Fall risk estimation device, fall risk estimation method, fall risk estimation program, artificial intelligence learning device, artificial intelligence learning method, and artificial intelligence learning program

Abstract

Using sleep sensor data, the risk of falls for those receiving care can be estimated with high accuracy. [Solution] The fall risk estimation device according to this disclosure inputs input data, including sleep data and basic data of a protected person, to a trained artificial intelligence, causing the artificial intelligence to calculate estimated data of the protected person's daily living activity data, including at least one item of transfer and mobility ability, problem-solving ability, and memory ability. The fall risk estimation device further divides each feature into multiple categories for a set of feature quantities, which includes sleep data including the distribution of nighttime and daytime sleep time and at least one item of daytime sleep time, basic data including at least one item of BMI, and data based on estimated data of daily living activity data. A predetermined evaluation value is assigned to each of the divided categories according to its correlation with falls, and the fall risk is calculated by summarizing the evaluation values ​​for the entire set of feature quantities.

Description

Technical Field

[0001] The present invention relates to a fall risk estimation device, a fall risk estimation method, a fall risk estimation program, an artificial intelligence learning device, an artificial intelligence learning method, and an artificial intelligence learning program for estimating the fall risk of protected persons such as patients and care recipients.

Background Art

[0002] The inventor of the present application has completed an invention for estimating the daily living activity data of a subject by using artificial intelligence based on the sleep data and basic data of the subject, which are objective data independent of an evaluator, in order to obtain an objective work therapy evaluation independent of an evaluator for the subject of work therapy evaluation (see Patent Document 1). Sleep data is data related to the sleep of the subject, such as respiration, pulse, sleep time, and sleep rhythm, and these data are acquired by a sleep sensor. Basic data is data related to the body of the subject, such as the age, gender, height, and weight of the subject, and is acquired through a doctor's diagnosis or the like. The daily living activity data is data related to the daily living activities of the subject, represented by the evaluation items defined in the Functional Independence Measure (FIM).

[0003] Patent Document 1 also reveals that, as one item of the daily living activity data, in addition to FIM, the fall risk of the subject can be estimated (paragraphs 0015, 0044, etc. of the specification of the same document; FIG. 3(c)). It can be said that the importance of routinely and simply evaluating the fall risk of protected persons such as the elderly has been increasing with the deepening of the aging society. In view of such a situation, the inventor of the present application has been searching for means to more accurately estimate the fall risk of protected persons while utilizing the patented invention according to Patent Document 1.

[0004] While Patent Documents 2-13 disclose techniques for estimating fall risk, their disclosures are as follows, and none of them utilize sleep sensors. Patent Document 2 uses a head-mounted display instead of a sleep sensor. Patent Document 3 uses a "gait measurement device 10" incorporating an acceleration sensor and an angular velocity sensor instead of a sleep sensor. Patent Document 4 uses a sensor related to "door opening and closing" instead of a sleep sensor. Patent Document 5 uses "walking sensing" and does not mention sleep data. Patent Document 6 estimates fall risk based on data from motion sensors attached to the waist and chest / back. It does not mention sleep sensors.

[0005] Patent Document 7 estimates the risk of falls based on data from incontinence detection pads, menstrual monitors, etc. There is no mention of sleep sensors. Patent Document 8 is identical to Patent Document 7. Patent Document 9 detects the fact of a fall, not the possibility of a fall. Sleep data is not used. Patent Document 10 analyzes the risk of falls based on questions posed to the user and video data. There is no mention of the use of sleep data. Patent Document 11 determines the risk, including falls, based on the user's answers to a questionnaire. There is no mention of the use of sleep data. Patent Document 12 detects the patient's condition based on microphones, load sensors, etc., and determines the risk of falls, etc. There is no mention of the use of sleep data. Patent Document 13 determines the risk of falls based on the subject's walking data measured by a gait measurement unit. There is no mention of the use of sleep data. [Prior art documents] [Patent Documents]

[0006] [Patent Document 1] Patent No. 6994262 [Patent Document 2] Patent No. 7540814 [Patent Document 3] Japanese Patent Publication No. 2024-101908 [Patent Document 4] Japanese Patent Publication No. 2024-047967 [Patent Document 5] Japanese Patent Publication No. 2023-155681 [Patent Document 6] Patent No. 7546904 [Patent Document 7] Patent No. 6896915 [Patent Document 8] Patent No. 6704076 [Patent Document 9] Patent No. 6739559 [Patent Document 10] Special Publication No. 2021-514087 [Patent Document 11] Japanese Patent Publication No. 2005-228305 [Patent Document 12] Japanese Patent Publication No. 2005-092440 [Patent Document 13] Japanese Patent Publication No. 2003-216743 [Overview of the project] [Problems that the invention aims to solve]

[0007] The present invention, resulting from the aforementioned exploration by the inventors, aims to provide a fall risk estimation device, a fall risk estimation method, and a fall risk estimation program that estimate the fall risk of a protected person with high accuracy using sleep sensor data. The present invention further aims to provide an artificial intelligence learning device, an artificial intelligence learning method, and an artificial intelligence learning program. [Means for solving the problem]

[0008] To achieve the above objective, a first aspect of the present invention is a fall risk estimation device comprising a first estimation unit, a fall risk calculation unit, and a fall risk output unit. The first estimation unit inputs input data, including first sleep data which is data relating to the protected person's sleep and first basic data which is data relating to the protected person's body, to a trained first artificial intelligence, causing the first artificial intelligence to calculate estimated data of daily living activity data which represents the level of the protected person's ability to perform daily living activities. The fall risk calculation unit calculates estimated fall risk data by referring to a data set including second sleep data which is data relating to the protected person's sleep and the estimated data of daily living activity data. The fall risk output unit outputs the estimated fall risk data calculated by the fall risk estimation unit. The first and second sleep data are data based on data measured by a sleep sensor, and each includes data expressed numerically. The first basic data also includes data expressed numerically. Furthermore, the daily living activity data includes data expressed numerically in steps.

[0009] With this configuration, the risk of falls is estimated by referring to a data set that includes sleep data and estimated data on daily living activities, thus providing highly accurate fall risk estimation results. Furthermore, since daily sleep data is used to obtain estimated data on daily daily living activities, and then the daily fall risk is estimated, it is possible to easily obtain fall risk estimation results that may change from day to day. In addition, since the input data and estimated data include data expressed numerically or data expressed numerically in steps, it is easy for artificial intelligence to calculate the estimated data. The trained artificial intelligence may be part of the fall risk estimation device in this configuration, or it may be an external device, such as one located on an external cloud server. The sleep data that the first estimation unit inputs to the first artificial intelligence and the sleep data that the fall risk estimation unit refers to do not necessarily have to be identical in content.

[0010] A second aspect of the present invention is a fall risk estimation device according to the first aspect, wherein the second sleep data referenced by the fall risk estimation unit includes data relating to sleep quality. With this configuration, the fall risk estimation unit refers to data on sleep quality, which has a high correlation with fall risk, as sleep data, thus obtaining fall risk estimation results with higher accuracy.

[0011] A third aspect of the present invention is a fall risk estimation device according to the second aspect, wherein the data relating to sleep quality includes data on at least one item from the distribution of nighttime and daytime sleep time and daytime sleep time. With this configuration, data with a high correlation to fall risk is used to estimate fall risk, resulting in more accurate fall risk estimation results.

[0012] A fourth aspect of the present invention is a fall risk estimation device according to any of the first to third aspects, wherein the set of data referenced by the fall risk estimation unit further includes second basic data which is data relating to the physical condition of the person being protected. Furthermore, the second basic data includes data expressed numerically. With this configuration, the fall risk estimation unit estimates the fall risk by referring to basic data in addition to estimated data from sleep data and daily living activity data, thus obtaining fall risk estimation results with higher accuracy. It should be noted that the basic data input by the first estimation unit to the first artificial intelligence and the basic data referenced by the fall risk estimation unit do not necessarily have to be identical in content.

[0013] A fifth aspect of the present invention is a fall risk estimation device according to the fourth aspect, wherein the second basic data included in the data set includes BMI (Body Mass Index). With this configuration, data with a high correlation to fall risk is used to estimate fall risk, resulting in more accurate fall risk estimation results.

[0014] A sixth aspect of the present invention is a fall risk estimation device according to any one of the first to fifth aspects, wherein the first estimation unit causes the first artificial intelligence to calculate estimation data including at least one item of transfer and movement ability, problem-solving ability, and memory ability as the daily living operation data. According to this configuration, data having a high correlation with the fall risk is used for estimating the fall risk, so that an estimation result of the fall risk can be obtained with higher accuracy.

[0015] A seventh aspect of the present invention is a fall risk estimation device according to any one of the first to sixth aspects, wherein the set of data referred to by the fall risk estimation unit further includes data of changes in the estimation data in a predetermined past period of the daily living operation data. According to this configuration, for example, the fall risk estimation unit calculates data of changes from the previous day of the estimation data of the daily living operation data or data of changes in the estimation data over the past several days, and estimates the fall risk by also referring to the calculated data. Therefore, an estimation result of the fall risk can be obtained with higher accuracy. The inventors of the present application have confirmed that such changes show a strong correlation with the fall risk.

