Pre-training model generation method, generation device, and program

Abstract

Provided is a method for generating a pre-training model that outputs an abnormality score indicating the degree of abnormality of the physical state of a subject using as input a feature amount based on activity data of the subject, a first pre-training model that is trained using first activity data of the subject acquired in a past first period, and a second pre-training model that is trained using second activity data of the subject acquired in a past second period, and that outputs an abnormality score indicating the degree of abnormality of the physical state of the subject. First activity data of the subject and second activity data of the subject acquired during a second period following the first period are acquired (S210), whether or not there is a difference between the first activity data and the second activity data is determined (S220), and if there is a difference between the first activity data and the second activity data (YES in S220), a second pre-training model is generated (S240) using third activity data of the subject acquired recently.

Description

Technical Field

[0001] The present disclosure relates to a method for generating a pre-training model, a device for generating a pre-training model, and a program. Background Art

[0002] The 2025 problem is the problem of an aging society, in which 8 million people of the so-called "baby boomer generation" will be over 75 years old, and one in four people will be over 75 years old. One of these problems is the shortage of manpower due to the increasing demand for medical care and nursing care.

[0003] As a result, the number of patients and individuals being cared for by medical and nursing professionals has increased, leading to the possibility that even small changes in physical condition that could indicate an abnormality in the patient's health may go unnoticed. Consequently, if these small changes in physical condition are overlooked, the patient may develop a more serious condition.

[0004] In this regard, for example, the technology disclosed in Patent Document 1 is to notify the abnormality of the monitored person to an appropriate notification party when the monitored person is judged to be abnormal. According to this, the abnormality of the monitored person can be notified to an appropriate monitoring person according to the abnormal state of the monitored person.

[0005] (Prior art literature)

[0006] (Patent Document)

[0007] Patent Document 1 International Publication No. 2018 / 116830 Summary of the Invention

[0008] Problems to be solved by the invention

[0009] However, the technology of Patent Document 1 described above may cause a decrease in abnormality detection performance when a subject's physical condition undergoes chronic changes.

[0010] Therefore, the present disclosure provides a method for generating a pre-trained model, a device for generating a pre-trained model, and a program, wherein the pre-trained model can support high-precision detection of abnormalities in a subject when the subject undergoes chronic changes in physical condition.

[0011] Means for solving problems

[0012] Regarding a method for generating a pre-trained model involved in one embodiment of the present invention, the pre-trained model takes a feature value based on the activity data of a subject as input and outputs an abnormality score indicating the degree of abnormality in the physical condition of the subject. The first pre-trained model is trained using the first activity data of the subject acquired in the past first period. In the method for generating the pre-trained model, the first activity data of the subject and the second activity data of the subject acquired in the second period after the first period are acquired, and it is determined whether there is a difference between the first activity data and the second activity data. If there is a difference between the first activity data and the second activity data, the second pre-trained model is generated using the third activity data of the subject acquired most recently.

[0013] Regarding a pre-trained model generation device involved in one embodiment of the present invention, the pre-trained model takes a feature value based on the activity data of the subject as input and outputs an abnormality score indicating the degree of abnormality in the physical condition of the subject. The first pre-trained model is trained using the first activity data of the subject acquired in the past first period. The pre-trained model generation device comprises: an acquisition unit, which acquires the first activity data of the subject and the second activity data of the subject acquired in the second period after the first period; a judgment unit, which judges whether there is a difference between the first activity data and the second activity data; and a model updating unit, which uses the third activity data of the subject acquired most recently to generate a second pre-trained model when there is a difference between the first activity data and the second activity data.

[0014] A program according to one embodiment of the present disclosure is a program for causing a computer to execute the above-described method for generating a pre-trained model.

[0015] Effects of the Invention

[0016] According to one embodiment of the present disclosure, a method for generating a pre-trained model can be realized, wherein the pre-trained model can support high-precision detection of abnormalities in a subject when the subject experiences chronic changes in physical condition. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a diagram showing an example of the configuration of a body condition detection system according to an embodiment.

[0018] Figure 2 This is a block diagram showing an example of the functional configuration of the information management server according to the embodiment.

[0019] Figure 3 This is a flowchart illustrating the operation of generating a first pre-trained model in the information management server according to the embodiment.

[0020] Figure 4 This is a flowchart showing the work of detecting the precursor of abnormal physical condition of the subject in the information management server involved in the embodiment.

[0021] Figure 5 This is a diagram showing an example of five-stage staged scoring and its conditions according to the embodiment.

[0022] Figure 6 This is a flowchart illustrating the operation of generating the second pre-trained model in the information management server according to the embodiment.

[0023] Figure 7 This is a schematic timing chart showing the operation of generating the second pre-trained model in the information management server according to the embodiment.

[0024] Figure 8 The results of comparing the detection success rate with and without switching based on the pre-trained model are shown.

[0025] Figure 9 FIG. 1 is a first diagram for explaining problems in conventional methods.

[0026] Figure 10 FIG. 2 is a second diagram for explaining problems in conventional methods. DETAILED DESCRIPTION

[0027] (The process of discovering this disclosure)

[0028] Before explaining the present disclosure, first refer to Figure 9 as well as Figure 10 The process of discovering the present disclosure is described. Figure 9 FIG. 1 is a first diagram for explaining problems in conventional methods. Figure 10 FIG. 2 is a second diagram for explaining problems in conventional methods.

[0029] The details will be described later, but here we discuss the generation of a machine learning model that detects small abnormal changes in the physical state of the subject that are related to abnormal physical state of the subject based on the subject's activity data, nursing records, etc. The machine learning model can be generated by supervised learning or by unsupervised learning. For example, when supervised learning is used, compared with the case of using unsupervised learning, a pre-trained model that can detect small abnormal changes in physical state with high precision can be generated. Since supervised learning is used, training data for supervised learning is collected, and the pre-trained model is generated through supervised learning using the collected training data. For example, in order to generate training data, activity data such as the breathing rate and heart rate of the subject are collected.

[0030] In this context, elderly individuals requiring care may experience chronic changes in their physical condition (e.g., activity data) due to illness or aging. Conventional model building methods use all collected and accumulated activity data as training data. Consequently, the training data may contain activity data from both before and after a chronic physical condition change occurs, or only from before the chronic physical condition change occurs. Using such activity data to generate a pre-trained model could potentially affect detection performance (specifically, reduce detection performance).

[0031] Chronic changes in physical condition refer to gradual, continuous changes in physical condition, such as those caused by age, lifestyle, genetic factors, obesity, and other factors. Chronic changes in physical condition can occur gradually over weeks or months. Chronic changes in physical condition do not include sudden changes in physical condition (e.g., a sudden headache or high fever).

[0032] like Figure 9 As shown, the training data accumulated during the training period (the first period) in which data is continuously accumulated is used to generate a pre-trained model, and the pre-trained model is used to detect small abnormal changes in the physical state of the subject. At this time, the range between the double-dotted lines is set as the range of normal values ​​in the training data. In addition, the first period is considered to include the period when the subject has a chronic change in physical state (during Figure 9 In the example, the period between the time when the activity data value continuously decreases (physical state changes) and the time before and after. That is, the training data includes activity data before and after the subject's chronic physical state changes. It can be considered that the normal value ranges of activity data before and after the chronic physical state changes are different.

[0033] Even after the first period, the subject's chronic physical condition changes continue to occur. For example, during the period of time outlined by the dotted circle, a small abnormal change in the physical condition occurs ( Figure 9 Although the activity data is outstanding, the normal value range is set based on the activity data of the first period. Therefore, when using a pre-trained model trained using the activity data of the first period, there is a possibility that the outstanding moment will be judged as normal.

