Estimation device, learning model, estimation method, learning model generation method, and computer program

Abstract

The present disclosure provides an estimation device, a learning model, an estimation method, a learning model generation method, and a computer program. The estimation device estimates rheumatoid arthritis index information of a subject. The estimation device comprises: an estimation unit that inputs input information to a learning model constructed by learning using a learning data set, and acquires output information from the learning model; and a storage unit that stores the output information. The learning data set includes vital information of a learning subject, behavior information of the learning subject, and rheumatoid arthritis index information of the learning subject. The input information includes vital information of the subject and behavior information of the subject. The output information includes rheumatoid arthritis index information of the subject.

Description

Estimation device, learning model, estimation method, learning model generation method, and computer program

[0001] The present disclosure relates to an estimation device, a learning model, an estimation method, a learning model generation method, and a computer program.

[0002] The learning device described in Patent Document 1 generates a pain estimation learning model for estimating the pain of a subject. The pain estimation learning model is a learned model for estimating the pain situation based on time-series biological data. The learning device acquires, for each of a plurality of first subjects, first biological data that is time-series biological data and pain data that is time-series data representing the pain situation.

[0003] The first biological data is a heartbeat or the like. The pain data is an observation result by CPOT (Critical-Care Pain Observation tool).

[0004] The learning device determines a stable period for each first subject based on the first biological data. The learning device determines a baseline during the stable period for each first subject based on the first biological data, and determines first difference data for each first subject. The first difference data is time-series data representing the difference between the first biological data and the baseline during the stable period. The learning device generates a pain estimation learning model by machine learning based on the first difference data and the pain data of each first subject.

[0005] Japanese Unexamined Patent Application Publication No. 2023-96312

[0006] However, the pain estimation learning model described in Patent Document 1 only estimates the pain of a patient lying in a hospital bed. That is, the pain estimation learning model does not estimate the pain related to a specific disease.

[0007] An object of the present disclosure is to provide an estimation device, a learning model, an estimation method, a learning model generation method, and a computer program capable of estimating rheumatoid arthritis index information.

[0008] According to this disclosure, an estimation device for estimating rheumatoid arthritis index information of a subject is provided, comprising: an estimation unit that inputs input information to a learning model constructed by learning using a learning dataset and obtains output information from the learning model; and a storage unit that stores the output information, wherein the learning dataset includes vital information of the learning subject, behavioral information of the learning subject, and rheumatoid arthritis index information of the learning subject; the input information includes the vital information of the subject and behavioral information of the subject; and the output information includes the rheumatoid arthritis index information of the subject.

[0009] Furthermore, according to this disclosure, a learning model can be provided which is constructed by learning using a learning dataset and causes a computer to function to estimate rheumatoid arthritis index information of a subject, wherein the learning dataset includes vital information of the learning subject, behavioral information of the learning subject, and rheumatoid arthritis index information of the learning subject, and the computer is caused to function to take input information as input and output information as output, wherein the input information includes the vital information of the subject and behavioral information of the subject, and the output information includes the rheumatoid arthritis index information of the subject.

[0010] Furthermore, the present disclosure provides an estimation method for estimating rheumatoid arthritis index information of a subject, comprising the steps of: inputting input information into a learning model constructed by learning using a learning dataset; and obtaining output information from the learning model into which the input information has been input, wherein the learning dataset includes vital information of the learning subject, behavioral information of the learning subject, and rheumatoid arthritis index information of the learning subject, the input information includes the vital information of the subject and behavioral information of the subject, and the output information includes the rheumatoid arthritis index information of the subject.

[0011] Furthermore, this disclosure provides a computer program that causes a computer to execute the above estimation method.

[0012] Furthermore, according to this disclosure, a method for generating a learning model is provided, which includes the steps of: acquiring a learning dataset; and generating a learning model that outputs output information when input information is input by performing learning using the learning dataset, wherein the learning dataset includes vital information of a learning subject, behavioral information of the learning subject, and rheumatoid arthritis index information of the learning subject; the input information includes vital information of the subject and behavioral information of the subject; and the output information includes the rheumatoid arthritis index information of the subject.

[0013] Furthermore, this disclosure provides a computer program that causes a computer to execute the above-described learning model generation method.

[0014] According to this disclosure, an estimation device, a learning model, an estimation method, a learning model generation method, and a computer program capable of estimating rheumatoid arthritis index information can be provided.

[0015] Figure 1 is a block diagram showing an example configuration of an information processing system according to Embodiment 1 of this disclosure. Figure 2 is a diagram showing an example of a learning dataset according to Embodiment 1. Figure 3 is a diagram showing an example of input information and output information according to Embodiment 1. Figure 4 is a block diagram showing an example configuration of an estimation system according to Embodiment 1. Figure 5 is a flowchart showing an example of an estimation method executed by the estimation system according to Embodiment 1. Figure 6 is a block diagram showing an example configuration of an estimation device according to Embodiment 1. Figure 7 is a graph showing an example of multiple unit data arranged in time series according to Embodiment 1. Figure 8 is a graph for explaining the classification execution process according to Embodiment 1. Figure 9 is a graph showing an example of first state data and second state data arranged in time series according to Embodiment 1. Figure 10 is a graph showing an example of correction data arranged in time series according to Embodiment 1. Figure 11 is a graph showing an example of cumulative data arranged in time series according to Embodiment 1. Figure 12 is a diagram schematically showing an example of a deep neural network according to Embodiment 1. Figure 13 is a diagram schematically showing another example of a deep neural network according to Embodiment 1. Figure 14 is a diagram schematically showing yet another example of a deep neural network according to Embodiment 1. Figure 15 is a flowchart showing an example of estimation processing according to Embodiment 1. Figure 16 is a block diagram showing an example configuration of a learning device according to Embodiment 1. Figure 17 is a flowchart showing an example of a learning method using the learning device according to Embodiment 1. Figure 18 is a block diagram showing an example configuration of a learning data creation device according to Embodiment 1. Figure 19 is a flowchart showing an example of a learning data creation method using the learning data creation device according to Embodiment 1. Figure 20 is a diagram showing an example of a learning dataset according to Embodiment 2 of this disclosure. Figure 21 is a diagram showing an example of input information and output information according to Embodiment 2. Figure 22 is a graph showing the ROC curve calculated for the learning model according to Embodiment 1 of this disclosure. Figure 23 is a graph showing the SHAP value calculated for the learning model according to Embodiment 1. Figure 24 is a graph showing the SHAP value calculated for the learning model according to Embodiment 9 of this disclosure.

[0016] Preferred embodiments of this disclosure will be described in detail below with reference to the attached drawings. In this specification and the drawings, components having substantially the same functional configuration are denoted by the same reference numerals, and redundant descriptions will be omitted.

[0017] (Embodiment 1) The information processing system according to Embodiment 1 of the present disclosure uses a learning model to estimate rheumatoid arthritis index information of a subject from at least the subject's biological state information (vital information and / or behavioral information). The rheumatoid arthritis index information may be an index representing the symptoms or activity of rheumatoid arthritis, or it may be information relating to an index representing the symptoms or activity of rheumatoid arthritis. Activity refers to disease activity. Disease activity refers to the degree of the severity of the disease or the intensity of the symptoms.

[0018] Rheumatoid arthritis index information includes, for example, at least one of the following: information on CDAI, information on SDAI, information on DAS (Disease Activity Score) 28, and information on joint pain. CDAI, SDAI, and DAS 28 are indices for assessing disease activity in rheumatoid arthritis. Information on CDAI is information that directly or indirectly represents CDAI. Information on SDAI is information that directly or indirectly represents SDAI. Information on DAS 28 is information that directly or indirectly represents DAS 28. Information on joint pain is information that directly or indirectly represents joint pain. CDAI is calculated by formula (1).

[0019] CDAI = Number of tender joints + Number of swollen joints + Subject's VAS score + Physician's VAS score ... (1)

[0020] The number of tender joints indicates the number of painful joints out of 28 joints. The number of swollen joints indicates the number of swollen joints out of 28 joints. The 28 joints are the shoulder (2), elbow (2), wrist (2), knee (2), metacarpophalangeal joints (10), and proximal interphalangeal joints (10).

[0021] The subject's VAS (visual analogue scale) value indicates the subject's own assessment of their overall health using the VAS scale. The physician's VAS value indicates the physician's assessment of the subject's overall health using the VAS scale. The VAS value is a value (e.g., 10 mm) that indicates where the current health (symptoms) is located on a horizontal line (VAS) with a length of 100 mm, where 0 mm (left end) represents very good health (no symptoms) and 100 mm (right end) represents very poor health (worst symptoms ever experienced). The unit of the VAS value is typically "mm". The subject is, for example, a patient. SDAI is calculated by formula (2).

[0022] SDAI = Number of tender joints + Number of swollen joints + Subject's VAS score + Physician's VAS score + CRP ... (2)

[0023] CRP indicates the mass of C-reactive protein per unit volume of blood. The unit is typically "mg / dl". C-reactive protein appears in the blood when inflammatory responses and / or tissue destruction are occurring in the body. C-reactive protein is a type of acute-phase reaction protein. C-reactive protein indicates the degree of rheumatoid arthritis.