[0016] An eighth aspect of the present invention is a fall risk estimation device according to any one of the first to sixth aspects, wherein the fall risk estimation unit includes a second estimation unit that inputs data including time-series data in a predetermined past period of the second sleep data to a learned second artificial intelligence, and causes the second artificial intelligence to calculate estimation data of changes in the daily living operation data in the predetermined past period. The set of data referred to by the fall risk estimation unit further includes the estimated data of the change. According to this configuration, data that captures finer changes in the estimated data of changes in the daily living operation data in a predetermined past period, such as the current day and the previous day or the past few days, can be obtained. The learned artificial intelligence may be a part of the fall risk estimation device of this configuration, or may be an external device, such as one placed in an external cloud server.

[0017] According to the ninth aspect of the present invention, there is provided a fall risk estimation device according to any one of the first to eighth aspects, wherein the fall risk estimation unit includes a third estimation unit that causes the learned third artificial intelligence to calculate estimation data of the fall risk by inputting the set of data to be referred to. According to this configuration, since the fall risk is estimated using the learned artificial intelligence, as learning accumulates, an estimation result of the fall risk can be obtained with higher accuracy. The learned artificial intelligence may be a part of the fall risk estimation device of this configuration, or may be an external device, such as one placed in an external cloud server.

[0018] According to the tenth aspect of the present invention, there is provided a fall risk estimation device according to any one of the first to seventh aspects, wherein the fall risk estimation unit divides each of the sets of data to be referred to into a plurality of ways, assigns a predetermined evaluation value corresponding to the correlation with a fall to each of the divided sections, and calculates estimation data of the fall risk by synthesizing the evaluation values for the entire set of data. According to this configuration, an evaluation value corresponding to the correlation with the fall risk is assigned, and by synthesizing them, the fall risk is estimated. Therefore, an estimation result of the fall risk can be obtained with high accuracy without using an artificial intelligence. Synthesizing the evaluation values corresponds to, for example, calculating the sum or average of the evaluation values.

[0019] According to the eleventh aspect of the present invention, there is provided a fall risk estimation device according to any one of the first to tenth aspects, further including a bed leaving signal receiving unit and an alert output unit. The bed leaving signal receiving unit receives a signal from a bed leaving sensor that detects the bed leaving of the protected person. The alert output unit outputs a fall alert when the bed leaving signal receiving unit receives the signal and the estimation data of the fall risk calculated by the fall risk calculation unit exceeds a predetermined reference value. With this configuration, caregivers can be notified when a person under care who is at high risk of falling gets out of bed, and can immediately take protective or assistive action to prevent falls. The bed exit sensor may also be used in conjunction with a sleep sensor.

[0020] A twelfth aspect of the present invention is a fall risk estimation method comprising a first estimation process, a fall risk calculation process, and a fall risk output process. The first estimation process inputs input data, including first sleep data which is data relating to the protected person's sleep and first basic data which is data relating to the protected person's physical condition, to a trained first artificial intelligence, causing the first artificial intelligence to calculate estimated data of daily living activity data which represents the level of the protected person's ability to perform daily living activities. The fall risk calculation process calculates estimated fall risk data by referring to a data set including second sleep data which is data relating to the protected person's sleep and the estimated data of daily living activity data. The fall risk output process outputs the estimated fall risk data calculated by the fall risk estimation process. The first and second sleep data are data based on data measured by a sleep sensor, and each includes data represented numerically. The first basic data also includes data represented numerically. Furthermore, the daily living activity data includes data represented numerically in stages.

[0021] With this configuration, the risk of falls is estimated by referring to a data set that includes sleep data and estimated data on daily living activities, thus obtaining highly accurate fall risk estimation results. Furthermore, since daily sleep data is used to obtain estimated data on daily daily living activities, and then the daily fall risk is estimated, it is possible to easily obtain fall risk estimation results that may change from day to day. In addition, since the input data and estimated data include data expressed numerically or data expressed numerically in steps, it is easy for artificial intelligence to calculate the estimated data. It should be noted that the sleep data that the first estimation unit inputs to the first artificial intelligence and the sleep data that the fall risk estimation unit refers to do not necessarily have to be identical in content.

[0022] A thirteenth aspect of the present invention is a fall risk estimation method according to the twelfth aspect, further comprising a bed exit signal reception process and an alert output process. The bed exit signal reception process receives a signal from a bed exit sensor that detects when the protected person leaves the bed. The alert output process outputs a fall alert when the bed exit signal reception process receives the signal, in which case the fall risk estimated by the fall risk calculation process exceeds a predetermined threshold value. With this configuration, caregivers and other personnel can become aware when a person under care who is at high risk of falling gets out of bed, and can immediately take protective and assistive measures to prevent falls.

[0023] A fourteenth aspect of the present invention is a fall risk estimation program, which, when read by a computer, causes the computer to function as a fall risk estimation device according to any of the first to eleventh aspects. By having a computer read a program with this configuration, a fall risk estimation device according to one of the first to eleventh embodiments can be realized by the computer.

[0024] A fifteenth aspect of the present invention is an artificial intelligence learning device for training an artificial intelligence that functions as the second artificial intelligence used in the fall risk estimation device according to the eighth aspect, comprising an input data receiving unit, a teacher data receiving unit, and a learning unit. The input data receiving unit receives data as input data, which includes time-series data of sleep data over a predetermined past period, which is data relating to the sleep of a protected person. The teacher data receiving unit receives teacher data corresponding to the input data, which includes data showing the actual changes in daily living activity data for the protected person over a predetermined past period. The learning unit inputs the input data received by the input data receiving unit and the teacher data received by the teacher data receiving unit to the artificial intelligence, thereby training the artificial intelligence to estimate the teacher data from the input data. The sleep data includes numerically represented data based on data measured by a sleep sensor. The daily living activity data includes numerically represented data in steps. According to this configuration, the second artificial intelligence used by the fall risk estimation device according to the eighth aspect is constructed through learning. The second artificial intelligence may be part of the artificial intelligence learning device according to this configuration, or it may be an external device, such as one located on an external cloud server.

[0025] A sixteenth aspect of the present invention is an artificial intelligence learning device for training an artificial intelligence that functions as the third artificial intelligence used in the fall risk estimation device according to the ninth aspect, comprising an input data receiving unit, a teacher data receiving unit, and a learning unit. The input data receiving unit receives a set of data as input data, which includes sleep data, which is data relating to the sleep of a protected person, and daily living activity data, which represents the level of the protected person's ability to perform daily living activities. The teacher data receiving unit receives teacher data, which corresponds to the input data and includes data indicating the risk of falling for the protected person. The learning unit inputs the input data received by the input data receiving unit and the teacher data received by the teacher data receiving unit to the artificial intelligence, thereby training the artificial intelligence to estimate the teacher data from the input data. The sleep data includes numerically represented data based on data measured by a sleep sensor. The daily living activity data includes numerically represented data in steps. According to this configuration, the third artificial intelligence used by the fall risk estimation device according to the ninth aspect is constructed through learning. The third artificial intelligence may be part of the artificial intelligence learning device according to this configuration, or it may be an external device, such as one located on an external cloud server.

[0026] A 17th aspect of the present invention is an artificial intelligence learning method for training an artificial intelligence that functions as the second artificial intelligence used in a fall risk estimation device according to the 8th aspect, comprising an input data reception process, a teacher data reception process, and a learning process. The input data reception process receives data as input data, which includes time-series data of sleep data over a predetermined past period, which is data relating to the sleep of a protected person. The teacher data reception process receives teacher data corresponding to the input data, which includes data showing the actual changes in daily living activity data for the protected person over a predetermined past period. The learning process inputs the input data received by the input data reception process and the teacher data received by the teacher data reception process to the artificial intelligence, thereby training the artificial intelligence to estimate the teacher data from the input data. The sleep data includes numerically represented data based on data measured by a sleep sensor. The daily living activity data includes numerically represented data in steps. According to this configuration, the second artificial intelligence used by the fall risk estimation device according to the eighth aspect is constructed through learning.

[0027] An 18th aspect of the present invention is an artificial intelligence learning method for training an artificial intelligence that functions as the third artificial intelligence used in a fall risk estimation device according to the 9th aspect, comprising an input data reception process, a teacher data reception process, and a learning process. The input data reception process receives a set of data as input data, which includes sleep data, which is data relating to the sleep of a protected person, and daily living activity data, which represents the level of the protected person's ability to perform daily living activities. The teacher data reception process receives teacher data, which corresponds to the input data and includes data indicating the risk of falling for the protected person. The learning process inputs the input data received by the input data reception process and the teacher data received by the teacher data reception process to the artificial intelligence, thereby training the artificial intelligence to estimate the teacher data from the input data. The sleep data includes numerically represented data based on data measured by a sleep sensor. The daily living activity data includes numerically represented data in steps. According to this configuration, the third artificial intelligence used by the fall risk estimation device according to the ninth aspect is constructed through learning.

