JP7517433B2 - Stress estimation device, estimation method, program, and storage medium - Google Patents

Description

The present invention relates to stress estimation using biological signals.

In recent years, it has become an issue that mental health is impaired by prolonged exposure to stressors that can lead to prolonged sympathetic nervous activity and autonomic nervous system imbalance. For this reason, biosignals that reflect the subject's biological information (such as sweat rate, skin surface temperature, and body movement) are measured over the long term from a wearable device worn by the subject on a daily basis, and the subject's long-term stress (chronic stress) is monitored. Technology has been proposed that aims to encourage stress reduction based on the results of such monitoring.

For example, Non-Patent Document 1 discloses a technique for identifying three physical activity states, sitting, walking, and running, from 30 days of data from 20 people using a common Activity Magnitude (moving average of the change in the RMS (Rooted Mean Square) of the three-axis acceleration). Non-Patent Document 2 discloses a technique for automatically deriving and applying a threshold for distinguishing between the three activity states, sitting, walking, and running, for each individual from the histogram of the individual's Activity Magnitude. Non-Patent Document 3 discloses a technique for estimating the cognitive stress scale of a subject with a certain degree of accuracy. Non-Patent Document 4 discloses a method for obtaining a domain-invariant model by learning a gradient for labels and simultaneously multiplying a negative coefficient by the gradient for classifying domains, and performing "adversarial" learning on the domain classification. Non-Patent Document 5 discloses a method for selecting features using Random Forest and Decision Tree.

A. Sano, “Measuring College Students’ Sleep, Stress, Mental Health and Wellbeing with Wearable Sensors and Mobile Phones”, Massachusetts Institute of Technology, 2015. Y. Nakashima et al., “An Effectiveness Comparison between the Use of Activity State Data and That of Activity Magnitude Data in Chronic Stress Recognition,” ACII workshop, 2019. S. Cohen, R. C. Kessler, and L. U. Gordon, “Measuring Stress: A Guide for Health and Social Scientists,” Oxford University Press, 1997. Iwasawa et al, “Learning Domain-Invariant Representations by Adversarial Training,” The 31st Annual Conference of the Japanese Society for Artificial Intelligence, 2017. “The Mathematics of Decision Trees, Random Forest and Feature Importance in Scikit-learn and Spark”, URL:https://towardsdatascience.com/the-mathematics-of-decision-trees-random-forest-and-feature-importance- in-scikit-learn-and-spark-f2861df67e3

Non-Patent Documents 1 and 2 attempt to estimate the stress level of a person using features such as statistics of biological signals such as sweating, heart rate, body movement, and body temperature. However, especially in the case of estimating chronic stress, the relationship between the features used to estimate stress, such as sweating, heart rate, body movement, and body temperature, and the stress level varies depending on the subject's life pattern, work pattern, or gender and age bias. Therefore, unless a database that incorporates these conditions evenly is constructed, a model that can be applied to all people cannot be created. However, creating such an unbiased database is costly.

In view of the above-mentioned problems, the object of the present disclosure is to provide a stress estimation device, estimation method, program, and storage medium that can appropriately estimate stress.

One aspect of the stress estimation device is
A biosignal acquiring means for acquiring a biosignal of a subject;
a biosignal feature amount calculation means for calculating a biosignal feature amount which is a feature amount of the biosignal related to stress;
an attribute/pattern data acquiring means for acquiring attribute/pattern data indicating at least one of the attributes and patterns of the subject;
an attribute/pattern common feature selection means for excluding, from the biological signal feature, an adversarial feature that distinguishes the attribute/pattern based on the attribute/pattern data , and selecting the biological signal feature from which the adversarial feature has been excluded as an attribute /pattern common feature that is a feature independent of the attribute/pattern indicated by the attribute/pattern data;
a stress estimation means for estimating a stress of the subject based on the attribute/pattern common feature amount;
The stress estimation device includes:

One aspect of the estimation method is
By computer,
Acquire a biological signal of the subject,
Calculating a biosignal feature amount, which is a feature amount of the biosignal related to stress;
Acquire attribute/pattern data indicating at least one of the attributes or patterns of the subject;
excluding, from the biological signal features , adversarial features that distinguish the attributes and patterns based on the attribute/pattern data , and selecting the biological signal features from which the adversarial features have been excluded as attribute/pattern common features that are features that are not dependent on the attributes and patterns indicated by the attribute/pattern data;
Estimating the stress of the subject based on the attribute/pattern common feature amount.
This is an estimation method.

One aspect of the program is
Acquire a biological signal of the subject,
Calculating a biosignal feature amount, which is a feature amount of the biosignal related to stress;
Acquire attribute/pattern data indicating at least one of the attributes or patterns of the subject;
excluding, from the biological signal features , adversarial features that distinguish the attributes and patterns based on the attribute/pattern data , and selecting the biological signal features from which the adversarial features have been excluded as attribute/pattern common features that are features that are not dependent on the attributes and patterns indicated by the attribute/pattern data;
The program causes a computer to execute a process of estimating the stress of the subject based on the attribute/pattern common features.