[0024] DAS28 has two components: DAS28_CRP and DAS28_ESR. DAS28_CRP is calculated by formula (3), where "Ln" is the natural logarithm.

[0025] DAS28_CRP = 0.566 × √(number of tender joints) + 0.28 × √(number of swollen joints) + 0.36 × Ln(CRP × 10 + 1) + 0.014 × subject's VAS value + 0.96 … (3)

[0026] DAS28_ESR is calculated by formula (4).

[0027] DAS28_ESR = 0.56 × √(number of tender joints) + 0.28 × √(number of swollen joints) + 0.7 × Ln(ESR) + 0.014 × subject's VAS value ... (4)

[0028] ESR indicates the rate at which red blood cells settle in a reagent. The unit is typically mm / h. ESR represents the degree of arthritis caused by rheumatoid arthritis.

[0029] Joint pain is indicated, for example, by a symptom rating scale value. The symptom rating scale value refers to the value on the symptom rating scale. Examples of symptom rating scales include VAS, NRS (Numerical Rating Scale), VRS (Verbal Rating Scale), or FRS (Face Rating Scale). In particular, the symptom rating scale and symptom rating scale value related to pain may be referred to as the pain rating scale and pain rating scale value, respectively.

[0030] Next, an information processing system 1 according to Embodiment 1 of the present disclosure will be described with reference to Figures 1 to 19. Figure 1 is a block diagram showing an example configuration of the information processing system 1. As shown in Figure 1, the information processing system 1 comprises a learning data creation device 2, a learning device 3, and an estimation device 4. Each of the learning data creation device 2, the learning device 3, and the estimation device 4 is a computer. The estimation device 4 comprises a pre-processing unit 41, an estimation unit 42, and a learning model TM1. The estimation device 4 may further comprise a post-processing unit 43.

[0031] The learning data creation device 2 acquires raw biological state data A1 and correct answer information Z1 of the learning subject. Raw biological state data A1 is raw data indicating the biological state of the learning subject. A learning subject is a subject from which learning raw data is acquired. Typically, a learning subject is a human being. Raw biological state data A1 includes raw vital data and behavioral data of the learning subject. Raw vital data is raw data indicating the vital signs of the organism. Raw behavioral data is raw data indicating the behavior of the organism. Correct answer information Z1 indicates rheumatoid arthritis index information.

[0032] The training data creation device 2 creates a training dataset F1 based on the raw biological state data A1 and the ground truth information Z1. The training dataset F1 includes feature information G1 and the ground truth label B1 of the training subject. Feature information G1 includes at least one of the vital information H1 and behavioral information J1 of the training subject. In Embodiment 1, feature information G1 includes at least the vital information H1 and behavioral information J1 of the training subject. Feature information G1 is an explanatory variable. Feature information G1 is created based on the raw biological state data A1. The ground truth label B1 is the target variable. The ground truth label B1 is created based on the ground truth information Z1. The ground truth label B1 indicates rheumatoid arthritis index information M1.

[0033] The learning device 3 performs learning using the learning dataset F1 by executing a machine learning algorithm. The learning device 3 generates a learning model TM1 by repeatedly performing learning using multiple learning datasets F1. The learning device 3 performs supervised learning. The learning model TM1 outputs output information L2 when input information K2 is input. Typically, the learning model TM1 is a pre-trained model. Also, the learning model TM1 is a computer program.

[0034] Meanwhile, the estimation device 4 uses the learning model TM1 to estimate the subject's rheumatoid arthritis index information M2. The subject is typically a human.

[0035] Specifically, the preprocessing unit 41 performs preprocessing on the subject's biological state raw data A2 and generates input information K2, which is the result of the preprocessing. The biological state raw data A2 is raw data indicating the subject's biological state. The biological state raw data A2 includes the subject's vital raw data and behavioral raw data. The vital raw data is raw data indicating the vital signs of the organism. The behavioral raw data is raw data indicating the behavior of the organism. The input information K2 includes at least one of the vital information H2 and behavioral information J2. In Embodiment 1, the input information K2 includes at least the vital information H2 and behavioral information J2.

[0036] The estimation unit 42 inputs the input information K2 to the learning model TM1. As a result, the learning model TM1 outputs output information L2. The input information K2 is an explanatory variable. The output information L2 is an objective variable. The output information L2 includes the subject's rheumatoid arthritis index information M2. The estimation unit 42 obtains the output information L2 from the learning model TM1. The post-processing unit 43 may perform post-processing on the output information L2 and generate output information N2, which is the result of the post-processing. The output information N2 includes the post-processed rheumatoid arthritis index information M3.

[0037] In the following, unless it is necessary to distinguish between the rheumatoid arthritis index information M2 and M3, the subject's rheumatoid arthritis index information M2 and M3 may be referred to as "rheumatoid arthritis index information MX".

[0038] In Embodiment 1, the estimation device 4 can output rheumatoid arthritis index information MX with high estimation accuracy by inputting at least vital information H2 and behavioral information J21 to the learning model TM1. In particular, the estimation device 4 does not require diagnosis and evaluation by a healthcare professional, nor blood tests. Healthcare professionals are, for example, doctors or nurses. In addition, the estimation device 4 automatically and continuously acquires biological state raw data A2 from a wearable device and / or mobile terminal. Therefore, rheumatoid arthritis index information MX can be obtained while reducing the burden on the subject.

[0039] Next, we will explain the correlation between the vital information H1 and behavioral information J1 of the subjects and the rheumatoid arthritis index information M1 in the training dataset F1. Figure 2 shows an example of the training dataset F1. As shown in Figure 2, the training dataset F1 includes feature information G1 and ground truth labels B1 of the training subjects.

[0040] Feature information G1 includes at least vital information H1 and behavioral information J1 of the training subject. Vital information H1 is information that directly or indirectly indicates the vital signs of the organism. Behavioral information J1 is information that directly or indirectly indicates the behavior of the organism. In this specification, the organism is typically a human being. On the other hand, in the ground truth label B1, the rheumatoid arthritis index information M1 is information that directly or indirectly indicates the symptoms or activity of rheumatoid arthritis. The rheumatoid arthritis index information M1 includes at least one of the following: information on CDAI, information on SDAI, information on DAS28, and information on joint pain. For example, in the ground truth label B1, the values ​​for CDAI, SDAI, DAS28, and joint pain are measured values.

[0041] Rheumatoid arthritis is a chronic disease characterized by systemic inflammatory responses. Therefore, physical and behavioral changes associated with these inflammatory responses are reflected in vital information H1 and behavioral information J1. In addition, joint symptoms such as deformation, pain, and stiffness (difficulty moving) of some joints occur as one of the main symptoms of rheumatoid arthritis. Therefore, physical and behavioral changes resulting from these joint symptoms are reflected in behavioral information J1. As a result, it can be inferred that there is a correlation between vital information H1 and behavioral information J1 and rheumatoid arthritis index information M1. Accordingly, according to Embodiment 1, the learning model TM1 constructed by learning using the learning dataset F1 (vital information H1, behavioral information J1, and rheumatoid arthritis index information M1) can output the subject's rheumatoid arthritis index information M2 when the subject's vital information H2 and behavioral information J2 are input. In other words, the estimation device 4 can estimate rheumatoid arthritis index information M2 by using the learning model TM1.

[0042] Preferably, vital information H1 includes information about heart rate. In this specification, heart rate is not limited to heart rate based on electrocardiogram waveforms obtained by electrocardiography (ECG), but also includes pulse rate based on pulse wave waveforms. In other words, heart rate based on electrocardiogram waveforms and pulse rate based on pulse wave waveforms are treated as "heart rate" in this specification. The method for obtaining the pulse wave waveform is not particularly limited and may be obtained, for example, by photoplethysmography (PCG).

[0043] Information related to heart rate includes, for example, heart rate (HR), R-R interval, time-domain indicators related to heart rate, frequency-domain indicators related to heart rate, or nonlinear indicators related to heart rate. Heart rate is the number of times the heart beats in a given period of time, and is expressed, for example, as beats per minute (bpm). A given period of time in heart rate may be referred to as the first given period. The R-R interval is the time interval from one QRS wave to the next in the electrocardiogram waveform. Time-domain indicators related to heart rate, frequency-domain indicators related to heart rate, and nonlinear indicators related to heart rate are collectively called heart rate variability (HRV) indicators. In this specification, for example, pulse rate (PR) is treated as heart rate, pulse interval (PI) is treated as R-R interval, and pulse variability is treated as heart rate variability.

[0044] The time domain indicators are, for example, SDNN (Standard deviation of NN intervals), RMSSD (Root Mean Square of Successive Differences), CVRR (Coefficent of Variation of RR intervals), SDRR (Standard deviation of RR intervals), SDANN (Standard Deviation of the Average NN intervals for each 5min segment of a 24h HRV recording), SDNN index, NN50 (the number of pairs of successive NN intervals that differ by more than 50 ms), pNN50 (the proportion of NN50 divided by the total number of NN intervals), HR Max, HR Min, (HR Max - HR Min), HTI (HRV Triangular Index), or TINN (Triangular Interpolation of the NN Interval Histogram). HRV is heart rate variability. RMSSD is, with respect to heartbeats, the square root of the average value of the squares of the differences between successive adjacent R-R intervals.