[0028] A 19th aspect of the present invention is an artificial intelligence learning program which, when read by a computer, causes the computer to function as an artificial intelligence learning device according to the 15th or 16th aspect. By having a computer read a program with this configuration, an artificial intelligence learning device according to the 15th or 16th embodiment is realized by the computer. [Effects of the Invention]

[0029] As described above, the present invention realizes a fall risk estimation device, a fall risk estimation method, and a fall risk estimation program that estimate the fall risk of a protected person with high accuracy using data from a sleep sensor. Furthermore, the present invention realizes an artificial intelligence learning device, an artificial intelligence learning method, and an artificial intelligence learning program. [Brief explanation of the drawing]

[0030] [Figure 1] This is a schematic diagram illustrating the overview of data processing by a fall risk estimation device according to one embodiment of the present invention. [Figure 2] This figure illustrates the configuration of a fall risk estimation system, including a fall risk estimation device that performs the data processing shown in Figure 1. [Figure 3] Figure 2 is a block diagram illustrating the configuration of a fall risk estimation device. [Figure 4] Figure 3 is a flowchart illustrating the processing flow of the fall risk estimation method implemented by the fall risk estimation device shown as an example. [Figure 5] Figure 3 is a flowchart illustrating the processing flow of the artificial intelligence learning method implemented by the fall risk estimation device shown as an example. [Figure 6] Figure 3 is a tabular diagram illustrating the input and output data of the artificial intelligence used by the fall risk estimation device shown as an example. [Figure 7] Figure 3 is a schematic diagram illustrating the conceptual structure of the artificial intelligence used in the fall risk estimation device shown as an example. [Figure 8] This is a screenshot illustrating an example of an image displayed on the screen of computer 10. [Figure 9] This is a schematic diagram illustrating the overview of data processing by a fall risk estimation device according to another embodiment of the present invention. [Figure 10] This is a block diagram illustrating the configuration of the fall risk estimation device that performs the data processing shown in Figure 9. [Figure 11] This is a schematic diagram illustrating the overview of data processing by a fall risk estimation device according to yet another embodiment of the present invention. [Figure 12] This is a schematic diagram illustrating the overview of data processing by a fall risk estimation device according to yet another embodiment of the present invention. [Figure 13] This is a block diagram illustrating the configuration of the fall risk estimation device that performs the data processing shown in Figure 12. [Modes for carrying out the invention]

[0031] (1. Outline of one embodiment) Figure 1 is a schematic diagram illustrating the data processing performed by a fall risk estimation device according to one embodiment of the present invention. The fall risk estimation device operates roughly as follows: Based on sleep data and basic data of a person under care, such as a patient or care recipient, artificial intelligence (AI) estimates the person's daily living activity data. Daily living activity data is data that numerically represents the level of the person's ability to perform daily living activities, and the AI ​​calculates estimated data that includes at least one item from transfer and mobility ability, problem-solving ability, and memory ability. Sleep data is acquired by a sleep sensor and includes, for example, heart rate and respiratory rate. Basic data includes, for example, age, sex, and BMI (Body Mass Index). The operation of this AI is disclosed by the applicant in Patent Document 1.

[0032] The risk of falling is calculated by comprehensively evaluating a set of features, which includes sleep data (at least one item including the distribution of nighttime and daytime sleep time and daytime sleep time), basic data (at least one item including BMI), and estimated data calculated by AI, according to a predetermined procedure. The calculated risk of falling is output, for example, to a display screen. When the calculated risk of falling exceeds a predetermined threshold, a fall alert is issued when a detection signal is received from a bed exit sensor that detects when the protected person leaves their bed. The fall alert is issued, for example, by image, light, or sound. "Leaving bed" means getting up or lifting half of the body off the bed.

[0033] (2. Fall risk estimation device / system according to one embodiment) Figure 2 illustrates the configuration of a fall risk estimation system, including a fall risk estimation device that performs the data processing shown in Figure 1. In addition to the fall risk estimation device 101, this fall risk estimation system 100 includes a sleep sensor 1, a network 5, and servers 7 and 9. The sleep sensor 1, network 5, and servers 7 and 9 are connected to the fall risk estimation device 101 and cooperate with it.

[0034] The fall risk estimation device 101 is a device that contributes to the appropriate protection of the protected person 11, such as a patient or care recipient, by outputting an estimated fall risk based on the patient's sleep data and basic data. In the illustrated example, the fall risk estimation device 101 is incorporated into the computer 10. That is, by installing and starting a specific application on the computer 10, the processing unit (processor), such as the central processing unit (CPU), of the computer 10 functions as the fall risk estimation device 101.

[0035] In the illustrated example, the protected persons 11 are assumed to be a large number of people living in a nursing home or other care facility, and the computer 10 is assumed to be managed by the care facility. In the illustrated example, the computer 10 is assumed to be a personal computer (PC), but it could also be a small portable device such as a smartphone. Alternatively, the functions of the computer 10 could be shared between the PC and the smartphone.

[0036] The sleep sensor 1 is a sensor that automatically acquires sleep data from the person being cared for 11 and has a communication function to transmit the acquired data wirelessly or by other means. The sleep data includes data related to the person's sleep, such as sleep duration, number of times turning over in bed, breathing while sleeping, and pulse rate. The sleep sensor 1 also functions as a bed exit sensor to detect when the person being cared for 11 gets out of bed. In the illustrated example, the sleep sensor 1 is a mat-shaped sensor used by placing it under the bedding where the person being cared for 11 lies down. Sleep sensors 1 of this form are already commercially available and well known. The individual items of sleep data will be described later.

[0037] The sleep sensor 1 is connected to, for example, a Wi-Fi router 3. The Wi-Fi router 3 can connect the sleep sensors 1 used by multiple protected persons 11 to the fall risk estimation device 101 via the facility's LAN (Local Area Network; e.g., wireless LAN), or to an external network 5. In the latter case, the fall risk estimation device 101 can receive the sleep data transmitted by the sleep sensor 1 via the network 5. In the illustrated example, the network 5 is the internet. If the sleep sensor 1 has the functionality of, for example, a built-in Wi-Fi router 3, it can also be connected directly to the network 5.

[0038] Server 7 is a server owned by a facility such as a hospital, which holds basic data such as the medical records of the protected person 11, and is connected to network 5. Server 7 may also be owned by an external vendor and used by a facility such as a hospital. The fall risk estimation device 101 can obtain basic data such as the age and medical history of the protected person 11 by accessing server 7. Information leakage can be prevented by requiring, for example, the input of an identification code (ID) and password during communication between the fall risk estimation device 101 and server 7. Basic data may also be entered by the administrator or operator of the fall risk estimation device 101, rather than through communication with server 7. Each item of basic data will be described later.

[0039] Server 9 is connected to network 5 and builds artificial intelligence that is available through network 5. The fall risk estimation device 101 uses artificial intelligence to estimate daily living activity data based on the sleep data and basic data of the protected person 11. Daily living activity data includes data on items related to daily living activities such as eating, toileting, defecation, toilet transfer, mobility and walking, comprehension, problem solving, and memory. Each item of daily living activity data will be described later.

[0040] The artificial intelligence may be built into the computer 10 as part of the fall risk estimation device 101, or it may be built into the computer 10 separately from the fall risk estimation device 101 so that the fall risk estimation device 101 can access it, or it may be an external artificial intelligence, such as the artificial intelligence provided by the server 9. Furthermore, information leakage can be prevented in communication between the fall risk estimation device 101 and the server 9 by requiring, for example, the input of an identification code (ID) and a password.

[0041] Figure 3 is a block diagram illustrating the configuration of the fall risk estimation device 101. Figure 4 is a flowchart illustrating the processing flow of the fall risk estimation method implemented by the fall risk estimation device 101. Note that the processing procedure of the fall risk estimation method illustrated in Figure 4 can also be performed manually. The fall risk estimation device 101 includes an interface 13, an input data receiving unit 15, a training data receiving unit 17, an estimation unit 19, a learning unit 21, artificial intelligence 23, a fall risk calculation unit 41, a bed exit signal receiving unit 43, an alert output unit 45, and a fall risk output unit 47.

[0042] Interface 13 is a device component that enables communication between the fall risk estimation device 101 itself and external devices according to a predetermined protocol for each external device. Communication between the sleep sensor 1, Wi-Fi router 3, servers 7 and 9, input devices such as keyboards 27, output devices such as printers or displays 29, storage media such as USB memory or CD-ROMs 31 and the fall risk estimation device 101 is performed through interface 13.

[0043] The input data receiving unit 15 receives input data including the sleep data of the protected person 11 and the basic data of the protected person 11 (step S1). The estimation unit 19 inputs the input data received by the input data receiving unit 15 to the artificial intelligence 23, causing the artificial intelligence 23 to calculate estimated data about the daily living activity data of the protected person 11 (step S3). If the artificial intelligence 23 has already been trained, it outputs highly accurate estimated data about the daily living activity data.

[0044] The fall risk calculation unit 41 calculates estimated fall risk data for the protected person 11 using the aforementioned set of features, which consists of the protected person 11's specific sleep data, the protected person 11's specific basic data, and the estimated data calculated by the artificial intelligence 23 (step S5). The sleep data and basic data used by the fall risk calculation unit 41 may overlap with the sleep data and basic data used by the estimation unit 19 in part or in whole, or they may not overlap at all. The fall risk output unit 47 outputs the calculated fall risk (step S7). The output fall risk is transmitted to, for example, the output device 29 via the interface 13. The fall risk is then printed or displayed on the output device 29.

[0045] The alert output unit 45 compares the fall risk calculated by the fall risk calculation unit 41 with a predetermined reference value 46 (step S9). If the fall risk does not exceed the reference value 46 (No in step S9), the process returns to step S1, and the input data receiving unit 15 receives new input data (step S1). The reference value 46 is pre-entered, for example, by the input device 27, and is stored in the memory (not shown) built into the fall risk estimation device 101 via the interface 13. When the fall risk exceeds the reference value 46 (Yes in step S9), and a signal detecting bed exit is received by the bed exit signal receiving unit 43 (Yes in step S11), the alert output unit 45 outputs a fall alert (step S13).