[0034] So, if Figure 10 As shown, when the subject has a chronic change in physical state, a training period (i.e., a period of accumulating data for switching (reconstructing) the pre-trained model) is set. Figure 10The "training period when the model is switched" in the second period is the second period), and it is hoped that the pre-trained model will be reconstructed using the data from the second period. That is, when the subject has a chronic change in physical state, it is hoped that the pre-trained model will be regenerated using the data collected during the second period. The second period is a period of chronic physical state change, and by reconstructing the pre-trained model using the activity data obtained during the second period, it is possible to reconstruct the range of normal values ​​in the case of chronic physical state change ( Figure 10 The "range of normal values ​​within the updated training data" in the training data is set again. In addition, the length of the second period is not particularly limited and can be shorter than the first period, the same as the first period, or longer than the first period.

[0035] Occurs after the normal value range is updated Figure 9 The outliers shown ( Figure 10 In the case of a prominent condition (shown by a dotted circle), the updated normal value range is used, thereby increasing the accuracy of detecting small abnormal changes in physical condition occurring at the time of the prominent condition. In other words, it is conceivable that the updated pre-trained model can be used to detect small abnormal changes in physical condition in the case of chronic physical condition changes with higher accuracy.

[0036] Patent Document 1 does not disclose a technique for detecting small abnormal changes in the physical condition of a subject with higher accuracy when the subject's physical condition undergoes chronic fluctuations.

[0037] Therefore, the inventors of the present invention have conducted in-depth research on supporting the detection of small abnormal changes in the subject's physical condition (precursors of abnormal physical condition) with higher accuracy when the subject's physical condition undergoes chronic changes, and created a method for generating a pre-trained model to be described below.

[0038] Regarding the method for generating a pre-trained model involved in the first embodiment of the present invention, the pre-trained model takes a feature value based on the activity data of the subject as input and outputs an abnormality score indicating the degree of abnormality in the physical condition of the subject. The first pre-trained model is trained using the first activity data of the subject acquired in the past first period. In the method for generating the pre-trained model, the first activity data of the subject and the second activity data of the subject acquired in the second period after the first period are acquired, and it is determined whether there is a difference between the first activity data and the second activity data. If there is a difference between the first activity data and the second activity data, the second pre-trained model is generated using the third activity data of the subject acquired most recently.

[0039] Accordingly, when there is a difference between the first activity data and the second activity data, that is, when it is assumed that the subject has undergone a chronic change in physical state, the pre-trained model that outputs the abnormality score is updated. Since the updated pre-trained model (the second pre-trained model) is trained using the most recent third activity data, it will become a model corresponding to the subject's most recent state (that is, the state in which a chronic change in physical state has occurred). Therefore, based on the method for generating a pre-trained model, by using the updated model, it is possible to support the detection of abnormalities in the subject with higher accuracy when the subject has undergone a chronic change in physical state. In addition, the update includes regenerating the pre-trained model.

[0040] Furthermore, for example, the method for generating a pre-trained model according to the second aspect may be based on the method for generating a pre-trained model according to the first aspect, and the third activity data may include the second activity data.

[0041] Therefore, since the second pre-trained model is trained using the subject's current state (latest activity data), a model can be generated that can accurately detect whether chronic changes in the subject's physical state have occurred in the subject's current state. This can support further high-precision detection of abnormalities in the subject.

[0042] Furthermore, for example, the method for generating a pre-trained model according to the third embodiment may be based on the method for generating a pre-trained model according to the second embodiment, and the third activity data may further include a portion of the first activity data.

[0043] Thus, if the training data for generating the second pre-trained model is insufficient, it can be compensated by the training data of the first period. Thus, it is possible to suppress the reduction in detection performance of the second pre-trained model due to insufficient training data.

[0044] For example, the method for generating a pre-trained model involved in the fourth embodiment may be based on the method for generating a pre-trained model involved in any one of the first to third embodiments, and the difference includes a difference indicating that the subject has a chronic change in physical condition.

[0045] This allows the model to be updated at the timing when chronic changes in the physical condition occur.

[0046] And for example, the method for generating a pre-trained model involved in the fifth method may be based on the method for generating a pre-trained model involved in any one of the first to fourth methods, and the judgment of whether there is the difference between the first activity data and the second activity data is performed using the statistics of the first activity data and the statistics of the second activity data.

[0047] According to this, by using statistics, it is possible to easily determine whether there is a chronic change in physical condition.

[0048] For example, the method for generating a pre-trained model according to the sixth embodiment is based on the method for generating a pre-trained model according to any one of the first to fifth embodiments, and the standard deviation of the first activity data is set to sd B , the standard deviation of the third activity data is set to sd A When the following formula 1 is satisfied, it is determined that there is the difference.

[0049] (|sd A -sd B | / sd A )×100≥threshold value (Formula 1).

[0050] Therefore, by substituting the standard deviation into Formula 1, it is possible to easily determine whether there is a chronic change in physical condition. Furthermore, by using the standard deviation, it is possible to easily detect a chronic change in physical condition, that is, a chronic change in activity data.

[0051] And for example, the method for generating a pre-trained model involved in the 7th method may be based on the method for generating a pre-trained model involved in any one of the 1st to 6th methods, and the judgment of whether there is the difference between the 1st activity data and the 2nd activity data is performed according to the prescribed period.

[0052] According to this, since whether or not to update the model is determined at a predetermined period, even if chronic changes in the physical condition occur in the future, it is possible to support detection of abnormalities in the subject with higher accuracy.

[0053] For example, the method for generating a pre-trained model involved in the 8th method may be based on the method for generating a pre-trained model involved in any one of the 1st to 7th methods, and further generates the second pre-trained model when a specified unexpected event occurs to the subject.

[0054] Accordingly, by generating a second pre-trained model according to unexpected events that may cause changes in the subject's physical state, it is possible to support the detection of abnormalities in the subject with higher accuracy even when there are no chronic changes in the subject's physical state.

[0055] For example, the method for generating a pre-trained model involved in the 9th method is based on the method for generating a pre-trained model involved in any one of the 1st to 8th methods, and when the second pre-trained model is generated, the pre-trained model used in the output of the abnormality score for the subject is switched from the first pre-trained model to the second pre-trained model.

[0056] Accordingly, when a subject's physical condition undergoes chronic changes, switching to the second pre-trained model suitable for the physical condition changes can support detection of abnormalities in the subject with higher accuracy.

[0057] For example, the method for generating a pre-trained model involved in the tenth method may be based on the method for generating a pre-trained model involved in any one of the first to ninth methods, and when there is no difference between the first activity data and the second activity data, the first pre-trained model continues to be used.

[0058] Accordingly, when the subject does not experience chronic changes in physical condition, the first pre-trained model can be continued to be used to support the detection of abnormalities in the subject. Furthermore, when there are no chronic changes in physical condition, the model is not updated, thereby reducing the processing capacity of the device executing the method for generating the pre-trained model.

[0059] Regarding the pre-trained model generation device involved in the 11th embodiment of the present disclosure, the pre-trained model takes as input a feature value based on the activity data of the subject and outputs an abnormality score indicating the degree of abnormality of the subject's physical condition. The first pre-trained model is trained using the first activity data of the subject acquired in the past first period. The pre-trained model generation device comprises: an acquisition unit that acquires the first activity data of the subject and the second activity data of the subject acquired in the second period after the first period; a judgment unit that judges whether there is a difference between the first activity data and the second activity data; and a model updating unit that uses the third activity data of the subject acquired most recently to generate the second pre-trained model when there is a difference between the first activity data and the second activity data. Furthermore, the program involved in the 12th embodiment of the present disclosure is a program for causing a computer to execute the pre-trained model generation method involved in any one of the first to tenth embodiments.