[0045] The frequency domain indicators are, for example, for heart rate variability, LF (power or peak of the low frequency component), HF (power or peak of the high frequency component), LF / HF, Total Power, LF Norm, HF Norm, ULF (power of the very low frequency region), or VLF (power of the ultra low frequency region).

[0046] Nonlinear indices include, for example, entropy, SD1 (standard deviation in the direction orthogonal to y = x in a Poincaré plot), SD2 (standard deviation in the direction along y = x in a Poincaré plot), SD1 / SD2, ApEn (Approximate entropy), SampEn (Sample entropy), DFA α1 (Detrended Fluctuation Analysis, which describes short-term fluctuations), DFA α2 (Detrended Fluctuation Analysis, which describes long-term fluctuations), (DFA α1) / (DFA α2), CVI (Cardiac Vagal Index), or CSI (Cardiac Sympathetic Index).

[0047] The inflammatory reaction of rheumatoid arthritis is controlled by the autonomic nervous system, resulting in the activation state of the sympathetic nerves. As a physical reaction accompanying the activation of the sympathetic nerves, for example, there is an increase in heart rate. Therefore, it can be speculated that an increase in heart rate may lead to an enhancement of the inflammatory reaction and an increase in disease activity of rheumatoid arthritis. By the same logic, for example, it can be speculated that changes in heart rate variability indices that capture the activity of the autonomic nervous system may lead to an enhancement of the inflammatory reaction and an increase in disease activity. From these results, it can be speculated that there is a correlation between information related to heart rate and rheumatoid arthritis index information M1.

[0048] Note that the vital information H1 may include, for example, one or more of blood pressure information, respiratory information, body temperature information, and electroencephalogram information. These information may be related to or potentially related to the inflammatory reaction of rheumatoid arthritis. Respiratory information is, for example, the respiratory rate. Body temperature information includes, for example, at least one of skin temperature, body temperature, and deep body temperature. In this case, the body temperature information may be, for example, skin temperature during sleep, body temperature during sleep, or deep body temperature during sleep.

[0049] Preferably, the behavioral information J1 includes at least one of the following: information regarding the number of steps and information regarding exercise intensity. In Embodiment 1, the behavioral information J1 includes information regarding the number of steps and information regarding exercise intensity. Information regarding the number of steps and information regarding exercise intensity are types of exercise volume information. The number of steps and exercise intensity may be collectively referred to as exercise volume. Information regarding the number of steps indicates, for example, the number of steps (spm: steps per minute) within a certain period of time (e.g., 1 minute). A certain period of time in terms of steps may be referred to as a second certain period of time. Information regarding exercise intensity is indicated, for example, by METs (Metabolic equivalents). METs indicate, for example, METs over a certain period of time (e.g., 1 minute). A certain period of time in terms of METs may be referred to as a third certain period of time. METs is a unit that expresses the intensity of physical activity as how many times greater it is than the resting state. Furthermore, information regarding exercise intensity is not limited to METs; for example, it may be expressed as a relative value to maximum oxygen uptake (%VO2max) or by methods based on maximum heart rate (%HRmax, %MHR).

[0050] The systemic inflammatory response observed in rheumatoid arthritis can cause fatigue or malaise. Fatigue or malaise restricts bodily activity, leading to a decrease in step count and exercise intensity. Furthermore, joint symptoms caused by rheumatoid arthritis also restrict bodily activity, resulting in a decrease in step count and exercise intensity. Therefore, it can be inferred that a decrease in step count and exercise intensity may be causing an exacerbation of the inflammatory response and an increase in disease activity in rheumatoid arthritis. From these results, it can be inferred that there is a correlation between information on step count and exercise intensity and rheumatoid arthritis index information M1.

[0051] Furthermore, joint stiffness and pain caused by rheumatoid arthritis can also be caused by not using the joints. In other words, stiffness and pain can also be caused by insufficient exercise, such as low step count and exercise intensity. On the other hand, too much exercise can worsen joint stiffness and pain. Thus, stiffness and pain can occur even without using the joints, but they can also occur if the joints are subjected to too much stress. In other words, there is an optimal amount of exercise that maintains joint range of motion with minimal load. From these results, it can be inferred that there is a correlation between exercise, such as step count and exercise intensity, and the rheumatoid arthritis index information M1.

[0052] Furthermore, the behavioral information J1 may include, for example, one or more pieces of information from energy consumption and body movement information. Body movement information may be, for example, the amount or frequency of a specific body movement (e.g., turning over or struggling). In addition, the behavioral information J1 may include the time the learning subject wears a biological state detection device (e.g., a wearable device).

[0053] As explained above with reference to Figure 2, according to Embodiment 1, it can be inferred that information on heart rate, step count, and exercise intensity correlates with rheumatoid arthritis index information M1. Therefore, according to Embodiment 1, the learning model TM1 constructed by performing learning using the learning dataset F1 (information on heart rate, information on step count, information on exercise intensity, and rheumatoid arthritis index information M1) can output the subject's rheumatoid arthritis index information M2 with high estimation accuracy when information on the subject's heart rate, step count, and exercise intensity is input.

[0054] Furthermore, in Embodiment 1, the behavioral information J1 may preferably include the sleep information of the learning subject. The sleep information is information relating to the sleep of a living organism. The sleep information includes at least one of the following: information relating to sleep duration and information relating to sleep rhythm.

[0055] Joint pain caused by rheumatoid arthritis can make it difficult to fall asleep, potentially leading to shorter sleep durations and altered sleep rhythms. Therefore, it can be inferred that there is a correlation between sleep information and rheumatoid arthritis index information M1. As a result, according to Embodiment 1, by adding sleep information to the behavioral information J1 of the learning dataset F1, the learning model TM1 can output the subject's rheumatoid arthritis index information M2 with even higher estimation accuracy.

[0056] Furthermore, in Embodiment 1, the behavioral information J1 may preferably include the learning subject's drinking information. Drinking information is information about a living organism's drinking. Drinking information is a type of information about a living organism's lifestyle. Therefore, lifestyle information includes information about a living organism's preferences. Drinking information is, for example, information indicating whether or not a person drinks alcohol, information indicating the amount of alcohol consumed, or an indicator for estimating drinking. Details of the indicator for estimating drinking will be described later.

[0057] Studies have shown that moderate alcohol consumption is associated with lower disease activity in rheumatoid arthritis. Furthermore, studies have indicated that alcohol consumption inhibits the synthesis of inflammatory cytokines and lowers CRP levels. Therefore, alcohol consumption influences the inflammatory response and disease activity of rheumatoid arthritis. Additionally, studies have shown that habitual alcohol consumption affects the levels of blood inflammatory biomarkers (e.g., ESR, CRP). Thus, a correlation can be inferred between alcohol consumption information and rheumatoid arthritis index information M1. As a result, according to Embodiment 1, by adding alcohol consumption information to the behavioral information J1 of the learning dataset F1, the learning model TM1 can output the subject's rheumatoid arthritis index information M2 with even higher estimation accuracy.

[0058] Furthermore, in Embodiment 1, the vital information H1 may preferably include at least one of the following: physical and mental recovery capacity information and exercise responsiveness information. Physical and mental recovery capacity information is information indicating the degree of physical and mental recovery capacity from the end of exercise. Typically, physical and mental recovery capacity information is information indicating the degree of heart rate recovery capacity from the end of exercise. In this case, cardiopulmonary function, parasympathetic nervous system function (autonomic nervous system disorder), and exercise deficiency can be evaluated using physical and mental recovery capacity information. Physical and mental recovery capacity information is, for example, an index for estimating physical and mental recovery. Details of the index for estimating physical and mental recovery will be described later. On the other hand, exercise responsiveness information is information indicating the degree of physical and mental response to exercise. Typically, exercise responsiveness information is information indicating the degree of increase in heart rate after the start of exercise. In this case, the state of the sympathetic nervous system (autonomic nervous system disorder) and exercise deficiency can be evaluated using exercise responsiveness information. Note that exercise responsiveness information may also include information indicating the degree of decrease in heart rate at the end of exercise.

[0059] Rheumatoid arthritis can cause various accompanying symptoms or secondary diseases in addition to joint symptoms. For example, interstitial pneumonia as a secondary disease can adversely affect cardiopulmonary and respiratory function. For example, chronic lack of exercise due to joint pain and fatigue can lead to a decline in cardiopulmonary function. For example, severe joint symptoms can cause autonomic nervous system dysfunction. As a method to quantify the decline in cardiopulmonary and respiratory function, changes in heart rate at the start of exercise such as walking, and changes in heart rate at the end of exercise such as walking (when switching from exercise to rest) can be used. Therefore, it can be inferred that there is a correlation between the information on physical and mental recovery ability and exercise responsiveness, and the rheumatoid arthritis index information M1. As a result, according to Embodiment 1, by adding at least one piece of information from the information on physical and mental recovery ability and exercise responsiveness to the vital information H1 of the learning dataset F1, the learning model TM1 can output the subject's rheumatoid arthritis index information M2 with even higher estimation accuracy.