[0046] The output fall alert is transmitted through interface 13 to, for example, output device 29. The fall alert is then notified by output device 29, for example, as an image, light, or sound. This notification allows caregivers who are looking after the person under care 11 in the facility to recognize that the person under care 11, who is at a considerably high risk of falling, has left their bed, and to appropriately protect and assist the person under care 11.

[0047] Subsequently, if it is necessary to wait for a new bed exit detection signal (Yes in step S15), the process returns to step S11. If no bed exit is detected in step S11, the process proceeds to step S15 without outputting a fall alert. In step S15, if it is not necessary to wait for a new bed exit detection (No in step S15), the process returns to step S1 if it is not necessary to terminate the process (No in step S17), and the input data receiving unit 15 receives new input data (step S1). If the process should be terminated (Yes in step S17), the process terminates.

[0048] The artificial intelligence 23 can output highly accurate estimation data through machine learning. The fall risk estimation device 101 has a training data receiving unit 17 and a learning unit 21, which allows the fall risk estimation device 101 to train the artificial intelligence 23 itself without using an external artificial intelligence learning device. In other words, the fall risk estimation device 101 also has an artificial intelligence learning device built in to train the artificial intelligence 23.

[0049] Figure 5 is a flowchart illustrating the processing flow of the artificial intelligence learning method implemented by the fall risk estimation device 101. When the fall risk estimation device 101 performs machine learning, first the input data receiving unit 15 receives input data including sleep data and basic data (step S61). The training data receiving unit 17 receives training data, which is correct daily living activity data corresponding to this input data (step S63). Processes S61 and S63 may occur in any order, or simultaneously. Next, the learning unit 21 inputs the input data received by the input data receiving unit 15 and the training data received by the training data receiving unit 17 into the artificial intelligence 23, thereby training the artificial intelligence 23 to estimate the training data from the input data (step S65).

[0050] Next, the fall risk estimation device 101 returns to S61 if it should repeat the process based on user instructions (Yes in step S67). As a result, the input data receiving unit 15 receives new input data (step S61), and the training data receiving unit 17 receives new training data (step S63). The fall risk estimation device 101 terminates the process if it should not repeat the process (No in step S67). In this way, by inputting a large number of pairs of corresponding input data and training data into the fall risk estimation device 101, the artificial intelligence 23 learns and the accuracy of the estimation improves.

[0051] In the past, sleep data and basic data collected from various protected persons 11, along with daily living activity data obtained through actual measurements corresponding to this data, are associated with each other and recorded, for example, on a storage medium 31. The input data receiving unit 15 and the teacher data receiving unit 17 then sequentially read a large amount of data required for learning from the storage medium 31, and the learning unit 21 can repeat the learning of the artificial intelligence 23 for each piece of data read. In this way, the fall risk estimation device 101 can switch between two operating modes: an estimation mode in which estimation data is calculated and output using the artificial intelligence 23, and a learning mode in which the artificial intelligence 23 is trained through machine learning. Switching between operating modes can be instructed, for example, by the input device 27.

[0052] In the example shown in Figure 3, the artificial intelligence 23 is integrated into the computer 10 as part of the fall risk estimation device 101. Alternatively, as illustrated by the dotted line in Figure 3, an artificial intelligence built on an external server 9 or the like may be used. In this case, the estimation unit 19 and the learning unit 21 operate the external artificial intelligence through the network 5 or the like. The estimation unit 19 inputs the input data received by the input data reception unit 15 into the externally trained artificial intelligence to output estimated data. The estimated data is sent, for example, to the fall risk calculation unit 41 through the interface 13. The learning unit 21 also inputs the input data received by the input data reception unit 15 and the training data received by the training data reception unit 17 into the external artificial intelligence, thereby training the external artificial intelligence to estimate the training data from the input data. When using an external artificial intelligence, the artificial intelligence 23, which is part of the fall risk estimation device 101, becomes unnecessary.

[0053] Furthermore, as already mentioned, the computer 10 into which the fall risk estimation device 101 is incorporated may be a small portable device such as a smartphone (not shown), or some of its functions may be handled by the small portable device. For example, it is possible to have a PC (personal computer) handle the function of the fall risk estimation device 101 that executes the learning mode for machine learning of the artificial intelligence 23, and have the functions of the fall risk estimation device 101, excluding the learning function, handled by the small portable device. In this case, the guardian can understand the status of the person to be protected 11 by the fall risk or fall alert displayed on the screen of the small portable device they carry, or simultaneously output by voice. When the functions of the computer 10 are handled by the small portable device, the hardware and processing burden on the small portable device can be further reduced by using an external artificial intelligence built on a server 9 or the like instead of the artificial intelligence 23.

[0054] Figure 6 is a tabular diagram illustrating the input and output data of the artificial intelligence 23 used by the fall risk estimation device 101. Figure 6(a) illustrates sleep data, Figure 6(b) illustrates basic data, and Figure 6(c) illustrates daily living activity data. Sleep data and basic data are provided to the artificial intelligence 23 as input data. Daily living activity data is output data of the artificial intelligence 23. Below, an example of how each data is represented for handling by the fall risk estimation device 101 is given. This is merely an example, and it is self-evident that other ways of representing the data are possible.

[0055] Of the sleep data (see Figure 6(a)), sleep duration and toilet time are expressed in hours (h) as in 6.5. The sleep rhythm is represented by the time-series changes in sleep duration and wake duration, and is expressed as a data sequence that shows whether the person is asleep or awake at 15-minute intervals from the time they lie down until, for example, 9 hours later (awake, awake, sleep, sleep, sleep, sleep, awake, awake, sleep, ...). This allows us to understand the time it takes to fall asleep after lying down, which is an indicator of how easily a person falls asleep. "Sleep" and "wake" are represented by pre-assigned symbols, such as the numbers "1" and "0". The number of times a person turns over, moves their body, and goes to the toilet are represented by natural numbers such as 1, 2, and 3. The number of times a person turns over, excluding turning over, means the number of times they move their feet or reach out from under the covers. The number of times a person goes to the toilet means the number of times they leave the bed to go to the toilet during sleep. Respiration and pulse are expressed as the number of breaths per minute.

[0056] The above sleep data is acquired by the sleep sensor 1. Alternatively, the above sleep data may be generated by an application on the computer 10 (for example, by the input data receiving unit 15 and the estimation unit 19) based on the output data of the sleep sensor 1. That is, data such as lying down, sleeping, waking, turning over in bed, and body movement are generated by the sleep sensor 1 itself or by the application, based on pressure changes, heart rate, respiratory rate, etc., sensed by the sleep sensor 1. Regarding body movement, even if it is not possible to identify which part of the body has moved, it is possible to detect that it is a body movement other than turning over in bed.

[0057] Basic data (see Figure 6(b)) is obtained, for example, from the hospital's server 7. Alternatively, it may be entered into the computer 10, for example, by manual operation. Of the basic data, age, height, and weight are represented by numerical values ​​based on their respective units. Gender is represented by a code corresponding to male or female, such as "0" or "1". Medical history is represented by a code pre-assigned to various disease names, such as "0", "1", "2", etc. Alternatively, it may be represented by a code corresponding to "none" or "present", such as "0" or "1", for each disease name. The level of care required represents the level of care needed and is represented, for example, by an 8-level numerical scale.

[0058] The example of daily living activity data (see Figure 6(c)) includes 18 items based on the Functional Independence Measure (FIM), a well-known effective tool for occupational therapy assessment. For each daily living activity item, the extent to which the person receiving care 11 can perform it independently is evaluated on a 7-point scale from 1 to 7. It consists of cognitive and motor items. The cognitive items are five items: understanding, expression, social interaction, problem solving, and memory. The motor items are the other 13 items: eating, grooming, bathing, dressing, toileting, urination management, defecation management, transfer (getting up and sitting), and mobility. Each item is expressed by a numerical value corresponding to the score. For each FIM item, a numerical value other than the 7-point scale may also be used, for example, a 3-point scale (e.g., "0", "1", "2").

[0059] If the daily living activity data is expressed as, for example, seven discontinuous numerical values, the fall risk calculation unit 41 converts the estimated value calculated by the artificial intelligence 23 into the closest numerical value among the seven levels, for example by rounding, and uses it. Since the converted data is based on the estimated data of daily living activity data calculated by the artificial intelligence 23, it remains estimated data of daily living activity data.

[0060] One of the inventors of the present invention, having many years of experience as an occupational therapist, has completed an invention to estimate daily living activity data based on sleep data and basic data, which are objective data independent of the evaluator, in order to solve the problem of obtaining objective occupational therapy evaluations that are independent of the evaluator (see Patent Document 1). It was expected that there would be a complex correlation between the sleep data and basic data set and the daily living activity data. Therefore, the inventor thought that it would be possible in principle to estimate daily living activity data based on sleep data and basic data, even if it would be an excessive burden and not practical to do so by human intelligence. The inventor then came to the realization that obtaining such estimated data for occupational therapy evaluations that are beyond human intelligence could be made realistic by using artificial intelligence, and completed the invention (see paragraph 0046 of the specification of Patent Document 1). The present invention utilizes the above patent invention by the same applicant to obtain daily living activity data of the protected person 11.