According to the present disclosure, stress can be appropriately estimated regardless of the attributes or pattern of the person being measured.

1 shows a block configuration of a stress estimation system according to a first embodiment. 2 shows a hardware configuration of the stress estimation device in the first embodiment. 4 is an example of a flowchart illustrating a processing procedure executed by the stress estimation device of the first embodiment. FIG. 11 is a block diagram showing a specific configuration of a stress estimation system according to a second embodiment. 13 shows a block configuration of a stress estimation device according to a second embodiment. 13A and 13B are schematic diagrams illustrating a process for creating a stress estimation model in the second embodiment. 13 shows a graph relating to the stress estimation results for sales staff in the second embodiment. FIG. 11 is a block diagram of a stress estimation device according to a third embodiment. 13 is an example of a flowchart illustrating a processing procedure of a stress estimation device according to a third embodiment.

Below, with reference to the drawings, embodiments of a stress estimation device, estimation method, program, and storage medium are described.

First Embodiment
[Configuration Description]
Hereinafter, the configuration of the stress estimation system in the first embodiment will be described with reference to FIGS.

Figure 1 shows the configuration of a stress estimation system 150 in this embodiment. Referring to Figure 1, the stress estimation system 150 in this embodiment has a stress estimation device 100 and a wearable terminal 200.

The stress estimation device 100 is configured with one or more computers. In this embodiment, the stress estimation device 100 is capable of wired or wireless data communication with a wearable terminal 200 attached to a part of the body of the person being measured, for example, on the arm.

The wearable device 200 measures the subject's biosignal "Sb" and supplies the measured biosignal Sb to the stress estimation device 100. The biosignal Sb is, for example, a signal indicating the subject's sweat rate, skin surface temperature, body movement, pulse rate, heart rate, respiratory rate, etc. Note that the biosignal Sb is not limited to the above examples, and may be any information that can estimate the subject's mental state, such as stress, such as information reflecting the subject's autonomic nerve activity.

Note that the configuration of the stress estimation system 150 shown in FIG. 1 is an example, and various modifications may be made. For example, the stress estimation system 150 may further include a mobile device terminal such as a smartphone owned by the person being measured. In this case, the stress estimation device 100 and the wearable terminal 200 may communicate data via the mobile device terminal.

The stress estimation device 100 functionally includes a biosignal acquisition unit 101, a biosignal memory unit 102, a biosignal feature calculation unit 103, an attribute/pattern data acquisition unit 104, an attribute/pattern common feature selection unit 105, and a stress estimation unit 106.

The biosignal acquisition unit 101 acquires the biosignal Sb from the wearable device 200 through data communication. The wearable device 200 may immediately supply the generated biosignal Sb to the biosignal acquisition unit 101, or may accumulate the generated biosignal Sb and collectively supply the accumulated biosignal Sb to the biosignal acquisition unit 101 at a predetermined timing.

The biosignal memory unit 102 stores the biosignal Sb acquired by the biosignal acquisition unit 101.

The biosignal feature calculation unit 103 calculates a feature used for stress estimation (also called "biosignal feature Fb") from the biosignal Sb stored in the biosignal storage unit 102. The biosignal feature Fb is any feature (including statistics) of the biosignal Sb used for stress estimation, and is, for example, the mean value, variance, time series histogram, power spectral density histogram, etc. of the biosignal Sb, as disclosed in Non-Patent Document 1 and Non-Patent Document 2.

The attribute/pattern data acquisition unit 104 acquires data indicating at least one of the attributes or behavioral patterns of the person being measured (also referred to as "attribute/pattern data Da"). The attribute/pattern data Da is, for example, information indicating any behavioral pattern, such as the person being measured's life pattern or work pattern, and/or information indicating the person's attributes, such as occupation, gender, and age, which reflect gender and age bias.

Here, the attribute and pattern data acquisition unit 104 may acquire the attribute and pattern data Da by various methods. In a first example, the attribute and pattern data acquisition unit 104 generates the attribute and pattern data Da based on input information input by the subject through any user interface (including a voice input device) to the wearable terminal 200 or the stress estimation device 100. In a second example, the attribute and pattern data acquisition unit 104 generates the attribute and pattern data Da by estimating the attributes and patterns of the subject from the subject's biological signal Sb. In a third example, the attribute and pattern data Da of the subject is acquired from a storage device that stores the attribute and pattern data Da of the subject in advance. The above-mentioned storage device may be a memory in the stress estimation device 100, or may be an external device other than the stress estimation device 100 (for example, a server that manages a database related to attributes and patterns).