[0060] Furthermore, in Embodiment 1, preferably, at least one of the vital information H1 and behavioral information J1 may include vital information H11 or behavioral information J11 on which cumulative processing has been performed. Cumulative processing refers to the process of accumulating vital information H1 or behavioral information J1 while avoiding value divergence. Details of cumulative processing will be described later. As an example, cumulative processing is performed on each of vital information H1 and behavioral information J1. Vital information H11 includes information on heart rate on which cumulative processing has been performed. Vital information H11 may include at least one of the information on physical and mental recovery ability and exercise responsiveness on which cumulative processing has been performed. Behavioral information J11 includes at least one of the information on step count and exercise intensity on which cumulative processing has been performed. Behavioral information J11 may include at least one of the information on sleep and alcohol consumption on which cumulative processing has been performed.

[0061] The cumulative processing of vital information H1 allows the accumulated physical and mental state over time to be expressed quantitatively and chronologically from the perspective of the body's vital signs. The cumulative processing of behavioral information J1 allows the accumulated physical and mental state over time to be expressed quantitatively and chronologically from the perspective of the body's behavior. As a result, the vital information H11 and behavioral information J11 of the learning dataset F1 can reflect the homeostasis of the body. Homeostasis refers to the body's property of trying to maintain a constant physical and mental state in response to changes in the body's internal environment and external environment. In other words, the state of the body is not only affected by the state at a given point in time, but also by the accumulated state from the past. Therefore, the cumulative processing allows the vital information H11 and behavioral information J11 to have characteristics that approximate the mechanisms of the body that ensure homeostasis. As a result, according to Embodiment 1, the learning model TM1 can output rheumatoid arthritis index information M2 of the subject that appropriately reflects the mechanisms of the body and has even higher estimation accuracy.

[0062] In particular, by performing cumulative processing, it is possible to reflect vicious cycle and virtuous cycle trends related to rheumatoid arthritis in the training dataset F1. For example, each of the vicious and virtuous cycles can occur over periods ranging from several days to several weeks or more.

[0063] In rheumatoid arthritis, a vicious cycle of symptom exacerbation is anticipated. First, let's illustrate this vicious cycle from the perspective of vital information H1. For example, a vicious cycle occurs when inflammation and disease activity in rheumatoid arthritis gradually increase day by day. The tendency of a vicious cycle can be captured by performing cumulative processing on heart rate-related information such as heart rate and heart rate variability. For example, by performing cumulative processing on nocturnal heart rate or nocturnal heart rate variability, a persistent trend of inflammation can be captured, and consequently, the gradual worsening of pain and joint deformation can also be captured. Therefore, by performing cumulative processing on heart rate-related information in vital information H1, the vicious cycle caused by rheumatoid arthritis can be reflected in the training dataset F1. This is also true when performing cumulative processing on psychosomatic function recovery information and exercise responsiveness information related to heart rate.

[0064] Next, we will illustrate a vicious cycle from the perspective of behavioral information J1. For example, joint pain can lead to a decrease in physical activity such as walking, which can induce a depressive state. As a result, the perceived pain intensifies, further reducing activity, and cyclically worsening joint symptoms. For example, joint pain can lead to a decrease in physical activity such as walking, which narrows the range of motion of the joints. As a result, joint pain becomes more likely to occur, further reducing physical activity such as walking, and cyclically worsening joint symptoms. From these examples, by performing cumulative processing on the number of steps and exercise intensity in behavioral information J1, it is possible to reflect the trend of decreasing physical activity over, for example, two or more days, in the learning dataset F1, and consequently, the vicious cycle caused by rheumatoid arthritis. Also, for example, if pain prevents sleep, sleep duration is shortened, which can induce a depressive state or reduce activity. As a result, the amount of physical activity decreases, making it even more difficult to fall asleep, and cyclically worsening sleep quality. From these examples, by performing cumulative processing on the sleep information in behavioral information J1, it is possible to reflect the trend of worsening sleep, which is a vicious cycle caused by rheumatoid arthritis, in the learning dataset F1. Furthermore, for example, continuous or habitual alcohol consumption affects the inflammatory response and disease activity of rheumatoid arthritis. Therefore, by performing cumulative processing on the alcohol consumption information in behavioral information J1, the trend of symptom changes in rheumatoid arthritis caused by continuous or habitual alcohol consumption can be reflected in the training dataset F1.

[0065] Furthermore, just like with vicious cycles, positive trends in rheumatoid arthritis can be reflected in the training dataset F1 by performing cumulative processing on vital information H1 and / or behavioral information J1. In other words, the improvement trend in rheumatoid arthritis caused by vital information H1 and / or behavioral information J1 can be reflected in the training dataset F1.

[0066] Furthermore, as mentioned above, stiffness and pain can occur even without using the joints, but they can also occur if the joints are subjected to excessive strain. Such exercise-induced stiffness and pain may be related not only to the amount of exercise in a single day, but also to the cumulative amount of exercise over two or more days. Therefore, by performing cumulative processing on the number of steps and exercise intensity in the behavioral information J1, it is possible to reflect the trend of exercise-induced stiffness and pain over two or more days in the training dataset F1.

[0067] Furthermore, in Embodiment 1, preferably, at least one of the vital information H1 and behavioral information J1 may include vital information or behavioral information during a period indicated by specific conditions. The specific conditions indicate any of the following: conditions related to sleep, conditions related to wakefulness, and conditions related to the time of day caused by the sun. The period indicated by the conditions related to sleep is the period during sleep, a predetermined period before falling asleep (e.g., 30 minutes), or a predetermined period after falling asleep (e.g., 30 minutes). The period indicated by the conditions related to wakefulness is the period during wakefulness, a predetermined period before getting up (e.g., 30 minutes), or a predetermined period after getting up (e.g., 30 minutes). The period indicated by the conditions related to the time of day caused by the sun is the daytime period (e.g., from 5:00 to 20:00), or the nighttime period (from 0:00 to 4:00, and from 21:00 to 23:00).

[0068] In rheumatoid arthritis, systemic inflammation occurs. This inflammation exhibits diurnal variation. In particular, in rheumatoid arthritis, the response is heightened at night and during sleep. For example, inflammation can cause an increase in heart rate at night and during sleep. For example, inflammation can maintain a high heart rate immediately after falling asleep. That is, under normal circumstances, the heart rate decreases immediately after falling asleep, but when inflammation occurs, the heart rate does not decrease. For example, inflammation can cause the nighttime heart rate to be higher than the daytime heart rate (e.g., resting heart rate). Furthermore, if the inflammation in rheumatoid arthritis gradually progresses day by day, cumulative processing of nighttime heart rate or nighttime heart rate variability can capture the trend of sustained inflammation.

[0069] Furthermore, inflammation in rheumatoid arthritis (especially systemic inflammation) causes fatigue or malaise. As a result, daytime and nighttime physical activity may decrease. Rheumatoid arthritis also causes joint deformation, stiffness, and pain. Joint stiffness and pain are more likely to occur upon waking (for example, in the morning) because the viscosity of synovial fluid in the joint capsule increases. Other possible underlying factors include reduced joint flexion and extension during sleep, or reduced metabolism and circulation during sleep. Moreover, stiffness and pain can lead to decreased daytime physical activity. Furthermore, stiffness and pain can increase movements such as turning over and struggling during sleep, or increased nighttime awakenings. This can lead to increased exercise intensity during sleep or increased steps taken due to temporary awakenings. In particular, movements such as turning over and struggling increase movement during sleep before waking, while stiffness and pain are thought to reduce physical activity immediately after waking. Furthermore, if the inflammation of rheumatoid arthritis gradually progresses day by day, the tendency for persistent inflammation can be captured by performing cumulative processing on the intensity of exercise during sleep or the number of steps taken at night, as movements such as turning over and struggling during sleep, or the number of steps taken due to temporary waking, continue every day.

[0070] As a result of the above, according to Embodiment 1, by including at least one of the vital information H1 and behavioral information J1 in the learning dataset F1, changes in heart rate, step count, and / or exercise intensity that are significantly affected by rheumatoid arthritis during a specific period (specific period), the learning dataset F1 can reflect these changes. As a result, the estimation accuracy of rheumatoid arthritis index information M2 by the learning model TM1 can be further improved.

[0071] Furthermore, in Embodiment 1, the feature information G1 may preferably further include basic physical information Q1 of the learning subject. Basic physical information Q1 is basic information about the body. Basic physical information Q1 includes at least one of BMI (Body Mass Index) information and body type information. BMI information can be used as an indicator of the degree of obesity of a living organism. Body type information indicates the body type of the living organism and is, for example, weight information and / or height information. Basic physical information Q1 may also include age and / or sex information.

[0072] Disease activity and inflammatory responses in rheumatoid arthritis can be higher with increasing levels of obesity. For example, some studies have shown that obese individuals have higher CRP levels in blood tests. These studies suggest that obesity can be interpreted as a low degree of systemic inflammation. Therefore, it can be inferred that there is a correlation between basic physical information Q1 and rheumatoid arthritis index information M1. As a result, according to Embodiment 1, by adding basic physical information Q1 to the feature information G1 of the training dataset F1, the training model TM1 can output the subject's rheumatoid arthritis index information M2 with even higher estimation accuracy.