[0061] For estimating daily living activity data, it is desirable that sleep data include at least some of the following obtained by the sleep sensor 1: respiration, heart rate (used as a concept including pulse), sleep duration, sleep rhythm, number of times turning over in bed, number of body movements, number of times using the toilet, and toilet time. In particular, heart rate and respiratory rate are desirable features for estimation by artificial intelligence 23. Furthermore, it is desirable that basic data include at least some of the following: age, sex, height, weight, medical history, care level, and BMI (Body Mass Index). In particular, age, sex, and BMI are desirable features for estimation by artificial intelligence 23. The more items included in both sleep data and basic data, the higher the accuracy of the estimated daily living activity data.

[0062] Figure 7 is a schematic diagram illustrating the conceptual configuration of the artificial intelligence used by the fall risk estimation device 101. The artificial intelligence provided by the server 9 has a similar configuration as an example. The illustrated artificial intelligence 23 is a neural network and has an input layer 33 in which nodes that receive data input are arranged, an output layer 37 in which nodes that output the calculated data are arranged, and an intermediate layer 35 in which nodes that connect the input layer 33 and the output layer 37 are arranged. In the illustrated example, there is a single intermediate layer 35, but there may be multiple layers. The values ​​of the preceding nodes are transmitted to the next node, reflecting the parameters assigned to each node, i.e., the weights and bias values ​​of each node.

[0063] The input layer 33 receives input data received by the input data reception unit 15, namely pairs of sleep data and basic data items. The input data is transmitted to the output layer 37 via the intermediate layer 35, reflecting the parameters of each node. The data transmitted to the output layer 37 becomes estimated data for pairs of daily living activity data items. The estimation unit 19 (see Figure 3) inputs the pairs of sleep data and basic data items of the protected person 11 to the input layer 33 of the artificial intelligence 23, causing the output layer 37 to generate estimated data for the protected person 11's daily living activity data. The fall risk calculation unit 41 uses the generated estimated data to calculate the fall risk, either after rounding or otherwise transforming it.

[0064] For the estimated data appearing in the nodes of the output layer 37 to be an accurate estimate of daily living activity data, it is necessary to train the artificial intelligence 23 using measured daily living activity data. The training is performed by inputting a set of sleep data and basic data items of a certain protected person 11, received by the input data receiving unit 15, into the input layer 33, and simultaneously inputting training data for the same protected person 11, i.e., a set of measured daily living activity data items, received by the training data receiving unit 17, into the output layer 37 as training data. The learning unit 21 (see Figure 3) inputs this data into the artificial intelligence 23.

[0065] Artificial intelligence 23 calculates estimated data for daily living activity data based on input sleep data and basic data, generates it in the output layer 37, and calculates the error between the generated estimated data and the daily living activity data input as training data. Then, artificial intelligence 23 modifies the parameters of each node from the output layer 37 to the input layer 33, for example, using a well-known backpropagation algorithm, so that error-free estimated data is generated. Such functions are inherent in artificial intelligence 23 itself. By preparing many pairs of input data and training data and repeating the learning process, artificial intelligence 23 becomes able to calculate highly accurate estimated data. When training artificial intelligence 23, it is also possible to adjust the number of hidden layers 35 and the number of nodes in each layer to optimal values. Such techniques are also well known.

[0066] To obtain estimated data for the daily living activities of the protected person 11, it is possible to input not only the latest data for the protected person 11's sleep data and basic data into the artificial intelligence 23, but also data from multiple points in time, including data from earlier periods, along with their respective time data. This allows for the acquisition of estimated data for the daily living activities of the same protected person 11, taking into account past history regarding the protected person 11's sleep data and basic data. This results in more accurate estimated data. The time data may be expressed as, for example, the date and time of each point in time, or as the date and time difference from the most recent point in time. The input data receiving unit 15 receives data from multiple points in time along with their respective time data, and the estimation unit 19 inputs the data from multiple points in time and their respective time data into the input layer 33 of the artificial intelligence 23. The more points in time the input data is received, the more the number of nodes in the input layer 33 that receive the data increases proportionally.

[0067] To obtain estimated data based on data from multiple time points, the artificial intelligence 23 needs to be trained using data from multiple time points, their respective time data, and corresponding training data. For example, to obtain estimated data on daily living activity data from sleep data and basic data from three past time points, including the most recent time point, the artificial intelligence 23 can be trained by inputting sleep data, basic data, and their respective time data from three time points for various protected persons 11 into the input layer 33, and inputting the latest measured data of daily living activity data for each protected person 11 into the output layer 37. Since time data is input simultaneously, the multiple time points from which sleep data and basic data were collected may differ among different protected persons 11. For example, for one protected person 11, data from the most recent time point, one week ago, and five weeks ago may be input, while for another subject, data from the most recent time point, three weeks ago, and fifteen weeks ago may be input. Through learning with a large amount of data, the artificial intelligence 23 adjusts the node parameters so that it calculates estimated data that also reflects the effect of the temporal distance from the most recent time point.

[0068] Artificial intelligence 23 may use not only the neural network exemplified in Figure 6, but also other types of artificial intelligence, such as LGBM (Light GBM; manufactured by Microsoft), which is decision tree-based. LGBM is particularly useful in the process of building artificial intelligence 23 because it has the advantage of easily analyzing which variables in the input data play an important role in the output estimation data.

[0069] As already mentioned, the fall risk estimation device 101 is integrated into the computer 10 in the example shown in Figure 2. By installing and launching a specific application, i.e., a program, on the computer 10, the computer 10 functions as the fall risk estimation device 101. This program may be supplied via the network 5, or it may be supplied via a storage medium 31 such as a CD-ROM (see Figure 2).

[0070] (Example of calculating fall risk) Table 1 illustrates the set of features used by the fall risk calculation unit 41 to calculate the fall risk, and the method for calculating the fall risk using this set of features.

[0071] [Table 1]

[0072] In the example in Table 1, the set of features used as sleep data includes the distribution of nighttime and daytime sleep time (SD1) and daytime sleep time (SD2). These features are examples of sleep data related to sleep quality. These features may be included in the sleep data received by the input data receiving unit 15, or they may be calculated by the input data receiving unit 15 or the fall risk calculation unit 41 from the sleep data received by the input data receiving unit 15. In either case, these features are data based on the sleep data received by the input data receiving unit 15, and are therefore still sleep data.

[0073] In the example in Table 1, BMI(KD1) is used as the basic data among the set of features. This feature may also be included in the basic data received by the input data receiving unit 15, or it may be calculated by the input data receiving unit 15 or the fall risk calculation unit 41 from the basic data received by the input data receiving unit 15. In either case, this feature is data based on the basic data received by the input data receiving unit 15, and remains basic data. Here, the code (KD1) is also assigned for convenience for the same reasons as (SD1), (SD2), etc.

[0074] In the example in Table 1, the artificial intelligence 23 further uses transfer and mobility ability (ND1), problem-solving ability (ND2), and memory ability (ND3) as estimated data for daily living activities. For transfer and mobility ability (ND1), the average value of toilet transfer ability (ND11), bed transfer ability (ND12), and mobility ability (ND13) is used. The artificial intelligence 23 may directly output the estimated data for transfer and mobility ability (ND1), or it may output the estimated values ​​for toilet transfer ability (ND11), bed transfer ability (ND12), and mobility ability (ND13), and for example, the fall risk calculation unit 41 may calculate transfer and mobility ability (ND1) from these three estimated values. In any case, each feature of transfer and mobility ability (ND1), problem-solving ability (ND2), and memory ability (ND3) is data based on estimated data calculated by the artificial intelligence 23, and remains estimated data for daily living activities.

[0075] In addition to these features, the data based on the estimated daily living activity data calculated by artificial intelligence 23 is combined, and the change from the previous day (ND4) is one of the set of features (see "Change in FIM" in Table 1). In the example in Table 1, "sum" is calculated as the "total," but calculating the "average" would not make any difference. It is also possible to use other forms of "total," such as a "weighted average" which is weighted for each feature. Furthermore, in the example in Table 1, the evaluation items defined in the Functional Independence Measure (FIM), which has a long track record and is highly reliable, are used as each item of the daily living activity data estimated by artificial intelligence 23. Note that the codes (ND1), (ND11), etc. are also added for convenience for the same reason as (SD1), (KD1), etc.

[0076] The fall risk calculation unit 41 divides each feature in a set of features into multiple categories and assigns a predetermined evaluation value to each category according to its correlation with falls. In the example in Table 1, the distribution of nighttime and daytime sleep time (SD1) is divided into four categories, and an evaluation value is assigned to each category. For example, if nighttime sleep time is less than 6 hours and daytime sleep time is less than 4 hours, an evaluation value of "7" is assigned. Daytime sleep time (SD2) is also divided into four categories. For example, if daytime sleep time is 2 hours or more and less than 4 hours, an evaluation value of "1" is assigned. BMI (KD1) is divided into two categories. For example, if the BMI is less than 18.7, an evaluation value of "11" is assigned.

