-related combined data is combined with the working style data generated in step S2. This generates initial combined data in which the working style data, the previous health check assessment data and stress check assessment data, and the current health check assessment data and stress check assessment data are combined (see FIG. 18). In step S8, the working style data generation unit 31 saves the initial combined data in the initial combined data recording unit 43. The time range for obtaining working style data is as shown in FIG. 19.

In step S9, the classification data generation unit 33 classifies the initial combined data using the machine learning model for classification generated in step S4. The number of categories here is four, the same as in step S3. Furthermore, the work style data classification unit 32 generates item combined data of item (G) by combining data of six items (A) to (F) x four types (categories 1 to 4) of work style data (see FIG. 20(a)), which is further classified using the machine learning model for classification generated in step S4, and divides the initial combined data for each category (see FIGS. 20(b) and 20(c)). This allows the generation of classified data of a total of seven items ((A) to (G)) x four types ("1" to "4"). The classified data generation unit 33 stores the classified data obtained by the machine learning model for classification in the classified data recording unit 44.

In step S10, the teacher data generation unit 34 sets the objective variables for each classification data to the "overall judgment" of the current health check, and each judgment of the current stress check: "work burden," "work style and purpose," "health condition and satisfaction," and "work environment (department)." The explanatory variables are then set to the previous "health check judgment data" and the previous "stress check judgment data." This generates initial classification teacher data, which is then stored in the initial classification teacher data recording unit 45. The objective variable value at this time is judgment "A" to "E." The objective variable for the health check here may be a judgment of blood pressure, blood sugar level, etc., but here it is set to "overall judgment." Furthermore, the judgment of the stress check is not limited to this.

In step S11, the teacher data generation unit 34 generates teacher data according to the following conditions (1) to (3) in order to prevent biased teacher data from being generated. Specifically, teacher data is generated from the initial classification teacher data as follows.

(1) First, the teacher data generation unit 34 creates teacher data with consistent amounts of data for the explanatory variables ("work style data," and the previous "health check judgment data" and previous "stress check judgment data") for each value of the objective variable judgment (the current health check's "overall judgment," and the current stress check's "work burden," "work style and purpose," "health condition and satisfaction," and "work environment (department)").

For example, in the case of predicting the "overall judgment" of the current health check, the number of people whose "overall judgment" of the previous health check on the explanatory variable side is
A case will be described in which 50 people are graded A, 200 people are graded B, 300 people are graded C, 70 people are graded D, and 20 people are graded E. In this case, the following teacher data (a) to (e) are generated.

(a) Ensure that the number of participants is equal to A.
A judgment: 50 people, B judgment: 50 people, C judgment: 50 people, D judgment: 50 people, E judgment: 20 people

(b) The number of people is equalized by B judgment.
A rating: 50 people, B rating: 200 people, C rating: 200 people, D rating: 70 people, E rating: 20 people

(c) The number of people is made uniform by C judgment. In this case, all data is used as is, so the results are the same.
A rating: 50 people, B rating: 200 people, C rating: 300 people, D rating: 70 people, E rating: 20 people

(d) Ensure that the number of participants is equal by D rating.
A grade: 50 people, B grade: 70 people, C grade: 70 people, D grade: 70 people, E grade: 20 people (e) The number of people is made uniform at E grade. In this case, all data will be the same amount.
A judgment: 20 people, B judgment: 20 people, C judgment: 20 people, D judgment: 20 people, E judgment: 20 people

In (1), training data is generated under each of the conditions (a) to (e).
(2) Next, the teacher data generation unit 34 generates teacher data by performing principal component analysis such as PCA (Principal Correlation Analysis), KPCA (Kernel PCA), and ISOMAP (Isometric Mapping) on the data in (1) above. In other words, this is teacher data that has been narrowed down from the teacher data in (1).
The training data may be generated using a method other than the principal component analysis.

(3) The teacher data generation unit 34 generates teacher data by adding the teacher data obtained by performing the principal component analysis described above in (2) to the data described above in (1) as new explanatory variables.