The attribute/pattern common feature selection unit 105 selects a feature (also called "attribute/pattern common feature Fc") that does not differ depending on the attribute/pattern indicated by the attribute/pattern data Da from the biosignal feature Fb calculated by the biosignal feature calculation unit 103, that is, does not depend on the attribute/pattern. In this case, the attribute/pattern common feature selection unit 105 eliminates a feature (also called "hostile feature Fa") that can distinguish attributes or patterns from the biosignal feature Fb, and identifies the remaining biosignal feature Fb as the attribute/pattern common feature Fc. In this case, the attribute/pattern common feature selection unit 105 performs a process of selecting a feature common to the attribute/pattern, that is, selecting an adversarial feature Fa that can distinguish the attribute/pattern, and eliminating the adversarial feature Fa (hereinafter referred to as "adversarial feature selection"). In this case, for example, in the training data, the adversarial feature selection and the targeted stress estimation are alternately performed to obtain an optimal feature set as input and output to the adversarial feature selection model, thereby learning the adversarial feature selection model in advance. Parameters of the adversarial feature selection model obtained by learning are stored in a memory or the like that can be referenced by the attribute/pattern common feature selection unit 105. Then, the attribute/pattern common feature selection unit 105 configures the adversarial feature selection model by referring to the parameters, and selects the adversarial feature Fa from the biological signal feature Fb using the model.

The stress estimation unit 106 estimates the stress of the subject based on the attribute/pattern common feature Fc selected by the attribute/pattern common feature selection unit 105. In this case, the stress estimation unit 106 calculates a stress score estimated from the attribute/pattern common feature Fc based on a model (also called a "stress estimation model") obtained by machine learning, for example. The parameters of the stress estimation model obtained by learning are stored in a memory or the like that can be referenced by the stress estimation unit 106. After stress estimation, a score indicating the estimated stress is presented to the subject and used for purposes such as encouraging stress reduction.

Figure 2 shows the hardware configuration of the stress estimation device 100. The stress estimation device 100 includes, for example, as hardware, a processor 11, a memory 12, and an interface 13. The processor 11, the memory 12, and the interface 13 are connected via a data bus 19.

The processor 11 executes a predetermined process by executing a program stored in the memory 12. The processor 11 is a processor such as a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), or a quantum processor. The processor 11 executes a program stored in the memory 12, for example, to function as the biosignal acquisition unit 101, the biosignal feature calculation unit 103, the attribute/pattern data acquisition unit 104, the attribute/pattern common feature selection unit 105, and the stress estimation unit 106 shown in FIG. 1.

Note that each of the components of the biosignal acquisition unit 101, the biosignal feature calculation unit 103, the attribute/pattern data acquisition unit 104, the attribute/pattern common feature selection unit 105, and the stress estimation unit 106 may not necessarily be realized by software using a program, but may be a controller realized by any combination of hardware, firmware, and software. Furthermore, each of these components may be realized using a user-programmable integrated circuit, such as an FPGA (field-programmable gate array) or a microcomputer. In this case, a program composed of each of the above components may be realized using this integrated circuit.

The memory 12 is composed of various types of volatile and non-volatile memories, such as a RAM (Random Access Memory) and a ROM (Read Only Memory). The memory 12 also stores a program for executing the processes executed by the stress estimation device 100. For example, the memory 12 functions as a biosignal storage unit 102. The program executed by the stress estimation device 100 may be stored in a storage medium other than the memory 12. Similarly, the biosignal storage unit 102 may be stored in an external device other than the stress estimation device 100.

The interface 13 is an interface for electrically connecting the stress estimation device 100 to other devices. For example, the interface 13 may be a communication interface for the stress estimation device 100 to perform wired or wireless data communication with an external device such as the wearable terminal 200. In another example, the interface 13 may be a hardware interface conforming to USB (Universal Serial Bus), SATA (Serial AT Attachment), or the like. In addition, various devices other than the wearable terminal 200 may be electrically connected to the stress estimation device 100 via the interface 13. For example, an input device that accepts user input regarding the attribute/pattern data Da, a display device or sound output device that outputs the estimation result by the stress estimation unit 106, etc. may exchange data with the stress estimation device 100 via the interface 13.

Note that the hardware configuration of the stress estimation device 100 is not limited to the configuration shown in FIG. 2. For example, the stress estimation device 100 may be a tablet terminal or the like that incorporates at least one of an input device, a display device, and a sound output device.

[Operation Description]
Next, the operation of the stress estimation device 100 in the first embodiment will be described with reference to Fig. 3. Fig. 3 is an example of a flowchart showing the operation of the stress estimation device 100.

First, the biosignal acquisition unit 101 of the stress estimation device 100 acquires the biosignal Sb transmitted from the wearable device 200 (step S11). Then, the biosignal acquisition unit 101 stores the acquired biosignal Sb in the biosignal storage unit 102 (step S12).