[0073] Furthermore, in Embodiment 1, the feature information G1 may preferably further include environmental information R1 of the training subject. Environmental information R1 is information indicating the environment in which the organism lives. Environmental information R1 includes, for example, seasonal information. Seasonal information may be indicated by a season name such as winter, or by a month such as January.

[0074] Symptoms of rheumatoid arthritis, such as joint pain and stiffness, are presumed to worsen when living temperatures are low (e.g., in winter) due to increased viscosity of synovial fluid within the joint capsule. Therefore, the likelihood of experiencing joint pain and stiffness varies depending on the season. Living temperatures are indicated, for example, by air temperature or room temperature. Furthermore, when living temperatures fluctuate drastically (e.g., during transitional months between seasons), various stresses on the body can disrupt autonomic nervous system function, potentially affecting the level of inflammatory responses. Therefore, it can be inferred that there is a correlation between environmental information R1 and rheumatoid arthritis index information M1. As a result, according to Embodiment 1, by adding environmental information R1 to the feature information G1 of the learning dataset F1, the learning model TM1 can output the subject's rheumatoid arthritis index information M2 with even higher estimation accuracy.

[0075] Next, with reference to Figure 3, the input information K2 input to the learning model TM1 and the output information L2 output from the learning model TM1 will be explained. Figure 3 is a diagram showing an example of input information K2 and output information L2. As shown in Figure 3, the input information K2 includes at least the subject's vital information H2 and behavioral information J2. The vital information H2 and behavioral information J2 are the same as the vital information H1 and behavioral information J1 in Figure 2, respectively.

[0076] Vital information H2 includes at least information related to heart rate. Vital information H2 may also include at least one of the following: information on physical and mental recovery ability and information on exercise responsiveness. Behavioral information J2 includes at least one of the following: information on step count and information on exercise intensity. In Embodiment 1, behavioral information J2 includes information on step count and information on exercise intensity. Behavioral information J2 may also include at least one of the following: sleep information and alcohol consumption information. Behavioral information J2 may also include the time the subject wore the biological state detection device 103 (e.g., a wearable device) (Figure 4).

[0077] The information regarding heart rate, physical and mental recovery ability, exercise responsiveness, step count, exercise intensity, sleep, and alcohol consumption in Figure 3 is the same as the information regarding heart rate, physical and mental recovery ability, exercise responsiveness, step count, exercise intensity, sleep, and alcohol consumption in Figure 2, respectively.

[0078] At least one of the vital information H2 and behavioral information J2 may include vital information H21 or behavioral information J21 on which cumulative processing has been performed. Cumulative processing refers to the process of accumulating vital information H2 or behavioral information J2 while avoiding value divergence. As an example, cumulative processing is performed on each of vital information H2 and behavioral information J2. Vital information H21 includes information on heart rate on which cumulative processing has been performed. Vital information H21 may include at least one of the information on physical and mental recovery ability and exercise responsiveness on which cumulative processing has been performed. Behavioral information J21 includes at least one of the information on step count and exercise intensity on which cumulative processing has been performed. Behavioral information J21 may include at least one of the information on sleep and alcohol consumption on which cumulative processing has been performed.

[0079] The vital information H21 and behavioral information J21 in Figure 3 are the same as the vital information H11 and behavioral information J11 in Figure 2, respectively. The cumulative processing performed on the heart rate information, physical and mental recovery capacity information, exercise responsiveness information, step count information, exercise intensity information, sleep information, and alcohol consumption information in Figure 3 is the same as the cumulative processing performed on the heart rate information, physical and mental recovery capacity information, exercise responsiveness information, step count information, exercise intensity information, sleep information, and alcohol consumption information in Figure 2, respectively.

[0080] At least one of the vital information H2 and behavioral information J2 may include vital information or behavioral information during a period indicated by specific conditions. The specific conditions refer to any of the following: conditions related to sleep, conditions related to wakefulness, and conditions related to the time of day caused by the sun. The specific conditions for vital information H2 and behavioral information J2 are the same as the specific conditions for vital information H1 and behavioral information J1 in Figure 2.

[0081] Input information K2 may include at least one of the following: basic physical information Q2 and environmental information R2. Basic physical information Q2 and environmental information R2 in Figure 3 are the same as basic physical information Q1 and environmental information R1 in Figure 2, respectively. Basic physical information Q2 includes at least one of the following: BMI information and body shape information. Environmental information R2 includes seasonal information. BMI information, body shape information, and seasonal information in Figure 3 are the same as the BMI information, body shape information, and seasonal information in Figure 2, respectively.

[0082] On the other hand, output information L2 includes rheumatoid arthritis index information M2. Rheumatoid arthritis index information M2 includes at least one of the following: information on CDAI, information on SDAI, information on DAS28, and information on joint pain.

[0083] In the explanation of Figure 2, for example, the terms "training subject," "feature information G1," "vital information H1, H11," "behavioral information J1, J11," "basic physical information Q1," "environmental information R1," "correct label B1," and "rheumatoid arthritis index information M1" can be replaced with "subject," "input information K2," "vital information H2, H21," "behavioral information J2, J21," "basic physical information Q2," "environmental information R2," "output information L2," and "rheumatoid arthritis index information M2," respectively, thereby substituting for the explanation of input information K2 and output information L2.

[0084] Typically, vital information H2 and behavioral information J2 are information created based on raw biological state data A2 acquired from a wearable device (e.g., the biological state detection device 103 described later) and / or a mobile terminal (e.g., the first terminal 102 described later).

[0085] Next, the stages of using the learning model TM1 will be described with reference to Figures 4 to 15. Figure 4 is a block diagram showing an example configuration of the estimation system 40 according to Embodiment 1. As shown in Figure 4, the estimation system 40 is connected to a network NW. The network NW includes, for example, the Internet, a private network, a public telephone network, a LAN (Local Area Network), and a short-range wireless network. The estimation system 40 is part of the information processing system 1 in Figure 1.

[0086] At least one client system 100 is connected to the network NW. The client system 100 comprises a cloud server 101, a plurality of first terminals 102, and a plurality of biological state detection devices 103. The cloud server 101, the first terminals 102, and the biological state detection devices 103 are connected to the network NW. In addition, a plurality of second terminals 200 are connected to the network NW.

[0087] The biological state detection device 103 detects the biological state of the subject. The biological state detection device 103 has, for example, a sensor that detects the biological state by contact or non-contact. The biological state detection device 103 outputs biological state raw data A2, which is raw data indicating the biological state of the subject. The biological state raw data A2 includes the subject's vital raw data and behavioral raw data. The vital raw data is raw data indicating the vital signs of the organism. The behavioral raw data is raw data indicating the behavior of the organism. The sensor that detects the vital raw data includes, for example, an optical sensor (light-emitting element and light-receiving element) that performs PCG. In this case, for example, the subject's pulse wave waveform is detected. Therefore, the biological state detection device 103 calculates the pulse rate (bpm) based on the pulse wave waveform. The sensor that detects the vital raw data may include, for example, a sensor that detects the subject's body temperature. The sensor that detects the behavioral raw data includes, for example, an acceleration sensor and / or a gyroscope sensor. In this case, for example, the subject's step count and exercise intensity (METs) are detected. In other words, the biological state detection device 103 calculates the number of steps and exercise intensity (METs) based on the output of the acceleration sensor and / or gyroscope sensor. The sensors for detecting behavioral data may include, for example, an acceleration sensor and / or gyroscope sensor, as well as a microphone. In this case, for example, the subject's sleep state is detected.

[0088] The biological state detection device 103 is, for example, a wearable device. The wearable device is attached to the subject. The wearable device is, for example, a wristwatch, a ring, or a sticker. The biological state detection device 103 is synchronized with the first terminal 102 and transmits the raw biological state data A2 to the first terminal 102. The first terminal 102 is, for example, a mobile terminal such as a smartphone. The mobile terminal is carried by the subject. The biological state detection device 103 may also transmit the raw biological state data A2 to the cloud server 101 or the estimation system 40 via the network NW.

[0089] The biological state detection device 103 may be wearable, portable, or stationary, or it may be a dedicated detector (measuring instrument) for detecting the biological state. Furthermore, the first terminal 102 may have some or all of the functions of the biological state detection device 103. Also, for example, the first terminal 102 may be a personal computer (PC).

[0090] The first terminal 102 transmits the subject's raw biological state data A2, the subject's basic physical information Q2, and the subject's environmental information R2 to the cloud server 101 via the network NW. The cloud server 101 transmits the biological state data A2, basic physical information Q2, and environmental information R2 to the estimation system 40 via the network NW. The estimation system 40 processes the biological state data A2, basic physical information Q2, and environmental information R2.

[0091] Alternatively, without providing a cloud server 101, the first terminal 102 may transmit the biological state raw data A2, basic physical information Q2, and environmental information R2 to the estimation system 40 via the network NW. Furthermore, the basic physical information Q2 and environmental information R2 may be transmitted from the biological state detection device 103 to the first terminal 102, the cloud server 101, or the estimation system 40.

[0092] More specifically, the estimation system 40 comprises a relay server 44, a first database 45, an estimation device 4, a second database 46, and an information providing server 47. The relay server 44, the first database 45, the estimation device 4, the second database 46, and the information providing server 47 are connected to a network NW. The estimation device 4 is, for example, a server.