In step S12, the teacher data generating unit 34 separates data that is less than 10% to 50% in terms of amount from each teacher data created in step S11 as verification data.
In step S13, the teacher data generating unit 34 stores the teacher data generated in the above steps (1) to (3) in the teacher data recording unit 46.

In step S14, the learning unit 35 performs learning using multiple machine learning algorithms on the training data (1) to (3) generated in step S11 to create multiple predictive machine learning models.

In step S15, the learning unit 35 uses the validation data generated in step S12 to predict the objective variable for each judgment of the explanatory variable for the multiple predictive machine learning models generated in step S14. The learning unit 35 then measures the "correct answer rate", which is the ratio at which the prediction is correct for the objective variable of the validation data. Next, the multiple predictive machine learning models are combined to make a prediction, and the correct answer rate of the prediction when a majority vote is taken among the combinations is measured. Multiple combinations of the predictive machine learning models are also created, and the correct answer rate is measured. Then, from among these correct answer rates, a single predictive machine learning model or a combination of predictive machine learning models with the highest correct answer rate is selected as the predictive machine learning model, and in step S16, the learning unit 35 stores the predictive machine learning model in the predictive machine learning model storage unit 48. The learning unit 35 generates predictive machine learning models for a total of five types, namely, the "overall judgment" of the health check, and the "work burden", "work style and purpose", "health condition and satisfaction", and "work environment (department)" of the stress check.

The value selected from the combination of a single predictive learning model and a predictive machine learning model is not limited to the accuracy rate, but may be the "precision rate" indicating the proportion of correct predicted values for each value of the objective variable, the "recall rate" indicating the proportion of correct answers among the values of the objective variable, or the "F-score" which is an evaluation index for machine learning. Furthermore, the method of determining predictions by combining predictive machine learning models is not limited to majority voting, and may be the average, weighted average, harmonic mean, etc.

In the above learning phase, as shown in FIG. 21, first, a classification machine learning model is generated. Then, the initial combined data is classified using the prediction machine learning model, and for each classification, the "overall judgment" of the current health check is set as the objective variable, and the previous "health check judgment data" is set as the explanatory variable to generate teacher data, and a learning process is performed to generate multiple prediction machine learning models.

<Prediction Phase>
Next, a procedure for predicting a future judgment regarding the health of an employee who is the subject of prediction (for example, one year after the date of the current health check or stress check) will be described with reference to FIG. 22.

In step S21, the work style data generation unit 31 reads data from the data accumulation unit 20. Before the data is read, the employee who is the prediction target inputs the current health check assessment data and the current stress check assessment data from the prediction target data input unit 50, and the prediction target data input unit 50 stores the input data in the data accumulation unit 20.

In step S22, the predicted target data generation unit 36 generates working style data consisting of items (A) to (F) starting from the implementation date of the predicted target's current health check or stress check. The time range for acquiring the working style data is as shown in FIG. 19. Here, the working style data (A) to (C) is data on the company to which the predicted target identified by the personal identification data and company identification data belongs, and (D) to (F) is data on the predicted target identified by the personal identification data.

In step S23, the predicted subject data generation unit 36 generates predicted subject data by combining the work style data generated in step S22 with the current health check assessment data and stress check assessment data of the predicted subject.

In step S24, the predicted target data generation unit 36 inputs the predicted target data to the prediction unit 37. In step S25, the prediction unit 37 classifies each of the items (A) to (G) of the predicted target data into "1" to "4" using a machine learning model for classification. In addition, it generates item combination data for item (G) that combines data from the six items (A) to (F) x four types (classifications 1 to 4) of work style data.

The prediction unit 37 then selects a machine learning model for prediction to use depending on the classification of the work style data of the prediction target person and the "overall judgment" of the prediction target person's current health check, and the judgments of the "work burden," "work style and purpose," "health condition and satisfaction," and "work environment (department)" of the stress check, and performs a prediction, for example, one year from the implementation date of the current health check or stress check. The prediction result output unit 38 outputs the prediction results to the prediction result display unit 61. These predictions are performed for five types of objective variables: the "overall judgment" of the health check, and the "work burden," "work style and purpose," "health condition and satisfaction," and "work environment (department)" of the stress check.