[0077] Transfer and mobility ability (ND1) is divided into three categories. For example, if transfer and mobility ability is 4 or higher and less than 6, an evaluation score of "5" is assigned. Problem-solving ability (ND2) is divided into three categories. For example, if problem-solving ability is 3 or higher and 5 or less, an evaluation score of "7" is assigned. Memory ability (ND3) is divided into three categories. For example, if memory ability is 3 or higher and 5 or less, an evaluation score of "5" is assigned. Variability (ND4) is divided into three categories. For example, if variability is either an increase or a decrease from the previous day, an evaluation score of "8" is assigned.

[0078] The fall risk calculation unit 41 further calculates the fall risk by summing the evaluation values ​​for the entire set of features. As in the example above, if the distribution of nighttime and daytime sleep time (SD1) is assigned an evaluation value of "7", daytime sleep time (SD2) is assigned an evaluation value of "1", BMI (KD1) is assigned an evaluation value of "11", transfer and mobility ability (ND1) is assigned an evaluation value of "5", problem-solving ability (ND2) is assigned an evaluation value of "7", memory ability (ND3) is assigned an evaluation value of "5", and variability (ND4) is assigned an evaluation value of "8", then, for example, by calculating the sum of these, an overall evaluation value of "44" can be obtained. In the example in Table 1, the sum of the highest evaluation values ​​for each feature is "58".

[0079] Calculating the "average" as the "overall" result makes virtually no difference. It is also possible to use other forms of "overall" calculations, such as a "weighted average" which multiplies each feature by a weight, but the "weighted average" of the evaluation values ​​is essentially the same as scaling the "evaluation values" defined for each feature and then calculating the "average".

[0080] The calculated overall evaluation value represents the level of fall risk. The fall risk output unit 47 outputs either the calculated overall evaluation value itself, or the overall evaluation value converted into multiple levels (e.g., 3 levels) corresponding to the level of fall risk. For example, if the overall evaluation value is "29 or more but less than 32", "32 or more but less than 58", and "58" for a maximum value of "58", the fall risk will be converted to "Level 1" (requires attention), "Level 2" (requires caution), and "Level 3" (dangerous), respectively, and the fall risk will be output. In any case, the fall risk will be output. The outputted fall risk is displayed, for example, on the output device 29 (see Figure 3). As already mentioned, the alert output unit 45 outputs a fall alert when bed exit is detected when the fall risk exceeds the standard value 46. The fall alert is notified, for example, by the output device 29, using images, light, or sound.

[0081] The algorithm (calculation procedure) for estimating the risk of falling from a set of features, as illustrated above, was obtained by preparing a large number of sets of features and historical data on whether or not falls occurred, training a decision tree-based artificial intelligence called LGBM (Light GBM; manufactured by Microsoft), and then using the results as a reference. In the historical real data collected to prepare the training data, there is overwhelmingly more data without falls (so-called "majority data" or "negative data") than data with falls (so-called "minority data" or "positive data"). Several techniques for efficiently training artificial intelligence based on such so-called "imbalanced data" are already widely known and can be adopted. For example, the technique of "undersampling," which reduces the number of negative data points used for training, and the technique of "oversampling," which increases the number of positive data points used for training, are known and can be adopted.

[0082] (Example of a computer screen) Figure 8 is a screen diagram illustrating an image displayed on the screen of computer 10 (see Figure 2). In this example, computer 10 is a smartphone. The screen in this example simultaneously displays the status of multiple protected persons 11, such as whether they are lying down or out of bed. By scrolling the screen as needed, the status of more protected persons 11 than can fit on a single screen can be visually grasped.

[0083] In the illustrated example, the background color of the name of the person being cared for changes in three ways depending on the level of fall risk output by the fall risk output unit 47, displaying "Danger," "Caution," and "Consideration." "Danger" corresponds to a fall risk exceeding the standard value of 46. If the person being cared for is in "Danger" and the person being cared for is detected to have left the bed, the alert output unit 45 outputs a fall alert. Based on the outputted fall alert, the alarm symbol for the person being cared for is activated on the screen, and an alarm sound is emitted.

[0084] As previously mentioned, the illustrated smartphone may be an example of computer 10, or it may be part of computer 10. This would reduce the hardware and processing load on the smartphone.

[0085] The inventors of this application have constructed and tested a fall risk estimation device 101 and a fall risk estimation system 100. As a result, it has been confirmed that fall risk estimation can be performed at a practical level.

[0086] The features based on sleep data, basic data, and daily living activity data estimated by artificial intelligence 23 do not have to be all of those exemplified in Table 1; some may be used, for example, just one item from each category. Even in this case, estimated fall risk data can be obtained with a reasonable degree of accuracy.

[0087] (3. Fall risk estimation device / system according to another embodiment) Figure 9 is a schematic diagram illustrating an overview of data processing by a fall risk estimation device according to another embodiment of the present invention. In this embodiment, the process of estimating fall risk by the fall risk calculation unit 41 illustrated in Figure 3 is performed by artificial intelligence. The procedure for the process can be illustrated by the flowchart in Figure 4. In the process of "calculating fall risk" (step S5) in Figure 4, artificial intelligence is used.

[0088] Figure 10 is a block diagram illustrating the configuration of a fall risk estimation device that performs the data processing shown in Figure 9. The fall risk estimation device 102 differs from the fall risk estimation device 101 (see Figure 3) in that the fall risk calculation unit 51 estimates the fall risk using artificial intelligence. In the illustrated example, the fall risk calculation unit 51 includes an estimation unit 53, artificial intelligence 55, and a learning unit 57.

[0089] The estimated data calculated by the artificial intelligence 23, that is, the estimated data for the daily living activities of the protected person 11, is read out by the estimation unit 53. The estimation unit 53 inputs the sleep data and basic data received by the input data receiving unit 16, along with the estimated data read out by the artificial intelligence 23, either as is or converted into specific features, to the artificial intelligence 55, causing the artificial intelligence 55 to calculate estimated data for the fall risk of the protected person 11. Even after conversion, the data remains the estimated data for daily living activities, sleep data, and basic data, respectively. For example, sleep data is data related to sleep quality that has a high correlation with fall risk, and the features exemplified in Table 1 are examples of this. Basic data is, for example, BMI, a feature exemplified in Table 1 that has a high correlation with fall risk. Activities of daily living data is, for example, data related to motor ability and cognitive ability, and the features exemplified in Table 1 that have a high correlation with fall risk. Furthermore, the estimated data of the artificial intelligence 23, the sleep data and basic data received by the input data receiving unit 16 may be input to the artificial intelligence 55 as features without any conversion. If conversion is performed, it may be done, for example, in the input data receiving unit 16 or the estimation unit 53.

[0090] The estimation unit 53 inputs the estimated data read from the artificial intelligence 23 to the artificial intelligence 55, either after performing conversions such as rounding, similar to the fall risk calculation unit 41 (see Figure 3), or without conversion. If the artificial intelligence 55 has already been trained, it calculates highly accurate estimated data for the fall risk. For example, if the fall risk is at the "caution required" level, it is represented by the number "1", if it is at the "caution required" level, it is represented by the number "2", and if it is at the "dangerous" level, exceeding the standard value of 46, it is represented by the number "3".

[0091] The fall risk calculation unit 51 inputs the estimated data calculated by the artificial intelligence 55 to the alert output unit 45 and the fall risk output unit 47 as estimated results for the fall risk of the protected person 11, either by rounding or otherwise transforming the data. In this way, the trained artificial intelligence 55 is made to estimate the fall risk, making it easy to obtain estimated results for the fall risk of the protected person 11. Furthermore, similar to the artificial intelligence 23, the artificial intelligence 55 can output highly accurate estimated data by undergoing machine learning.

[0092] The fall risk estimation device 102 has an input data receiving unit 16, a training data receiving unit 18, and a learning unit 57, which allows the fall risk estimation device 102 to train the artificial intelligence 55 itself without using an external artificial intelligence learning device. In other words, the fall risk estimation device 102 also has an artificial intelligence learning device built in to train the artificial intelligence 55. The processing procedure for training the artificial intelligence 55 can be illustrated by the flowchart in Figure 5.

[0093] When the fall risk estimation device 102 instructs the artificial intelligence 55 to perform machine learning, the input data receiving unit 16 receives input data including daily living activity data, sleep data, and basic data related to the motor and cognitive functions of the protected person 11, and the teacher data receiving unit 18 receives input of teacher data, which includes data indicating the level of fall risk of the protected person 11, corresponding to this input data. The learning unit 57 inputs the input data received by the input data receiving unit 16 and the teacher data received by the teacher data receiving unit 18 into the artificial intelligence 55, thereby training the artificial intelligence 55 to estimate the teacher data from the input data. By inputting many pairs of corresponding input data and teacher data into the fall risk estimation device 102, the learning of the artificial intelligence 55 progresses and the accuracy of the estimation improves. It is also possible to have the artificial intelligence 55 estimate the reference value 46 by adding actual values ​​of the reference value 46 to the teacher data.

[0094] In the past, data on daily living activities related to motor and cognitive functions, sleep data, and basic data collected from various protected persons 11, as well as historical data on the high risk of falls for each protected person 11, are associated with each other and recorded, for example, on a storage medium 31. The input data receiving unit 16 and the teacher data receiving unit 18 then sequentially read a large amount of data required for learning from the storage medium 31, and the learning unit 57 can repeat the learning of the artificial intelligence 55 for each piece of data read. In this way, the fall risk estimation device 102 can also switch between two operating modes: an estimation mode in which estimation data is calculated using the artificial intelligence 55, and a learning mode in which the artificial intelligence 55 is trained through machine learning. Switching between operating modes can be instructed, for example, by the input device 27.