[0093] Each of the relay server 44 and the information providing server 47 may include a processing unit, a communication unit, and a storage unit, and may also include an input unit and an output unit. The hardware configuration of the processing unit, communication unit, storage unit, input unit, and output unit is the same as the hardware configuration of the processing unit 400, communication unit 401, storage unit 402, input unit 403, and output unit 404 of the estimation device 4 shown in Figure 6, which will be described later. In addition, the first database 45 and the second database 46 include at least a storage device such as a hard disk drive. The first database 45 and the second database 46 may have the same hardware configuration as the relay server 44 or the information providing server 47. The estimation system 40 may not include all or part of the relay server 44, the first database 45, the second database 46, and the information providing server 47.

[0094] Figure 5 is a flowchart showing an example of an estimation method performed by the estimation system 40. The estimation method estimates the subject's rheumatoid arthritis index information MX. The estimation method includes steps S1 to S5.

[0095] As shown in Figures 4 and 5, first, in step S1, the relay server 44 receives the subject's biological state raw data A2, the subject's basic physical information Q2, and the subject's environmental information R2 from the cloud server 101. The relay server 44 implements, for example, an API (Application Programming Interface).

[0096] Next, in step S2, the first database 45 stores the biological state raw data A2, basic body information Q2, and environmental information R2 received by the relay server 44.

[0097] Next, in step S3, the estimation device 4 uses the biological state raw data A2, basic physical information Q2, and environmental information R2 stored in the first database 45, along with the learning model TM1, to estimate the subject's rheumatoid arthritis index information MX. Details of the estimation process will be described later.

[0098] Next, in step S4, the second database 46 stores the rheumatoid arthritis index information MX of the subject estimated by the estimation device 4.

[0099] Next, in step S5, the information provision server 47 transmits the subject's rheumatoid arthritis index information MX, stored in the second database 46, to the second terminal 200 and / or the first terminal 102 via the network NW. Specifically, the information provision server 47 displays the subject's rheumatoid arthritis index information MX on the second terminal 200 and / or the first terminal 102. The second terminal 200 is, for example, a terminal of a medical institution, a health management institution, or a research institution. The first terminal 102 is, for example, the subject's terminal. Once step S5 is completed, the estimation method is finished.

[0100] In the example shown in Figure 4, the rheumatoid arthritis index information MX is transmitted to the first terminal 102 via the cloud server 101, but it may also be transmitted directly to the first terminal 102.

[0101] Figure 6 is a block diagram showing an example configuration of the estimation device 4 shown in Figure 4. As shown in Figure 6, the estimation device 4 comprises a processing unit 400, a communication unit 401, and a storage unit 402. The estimation device 4 may further comprise an input unit 403 and an output unit 404.

[0102] The input unit 403 is an input device for inputting various types of information to the processing unit 400. For example, the processing unit 400 may be a keyboard and pointing device, or a touch panel.

[0103] The output unit 404 outputs various types of information. The output unit 404 includes, for example, a display unit that displays various types of information. The display unit is, for example, a liquid crystal display or an organic electroluminescent display.

[0104] The communication unit 401 is connected to a network NW. The communication unit 401 communicates with external devices connected to the network NW. The communication unit 401 is a communication device that performs communication according to a predetermined communication protocol, and includes, for example, a network interface controller. The predetermined communication protocol is, for example, a protocol compliant with Ethernet® and an Internet Protocol Suite. The external devices are, for example, a first terminal 102, a second terminal 200, and a biological state detection device 103.

[0105] The storage unit 402 includes a storage device and stores data and computer programs. The storage unit 402 includes a main storage device such as a semiconductor memory and an auxiliary storage device such as a semiconductor memory and a hard disk drive. The storage unit 402 may also include a removable medium such as an optical disc. The storage unit 402 may be, for example, a non-temporary computer-readable storage medium.

[0106] The memory unit 402 stores the learning model TM1. The learning model TM1 is a trained model. The learning model TM1 is a computer program. The learning model TM1 causes the computer to function in order to estimate the subject's rheumatoid arthritis index information M2. Specifically, the learning model TM1 takes input information K2 as input and causes the computer to output output information L2.

[0107] The processing unit 400 performs various calculations. The processing unit 400 includes processors such as a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit).

[0108] Specifically, the processing unit 400 includes a pre-processing unit 41 and an estimation unit 42. The processing unit 400 may further include a post-processing unit 43. For example, the processing unit 400 functions as the pre-processing unit 41, estimation unit 42, and post-processing unit 43 by executing a computer program stored in the memory unit 402. The pre-processing unit 41 preferably includes a statistical processing unit 410, a condition processing unit 411, a sleep processing unit 412, and an accumulation processing unit 416. The pre-processing unit 41 may further include at least one of an alcohol consumption estimation unit 413, a mental and physical recovery ability estimation unit 414, and an exercise responsiveness estimation unit 415.

[0109] The preprocessing unit 41 acquires biological state raw data A2, basic physical information Q2, and environmental information R2 from the first database 45. The storage unit 402 stores the biological state raw data A2, basic physical information Q2, and environmental information R2. The vital raw data of biological state raw data A2 includes, for example, heart rate data. The behavioral raw data of biological state raw data A2 includes, for example, step count data, exercise intensity data, and sleep data.

[0110] Heart rate data includes information on heart rate (bpm). Heart rate is measured or aggregated, for example, at a sampling rate of at least once per minute. Heart rate data is accompanied by the corresponding measurement date, time, and subject identification information. Heart rate data includes heart rate information arranged in a time series.

[0111] Step count data includes information on the number of steps (spm). Step counts are measured or aggregated, for example, at a sampling rate of at least once per minute. Step count data is accompanied by the corresponding measurement date, time, and subject identification information. Step count data includes step count information arranged in chronological order.

[0112] Exercise intensity data includes METs information per unit of time (1 minute). In other words, exercise intensity is typically expressed in METs. Exercise intensity data is measured or aggregated, for example, at a sampling rate of at least once per minute. Exercise intensity data is accompanied by the corresponding measurement date, time, and subject identification information. Exercise intensity data includes METs information arranged in a time series.

[0113] The sleep data includes information on the start time (day, hour, minute) and end time (day, hour, minute, second) for each sleep period, as detected by the algorithm of the biological state detection device 103. The sleep data is accompanied by the date on which the sleep start time and end time were measured, and the subject's identification information.

[0114] The preprocessing unit 41 performs preprocessing on the biological state raw data A2 and generates input information K2, which is the result of the preprocessing. The storage unit 402 stores the input information K2.

[0115] As an example, the statistical processing unit 410 of the preprocessing unit 41 performs statistical processing on each of the heart rate data, step count data, and exercise intensity data for each unit period. As a result, the statistical processing unit 410 outputs multiple statistical indicators, which are the results of the statistical processing, for each of the heart rate data, step count data, and exercise intensity data. The statistical processing by the statistical processing unit 410 represents mathematical processing to quantitatively express the characteristics and trends of the data being processed. The unit period is, for example, one day. The heart rate data, step count data, and exercise intensity data are, in detail, extracted heart rate data, extracted step count data, and extracted exercise intensity data, which will be described later. In addition, the statistical processing unit 410 may perform difference processing and percentage processing for each unit period. Details of this will be described later.

[0116] Statistical indicators include, for example, the mean, sum, standard deviation, xth percentile, median, minimum, maximum, RMSSD, coefficient of variation CV (= standard deviation / mean), reciprocal of the coefficient of variation CV, mean change, percentage of changes greater than or equal to a specified value, and statistically processed values ​​of the moving median over the time window. Furthermore, RMSSD is calculated not only for heart rate (bpm) but also for steps per minute (SPM) and exercise intensity (METs / min). In this case, RMSSD is the square root of the mean of the squared differences between consecutive adjacent values. Consecutive adjacent values ​​are, for example, the reciprocal of heart rate (bpm), consecutive adjacent steps per minute (SPM), or consecutive adjacent exercise intensity (METs / min).

[0117] The statistical processing unit 410 may generate all or some of the multiple types of statistical indicators.

[0118] As an example, the condition processing unit 411 of the preprocessing unit 41 performs processing (condition processing) according to specific conditions for each of the vital raw data and behavioral raw data of the biological state raw data A2 for each unit period.

[0119] Specifically, the condition processing unit 411 performs the process of extracting vital data from the subject's raw vital data for a period indicated by specific conditions, and / or the process of extracting behavioral data from the subject's raw behavioral data for a period indicated by specific conditions. This will be explained in detail below.

[0120] First, let's explain the processing of heart rate data from the vital data. The condition processing unit 411 extracts heart rate data for a period indicated by a specific condition from the heart rate data for a unit period. The specific condition includes at least one of the following conditions: conditions related to sleep, conditions related to wakefulness, and conditions related to the time of day caused by the sun.

[0121] For example, the condition processing unit 411 extracts heart rate data during wakefulness, during sleep, during the day, at night, at a predetermined time before falling asleep, at a predetermined time after falling asleep, at a predetermined time before waking up, and at a predetermined time after waking up from the heart rate data for a unit period. Hereinafter, the heart rate data extracted according to specific conditions may be referred to as "extracted heart rate data".