That is, as shown in FIG. 23, in the prediction phase, the predicted subject data generation unit 36 adds item (G) to the work style data consisting of items (A) to (F) starting from the implementation date of the current health check and stress check of the predicted subject, and generates predicted subject data that combines the current health check judgment data and stress check judgment data of the predicted subject. Then, the work style data classification unit 32 classifies each of the items (A) to (G) of the predicted subject data. Then, FIG. 24 shows the case of the "overall judgment" of the health check. If the "overall judgment" of the current health check is judgment A, a prediction is made using a predictive machine learning model for the case of judgment A, and the "overall judgment" of the future health check is predicted. Then, in step S26, the prediction result output unit 38 outputs to the prediction result display unit 61.

Figure 25 shows an example of the display of prediction results in the prediction result display unit 61. The prediction results show the prediction target's future "overall health checkup judgment," as well as judgments for the stress check's "work burden," "work style and purpose," "health condition and satisfaction," and "work environment (department)."

Incidentally, the prediction unit 37 can also make predictions using work style data and/or predictive machine learning models that are different from the original classification of the prediction target, such as the operator of the prediction target. For example, assuming that the prediction target will have a change in the workplace environment in the future due to promotion or transfer to another department, it can predict the "overall health checkup judgment" and the stress check's "work burden," "work style and purpose," "health condition and satisfaction," and "work environment (department)" when the workplace environment changes from the current workplace to the assumed workplace environment.

As shown in FIG. 26, in step S31, the work style data generation unit 31 reads out the current health check assessment data and stress check assessment data of the prediction target from the data storage unit 20. In step S32, the prediction target data generation unit 36 generates work style data consisting of items (A) to (F) starting from the implementation date on which the prediction target's current health check and stress check were implemented.

In step S33, the predicted target data generation unit 36 classifies each of the items (A) to (G) of the work style data of the predicted target into "1" to "4" according to the machine learning model for classification. It also generates and classifies item combination data for item (G) that combines data on six items (A) to (F) of work style data x four types (classifications 1 to 4).

In step S34, the predicted target data generation unit 36 generates predicted target data by combining the current health check assessment data and stress check assessment data with the working style data of the predicted target. At this time, the working style data is set to data different from the working style data of the predicted target. For example, working style data when promoted or transferred to another department is set. Specifically, the data in items (D) to (F) is changed.

In step S35, the predicted subject data generation unit 36 inputs the changed predicted subject data to the prediction unit 37. In step S36, the prediction unit 37 classifies each of the items (A) to (G) of the predicted subject data into "1" to "4" according to the classification machine learning model. In addition, the prediction unit 37 generates item combination data for item (G) that combines data of six items (A) to (F) x four types (classifications 1 to 4) of work style data. The prediction unit 37 then selects a predictive machine learning model to use according to the classification of the work style data of the predicted subject and the "overall judgment" of the predicted subject's current health check, and the judgment of the "work burden," "work style and purpose," "health condition and satisfaction," and "work environment (department)" of the stress check, and performs a prediction, for example, one year after the implementation date of the current health check or stress check.

The prediction result output unit 38 outputs the prediction results to the prediction result display unit 61. These predictions are made for five types of objective variables: the "overall assessment" of the health check, and the "work burden," "work style and purpose," "health condition and satisfaction," and "work environment (department)" of the stress check. Note that it is not necessary to make predictions for all of these items, and it is possible to predict only the items desired by the person being predicted.

Then, in step S37, the prediction result output unit 38 outputs to the prediction result display unit 61. The prediction result shows the future "overall health checkup assessment" of the person being predicted, as well as the assessment results for the "work burden," "work style and purpose," "health condition and satisfaction," and "work environment (department)" of the stress check.