The biosignal feature calculation unit 103 calculates the biosignal feature Fb based on the biosignal Sb stored in the biosignal storage unit 102 (step S13). The attribute/pattern data acquisition unit 104 acquires attribute/pattern data Da in parallel with the processing of steps S11 to S13 or before or after these steps (step S14).

Then, the attribute/pattern common feature selection unit 105 determines whether or not the biosignal feature Fb calculated in step S13 corresponds to an adversarial feature Fa, which is a feature that can distinguish the attribute/pattern indicated by the attribute/pattern data Da, based on the attribute/pattern data Da (step S15). If the biosignal feature Fb corresponds to an adversarial feature Fa that can distinguish the attribute/pattern (step S15; Yes), the attribute/pattern common feature selection unit 105 determines that the target biosignal feature Fb does not correspond to the attribute/pattern common feature Fc, and proceeds to step S17 without performing stress estimation in step S16.

On the other hand, if the biosignal feature Fb does not correspond to the hostile feature Fa (step S15; No), the attribute/pattern common feature selection unit 105 regards the target biosignal feature Fb as an attribute/pattern common feature Fc. In this case, the stress estimation unit 106 estimates the stress of the subject based on the attribute/pattern common feature Fc (step S16).

Thereafter, the stress estimation device 100 determines whether or not the stress estimation process should be terminated (step S17). For example, the stress estimation device 100 determines that the stress estimation process should be terminated when it is no longer possible to acquire the biological signal Sb, when a user input indicating that the stress estimation process should be terminated is detected, or when other predetermined conditions for terminating the stress estimation process are satisfied. Then, when the stress estimation device 100 determines that the stress estimation process should be terminated (step S17; Yes), it terminates the processing of the flowchart. On the other hand, when the stress estimation device 100 determines that the stress estimation process should not be terminated (step S17; No), it returns the processing to step S11.

[Effects]
Next, the effects of the first embodiment will be supplementarily described.

In the first embodiment, by using an adversarial feature selection method to eliminate the adversarial feature Fa, which is a feature that distinguishes attributes and patterns, stress can be estimated using only the attribute/pattern common feature Fc, which is a feature that does not differ depending on the attribute/pattern. Therefore, the stress of the person being measured can be estimated with high accuracy without depending on the attribute/pattern of the person being measured.

For example, in Non-Patent Document 1, only a database with bias in the life patterns and age groups of subjects can be obtained, such as students in Non-Patent Document 2, and office workers in an IT company in Non-Patent Document 2. In addition, Non-Patent Document 4 is a method for obtaining a domain-invariant model, which can handle biased databases, but since it is based on deep learning that assumes a large amount of data, it cannot be applied to the analysis of databases such as long-term stress, where it is difficult to obtain a large amount of data. In contrast, in the first embodiment, even if there is no large amount of training data and the stress estimation system is created from a biased database, it is possible to appropriately estimate the stress of the subject without relying on attributes or patterns.

Second Embodiment
Next, a second embodiment, which is a more specific embodiment of the first embodiment, will be described with reference to FIGS.

Figure 4 shows the configuration of a stress estimation system 150A according to the second embodiment. The stress estimation system 150A has a plurality of wearable terminals 400 and a computer 600. The computer 600 and each wearable terminal 400 communicate with each other via communication means 500 (501, 502, 503, 504).

The wearable terminal 400 acquires a biosignal "Sb" of the subject 300. The biosignal Sb of the subject 300 may be electrodermal activity reflecting the subject's sweating. Other examples of the biosignal Sb may be signals reflecting various bioinformation such as body temperature, pulse waves, heart rate, voice, brain waves, breathing, electromyography, electrocardiography, and body movements. In this way, the biosignal Sb may be a signal reflecting any bioinformation that is affected by the subject's mental activity.

The wearable terminal 400 is a terminal that can be worn by the subject 300 and measures at least one of the biosignals that reflect the bioinformation listed above. For example, the wearable terminal 400 acquires skin conductivity at a constant sampling rate and stores it in an internal memory. The wearable terminal 400 may be in various forms, such as a wristband type, a badge type, an employee ID type, an earphone type, or a shirt type.

The communication means 500 (501, 502, 503, 504) transmits the biosignal Sb (which may include an acceleration signal) acquired by the wearable terminal 400 to the computer 600. Specifically, for example, the wearable terminal 400 connects to the smartphone 502 by short-range communication 501 such as Bluetooth (registered trademark) and transmits the biosignal Sb to the smartphone 502. The smartphone 502 then transmits the biosignal Sb to the Internet 504 by packet communication 503 using an installed application. As a result, the biosignal Sb is uploaded to the computer 600 connected to the Internet 504.

The computer 600 corresponds to the stress estimation device 100 of the first embodiment and has, for example, a hardware configuration as shown in FIG. 2.