[0095] For the protected persons 11, historical data on fall risk can be used, for example, data on the frequency of falls or the severity level of fall accidents. As already mentioned, historical data overwhelmingly consists of data without falls (so-called "majority data" or "negative data") compared with data with falls (so-called "minority data" or "positive data"). To address this so-called "imbalanced data," techniques such as "undersampling" and "oversampling," as already mentioned, can be used.

[0096] In this case, the distinction between "positive data" and "negative data" may be made for each level of fall with different frequencies or severity. For example, to estimate the risk of falls corresponding to risk level 3, data in which falls corresponding to risk level 3 occurred may be designated as "positive data," and other data as "negative data." After performing an operation to increase the proportion of positive data, the input data and training data pair may be input into the artificial intelligence 55. By training the artificial intelligence 55 with data that has undergone such an operation, the artificial intelligence 55 will be able to estimate the risk of falls with even higher accuracy.

[0097] The learning unit 57 may perform an operation on the pair of input data received by the input data receiving unit 16 and the training data received by the training data receiving unit 18, such as "undersampling," to increase the ratio of "positive data" to "negative data," and then input the data pair to the artificial intelligence 55. Alternatively, the data pair that has been operated in this way may be recorded on a storage medium 31, for example, so that the data that has already been operated is received by the input data receiving unit 16 and the training data receiving unit 18.

[0098] Even if techniques such as "undersampling" are used, the learning of the artificial intelligence 55 remains the same, as already described, by inputting a set of daily living activity data, sleep data, and basic data related to the motor and cognitive functions of the protected person 11, received by the input data receiving unit 16, into the input layer 33 of the artificial intelligence 55, and inputting training data for the same protected person 11, i.e., data indicating the high risk of falling, received by the training data receiving unit 17, into the output layer 37 of the artificial intelligence 55.

[0099] Artificial intelligence 55 can use not only the neural network exemplified in Figure 7, but also other types of artificial intelligence, such as LGBM (Light GBM; manufactured by Microsoft), which is decision tree-based. LGBM is particularly useful in the process of building artificial intelligence 55 because it has the advantage of easily analyzing which variables in the input data play an important role in the output estimation data.

[0100] In the example shown in Figure 10, the artificial intelligence 55 is incorporated into the computer 10 (see Figure 2) as part of the fall risk estimation device 102. Alternatively, as illustrated by the dotted line in Figure 10, an artificial intelligence built on an external server 59 or the like may be used. In this case, the estimation unit 53 and the learning unit 57 operate the external artificial intelligence through the network 5 or the like. The estimation unit 53 inputs the input data received by the input data reception unit 16 into the externally trained artificial intelligence to output estimated data. The estimated data is sent, for example, to the alert output unit 45 and the fall risk output unit 47 through the interface 13. The learning unit 57 also inputs the input data received by the input data reception unit 16 and the training data received by the training data reception unit 18 into the external artificial intelligence, thereby training the external artificial intelligence to estimate the training data from the input data. When using an external artificial intelligence, the artificial intelligence 55 that forms part of the fall risk estimation device 102 becomes unnecessary.

[0101] (4. A fall risk estimation device / system according to yet another embodiment) Figure 11 is a schematic diagram illustrating an overview of data processing by a fall risk estimation device according to yet another embodiment of the present invention. In Figure 11, the artificial intelligence that estimates the fall risk in the data processing exemplified in Figure 9 adds data on changes over a predetermined period in the past, such as changes in the daily living activity data of the protected person 11 from the previous day, to the input data. Similar to Figure 9, the input data may also include basic data such as BMI.

[0102] The procedure for the process illustrated in Figure 11 can be illustrated by the flowchart in Figure 4. In the "calculating fall risk" process (step S5) in Figure 4, as described above, data on changes in daily living activity data over a predetermined period is added to the input data of the artificial intelligence that estimates fall risk. The configuration of the fall risk estimation device that executes the procedure illustrated in Figure 11 can be illustrated in the same way as the fall risk estimation device 102 in Figure 10.

[0103] In the data processing illustrated in Figure 11, the estimation unit 53 of the fall risk estimation device 102 illustrated in Figure 10 adds, for example, data on changes from the previous day to the input data of the artificial intelligence 55, as estimated data calculated by the artificial intelligence 23, i.e., estimated data on the daily living activity data of the protected person 11. The estimation unit 53 stores, for example, the estimated data calculated by the artificial intelligence 23 in the memory (not shown) of the fall risk estimation device 102 for a certain period of time in the past, and can calculate change data by reading the stored data. The inventors of the present invention have confirmed that such changes show a strong correlation with the risk of falling. Therefore, the fall risk estimation data by the artificial intelligence 55 can be obtained with higher accuracy.

[0104] The feature sets (SD1), (SD2), (KD1), (ND1) to (ND4) exemplified in Table 1 are examples of input data that the estimation unit 53 provides to the artificial intelligence 55 in the processing procedure exemplified in Figure 11. Among these, the change from the previous day in the daily living activity data (ND4) is an example of the data on the change in the daily living activity data over a predetermined past period as described above. The processing procedure for training the artificial intelligence 55 can be illustrated by the flowchart in Figure 5. (5. A fall risk estimation device / system according to yet another embodiment)

[0105] Figure 12 is a schematic diagram illustrating an overview of data processing by a fall risk estimation device according to yet another embodiment of the present invention. In the example of Figure 12, new artificial intelligence is introduced to obtain data on changes over a predetermined past period, such as changes from the previous day in the daily living activity data of the protected person 11, which are added as input data for the artificial intelligence for fall risk estimation in the data processing of Figure 11. The new artificial intelligence estimates data on changes over a predetermined past period of daily living activity data from data on changes over a predetermined past period of sleep data. As with Figure 9, basic data such as BMI may also be included as input data for the artificial intelligence for fall risk estimation.

[0106] The procedure for the process illustrated in Figure 12 can be illustrated by the flowchart in Figure 4. In the example in Figure 12, in the process of "calculating fall risk" (step S5) in Figure 4, as described above, data on changes in daily living activity data, derived from data on changes in sleep data, is added to the input data of the artificial intelligence that estimates fall risk.

[0107] Figure 13 is a block diagram illustrating the configuration of a fall risk estimation device that performs the data processing shown in Figure 11. This fall risk estimation device 103 differs from the fall risk estimation device 102 (Figure 10) in that, as described above, it estimates data on changes in daily living activity data from data on changes in sleep data using artificial intelligence. In the illustrated example, the fall risk calculation unit 71 includes an estimation unit 73, artificial intelligence 75, and a learning unit 77.

[0108] The estimation unit 73 inputs data on changes in the protected person's sleep data over a predetermined period in the past, such as changes from the previous day to the current day, which has been received by the input data receiving unit 79, into the trained artificial intelligence 75. As a result, the artificial intelligence 75 calculates estimated data on changes in the protected person's daily living activity data over a predetermined period in the past, such as changes from the previous day to the current day. The calculated estimated data is sent to the estimation unit 53. The sleep data received by the input data receiving unit 79 includes, for example, heart rate and respiratory rate (see Figure 12). In this way, since the original sleep data change data is used to obtain estimated data on changes in daily living activity data over a predetermined period in the past, which is expressed in stages, estimated data that captures more subtle changes can be obtained.

[0109] Instead of receiving data on changes in sleep data in the input data receiving unit 79, the input data receiving unit 79 may receive sleep data before it is converted into change data, and the input data receiving unit 79 or the estimation unit 73 may convert it into change data for a predetermined past period. The converted sleep data can be said to be based on the sleep data received by the input data receiving unit 79.

[0110] The fall risk estimation device 103 has an input data receiving unit 79, a training data receiving unit 81, and a learning unit 77, which allows the fall risk estimation device 103 to train the artificial intelligence 75 itself without using an external artificial intelligence learning device. In other words, the fall risk estimation device 103 also has an artificial intelligence learning device built in to train the artificial intelligence 75. The processing procedure for training the artificial intelligence 75 can be illustrated by the flowchart in Figure 5.

[0111] When the fall risk estimation device 103 instructs the artificial intelligence 75 to perform machine learning, the input data receiving unit 19 receives input data including data on changes in the protected person's sleep data over a predetermined period in the past, and the teacher data receiving unit 18 receives teacher data including data on changes in the protected person's daily living activity data over a predetermined period in the past, corresponding to this input data. The learning unit 77 inputs the input data received by the input data receiving unit 19 and the teacher data received by the teacher data receiving unit 81 to the artificial intelligence 75, thereby training the artificial intelligence 75 to estimate the teacher data from the input data. By inputting a large number of pairs of corresponding input data and teacher data into the fall risk estimation device 103, the learning of the artificial intelligence 75 progresses, and the accuracy of the estimation improves.

[0112] Similar to the learning of artificial intelligences 23 and 55, the learning of artificial intelligence 75 can be efficiently performed by associating numerous sets of input data and training data with each other and recording them, for example, in a storage medium 31. The fall risk estimation device 103 can also be switched between two operating modes: an estimation mode in which estimation data is calculated using artificial intelligence 75, and a learning mode in which artificial intelligence 75 is trained using machine learning. As for artificial intelligence 75, in addition to the neural network exemplified in Figure 7, other types of artificial intelligence may be used, such as LGBM (Light GBM; manufactured by Microsoft), which is decision tree-based.