[0060] Based on this, the same effect as the above-mentioned method of generating the pre-training model is achieved.

[0061] Furthermore, these general or specific aspects may be implemented by a system, method, integrated circuit, computer program, or non-transitory recording medium such as a computer-readable CD-ROM, or by any combination of these systems, methods, integrated circuits, computer programs, or recording media. The program may be stored in advance on a recording medium or provided to the recording medium via a wide area communication network, including the Internet.

[0062] Hereinafter, the embodiments will be described in detail with reference to the accompanying drawings.

[0063] In addition, the embodiments described below are general or specific examples. The numerical values, components, configuration positions of components, connection methods, steps, and the order of steps shown in the following embodiments are examples and are not intended to limit the present disclosure. In addition, for the components of the following embodiments that are not described in the independent technical solutions, they will be described as arbitrary components.

[0064] Furthermore, each figure is a schematic diagram and is not a strict illustration. Therefore, for example, the scales in each figure are not consistent. In addition, in each figure, substantially the same components are given the same number, and repeated descriptions are omitted or simplified.

[0065] In this specification, numerical values ​​and numerical ranges are not strictly defined but are intended to include substantially equivalent ranges, for example, including differences of approximately several percent (eg, approximately 10%).

[0066] Furthermore, in this specification, ordinal numbers such as “first” and “second” do not refer to the number or order of components unless otherwise specified, but are used for the purpose of distinguishing components to avoid confusion between the same components.

[0067] (Implementation Method)

[0068] The following reference Figures 1 to 8 The method for generating the pre-trained model according to this embodiment is described.

[0069] [1. Configuration of the physical condition detection system]

[0070] First, refer to Figure 1 as well as Figure 2 The structure of a physical condition detection system including an information management server that executes a method for generating a pre-trained model is described. Figure 1 An example of the configuration of the body condition detection system 100 according to this embodiment is shown.

[0071] The physical condition detection system 100 involved in this embodiment is a system configured so that the information management server 10 can detect small abnormal changes in the physical condition of the subject 50 being cared for or looked after (i.e., precursors to abnormal physical condition).

[0072] like Figure 1 As shown, the body state detection system 100 includes an information management server 10, a sensor unit 20, and a display terminal unit 30. These are connected to each other via a communication network 40 so as to be able to communicate with each other. The communication network 40 can be a wired network, a wireless network, or both a wired network and a wireless network. Figure 1 3 shows: a subject 50 who is the subject of nursing or care; a user 60 who is a field worker such as a medical professional providing nursing or care for the subject 50; a field worker, i.e., a user 61 who is a monitor of the subject 50 and can confirm the display terminal unit 30; and recorded data 25 that records the details of the nursing or care provided by the user 60 to the subject 50. The recorded data 25 includes, for example, the amount of food consumed by the subject 50 in the morning, afternoon, and evening, i.e., the amount of food consumed, and the body temperature, input by the field worker, i.e., the user 60.

[0073] In addition, Figure 1 Although the example in which the body condition detection system 100 includes one sensor unit 20 is shown, the present invention is not limited thereto and the body condition detection system 100 may include the same number of sensor units 20 as the number of the subjects 50 being cared for or looked after.

[0074] The sensing unit 20 acquires activity data by sensing the activity data of the subject 50 during a specified period. The activity data includes at least one of heart rate, respiratory rate, whether or not the subject is in bed, body temperature, and food intake. For example, it may also include at least two of heart rate, respiratory rate, whether or not the subject is in bed, body temperature, and food intake. For example, the sensing unit 20 can acquire data such as the heart rate, respiratory rate, body movement, etc. of the subject 50 while the subject is in bed per second (hereinafter also referred to as sensing data). Furthermore, the sensing unit 20 can sense whether the subject 50 is in bed by whether it can sense the heart rate, respiratory rate, body movement, etc.

[0075] Furthermore, the interval for acquiring sensing data such as heart rate, respiratory rate, and body motion is not limited to one second, but may be two seconds, or may be an interval based on changes in sensing data of the subject 50. Furthermore, the sensing unit 20 may further detect sleep patterns and other life patterns by detecting heart rate, respiratory rate, and body motion.

[0076] In addition, the following description will mainly focus on the case where activity data includes respiratory rate and heart rate.

[0077] Furthermore, the sensing unit 20 may be a sensor device including a pressure sensor, for example, and may be placed on a bed to sense the condition of the subject 50 on a per-second basis. In this case, the sensing unit 20 may output, for example, a value of "1" indicating that the subject 50 is not in bed as sensing data indicating that the subject 50 is not in bed, per second. Furthermore, the sensing unit 20 may also output sensing data such as the subject's 50 respiratory rate, per second, for example.

[0078] Furthermore, the sensor unit 20 may be, for example, an imaging device such as a camera, configured to capture images of the subject 50 lying in bed or while eating. The camera may be a thermal imaging camera for detecting the body temperature of the subject 50, or a conventional camera (e.g., a CCD (Charge Coupled Device) camera). The amount of food consumed can be determined through image analysis of the images.

[0079] The information management server 10 is implemented, for example, by a computer having a processor (microprocessor), a memory, a communication interface, and the like. The information management server 10 may also be partially included in a cloud server to operate. The information management server 10 generates a pre-trained model for detecting small abnormal changes in the physical state of the subject 50 (i.e., precursors to abnormal physical state) related to abnormal physical state, and uses the generated pre-trained model to detect small abnormal changes in the physical state of the subject 50 related to abnormal physical state.

[0080] Figure 2 This is a block diagram showing an example of the functional configuration of the information management server 10 according to the present embodiment.

[0081] like Figure 2 As shown, the information management server 10 includes a transceiver 11, an information recording unit 12, a feature value calculation unit 13, a model generation unit 14, a model update determination unit 15, and a body condition detection unit 16. The pre-trained model generation device is composed of at least the model update determination unit 15. Alternatively, the pre-trained model generation device can be implemented as a separate device.

[0082] The transceiver 11 includes, for example, a communication interface, and transmits and receives various information with the sensor unit 20 or the display terminal 30 via the communication network 40. For example, the transceiver 11 acquires activity data including the respiratory rate and heart rate of the subject 50 during a predetermined period. Here, activity data may include at least the respiratory rate and heart rate, such as the subject's body temperature, food intake, respiratory rate, heart rate, and bed ambulation rate during the predetermined period. Furthermore, the transceiver 11 outputs the staged scores calculated by the physical condition detection unit 16 to a terminal of a user 61, such as a monitor of the subject 50.

[0083] In this embodiment, the transceiver 11 obtains the sensing data of the subject 50 lying on the bed, such as the heart rate, respiratory rate, and body movement, from the sensor 20 via the communication network 40 at a predetermined interval of, for example, 1 minute. Figure 1 The recorded data 25 shown is that the on-site staff, namely the user 60, provides care or supervision to the subject 50. In this way, the transceiver 11 acquires activity data including the sensed data and the recorded data 25 via the communication network 40, that is, acquires activity data obtained daily on-site.