[0122] The condition processing unit 411 passes multiple types of extracted heart rate data for a unit period to the statistical processing unit 410. The statistical processing unit 410 performs different statistical processing for each of the extracted heart rate data and outputs multiple different statistical indicators.

[0123] Furthermore, the statistical processing unit 410 calculates the difference (hereinafter referred to as the day-night difference) between the statistical indicator obtained by statistically processing the "daytime heart rate data" and the statistical indicator obtained by statistically processing the "nighttime heart rate data" for each different statistical indicator (difference processing).

[0124] As described above, multiple types of statistical indicators and multiple types of diurnal differences are calculated for each of the multiple types of extracted heart rate data for each unit period, and these constitute a part of the input information K2. In this case, the statistical indicators and diurnal differences are sometimes collectively referred to as heart rate statistical indicators. The storage unit 402 stores the multiple heart rate statistical indicators as part of the input information K2.

[0125] The heart rate statistical index is an example of vital information H2 from the input information K2. The extracted heart rate data is an example of "extracted raw vital data." Therefore, the statistical processing unit 410 generates vital information H2 by performing statistical processing on the extracted raw vital data.

[0126] The statistical processing unit 410 and the condition processing unit 411 may generate all or some of the multiple types of heart rate statistical indices.

[0127] Next, we will explain the processing of step count data from the raw behavioral data. The condition processing unit 411 extracts step count data for a period indicated by a specific condition from the step count data for a unit period. The specific condition in this case is the same as the specific condition for heart rate data.

[0128] For example, the condition processing unit 411 extracts step count data during wakefulness, step count data during sleep, step count data during the day, step count data at night, step count data for a predetermined time before falling asleep, step count data for a predetermined time after falling asleep, step count data for a predetermined time before waking up, and step count data for a predetermined time after waking up from the step count data for a unit period. Hereinafter, the step count data extracted according to specific conditions may be referred to as "extracted step count data".

[0129] The condition processing unit 411 passes multiple types of extracted step count data for a unit period to the statistical processing unit 410. The statistical processing unit 410 performs different statistical processing for each of the extracted step count data and outputs multiple statistical indicators.

[0130] Furthermore, the statistical processing unit 410 calculates the difference (hereinafter referred to as the day-night difference) between the statistical indicator obtained by statistically processing the "daytime step count data" and the statistical indicator obtained by statistically processing the "nighttime step count data" for each different statistical indicator (difference processing).

[0131] Furthermore, the statistical processing unit 410 calculates the ratio of the period in which the number of steps is zero to the period indicated by the specific conditions (hereinafter referred to as the occurrence ratio) (ratio processing).

[0132] For example, the statistical processing unit 410 calculates the incidence rate during wakefulness, during sleep, during the day, at night, a predetermined time before falling asleep, a predetermined time after falling asleep, a predetermined time before waking up, and a predetermined time after waking up.

[0133] As described above, multiple types of statistical indicators, multiple types of day-night differences, and multiple types of occurrence rates are calculated for each of the multiple types of extracted step count data for each unit period, and constitute a part of the input information K2. In this case, the statistical indicators, day-night differences, and occurrence rates are sometimes collectively referred to as step count statistical indicators. The storage unit 402 stores the multiple step count statistical indicators as part of the input information K2.

[0134] The step count statistics index is an example of behavioral information J2 from input information K2. The extracted step count data is also an example of "extracted raw behavioral data." Therefore, the statistical processing unit 410 generates behavioral information J2 by ​​performing statistical processing on the extracted raw behavioral data.

[0135] The statistical processing unit 410 and the condition processing unit 411 may generate all or some of the multiple types of step count statistical indicators.

[0136] Next, we will explain the processing of exercise intensity data from the raw behavioral data. The condition processing unit 411 extracts exercise intensity data for a period indicated by a specific condition from the exercise intensity data for a unit period. The specific condition in this case is the same as the specific condition for heart rate data.

[0137] For example, the condition processing unit 411 extracts exercise intensity data from exercise intensity data for a unit period, including exercise intensity data during wakefulness, exercise intensity data during sleep, daytime exercise intensity data, nighttime exercise intensity data, exercise intensity data for a predetermined time before falling asleep, exercise intensity data for a predetermined time after falling asleep, exercise intensity data for a predetermined time before waking up, and exercise intensity data for a predetermined time after waking up. Hereinafter, the exercise intensity data extracted according to specific conditions may be referred to as "extracted exercise intensity data".

[0138] The condition processing unit 411 passes multiple types of extracted exercise intensity data for a unit period to the statistical processing unit 410. The statistical processing unit 410 performs different statistical processing for each of the extracted exercise intensity data and outputs multiple statistical indicators.

[0139] Furthermore, the statistical processing unit 410 calculates the difference (hereinafter referred to as the day-night difference) between the statistical indicator obtained by statistically processing the "daytime exercise intensity data" and the statistical indicator obtained by statistically processing the "nighttime exercise intensity data" for each different statistical indicator (difference processing).

[0140] Furthermore, the statistical processing unit 410 calculates the ratio of the period in which METs is less than or equal to a first specified value (hereinafter referred to as the first occurrence ratio) to the period indicated by the specific conditions (ratio processing). The condition processing unit 411 calculates the ratio of the period in which METs is greater than or equal to a second specified value (hereinafter referred to as the second occurrence ratio) to the period indicated by the specific conditions (ratio processing).

[0141] For example, the statistical processing unit 410 calculates the first incidence rate and the second incidence rate for each of the following periods: while awake, while asleep, during the day, at night, a predetermined time before falling asleep, a predetermined time after falling asleep, a predetermined time before waking up, and a predetermined time after waking up.

[0142] As described above, multiple types of statistical indicators, multiple types of diurnal differences, and multiple types of occurrence rates are calculated for each of the multiple types of extracted exercise intensity data for each unit period, and constitute a part of the input information K2. In this case, the statistical indicators, diurnal differences, and occurrence rates are sometimes collectively referred to as exercise intensity statistical indicators. The storage unit 402 stores the multiple exercise intensity statistical indicators as part of the input information K2.

[0143] The exercise intensity statistical index is an example of behavioral information J2 from the input information K2. The extracted exercise intensity data is also an example of "extracted raw behavioral data." Therefore, the statistical processing unit 410 generates behavioral information J2 by ​​performing statistical processing on the extracted raw behavioral data.

[0144] The statistical processing unit 410 and the condition processing unit 411 may generate all or some of the multiple types of exercise intensity statistical indices.

[0145] As described above, when calculating the step count statistical index and the exercise intensity statistical index, statistical processing is performed by the statistical processing unit 410 after processing by the condition processing unit 411. Therefore, according to Embodiment 1, statistical processing can be performed during a period that accurately reflects the activity or symptoms of rheumatoid arthritis. As a result, the estimation accuracy of the rheumatoid arthritis index information M2 by the learning model TM1 can be further improved.

[0146] As described above, the statistical processing unit 410 generates vital information H2 by performing statistical processing on the raw vital data extracted by the condition processing unit 411, and / or generates behavioral information J2 by ​​performing statistical processing on the raw behavioral data extracted by the condition processing unit 411.

[0147] Next, we will explain the processing of sleep data from behavioral data. As an example, the sleep processing unit 412 of the preprocessing unit 41 calculates several types of sleep indices based on sleep data over a unit period. Sleep indices are indicators related to the sleep of living organisms.

[0148] For example, the sleep processing unit 412 calculates sleep indicators from the first to the fifth sleep indicator. The first sleep indicator is the sleep duration during the target unit period. The second sleep indicator is the time of falling asleep during the target unit period. The third sleep indicator is the time of waking up during the target unit period. The fourth sleep indicator is the midpoint time during the target unit period. The midpoint time is the time midway between the time of falling asleep and the time of waking up. The fifth sleep indicator is information on the number of sleep cycles during the target unit period.

[0149] As described above, multiple types of sleep indicators are calculated for each unit period and constitute a part of the input information K2. The memory unit 402 stores the multiple sleep indicators as part of the input information K2. The sleep indicators are an example of the behavioral information J2 of the input information K2.

[0150] The sleep processing unit 412 may generate all or some of the multiple types of sleep indicators.

[0151] As an example, the alcohol consumption estimation unit 413 of the preprocessing unit 41 generates alcohol consumption information for each unit period based on the heart rate data and step count data of the vital raw data. For example, the alcohol consumption estimation unit 413 calculates an alcohol consumption estimation index. The alcohol consumption estimation index is an index for estimating alcohol consumption, calculated based on heart rate excluding heart rate during walking.

[0152] Specifically, the alcohol consumption estimation unit 413 calculates a unit alcohol consumption index. The unit alcohol consumption index is defined using the total heart rate change, a first function, and a second function.

[0153] The total heart rate change is calculated by summing up the heart rate changes within a third time period (e.g., within 80 minutes). The heart rate change is the difference between the subject's heart rate within a first time period (e.g., within 1 minute) and the subject's heart rate within a second time period following the first time period (e.g., within 1 minute), calculated based on the subject's heart rate (bpm) measured at first fixed time intervals (e.g., every 1 minute). The heart rate change is calculated at predetermined time intervals (e.g., every 1 minute).