[0113] In the example shown in Figure 13, the artificial intelligence 75 is incorporated into the computer 10 (see Figure 2) as part of the fall risk estimation device 103. Alternatively, as illustrated by the dotted line in Figure 13, an artificial intelligence built on an external server 61 or the like may be used. In this case, the estimation unit 73 and the learning unit 77 operate the external artificial intelligence through the network 5 or the like. The estimation unit 73 inputs the input data received by the input data receiving unit 79 into the externally trained artificial intelligence to output estimated data. The estimated data is sent to the estimation unit 53, for example, through the interface 13. The learning unit 77 also inputs the input data received by the input data receiving unit 79 and the training data received by the training data receiving unit 81 into the external artificial intelligence, thereby training the external artificial intelligence to estimate the training data from the input data. When using an external artificial intelligence, the artificial intelligence 75 that forms part of the fall risk estimation device 103 becomes unnecessary. [Explanation of symbols]

[0114] 1 Sleep sensor, Wi-Fi router 3, 5 Network, 7,9 Server, 10 Computer, 11 Protected person, 13 Interface, 15 Input data reception unit, 16 Input data reception unit, 17 Training data reception unit, 18 Training data reception unit, 19 Estimation unit, 21 Learning unit, 23 Artificial intelligence, 27 Input device, 29 Output device, 31 Storage medium, 33 Input layer, 35 Intermediate layer, 37 Output layer, 41 Fall risk calculation unit, 45 Alert output unit, 46 Reference value, 47 Fall risk output unit, 51 Fall risk calculation unit, 53 Estimation unit, 55 Artificial intelligence, 57 Learning unit, 71 Fall risk calculation unit, 73 Estimation unit, 75 Artificial intelligence, 77 Learning unit, 79 Input data reception unit, 81 Teacher data reception unit, 100 Fall risk estimation system, 101, 102, 103 Fall risk estimation device.

Claims

1. A first estimation unit inputs input data, including first sleep data which is data related to the protected person's sleep and first basic data which is data related to the protected person's physical condition, into a first artificial intelligence that has been trained, thereby causing the first artificial intelligence to calculate estimated data of daily living activities that represents the level of the protected person's ability to perform daily living activities. A fall risk calculation unit calculates estimated fall risk data by referring to a set of data including second sleep data, which is data on the sleep of the protected person, and the estimated data of the daily living activity data. The system includes a fall risk output unit that outputs estimated data of the fall risk calculated by the fall risk calculation unit, The first and second sleep data are data based on data measured by a sleep sensor, and each includes data expressed numerically. The aforementioned basic data includes data expressed in numerical form, The aforementioned daily living activity data includes data that is represented numerically in stages, and is used to estimate the risk of falls.

2. The fall risk estimation device according to claim 1, wherein the second sleep data referenced by the fall risk calculation unit includes data relating to sleep quality.

3. The fall risk estimation device according to claim 2, wherein the data relating to sleep quality includes data on at least one item from the distribution of nighttime and daytime sleep time and daytime sleep time.

4. The set of data referenced by the fall risk calculation unit further includes second basic data which is data relating to the physical condition of the person being protected. The fall risk estimation device according to claim 1, wherein the second basic data includes data expressed numerically.

5. The fall risk estimation device according to claim 4, wherein the second basic data included in the set of data includes BMI (Body Mass Index).

6. The fall risk estimation device according to claim 1, wherein the first estimation unit causes the first artificial intelligence to calculate estimation data, which includes at least one item of transfer and mobility ability, problem-solving ability, and memory ability, as the daily living activity data.

7. The fall risk estimation device according to claim 1, wherein the set of data referenced by the fall risk calculation unit further includes data on the changes in the estimated data over a predetermined past period of the daily living activity data.

8. The fall risk calculation unit includes a second estimation unit that inputs data including time-series data of the second sleep data over a predetermined past period into a trained second artificial intelligence, thereby causing the second artificial intelligence to calculate estimated data of changes in the daily living activity data over the predetermined past period. The fall risk estimation device according to claim 1, wherein the set of data referenced by the fall risk calculation unit further includes the change estimation data.

9. The fall risk estimation device according to claim 1, wherein the fall risk calculation unit includes a third estimation unit that inputs the set of referenced data into a trained third artificial intelligence to calculate estimated data of the fall risk.

10. The fall risk estimation device according to claim 1, wherein the fall risk calculation unit divides each of the referenced data sets into multiple categories, assigns a predetermined evaluation value to each of the divided categories according to its correlation with falls, and calculates estimated fall risk data by combining the evaluation values ​​for the entire data set.

11. A bed exit signal receiving unit that receives a signal from a bed exit sensor that detects when the person being protected leaves the bed, The fall risk estimation device according to claim 1, further comprising: an alert output unit that outputs a fall alert when the bed exit signal receiving unit receives a signal when the estimated data of the fall risk calculated by the fall risk calculation unit exceeds a predetermined standard value.

12. A first estimation process involves inputting input data, including first sleep data which is data related to the protected person's sleep and first basic data which is data related to the protected person's physical condition, into a first artificial intelligence that has been trained, thereby causing the first artificial intelligence to calculate estimated data of daily living activities that represent the level of the protected person's ability to perform daily living activities. A fall risk calculation process calculates estimated fall risk data by referring to a data set that includes second sleep data, which is data on the sleep of the protected person, and the estimated data of the activities of daily living. The system includes a fall risk output process that outputs estimated data of the fall risk calculated by the fall risk calculation process, The first and second sleep data are data based on data measured by a sleep sensor, and each includes data expressed numerically. The aforementioned basic data includes data expressed in numerical form, The aforementioned daily living activity data includes data that is expressed numerically in stages, and is a method for estimating the risk of falling.

13. A bed exit signal reception process that receives a signal from a bed exit sensor that detects when the protected person leaves the bed, The fall risk estimation method according to claim 12, further comprising: an alert output process that outputs a fall alert when the bed exit signal reception process receives a signal when the fall risk estimated by the fall risk calculation process exceeds a predetermined standard value.

14. A fall risk estimation program that, when read by a computer, causes the computer to function as a fall risk estimation device according to any one of claims 1 to 11.

15. An artificial intelligence learning device for training an artificial intelligence that functions as the second artificial intelligence used in the fall risk estimation device described in claim 8, An input data receiving unit accepts data, including time-series data of sleep data for a predetermined past period, which is data related to the sleep of protected persons, as input data. A training data receiving unit receives training data that corresponds to the input data and includes data showing the historical changes in the daily living activity data of the protected person over a predetermined past period, The system includes a learning unit that inputs the input data received by the input data receiving unit and the training data received by the training data receiving unit to the artificial intelligence, thereby training the artificial intelligence to estimate the training data from the input data. The aforementioned sleep data includes numerical data based on data measured by a sleep sensor. The aforementioned daily living activity data includes data represented numerically in stages, and is an artificial intelligence learning device.

16. An artificial intelligence learning device for training an artificial intelligence that functions as the third artificial intelligence used in the fall risk estimation device described in claim 9, An input data receiving unit accepts a set of data as input data, which includes sleep data, which is data related to the sleep of the protected person, and daily living activity data, which represents the level of the protected person's ability to perform daily living activities. A training data receiving unit that receives training data including data indicating the risk of falling for the protected person, corresponding to the input data, The system includes a learning unit that inputs the input data received by the input data receiving unit and the training data received by the training data receiving unit to the artificial intelligence, thereby training the artificial intelligence to estimate the training data from the input data. The aforementioned sleep data includes numerical data based on data measured by a sleep sensor. The aforementioned daily living activity data includes data represented numerically in stages, and is an artificial intelligence learning device.

17. An artificial intelligence learning method for training an artificial intelligence that functions as the second artificial intelligence used in the fall risk estimation device described in claim 8, An input data reception process that accepts data including time-series data of sleep data for a specified past period, which is data related to the sleep of protected persons, as input data, A training data reception process that receives training data corresponding to the input data, including data showing the historical changes in daily living activity data for the protected person over a predetermined past period, The system includes a learning process which involves inputting the input data received by the input data receiving process and the training data received by the training data receiving process into the artificial intelligence, thereby training the artificial intelligence to estimate the training data from the input data. The aforementioned sleep data includes numerical data based on data measured by a sleep sensor. The aforementioned daily living activity data includes data that is represented numerically in stages, and is an artificial intelligence learning method.

18. An artificial intelligence learning method for training an artificial intelligence that functions as the third artificial intelligence used in the fall risk estimation device described in claim 9, An input data reception process that accepts a set of data as input data, which includes sleep data, which is data related to the sleep of the protected person, and daily living activity data, which represents the level of the protected person's ability to perform daily living activities. A training data reception process that receives training data including data indicating the risk of falling for the protected person, corresponding to the input data, The system includes a learning process which involves inputting the input data received by the input data receiving process and the training data received by the training data receiving process into the artificial intelligence, thereby training the artificial intelligence to estimate the training data from the input data. The aforementioned sleep data includes numerical data based on data measured by a sleep sensor. The aforementioned daily living activity data includes data that is represented numerically in stages, and is an artificial intelligence learning method.

19. An artificial intelligence learning program that, when read by a computer, causes the computer to function as the artificial intelligence learning device described in claim 15 or 16.

Images