5 shows the functional configuration of a computer 600 in the second embodiment. The computer 600 functionally includes a communication interface (I/F) 601, a biosignal acquisition unit 602, a biosignal storage unit 603, an attribute/pattern data acquisition unit 604, an attribute/pattern data storage unit 605, a biosignal feature calculation unit 606, a biosignal feature storage unit 607, an attribute/pattern common feature selection unit 608, an attribute/pattern common feature storage unit 609, a stress estimation unit 610, a stress estimation result storage unit 611, and a stress estimation result output unit 612.

First, the biosignal acquisition unit 602 stores the biosignal Sb obtained from the communication interface 601 in the biosignal storage unit 603. Next, the attribute/pattern data acquisition unit 604 acquires the attribute/pattern data "Da". In the second embodiment, the attribute/pattern data Da indicates the subject's occupation (sales or technical). The acquired attribute/pattern data Da is stored in the attribute/pattern data storage unit 605. Note that the attribute/pattern data acquisition unit 604 may acquire attribute/pattern data Da indicating the subject's attributes and patterns from the wearable terminal 400 or other external device, or may generate the attribute/pattern data Da by estimating the subject's attributes and patterns from the biosignal Sb.

Next, the biosignal feature calculation unit 606 calculates the biosignal feature "Fb", which is a feature related to stress, based on the biosignal Sb extracted from the biosignal storage unit 603. In this embodiment, the biosignal feature Fb is a time series histogram, power spectral density, and statistics (average, median, variance) of sweating and body movement. The calculated biosignal feature Fb is stored in the biosignal feature storage unit 607.

Next, the attribute/pattern common feature selection unit 608 selects the attribute/pattern common feature "Fc", which is a feature that does not differ depending on the attribute/pattern, from the biological signal feature Fb by eliminating the adversarial feature "Fa", which is a feature that distinguishes the attributes/patterns.

In this embodiment, the attribute/pattern common feature selection unit 608 performs adversarial feature selection using a random forest and eliminates the selected adversarial feature Fa from the biosignal feature Fb. Then, the attribute/pattern common feature selection unit 608 selects the remaining biosignal feature Fb that does not depend on the attribute/pattern as the attribute/pattern common feature Fc.

Specifically, the attribute/pattern common feature selection unit 608 first classifies data with different attributes and patterns using a random forest method. In the second embodiment, the attribute/pattern common feature selection unit 608 performs two-class classification to distinguish between technical jobs and sales jobs. In this classification, the attribute/pattern common feature selection unit 608 evaluates the importance of each biosignal feature Fb. Then, the attribute/pattern common feature selection unit 608 regards biosignal features Fb with importance equal to or greater than a predetermined threshold as adversarial features Fa and excludes them from the feature set used for stress estimation. Note that parameters required for adversarial feature selection using random forest are stored in the memory of the computer 600 or an external storage device.

The attribute/pattern common feature selection unit 608 stores the attribute/pattern common feature Fc common to the attributes/patterns selected through the adversarial feature selection in the attribute/pattern common feature memory unit 609.

Next, the stress estimation unit 610 estimates stress based on the attribute/pattern common features Fc stored in the attribute/pattern common feature memory unit 609.

Specifically, in the second embodiment, the stress estimation unit 610 uses the score of the 10-item version of the Perceived Stress Scale (PSS) (hereinafter referred to as "PSS10") as the correct value of stress, and creates a stress estimation model that estimates the score of this PSS10 by regression analysis. At this time, the PSS10 score calculated from the PSS questionnaire administered to the subject at the end of the experimental period (for example, 4 weeks) is used as teacher data (correct value), and training data is prepared in which the attribute/pattern common feature Fc based on the biological signal obtained from the subject is used as corresponding input data. Then, the stress estimation unit 610 uses the stress estimation model obtained by learning a machine learning model such as an SVM model based on this training data to estimate the score of PSS10 from the attribute/pattern common feature Fc. Then, the stress estimation unit 610 generates the estimated score of PSS10 as a stress estimation result. A more detailed explanation of the above-mentioned learning will be described later with reference to FIG. 6 and FIG. 7. The parameters of the stress estimation model obtained by learning the above-mentioned machine learning model are stored in the memory of the computer 600 or an external storage device so that they can be referenced by the stress estimation unit 610. The estimated stress estimation result is stored in the stress estimation result storage unit 611.

Next, in response to a request from the subject 300, the stress estimation result output unit 612 outputs the stress estimation result stored in the stress estimation result memory unit 611.