[0084] Furthermore, the transceiver 11 transmits the staged score calculated by the physical condition detection unit 16 to the display terminal 30 via the communication network 40. Furthermore, the transceiver 11 may also transmit information for display on the user interface of the display terminal 30. Such displays include the staged score display, vital sign change chart display, or list group display (described later) that allow the user 61 to take measures to address abnormal physical conditions of the subject 50. Furthermore, the transceiver 11 may also obtain at least one of the subject 50's body temperature and food intake included in the nursing record.

[0085] The information recording unit 12 records information sent and received by the transceiver 11. The information recording unit 12 is a recording medium capable of recording information, and is comprised of a rewritable, nonvolatile memory such as a hard disk drive or solid-state drive. Furthermore, the information recording unit 12 may also record multiple feature quantities calculated by the feature quantity calculation unit 13.

[0086] The characteristic quantity calculation unit 13 comprises, for example, a computer including a memory and a processor (microprocessor), and the processor executes a control program stored in the memory to implement the function of calculating multiple characteristic quantities. The characteristic quantity calculation unit 13 calculates multiple characteristic quantities based on the activity data, including the respiratory rate and heart rate, of the subject 50, acquired by the transceiver 11. For example, the characteristic quantity calculation unit 13 acquires sensory data for a time period including the date and time of the physical condition detection based on the activity data acquired by the transceiver 11 or recorded in the information recording unit 12, and calculates the characteristic quantity for each hour based on the sensory data, such as the respiratory rate.

[0087] Here, the characteristic quantity calculation unit 13 may calculate, for example, at least the average value and maximum value of the respiratory frequency of the subject 50 and the average value and maximum value of the heart rate of the subject 50 as multiple characteristic quantities per hour. In this embodiment, the characteristic quantity calculation unit 13 calculates, based on at least the respiratory frequency and heart rate, the average value, maximum value, standard deviation, skewness, kurtosis in the differential data of the respiratory frequency, the heart rate, and the average value and maximum value of the respiratory frequency and heart rate in the pulse factor as multiple characteristic quantities. Here, the pulse factor is obtained by subtracting the average value from the maximum value. In this way, the characteristic quantity calculation unit 13 performs statistical processing on the activity data to calculate multiple characteristic quantities.

[0088] More specifically, feature quantity calculation unit 13 calculates, for example, feature quantities related to the respiratory rate and heart rate of subject 50 on an hourly basis. For example, feature quantity calculation unit 13 obtains sensing data indicating the respiratory rate of subject 50 during a time period including the target date and time for physical condition detection from the activity data recorded in information recording unit 12 or the sensing data acquired from sensing unit 20, and calculates statistical feature quantities for each hour of that time period.

[0089] More specifically, the characteristic quantity calculation unit 13 obtains respiratory rate data from the activity data, for example, data showing a respiratory rate that is not zero for a certain hour. Based on the obtained respiratory rate data, the characteristic quantity calculation unit 13 calculates the average value, maximum value, minimum value, standard deviation, skewness, kurtosis, and impulse factor for that hour as statistical characteristic quantities. The impulse factor can be calculated from the difference between the maximum value and the average value (maximum value minus average value) in the respiratory rate data for that hour. Furthermore, based on the differential data of the obtained respiratory rate data, the characteristic quantity calculation unit 13 calculates the average value, maximum value, minimum value, standard deviation, skewness, kurtosis, and impulse factor for that hour as statistical characteristic quantities. The differential data of the obtained respiratory rate data, for example, is data showing the difference between the respiratory rate at time t and the respiratory rate at time t+1, one second after time t, that is, data showing the difference in respiratory rate data per second. Furthermore, the characteristic quantity calculation unit 13 only needs to calculate at least the average value and maximum value for that hour as statistical characteristic quantities based on the obtained respiratory rate data.

[0090] Furthermore, for example, the characteristic quantity calculation unit 13 obtains heart rate data showing the heart rate of the subject 50 in a time period including the object date and time of the physical condition detection from the activity data recorded in the information recording unit 12 or the sensing data obtained from the sensing unit 20, and calculates the statistical characteristic quantity for each hour of the time period.

[0091] Here, the feature quantity calculation unit 13 obtains heart rate data, for example, indicating a non-zero heart rate for a certain hour, from the activity data. Based on the obtained heart rate data, the feature quantity calculation unit 13 calculates the average value, maximum value, minimum value, standard deviation, skewness, kurtosis, and impulse factor for that hour as statistical feature quantities. Furthermore, based on the differential data of the obtained heart rate data, the feature quantity calculation unit 13 calculates the average value, maximum value, minimum value, standard deviation, skewness, kurtosis, and impulse factor for that hour as statistical feature quantities. The differential data of the obtained heart rate data, like the differential data of the respiratory rate data, is data showing, for example, the difference between the heart rate at time t and the heart rate at time t+1, one second after time t, that is, data showing the difference in heart rate data per second. Furthermore, the feature quantity calculation unit 13 only needs to calculate, based on the obtained heart rate data, at least the average value and maximum value for that hour as statistical feature quantities.

[0092] Furthermore, the feature quantity calculation unit 13 may calculate the food intake, bed leaving rate, and body temperature of the measurement subject 50 as one of the plurality of feature quantities.

[0093] For example, the characteristic quantity calculation unit 13 can calculate the amount of food consumed by the subject 50 as one of the multiple characteristic quantities based on the nursing records included in the activity data. In this case, the characteristic quantity calculation unit 13 can calculate the total amount of food consumed in the past day based on the nursing records, and then calculate the sum of the food consumed during the time period including the target date and time of the physical condition detection. Here, when the target date and time are the time periods of morning, noon, and evening, the characteristic quantity calculation unit 13 can, for example, calculate the amount of food consumed from the morning of the previous day to the morning of the current day, the amount of food consumed from noon of the previous day to noon of the current day, and the amount of food consumed from the evening of the previous day to the evening of the current day.

[0094] Furthermore, for example, the characteristic quantity calculation unit 13 can calculate the bed leaving rate as one of the multiple characteristic quantities based on the activity data acquired by the transceiver unit 11 and recorded in the information recording unit 12. In this case, the characteristic quantity calculation unit 13 can obtain presence / absence data indicating whether the subject 50 was in bed during a time period including the target date and time for physical condition detection from the activity data recorded in the information recording unit 12 or the sensing data acquired from the sensing unit 20, and calculate the bed leaving rate for each hour of that time period. More specifically, the characteristic quantity calculation unit 13 can calculate the bed leaving rate for each hour by, for example, counting the number of "1" values ​​indicating periods of bed leaving in a given hour and dividing this by the total number of times the subject was in bed in that hour (i.e., the sum of the number of "1" values ​​indicating periods of bed leaving in that hour and the number of "0" values ​​indicating being in bed).

[0095] Furthermore, for example, the characteristic quantity calculation unit 13 may also calculate the body temperature of the subject 50 as one of the multiple characteristic quantities based on the nursing records included in the activity data. In this case, the characteristic quantity calculation unit 13 may calculate the body temperature (e.g., average body temperature) for the past day based on the nursing records, and then calculate the body temperature (e.g., average body temperature) for the time period including the target date and time of the physical condition detection. Here, when the target date and time is the time period of morning, noon, or evening, the characteristic quantity calculation unit 13 may, for example, calculate the body temperature from the morning of the previous day to the morning of the current day, the body temperature from noon of the previous day to noon of the current day, and the body temperature from the evening of the previous day to the evening of the current day.