That is, as shown in FIG. 27, by changing the working style data, the prediction unit 37 can use a different predictive machine learning model than that in FIG. 24 to obtain a prediction result according to the changed working style data. As an example, in FIG. 23, an A rating is selected in the "overall rating" of the health check, but by changing the working style data, a D rating may be obtained. In this case, the person to be predicted can indicate what kind of working style will result in a D rating by understanding the working style data that would result in a D rating.

According to the above-described healthcare system, the following effects can be obtained.
(1) Using a predictive machine learning model tailored to the working style of the person being predicted, it is possible to predict the outcome of health checkups and stress checks for that person, thereby helping to determine the outcome of health-related items for that person.

Until now, when industrial physicians/public health nurses selected employees for guidance/interviews, they did so based on the employee's current assessment, but this system makes it possible to make future predictions, so it can be used to support industrial physicians/public health nurses by interviewing people who are doing well now but who may deteriorate in the future.

(2) In the learning phase, work style data can be classified and a predictive machine learning model can be generated for each classification. As a result, in the prediction phase, the results of items related to health checkups and stress checks can be predicted using a predictive machine learning model that is tailored to the work style of the person being predicted, thereby improving prediction accuracy.

(3) It is possible to predict the results of a health checkup and a stress check of a person to be predicted, i.e., it is possible to predict the physical and mental health of a person to be predicted.
(4) The work style data includes data from the company's perspective (A) to (C) and data from the individual's perspective (D) to (F), which allows for prediction of the results of items related to employee health checkups and stress checks by taking into account various factors. Furthermore, by including (C) in the work style data, it is possible to predict the results of items related to employee health checkups and stress checks by taking into account the performance of the company and industry to which the employee belongs.

(5) Item-combined data is generated by combining the classified item data in the work style data (section (G)), and this data is classified. The item-combined data is a combination of the classified item data that make up the work style data, and is a classification from a holistic perspective across the entire work style data (such as trends in work styles in industries or at each organization), rather than a classification focused on an item. Therefore, by using a predictive machine learning model that reflects such a holistic perspective, it is possible to improve the accuracy of predictions for items related to the health of the person being predicted.

(6) The person being predicted can learn the characteristics of the classification of their own work style data.
(7) By changing the predictive machine learning model, it is possible to understand what kind of work style will worsen or improve the assessment of future health-related items. At this time, the feature data can be used as a reference for explaining what kind of work style will improve or worsen the health-related items according to the classification after the change in the work style data.

(8) By including attendance history data (section (D)) such as the employee's attendance and departure times for each individual in the work style data, it is possible to classify the work style data taking into account not only the length of work hours but also the working hours.

The above-described healthcare system can be modified as follows.
- It is possible to avoid including employees' individual attendance history data in the work style data.
The function of changing the machine learning model for prediction selected by the prediction unit in the prediction phase may be omitted.

In the learning phase, the work style data may omit the item combination data (item (G)).
The health-related items may include items other than the health check result items and the stress check result items. Also, the health-related items may be only either the health check result items or the stress check result items.

- The assessment of predicted health-related items will not be limited to one year after the date of the current health check or stress check.
The organization may be a for-profit organization such as a corporation, or a non-profit organization.

DESCRIPTION OF SYMBOLS 10...Learning data input section 11...Health check data input section 12...Stress check data input section 13...Attendance data input section 14...Personnel data input section 15...Salary data input section 16...Lifestyle data input section 17...Medical data input section 18...Company data input section 21...Health check data storage section 22...Stress check data storage section 23...Attendance data storage section 24...Personnel data storage section 25...Salary data storage section 26...Lifestyle data storage section 28...Company data storage section 31...Work style data generation section 32...Work style data classification section 33...Classification data generation section 34...Teacher data generation section 35...Learning section 36...Prediction target data generation section 37...Prediction section