Specific examples of output methods include, but are not limited to, screen output and print output. Output timing can be constant or at the request of the person being measured. Specifically, in the case of screen output, the stress estimation result output unit 612 transmits the stress estimation result stored in the stress estimation result storage unit 611 to the wearable terminal 400 or smartphone 502 using the communication means 500 via the communication interface 601. The wearable terminal 400 or smartphone 502 then outputs the stress estimation result on an associated screen. This allows the computer 600 to present the stress estimation result to the person being measured 300 in an appropriate manner.

Here, we will provide additional explanation on the creation process (learning) of the above-mentioned stress estimation model using Figures 6 and 7.

Figure 6 shows a schematic diagram of the stress estimation model creation process. In the following, as an example, the stress estimation model creation process will be described as being executed by computer 600. Note that the stress estimation model creation process may be executed by a device other than computer 600.

First, the computer 600 generates biosignal features Fb of sales staff and biosignal features Fb of technicians from the biosignals Sb of sales staff and technicians, and searches for adversarial features Fa that can distinguish both attributes and patterns using the random forest method. Note that the "sales staff data" in FIG. 6 corresponds to the biosignals Sb of sales staff or the biosignal features Fb of sales staff based on the biosignals Sb, and the "technical staff data" in FIG. 6 corresponds to the biosignals Sb of technicians or the biosignal features Fb of technicians based on the biosignals Sb.

Then, the computer 600 creates a stress estimation model based on the attribute/pattern common feature Fc obtained by excluding the hostile feature Fa from each biosignal feature Fb of the sales and technical workers. In this case, the computer 600 first learns the stress estimation model based on the attribute/pattern common feature Fc obtained by excluding the hostile feature Fa from the biosignal feature Fb generated from the biosignal Sb of the technical workers, and the PSS10 score (i.e., the teacher data of the objective variable at the time of stress estimation) which is the result of the stress questionnaire of the technical workers. Then, the computer 600 estimates the PSS10 score of the sales workers (i.e., the result of the stress questionnaire of the sales workers) from the attribute/pattern common feature Fc of the sales workers by using the stress estimation model learned based on the data related to the technical workers. In this estimation, as an example, Support Vector Regression (SVR) is used.

Figure 7 shows a graph of the stress estimation results for sales staff using SVR. In the graph, the horizontal axis shows the number of adversarial features Fa selected by adversarial feature selection using random forest (i.e., the number of features to be excluded), and the vertical axis shows the error between the estimated score of PSS10 using SVR and the actual score of PSS10, which is the correct answer.

Here, in the adversarial feature selection using random forest, which is related to the horizontal axis, features that can easily distinguish between sales and technical jobs, that is, features that are highly dependent on the attribute/pattern of job type, are selected as the adversarial features Fa. As the "number of features to be excluded" shown on the horizontal axis increases, the number of features used in the stress estimation model decreases, and there is a strong tendency for the stress estimation model to be formed only with features that are not dependent on attributes/patterns (job type in this case). Therefore, even a stress estimation model trained only from data on technical jobs is more likely to be able to accurately estimate the stress of sales workers. This can be understood from the situation in which the error shown on the vertical axis gradually decreases as the number of features increases.

After that, if the number of adversarial features Fa selected by the random forest is further increased, the features used in the stress estimation model are eliminated accordingly, and the performance of the stress estimation model deteriorates due to an insufficient number of training samples, and the error indicated on the vertical axis increases. In Figure 7, the error on the vertical axis increases as the "number of features to be excluded" indicated on the horizontal axis increases beyond the number "Nopt" at which the error on the vertical axis is at a minimum.

Taking the above into consideration, in the second embodiment, the relationship between the estimation error corresponding to the vertical axis of FIG. 7 and the parameters for selecting adversarial features is found in the training data. Then, the computer 600 constructs a stress estimation model by eliminating the features selected by the parameters when the estimation error is at its minimum value. By using such a stress estimation model, the computer 600 can estimate the stress of the subject with high accuracy.

Third Embodiment
8 is a block diagram of a stress estimation device 100X according to the third embodiment. The stress estimation device 100X mainly includes a biosignal acquisition means 101X, a biosignal feature amount calculation means 103X, an attribute/pattern data acquisition means 104X, an attribute/pattern common feature amount selection means 105X, and a stress estimation means 106X. The stress estimation device 100X may be composed of a plurality of devices.

The biosignal acquisition means 101X acquires a biosignal of the subject. The biosignal acquisition means 101X can be the biosignal acquisition unit 101 of the first embodiment or the biosignal acquisition unit 602 of the second embodiment. In another example, the biosignal acquisition means 101X can be the biosignal feature amount calculation unit 103 that acquires the biosignal Sb from the biosignal storage unit 102 in the first embodiment.

The biosignal feature calculation means 103X calculates a biosignal feature that is a feature of a biosignal related to stress. The biosignal feature calculation means 103X can be the biosignal feature calculation unit 103 of the first embodiment or the biosignal feature calculation unit 606 of the second embodiment.