[0096] The model generation unit 14 generates a model using multiple feature quantities. The model generation unit 14 can also generate a model (supervised learning model) through supervised learning using training data including information based on multiple feature quantities and label data. In addition, the model generation unit 14 can also generate a model that has learned normality or abnormality (unsupervised learning model) through unsupervised learning using training data including an activity data group composed of multiple feature quantities. When multiple feature quantities of the subject 50 are input, the model generated in this way outputs an abnormality score indicating the degree of abnormality of the physical condition of the subject 50. In addition, the following mainly uses the example of using supervised learning to generate a model for explanation.

[0097] In this embodiment, the model generation unit 14 includes, for example, a computer including a memory and a processor (microprocessor). The processor executes a control program stored in the memory to implement various functions. The model generation unit 14 obtains activity data for the entire training period from the activity data recorded by the information recording unit 12 or the sensory data obtained by the sensor 20. Alternatively, the model generation unit 14 may obtain recorded data 25 for the entire training period and include it in the activity data for the entire training period.

[0098] Then, the model generation unit 14 causes the feature value calculation unit 13 to calculate the feature value for each hour based on the activity data for the entire training period. The model generation unit 14 generates a model by training the model using the feature value for each hour during the entire training period.

[0099] As supervised learning models, for example, AnoGAN (Anomaly Detection with Generative Adversarial Network), VAE, and DeepSVDD (Deep Support Vector Data Description) are used, but the present invention is not limited thereto.

[0100] Furthermore, as unsupervised learning models, a model that separates outliers using a decision tree (for example, an isolation forest model: a model based on an isolation forest algorithm) or a model based on a k-means algorithm is used, but the present invention is not limited thereto.

[0101] The model generating unit 14 may generate a model unique to the subject 50. In other words, there may be as many models as there are subjects 50.

[0102] Model update determination unit 15 determines whether a chronic change in the physical condition of subject 50 has occurred based on the acquired activity data. If a chronic change in the physical condition of subject 50 has occurred, model generation unit 14 is caused to generate a new model. For example, model update determination unit 15 may determine whether a chronic change in the physical condition has occurred by using at least activity data acquired after model generation unit 14 has generated the model.

[0103] The model update determination unit 15 is implemented by, for example, a computer including a processor (microprocessor), a memory, a communication interface, etc., and various functions are realized by the processor executing a control program stored in the memory.

[0104] The model update determination unit 15 includes a data aggregation unit 151 , a chronic change detection unit 152 , and a model updating unit 153 .

[0105] The data aggregation unit 151 aggregates the first activity data of the subject 50 obtained during the first period and the second activity data of the subject 50 obtained during the second period after the first period. The data aggregation unit 151 aggregates the first activity data and the second activity data, the first activity data being data that was used to generate the model currently used by the body state detection unit 16 (an example of the first pre-trained model), and the second activity data being data acquired during a period including the most recent prescribed period after the model was generated. Although the data aggregation unit 151 collects the first feature quantity based on the first activity data and the second feature quantity based on the second activity data from the feature quantity calculation unit 13, the first activity data and the second activity data themselves may also be collected. The data aggregation unit 151 is an example of an acquisition unit. Aggregation refers to the accumulation of data, for example, it means collecting data and storing it in a storage device.

[0106] The most recent prescribed period is a period that includes the current time and extends back a prescribed period to the past, and is, for example, the most recent past month.

[0107] Chronic change detection unit 152 detects whether a chronic change in the physical condition of subject 50 has occurred based on the first and second motion data collected by data collection unit 151. Chronic change detection unit 152 may, for example, determine whether a difference between the first and second motion data exceeds a threshold value. If a difference exceeding the threshold value is determined, chronic change in the physical condition of subject 50 may be detected. The threshold value may be pre-set to indicate the presence of a chronic change in the physical condition of subject 50.

[0108] Furthermore, whether the first activity data and the second activity data differ can be determined using, for example, the statistics of the first activity data and the statistics of the second activity data. For example, whether the first activity data and the second activity data differ can be determined using the difference in the mean values ​​of the statistics (e.g., Equation 1 shown below) or using a significant difference test.

[0109] When the standard deviation of the first activity data is set to sd B , set the standard deviation of the most recent second activity data to sd A When the following formula 1 is satisfied, the chronic change detection unit 152 determines that there is a difference in the activity data.

[0110] (|sd A -sd B | / sd A )×100≥threshold value (Formula 1)

[0111] Standard deviation sd A and sd B is an example of a statistic.

[0112] When the activity data is respiratory rate, the threshold is set to a higher value than when the activity data is heart rate, such as 10%, 15%, or 20%. Furthermore, when the activity data is heart rate, the threshold is set to a lower value than when the activity data is respiratory rate, such as 3%, 5%, or 7%. The threshold can be set in advance and recorded in the model update determination unit 15.

[0113] The chronic change detection unit 152 uses formula 1 to judge each of multiple activity data (multiple feature quantities). When at least one activity data (at least one feature quantity) does not satisfy formula 1, it can be detected that a chronic physical state change has occurred in the subject 50.

[0114] Furthermore, chronic change detection unit 152 determines whether a predetermined unexpected event has occurred in subject 50. If so, it may determine to generate a second pre-trained model. The predetermined unexpected event is an unexpected event related to the physical condition of subject 50, such as hospitalization or discharge. Chronic change detection unit 152 is an example of a determination unit.

[0115] When the chronic change detection unit 152 detects chronic changes in the physical condition of the subject 50, that is, when there is a difference between the first and second activity data, the model updating unit 153 causes the model generation unit 14 to generate a second pre-trained model trained using a third feature value based on the most recently acquired third activity data of the subject 50. The third activity data may include at least the second activity data. Furthermore, the third activity data may also include data from a portion of the first activity data. The portion of the first activity data is data from the closest side (new side) of the first period. For example, this portion of the first activity data can be used when the second period in which the second activity data was acquired is shorter than a prescribed period, that is, when the amount of second activity data is small. For example, the model updating unit 153 can determine whether the amount of second activity data is greater than a prescribed amount. If it is less than the prescribed amount, only a portion of the first activity data can be included in the third activity data. Alternatively, if the amount of second activity data is less than the prescribed amount, none of the first activity data can be included in the third activity data.

[0116] When the model generation unit 14 receives the model generation instruction from the model updating unit 153, it generates a new model by changing the training data used for model generation to a different one from the previous one. Specifically, it regenerates the model (second pre-trained model) using the third activity data.

[0117] Physical condition detection unit 16 is implemented, for example, by a computer equipped with a processor (microprocessor), memory, and a communication interface. The processor executes a control program stored in the memory to implement various functions. Physical condition detection unit 16 detects abnormalities in the physical condition of subject 50 using the model generated by model generation unit 14 and the multiple feature quantities calculated by feature quantity calculation unit 13.

[0118] The physical condition detection unit 16 includes an abnormality score calculation unit 161 , a staged score calculation unit 162 , a factor analysis unit 163 , and a calculation result recording unit 164 .

[0119] The abnormality score calculation unit 161 inputs the plurality of feature quantities calculated by the feature quantity calculation unit 13 into the model generated by the model generation unit 14 to obtain an abnormality score indicating the degree of abnormality of the physical condition for each predetermined period.

[0120] In this embodiment, the abnormality score calculation unit 161 outputs the plurality of feature values ​​calculated by the feature value calculation unit 13 for each hour of the target day for detecting the physical condition of the subject 50 to the supervised learning model generated by the model generation unit 14. The abnormality score calculation unit 161 records the calculated abnormality score for each hour of the target day for detecting the physical condition of the subject 50 in the calculation result recording unit 164.