Claims (6)

health checkup judgment aggregate data obtained by aggregating health checkup judgment data related to health checkup results of members of a plurality of organizations, stress check judgment aggregate data obtained by aggregating stress check judgment data related to stress check results of the members, attendance aggregate data obtained by aggregating attendance history related to attendance for each of the members, personnel history data related to personnel, and attendance history data related to attendance and lifestyle history data related to lifestyle for each member, each of which is item data, and a collection of the item data is work style data,
A work style data classification unit that generates a classification machine learning model that classifies each item data of the work style data into multiple types by machine learning;
a classification data generation unit which generates initial combined data by combining the health-related combined data and the working style data for a period going back from the implementation date of the previous assessment of items related to the health checkup and stress check of the member, the initial combined data being generated by combining the health-related combined data and the working style data for a period going back from the implementation date of the previous assessment of items related to the health checkup and stress check of the member, the working style data classification unit classifying the initial combined data into multiple types, and dividing the initial combined data for each classification;
a teacher data generating unit that generates teacher data for each of the classification data by setting a judgment related to the current health check judgment data and stress check judgment data of the member as an objective variable and each item data of the work style data and a judgment related to the previous health check judgment data and stress check judgment data of the member as explanatory variables;
A learning unit that generates a machine learning model for predicting the objective variable for each determination of the explanatory variable;
and a prediction unit that uses the classification machine learning model to classify the working style data for a period going back from the date of the current assessment of items related to health checkups and stress checks for the member who is the target of prediction, and prediction target data that combines the current health check assessment data and stress check assessment data of the member who is the target of prediction, and uses the predictive machine learning model in accordance with the classification result to set the prediction target data as an explanatory variable and predict the assessment of items related to future health checkups and stress checks, which is the objective variable. The work style data classification unit further classifies item combination data obtained by combining the classified item data in the work style data into a plurality of types by machine learning,
The healthcare system according to claim 1 , wherein the classified data generation unit further classifies, into a plurality of types, item combined data obtained by combining the classified item data in the initial combined data, by the work style data classification unit. The healthcare system according to claim 1 or 2, wherein the work style data classification unit generates feature points for each of the classifications. The work style data of the prediction subject is changeable,
The healthcare system according to any one of claims 1 to 3, wherein the prediction unit sets the prediction subject data including the changed working style data as an explanatory variable, and predicts a judgment related to future health check judgment data and stress check judgment data, which are objective variables, using a predictive machine learning model different from the selected predictive machine learning model. The healthcare system according to claim 1 , wherein the work style data includes work attendance history data for each of the members. health checkup judgment aggregate data obtained by aggregating health checkup judgment data related to health checkup results of members of a plurality of organizations, stress check judgment aggregate data obtained by aggregating stress check judgment data related to stress check results of the members, attendance aggregate data obtained by aggregating attendance history related to attendance for each of the members, personnel history data related to personnel, and attendance history data related to attendance and lifestyle history data related to lifestyle for each member, each of which is item data, and a collection of the item data is work style data,
Computer,
A work style data classification unit that generates a classification machine learning model that classifies each item data of the work style data into multiple types by machine learning;
a classification data generation unit which generates initial combined data by combining the health-related combined data and the working style data for a period going back from the implementation date of the previous assessment of items related to the health checkup and stress check of the member, the initial combined data being generated by combining the health-related combined data and the working style data for a period going back from the implementation date of the previous assessment of items related to the health checkup and stress check of the member, the working style data classification unit classifying the initial combined data into multiple types, and dividing the initial combined data for each classification;
a teacher data generating unit that generates teacher data for each of the classification data by setting a judgment related to the current health check judgment data and stress check judgment data of the member as an objective variable and each item data of the work style data and a judgment related to the previous health check judgment data and stress check judgment data of the member as explanatory variables;
A learning unit that generates a machine learning model for predicting the objective variable for each determination of the explanatory variable;
A healthcare program that functions as a prediction unit that classifies, using the classification machine learning model, the work style data for the period going back from the date of the current assessment of items related to health checkups and stress checks for the member who is the target of prediction, and prediction target data that combines the current health check assessment data and stress check assessment data of the member who is the target of prediction, and uses the predictive machine learning model in accordance with the classification results to set the prediction target data as an explanatory variable and predict the assessment of items related to future health checkups and stress checks, which are the objective variables.