The attribute/pattern data acquisition means 104X acquires attribute/pattern data indicating at least one of the attributes or patterns of the subject. The attribute/pattern data acquisition means 104X can be the attribute/pattern data acquisition unit 104 of the first embodiment or the attribute/pattern data acquisition unit 604 of the second embodiment.

The attribute/pattern common feature selection means 105X selects attribute/pattern common features, which are features independent of the attributes and patterns indicated by the attribute/pattern data, from the biosignal features. The attribute/pattern common feature selection means 105X can be the attribute/pattern common feature selection unit 105 of the first embodiment or the attribute/pattern common feature selection unit 608 of the second embodiment.

The stress estimation means 106X estimates the stress of the person being measured based on the attribute/pattern common features. The stress estimation means 106X can be the stress estimation unit 106 of the first embodiment or the stress estimation unit 610 of the second embodiment.

Figure 9 is an example of a flowchart executed by the stress estimation device 100X in the third embodiment. First, the biosignal acquisition means 101X acquires the biosignal of the subject (step S21). Next, the biosignal feature calculation means 103X calculates the biosignal feature, which is a feature of the biosignal related to stress (step S22). The attribute/pattern data acquisition means 104X acquires attribute/pattern data indicating at least one of the attributes or patterns of the subject (step S23). The attribute/pattern common feature selection means 105X selects an attribute/pattern common feature, which is a feature independent of the attribute/pattern indicated by the attribute/pattern data, from the biosignal feature (step S24). The stress estimation means 106X estimates the stress of the subject based on the attribute/pattern common feature (step S25).

The stress estimation device 100X of the third embodiment can appropriately estimate the stress of a person being measured regardless of the person's attributes or pattern.

In each of the above-mentioned embodiments, the program can be stored using various types of non-transitory computer readable media and supplied to a computer, such as a processor. The non-transitory computer readable medium includes various types of tangible storage media. Examples of the non-transitory computer readable medium include magnetic storage media (e.g., flexible disks, magnetic tapes, hard disk drives), magneto-optical storage media (e.g., magneto-optical disks), CD-ROMs (Read Only Memory), CD-Rs, CD-R/Ws, and semiconductor memories (e.g., mask ROMs, PROMs (Programmable ROMs), EPROMs (Erasable PROMs), flash ROMs, and RAMs (Random Access Memory)). The program may also be supplied to the computer by various types of transitory computer readable media. Examples of the transitory computer readable medium include electrical signals, optical signals, and electromagnetic waves. The transitory computer readable medium can supply the program to the computer via wired communication paths such as electric wires and optical fibers, or wireless communication paths.

In addition, some or all of the above embodiments may be described as, but are not limited to, the following notes:

[Appendix 1]
A biosignal acquiring means for acquiring a biosignal of a subject;
a biosignal feature amount calculation means for calculating a biosignal feature amount which is a feature amount of the biosignal related to stress;
an attribute/pattern data acquiring means for acquiring attribute/pattern data indicating at least one of the attributes and patterns of the subject;
an attribute/pattern common feature selection means for selecting an attribute/pattern common feature, which is a feature independent of an attribute/pattern represented by the attribute/pattern data, from the biological signal feature;
a stress estimation means for estimating a stress of the subject based on the attribute/pattern common feature amount;
A stress estimation device comprising:

[Appendix 2]
The stress estimation device according to claim 1, wherein the attribute/pattern common feature selection means selects the attribute/pattern common feature by excluding an adversarial feature that distinguishes the attribute/pattern from the biological signal feature.

[Appendix 3]
The stress estimation device according to claim 2, wherein the attribute/pattern common feature selection means calculates an importance of the biological signal feature in classification of the attribute/pattern, and selects the adversarial feature based on the importance.

[Appendix 4]
The stress estimation device according to claim 2 or 3, wherein the attribute/pattern common feature selection means selects the adversarial features based on a random forest.

[Appendix 5]
The stress estimation means calculates a score indicating an estimation result of the stress from the attribute/pattern common feature amount using a model based on regression analysis.

[Appendix 6]
The biological signal acquisition device further includes a biological signal storage device for storing the biological signal acquired by the biological signal acquisition device,
The stress estimation device according to any one of claims 1 to 5, wherein the feature calculation means calculates the biosignal feature from the biosignal stored in the biosignal storage means.

[Appendix 7]
The stress estimation device according to any one of claims 1 to 6, wherein the biological signal acquisition means receives the biological signal from a wearable device worn by the subject.

[Appendix 8]
By computer,
Acquire a biological signal of the subject,
Calculating a biosignal feature amount, which is a feature amount of the biosignal related to stress;
Acquire attribute/pattern data indicating at least one of the attributes or patterns of the subject;
selecting an attribute/pattern common feature, which is a feature independent of the attribute/pattern indicated by the attribute/pattern data, from the biological signal feature;
Estimating the stress of the subject based on the attribute/pattern common feature amount.
Estimation method.