[0121] The staged score calculation unit 162 calculates a staged score for expressing the degree of abnormality of the physical condition of the measurement subject 50 in stages based on the abnormality score calculated by the abnormality score calculation unit 161 .

[0122] In this embodiment, the staged score calculation unit 162 calculates the average abnormality score for the day based on the abnormality score for each hour of the target day of the physical condition examination recorded in the calculation result recording unit 164 or calculated by the abnormality score calculation unit 161. Similarly, the staged score calculation unit 162 calculates the average abnormality score for each of the previous and previous days based on the abnormality score for each hour of the previous and previous days of the target day of the physical condition examination recorded in the calculation result recording unit 164. The staged score calculation unit 162 sums the average abnormality score for the target day, the previous and previous days, and calculates a three-day total score. The three-day total score is an example of a method for calculating the staged score with high accuracy, but is not limited to this method. The calculation may also be performed within a range from a one-day total score to a five-day total score.

[0123] The staged score calculation unit 162 calculates a threshold value (also referred to as a staged threshold value) for the staged score based on the three-day aggregate score group for the past 90 days or so, recorded in the calculation result recording unit 164. More specifically, the staged score calculation unit 162 calculates the staged threshold value by calculating the average value and standard deviation of the three-day aggregate score group for the past 90 days or so.

[0124] The staged score calculation unit 162 outputs the calculated staged score value to the calculation result recording unit 164. Furthermore, if the staged score has five stages, for example, and the calculated staged score value is 1 to 3, the staged score calculation unit 162 may output the calculated staged score to the display terminal unit 30 via the communication network 40.

[0125] The processing of the staged score calculation unit 162 will be described later. Figure 5 To explain.

[0126] If the staged score is above a specified value, the factor analysis unit 163 performs factor analysis on each element included in the activity data to determine whether the element is a factor. Here, the elements include the subject's 50 food intake, respiratory rate, heart rate, body temperature, or bed rest rate during the specified period. Furthermore, the specified value is the value required to address the precursors of abnormal physical condition. For example, if the staged score has five stages, the specified value can be 4 or 5; if it has three stages, the specified value can be 3; if it has two stages, the specified value can be 2, and so on.

[0127] In this embodiment, when the staged score calculated by the staged score calculation unit 162 is 4 or 5, the factor analysis unit 163 performs factor analysis on the heart rate, respiratory rate, bed ambulation rate, food intake, and body temperature elements included in the activity data used to calculate the characteristic value. Alternatively, when the activity data used to calculate the characteristic value only includes heart rate and respiratory rate, factor analysis may be performed on the heart rate and respiratory rate elements.

[0128] For example, factor analysis unit 163 converts the multiple feature quantities of each factor over the entire period used to calculate the staged score into data on multiple feature quantities for each factor on each day, and calculates the average value and standard deviation for the entire period used to calculate the staged score. In this embodiment, factor analysis unit 163 converts the multiple feature quantities of each factor over three days into data on multiple feature quantities for each factor on each day, and calculates the average value and standard deviation for each factor over the three days.

[0129] Therefore, the factor analysis unit 163 analyzes that the element is not a factor when the following formula 2 holds true, and analyzes that the element is a factor when the following formula 2 does not hold true.

[0130] (Average value - 2*standard deviation) ≤ (feature value of the element at the target date and time) ≤ (Average value + 2*standard deviation) (Formula 2)

[0131] In addition, the property of the standard deviation used in Formula 2 is that 95.45% of all data are distributed within the range of the mean ± 2 times the standard deviation.

[0132] The factor analysis unit 163 outputs the elements analyzed as factors and the staged scores to the calculation result recording unit 164. The factor analysis unit 163 may also output the elements analyzed as factors and the staged scores to the display terminal unit 30 via the communication network 40.

[0133] The calculation result recording unit 164 is a recording medium capable of recording calculation results, and is comprised of, for example, a rewritable, nonvolatile memory such as a hard disk drive or solid-state drive. In this embodiment, the calculation result recording unit 164 records the anomaly score calculated by the anomaly score calculation unit 161 and the staged score calculated by the staged score calculation unit 162 as calculation results. Furthermore, the calculation result recording unit 164 may also record factors analyzed by the factor analysis unit 163 as calculation results.

[0134] The display terminal unit 30 is implemented as a computer equipped with a processor (microprocessor), memory, a communication interface, a user interface, and the like. The display terminal unit 30 is a terminal for a user 61, such as a monitor of the subject 50, and is, for example, a portable terminal device such as a tablet computer or a smartphone. The display terminal unit 30 may also be a mobile computer or a desktop computer connected to a display (a stationary terminal device).

[0135] In this embodiment, the display terminal unit 30 can be checked by a user 61, such as a monitor of the subject 50. The display terminal unit 30 is connected to the communication network 40. When the staged scores, etc., are obtained from the information management server 10, the display terminal unit 30 causes the user 61 to display information on a user interface for responding to abnormal physical conditions of the subject 50. The user interface can display information based on input from the user 61, etc.

[0136] [2. Operation of the information management server]

[0137] Next, refer to Figures 3 to 8 The operation of the information management server 10 having the above configuration will be described. Figure 3 The operation of generating a model (here, a supervised learning model) for detecting a precursor to abnormality in the physical condition of the subject 50 will be described.

[0138] Figure 3 This is a flowchart showing the operation of generating the first pre-trained model (a method of generating a pre-trained model) in the information management server 10 according to the present embodiment.

[0139] like Figure 3 As shown, the transceiver 11 obtains activity data for the first period from the sensor 20 through communication (S10). For example, the transceiver 11 obtains activity data every minute. The transceiver 11 obtains sensory data such as heart rate, respiratory rate, and body movement every minute. Alternatively, the transceiver 11 may obtain activity data for the first period all at once, or may obtain activity data each time the sensor 20 performs sensing.

[0140] Next, the feature quantity calculation unit 13 calculates feature quantities based on the acquired activity data (S20). The feature quantity calculation unit 13 calculates statistics for each hour of activity data. For example, the feature quantity calculation unit 13 can perform statistical processing on the heart rate and respiratory rate for each hour to calculate one or more statistics. Furthermore, the feature quantity calculation unit 13 can perform statistical processing on the results of whether or not the person was in bed for each hour to calculate one or more statistics including the bed leaving rate. The calculated statistics are examples of feature quantities.

[0141] Next, the feature quantity calculation unit 13 takes the statistics of activity data in every 24 hours as a group to generate multiple feature groups. The feature quantity calculation unit 13 takes the statistics of heart rate and / or respiratory rate in 24 hours and the in-bed / out-of-bed status as a group to perform training data for the data enhancement model.

[0142] Next, the model generation unit 14 generates a model using the calculated feature values ​​( S30 ). For example, in the case of supervised learning, the model generation unit 14 uses the feature values ​​as input data and the anomaly scores as label data (correct answer data) to train the model through machine learning.

[0143] Next, the model generation unit 14 outputs the generated model to the abnormality score calculation unit 161 ( S40 ) and may record the generated model in a recording unit (not shown) of the physical condition detection unit 16 .

[0144] Next, refer to Figure 4 as well as Figure 5 , the work of using the model generated as described above to detect the signs of abnormal physical condition of the subject 50 is explained.

[0145] Figure 4 This is a flowchart showing the operation of detecting a sign of abnormal physical condition of the measurement subject 50 in the information management server 10 according to this embodiment.

[0146] like Figure 4 As shown, the transmitter-receiver unit 11 acquires activity data (for example, activity data including respiratory rate and heart rate) of the measurement subject 50 during a predetermined period ( S110 ).

[0147] Next, the feature quantity calculation unit 13 calculates a feature quantity based on the activity data acquired in step S110 (S120). The feature quantity calculation unit 13 calculates, for example, a feature quantity (e.g., multiple feature quantities) for each hour. For example, the feature quantity calculation unit 13 can calculate multiple feature quantities for each hour of the target day for detecting the physical condition of the subject 50 based on the activity data including at least the respiratory rate and heart rate of the subject 50 acquired by the transceiver 11.

[0148] Next, the abnormality score calculation unit 161 inputs the feature values ​​calculated in step S120 into a pre-trained supervised learning model to obtain an abnormality score for each predetermined period (S130). For example, the abnormality score calculation unit 161 calculates an abnormality score for each hour based on a plurality of feature values ​​for each hour. The abnormality score calculation unit 161 inputs the feature values ​​calculated by the feature value calculation unit 13 into the model generated by the model generation unit 14 to obtain an abnormality score indicating the degree of abnormality in the physical condition for each hour of the predetermined period, including the target day.

[0149] Next, the staged score calculation unit 162 calculates a staged score for expressing the degree of abnormality in the physical condition of the subject 50 in stages based on the abnormality score acquired in step S130 (S140). For example, the staged score calculation unit 162 calculates the average daily average of the abnormality score based on the abnormality score for each hour. In this embodiment, the staged score calculation unit 162 calculates the average daily average of the abnormality score based on the abnormality score for each hour during a predetermined period, including the target day for the physical condition detection of the subject 50.

[0150] Then, the staged score calculation unit 162 calculates the average value and standard deviation of the past abnormality scores (for example, the abnormality scores of the past 90 days or so) for the target day to calculate the staged threshold value.

[0151] Figure 5 An example of five-stage staged scoring and conditions thereof according to this embodiment is shown.

[0152] like Figure 5 As shown in FIG, the staged score calculation unit 162 can calculate the threshold value based on the average value and the standard deviation when calculating the staged scores of five stages. Figure 5 The conditions shown are, for example, that the threshold value for a staged score of 1 is equal to or greater than the average value, and that the threshold value for a staged score of 2 is equal to the value obtained by subtracting half the standard deviation from the average value and the average value.

[0153] Then, the staged score calculation unit 162 calculates the staged score by applying the calculated staged threshold to the average value of the abnormality score of the target day. More specifically, the staged score calculation unit 162 calculates the staged score by using the Figure 5 The average value of the abnormality score of the target day is determined based on the threshold value calculated according to the conditions shown, thereby calculating the value of the staged score.

[0154] Alternatively, the staged score calculation unit 162 may further confirm whether the staged score calculated in step S140 indicates a value of 4 or 5, that is, whether a value indicating an abnormal physical condition is calculated. Therefore, if the staged score is 4 or 5, the factor analysis unit 163 may perform factor analysis on each element of the activity data used to calculate the characteristic value, including heart rate, respiratory rate, bed ambulation rate, body temperature, and food intake.

[0155] Next, the staged score calculation unit 162 outputs the staged score calculated in step S140 (S150). If the factor analysis performed by the factor analysis unit 163 finds an element that is a factor, the staged score calculation unit 162 may output the element analyzed as a factor and the staged score in step S150.

[0156] Next, refer to Figures 6 to 8 The work of updating the model generated above will be described.

[0157] Figure 6 This is a flowchart showing the operation of generating the second pre-trained model (a method of generating a pre-trained model) in the information management server 10 according to the present embodiment. Figure 7 This is a timing chart schematically showing the operation of generating the second pre-trained model in the information management server 10 according to the present embodiment.

[0158] Figure 6 The work shown in Figure 3 After the step S40 shown is executed or Figure 4 The detection work of the precursor of abnormal physical condition shown is performed during execution. Figure 6 The operations shown may be performed at a predetermined period or when a predetermined unexpected event occurs in the subject 50. Furthermore, the predetermined period may be, for example, the period required to collect enough activity data to determine whether a chronic change in the physical condition of the subject 50 has occurred. The predetermined period is pre-set and recorded in the recording unit (not shown) of the information management server 10.

[0159] Figure 7 The "originally used model" shown here means the currently used model (first model, first learning model). This is described as the "originally used model" assuming that the model is updated.

[0160] like Figure 6 as well as Figure 7As shown, the data aggregation unit 151 of the model update determination unit 15 obtains activity data for a second period following the first period from the information recording unit 12 or the feature quantity calculation unit 13 (S210). The second period may include the period after the start of anomaly detection using the generated first model. Furthermore, the data aggregation unit 151 may also obtain activity data for the first period. Furthermore, the first and second periods may not overlap in time, for example, but may partially overlap.

[0161] Next, the chronic change detection unit 152 determines whether there is a difference between the activity data of the first period and the activity data of the second period (S220). The chronic change detection unit 152 makes the determination in step S220 using, for example, the above-mentioned formula 1. The determination in step S220 is performed every predetermined period in step S210.

[0162] Next, if the chronic change detection unit 152 determines that there is a difference ("YES" in S220), the model updating unit 153 determines to update the model (S230). Since the subject 50 may have experienced chronic physical changes, the model updating unit 153 determines to regenerate a model suitable for the state in which chronic physical changes have occurred.

[0163] Next, the model updating unit 153 instructs the model generation unit 14 to reconstruct the model based on the activity data from the third period (S240). The model updating unit 153 instructs the model generation unit 14 to regenerate the model using the activity data from the third period, i.e., the most recent activity data. The third period is, for example, the period used to obtain activity data for model reconstruction. It can be the same period as the second period, a portion of the second period, or a period that includes the second period and a portion of the first period.

[0164] When the model generation unit 14 receives an instruction from the model update unit 153, it uses the recently acquired activity data to regenerate a model corresponding to the chronic physical state change (an example of the second pre-trained model, Figure 7 ). The regenerated model is output to the physical condition detection unit 16. Then, in the physical condition detection unit 16, the pre-trained model used in the output of the abnormality score of the subject 50 is switched from the first pre-trained model to the second pre-trained model. Accordingly, it can be expected that the accuracy of detecting the precursor of the abnormal physical condition of the subject 50 will be improved in the case where the subject 50 has undergone chronic changes in physical condition. That is, through the information management server 10, the precursor of the abnormal physical condition of the subject 50 in the case where the subject 50 has undergone chronic changes in physical condition can be detected with higher accuracy.

[0165] If the chronic change detection unit 152 determines that there is no difference ("No" in S220), the physical state detection unit 16 ends the model update process. If there is no difference between the first and second activity data, that is, if it is assumed that the subject 50 has not experienced chronic physical state changes, the first pre-trained model (first model) continues to be used.

[0166] Here, refer to Figure 8 The success rate of detecting abnormalities of the subject 50, which is determined by whether or not to update the model, is described. Figure 8 The results of comparing the detection success rate with and without switching based on the pre-trained model are shown. Figure 8 In the figure, the model is updated without Figure 8 "Sensor only one month" in the ) and with model updates ( Figure 8 "Only the sensor is switched for one month"), as an abnormal physical state, the detection success rate of low SPO2 (Saturation Pulse O2: transcutaneous arterial oxygen saturation) and fever are compared. No model update shows a previous example in which the model is not updated although a chronic physical state change occurs in the subject 50, and shows a case in which the model before the update is continued to be used in the case of model update despite the chronic physical state change occurs in the subject 50. Model update shows the method of the present invention, that is, the model is updated because a chronic physical state change occurs in the subject 50. In addition, Figure 8 The vertical axis shows the detection success rate (0 to 1).