[Appendix 9]
Acquire a biological signal of the subject,
Calculating a biosignal feature amount, which is a feature amount of the biosignal related to stress;
Acquire attribute/pattern data indicating at least one of the attributes or patterns of the subject;
selecting an attribute/pattern common feature, which is a feature independent of the attribute/pattern indicated by the attribute/pattern data, from the biological signal feature;
A program that causes a computer to execute a process of estimating the stress of the subject based on the attribute/pattern common features.

[Appendix 10]
A storage medium having the program described in appendix 9 stored thereon.

Although the present invention has been described above with reference to the embodiments, the present invention is not limited to the above-mentioned embodiments. Various modifications that can be understood by a person skilled in the art can be made to the configuration and details of the present invention within the scope of the present invention. In other words, the present invention naturally includes various modifications and amendments that a person skilled in the art could make in accordance with the entire disclosure, including the scope of the claims, and the technical ideas. Furthermore, the disclosures of the above cited patent documents, etc. are incorporated by reference into this document.

100 Stress estimation device 101 Biosignal acquisition unit 102 Biosignal storage unit 103 Biosignal feature calculation unit 104 Attribute/pattern data acquisition unit 105 Attribute/pattern common feature selection unit 106 Stress estimation unit 150 Stress estimation system 200 Wearable terminal 300 Subject 400 Wearable terminal 500 Communication means 501 Bluetooth communication 502 Smartphone 503 Mobile data communication or Wifi communication 504 Internet 600 Computer 601 Communication interface 602 Biosignal acquisition unit 603 Biosignal storage unit 604 Attribute/pattern data acquisition unit 605 Attribute/pattern data storage unit 606 Biosignal feature calculation unit 607 Biosignal feature storage unit 608 Attribute/pattern common feature selection unit 609 Attribute/pattern common feature storage unit 610 Stress estimation unit 611 Stress estimation result storage unit 612 Stress estimation result output unit

Claims (9)

A biosignal acquiring means for acquiring a biosignal of a subject;
a biosignal feature amount calculation means for calculating a biosignal feature amount which is a feature amount of the biosignal related to stress;
an attribute/pattern data acquiring means for acquiring attribute/pattern data indicating at least one of the attributes and patterns of the subject;
an attribute/pattern common feature selection means for excluding, from the biological signal feature, an adversarial feature that distinguishes the attribute/pattern based on the attribute/pattern data , and selecting the biological signal feature from which the adversarial feature has been excluded as an attribute /pattern common feature that is a feature independent of the attribute/pattern indicated by the attribute/pattern data;
a stress estimation means for estimating a stress of the subject based on the attribute/pattern common feature amount;
A stress estimation device comprising: 2. The stress estimation device according to claim 1 , wherein the attribute/pattern common feature selection means calculates an importance of the biological signal feature in classification of the attribute/pattern, and selects the adversarial feature based on the importance. The stress estimation device according to claim 1 , wherein the attribute/pattern common feature selection means selects the adversarial features based on a random forest. The stress estimation device according to claim 1 , wherein the stress estimation means calculates a score indicating an estimation result of the stress from the attribute/pattern common feature amount by using a model based on a regression analysis. The biological signal acquisition device further includes a biological signal storage device for storing the biological signal acquired by the biological signal acquisition device,
5. The stress estimation device according to claim 1 , wherein the feature amount calculation means calculates the biosignal feature amount from the biosignal stored in the biosignal storage means. The stress estimation device according to claim 1 , wherein the biological signal acquisition means receives the biological signal generated by a wearable device worn by the person being measured. By computer,
Acquire a biological signal of the subject,
Calculating a biosignal feature amount, which is a feature amount of the biosignal related to stress;
Acquire attribute/pattern data indicating at least one of the attributes or patterns of the subject;
excluding, from the biological signal features , adversarial features that distinguish the attributes and patterns based on the attribute/pattern data , and selecting the biological signal features from which the adversarial features have been excluded as attribute/pattern common features that are features that are not dependent on the attributes and patterns indicated by the attribute/pattern data;
Estimating the stress of the subject based on the attribute/pattern common feature amount.
Estimation method. Acquire a biological signal of the subject,
Calculating a biosignal feature amount, which is a feature amount of the biosignal related to stress;
Acquire attribute/pattern data indicating at least one of the attributes or patterns of the subject;
excluding, from the biological signal features , adversarial features that distinguish the attributes and patterns based on the attribute/pattern data , and selecting the biological signal features from which the adversarial features have been excluded as attribute/pattern common features that are features that are not dependent on the attributes and patterns indicated by the attribute/pattern data;
A program that causes a computer to execute a process of estimating the stress of the subject based on the attribute/pattern common features. A storage medium on which the program according to claim 8 is stored.

Images