[0167] like Figure 8 As shown, it can be seen that regardless of whether the SPO2 value is low or the patient has a fever, the detection success rate is improved when the model is updated.

[0168] (Other Embodiments)

[0169] Although the above description of the generation method of the pre-trained model involved in one or more embodiments is based on the embodiment, the present disclosure is not limited to these embodiments. The present disclosure may also include forms obtained by applying various modifications conceived by those skilled in the art to the present embodiment without departing from the scope of the present disclosure, and forms constructed by combining constituent elements in different embodiments.

[0170] For example, in the above-described embodiments, each component may be formed by dedicated hardware or implemented by executing a software program suitable for each component. Each component may also be implemented by a program execution unit such as a CPU or a processor reading and executing a software program recorded on a recording medium such as a hard disk or a semiconductor memory.

[0171] Furthermore, the order in which the steps in the flowchart are executed is an example given for the purpose of specifically illustrating the present disclosure, and may be an order other than the above. Furthermore, some of the steps may be executed simultaneously (in parallel) with other steps, or some of the steps may not be executed.

[0172] The functional block division in the block diagram is merely an example. Multiple functional blocks can be implemented as a single functional block, a single functional block can be divided into multiple blocks, or some functions can be transferred to other functional blocks. Furthermore, the functions of multiple functional blocks with similar functions can be processed in parallel or in a time-sharing manner by a single hardware or software.

[0173] Furthermore, the information management server involved in the above-mentioned embodiments and the like may be implemented by a single device or by multiple devices. When the information management server is implemented by multiple devices, the components of the information management server may be arbitrarily distributed across the multiple devices. When the information management server is implemented by multiple devices, the communication method between the multiple devices is not particularly limited and may be wireless or wired communication. Furthermore, communication between the devices may be a combination of wireless and wired communication.

[0174] Furthermore, each component described in the above embodiments and the like can be implemented as software, typically as an LSI (Large Scale Integrated Circuit). These can be individually made into a single chip, or a portion or all of them can be made into a single chip. Although referred to herein as LSI, it may also be referred to as IC, system LSI, super LSI, or ultra-large-scale integrated circuit, depending on the degree of integration. Furthermore, the method of integrated circuitization is not limited to LSI, and can also be implemented by a dedicated circuit (a general-purpose circuit that executes a dedicated program) or a general-purpose processor. An FPGA (Field Programmable Gate Array) that can be programmed after LSI manufacturing or a reconfigurable processor that can reconfigure the connections or settings of circuit cells within the LSI can also be utilized. Furthermore, with the advancement of semiconductor technology or the development of other technologies, if an integrated circuit technology that replaces LSI emerges, such technology can of course also be used to integrate the components.

[0175] A system LSI is a highly multifunctional LSI that integrates multiple processing units on a single chip. Specifically, it is a computer system composed of a microprocessor, ROM (Read Only Memory), and RAM (Random Access Memory). The ROM stores computer programs. The system LSI realizes its functions by operating the microprocessor according to the computer program.

[0176] Furthermore, one embodiment of the present disclosure may be to enable a computer to execute Figure 3 、 Figure 4 、 Figure 6 、 Figure 7 A computer program comprising the characteristic steps of any of the methods for generating a pre-trained model shown.

[0177] Furthermore, for example, the program may be a program for causing a computer to execute. Furthermore, one embodiment of the present disclosure may be a computer-readable, non-transitory recording medium recording these programs. For example, these programs may be recorded on a recording medium and distributed or circulated. For example, the distributed program may be installed on a device having another processor and caused to execute the program, thereby enabling the device to perform the aforementioned processes.

[0178] Industrial applicability

[0179] The present disclosure provides a method for generating a pre-trained model that can be utilized to generate a pre-trained model capable of supporting the detection of small changes in a subject's physical condition that are associated with abnormalities in the subject's physical condition.

[0180] Description of Reference Numerals

[0181] 10 Information Management Server

[0182] 11. Transceiver Department

[0183] 12 Information Recording Department

[0184] 13. Feature Calculation Unit

[0185] 14 Model Generation Department

[0186] 15 Model update judgment unit

[0187] 16. Physical condition detection unit

[0188] 20 Sensor

[0189] 25 Recording Data

[0190] 30 Display terminal

[0191] 40 Communication Network

[0192] 50 subjects

[0193] 60, 61 users

[0194] 100 Body Condition Detection System

[0195] 151 Data Aggregation Department (Acquisition Department)

[0196] 152 Chronic Change Detection Unit (Judgment Unit)

[0197] 153 Model Update Department

[0198] 161 Abnormality Score Calculation Department

[0199] 162 Staged Score Calculation Department

[0200] 163 Factor Analysis Department

[0201] 164 Calculation result recording unit

Claims

1. A method for generating a pre-trained model, wherein the pre-trained model takes as input a feature quantity based on activity data of a subject and outputs an abnormality score indicating the degree of abnormality of the subject's physical condition, The first pre-trained model is trained using the first activity data of the subject acquired during the first past period. In the method for generating the pre-training model, acquiring the first activity data of the subject and the second activity data of the subject acquired during a second period after the first period, determining whether the first activity data and the second activity data are different; When there is a difference between the first motion data and the second motion data, a second pre-trained model is generated using the most recently acquired third motion data of the subject.

2. The method for generating a pre-training model according to claim 1, The third activity data includes the second activity data.

3. The method for generating a pre-training model according to claim 2, The third activity data also includes data of a portion of the first activity data.

4. The method for generating a pre-training model according to any one of claims 1 to 3, The differences include differences indicating that the subject has chronic changes in physical condition.

5. The method for generating a pre-training model according to any one of claims 1 to 3, Whether the first activity data and the second activity data have the difference is determined by using the statistics of the first activity data and the statistics of the second activity data.

6. The method for generating a pre-training model according to any one of claims 1 to 3, When the standard deviation of the first activity data is set to sd B , the standard deviation of the third activity data is set to sd A When the following formula 1 is satisfied, it is determined that there is the difference. (|sd A -sd B | / sd A )×100≥threshold value (Formula 1).

7. The method for generating a pre-training model according to any one of claims 1 to 3, The determination of whether the first activity data and the second activity data have the difference is performed every predetermined period.

8. The method for generating a pre-training model according to any one of claims 1 to 3, Furthermore, when a predetermined unexpected event occurs to the subject, the second pre-trained model is generated.

9. The method for generating a pre-training model according to any one of claims 1 to 3, When the second pre-trained model is generated, the pre-trained model used to output the abnormality score for the subject is switched from the first pre-trained model to the second pre-trained model.

10. The method for generating a pre-training model according to any one of claims 1 to 3, When there is no difference between the first activity data and the second activity data, the first pre-trained model is continuously used.

11. A device for generating a pre-trained model, the pre-trained model taking as input a feature quantity based on activity data of a subject and outputting an abnormality score indicating the degree of abnormality of the subject's physical condition, The first pre-trained model is trained using the first activity data of the subject acquired during the first past period. The generating device of the pre-training model comprises: an acquisition unit that acquires the first activity data of the subject and second activity data of the subject acquired during a second period after the first period; a determination unit configured to determine whether the first activity data and the second activity data are different; as well as The model updating unit generates a second pre-trained model using the most recently acquired third activity data of the subject when there is a difference between the first activity data and the second activity data.

12. A program for causing a computer to execute the method for generating a pre-training model according to any one of claims 1 to 3.

Images