[0154] The total change in heart rate is calculated by shifting the starting point of the third time period by a predetermined amount of time (for example, one minute at a time). Therefore, the total change in heart rate is calculated at predetermined time intervals. Since heart rate increases with alcohol consumption, the total change in heart rate is a useful indicator for estimating alcohol consumption.

[0155] The first function is defined based on the number of steps (spm) measured at two fixed time intervals (for example, every minute) and indicates whether the subject is walking or not. For example, at predetermined time intervals (for example, every predetermined time interval), if the number of steps is 0, the first function is set to "1", and if the number of steps is 1 or more, the second function is set to "0". In this example, the value of the first function is obtained at predetermined time intervals.

[0156] The second function indicates whether the representative heart rate exceeds a baseline heart rate (e.g., 80) that is estimated to indicate possible intoxication. The representative heart rate is a representative heart rate selected within a fourth time period (e.g., within 40 minutes) based on the heart rate (bpm). The representative heart rate can be, for example, the minimum, mode, median, or mean. The representative heart rate is calculated by shifting the starting point of the fourth time period by a predetermined amount of time (e.g., 1 minute at a time). Therefore, the representative heart rate is calculated at predetermined time intervals. Thus, for example, at predetermined time intervals, if the representative heart rate exceeds the baseline heart rate, the second function is set to "1", and if the representative heart rate is less than or equal to the baseline heart rate, the second function is set to "0". In this example, the value of the second function is obtained at predetermined time intervals.

[0157] The "third time" and the "fourth time" are longer than the "first fixed time," the "second fixed time," the "first time," and the "second time."

[0158] Let the unit alcohol consumption index be "f(t)", the total change in heart rate be "SumΔHR(t)", the first function be "L_ST(t)", and the second function be "L_DR(t)". ​​t represents a unit of time, where one unit is a predetermined time (for example, 1 minute). In this case, the unit alcohol consumption index is given by equation (5).

[0159] f(t)=SumΔHR(t)×L_ST(t)×L_DR(t)…(5)

[0160] In equation (5), L_ST(t) is 0 during walking. Therefore, in the unit alcohol consumption index, it is possible to consider the potential immediate effect that the subject's walking (number of steps) may have on the subject's heart rate. Also, in equation (5), if SumΔHR(t) is a negative value, f(t) is set to 0.

[0161] In Embodiment 1, the alcohol consumption estimation unit 413 performs statistical processing on the unit alcohol consumption index for one or more target time periods within a unit period and calculates an alcohol consumption estimation index as a result of the statistical processing. In this case, the statistical processing is, for example, a process to calculate the sum of the unit alcohol consumption index, or a process to calculate the average value of the unit alcohol consumption index. In this case, the value of the alcohol consumption parameter may be changed for each target time period to calculate the alcohol consumption estimation index. The alcohol consumption parameter is, for example, one or more parameters selected from the third time, the fourth time, and the reference heart rate.

[0162] In Embodiment 1, as an example, the alcohol consumption estimation unit 413 calculates four types of alcohol consumption estimation indicators for each of two different target time periods, using two different values ​​of the same type of alcohol consumption parameter. The four types of alcohol consumption estimation indicators are calculated for each unit period and constitute a part of the input information K2. The storage unit 402 stores the multiple alcohol consumption estimation indicators as part of the input information K2. The alcohol consumption estimation indicators are an example of the behavioral information J2 of the input information K2.

[0163] The alcohol consumption estimation unit 413 may generate all or some of the multiple types of alcohol consumption estimation indicators.

[0164] As an example, the mental and physical recovery capacity estimation unit 414 of the preprocessing unit 41 generates mental and physical recovery capacity information for each unit period based on vital raw data (heart rate data) and behavioral raw data (e.g., exercise intensity data). For example, the mental and physical recovery capacity estimation unit 414 generates a mental and physical recovery estimation index. The mental and physical recovery estimation index indicates the degree of the subject's heart rate recovery capacity from the end of exercise.

[0165] The index for estimating physical and mental recovery includes a first parameter value and a second parameter value that realize a model function that approximates the relationship between the change in heart rate and the corresponding heart rate. The change in heart rate is the difference between the subject's heart rate (bpm) at first constant time intervals (e.g., every minute) during a fifth period of time (e.g., within one minute) in which the subject is exercising and the heart rate during a sixth unit time period (e.g., within one minute) following the fifth period in which the subject is at rest. The corresponding heart rate indicates the heart rate during the fifth period (e.g., within one minute). For example, if the subject's exercise intensity is 1 MET, the subject is judged to be in a resting state, and if the subject's exercise intensity exceeds 1 MET, the subject is judged to be in an exercise state.

[0166] The model function includes a regression line with a slope and an intercept. The magnitude of the slope represents the heart rate's recovery capacity. The intercept represents the magnitude of the heart rate requiring recovery. The first parameter value indicates the slope of the regression line. The second parameter value indicates the intercept of the regression line.

[0167] In Embodiment 1, the mental and physical recovery capacity estimation unit 414 calculates mental and physical recovery estimation indices (first parameter value and second parameter value) for each unit period, specifically for the nighttime period and the daytime period (periods other than the nighttime period). In addition, the mental and physical recovery capacity estimation unit 414 calculates the difference between the first parameter value during the nighttime period and the first parameter value during the daytime period (hereinafter referred to as the day-night difference), and the difference between the second parameter value during the nighttime period and the second parameter value during the daytime period (hereinafter referred to as the day-night difference).

[0168] As described above, two types of parameter values ​​(four types in total) for two different time periods, and two types of diurnal differences are calculated for each unit period and constitute a part of the input information K2. In this case, the parameter values ​​and diurnal differences are sometimes collectively referred to as the mental and physical recovery index. The memory unit 402 stores multiple mental and physical recovery indices as part of the input information K2. The mental and physical recovery index is an example of the vital information H2 of the input information K2.

[0169] The mental and physical recovery ability estimation unit 414 may generate all or some of the multiple types of mental and physical recovery indicators.

[0170] As an example, the exercise response estimation unit 415 of the preprocessing unit 41 generates exercise response information for each unit period based on vital raw data (heart rate data) and behavioral raw data (e.g., exercise intensity data). For example, the exercise response estimation unit 415 generates a first exercise response estimation index and a second exercise response estimation index. The first exercise response estimation index is an index that shows the degree of increase in heart rate after the start of exercise. The second exercise response estimation index is an index that shows the degree of decrease in heart rate at the end of exercise.

[0171] In Embodiment 1, the motor responsiveness estimation unit 415 calculates a first motor responsiveness estimation index and a second motor responsiveness estimation index for each unit period, for the nighttime period and the daytime period (periods other than the nighttime period). In addition, the motor responsiveness estimation unit 415 calculates the difference between the first motor responsiveness estimation index between the nighttime period and the daytime period (hereinafter referred to as the day-night difference), and the difference between the second motor responsiveness estimation index between the nighttime period and the daytime period (hereinafter referred to as the day-night difference).

[0172] As described above, two types of exercise responsiveness estimation indices (a total of four types) for two different time periods, and two types of diurnal differences, are calculated for each unit period and constitute part of the input information K2. In this case, the exercise responsiveness estimation indices and diurnal differences are sometimes collectively referred to as exercise responsiveness indices. The memory unit 402 stores multiple exercise responsiveness indices as part of the input information K2. The exercise responsiveness indices are an example of vital information H2 in the input information K2.

[0173] The motor responsiveness estimation unit 415 may generate all or some of the multiple types of motor responsiveness indices.

[0174] The accumulation processing unit 416 of the preprocessing unit 41 executes the accumulation process.

[0175] Specifically, the cumulative processing unit 416 performs cumulative processing on the heart rate statistical indicators for each different type of heart rate statistical indicator during the observation period, and outputs cumulative data for each heart rate statistical indicator.

[0176] The cumulative processing unit 416 performs cumulative processing on the step count statistics for each different type of step count statistics during the observation period and outputs cumulative data for each step count statistics.

[0177] The cumulative processing unit 416 performs cumulative processing on the exercise intensity statistical indicators for each different type of exercise intensity statistical indicator during the observation period, and outputs cumulative data for each exercise intensity statistical indicator.

[0178] The cumulative processing unit 416 performs cumulative processing on the sleep indicators for each different type of sleep indicator during the observation period and outputs cumulative data for each sleep indicator.

[0179] The cumulative processing unit 416 performs cumulative processing on each different type of alcohol consumption estimation indicator for the observation period and outputs cumulative data for each alcohol consumption estimation indicator.

[0180] The cumulative processing unit 416 performs cumulative processing on the mental and physical recovery indicators for each different type of mental and physical recovery indicator during the observation period, and outputs cumulative data for each mental and physical recovery indicator.

[0181] The cumulative processing unit 416 performs cumulative processing on the motor responsiveness index for each different type of motor responsiveness index during the observation period and outputs cumulative data for each motor responsiveness index.

[0182] The observation period subject to cumulative processing may be a fixed value or may be changed dynamically.

[0183] As described above, the multiple types of cumulative data calculated for multiple types of indicators constitute a part of the input information K2. The memory unit 402 stores the multiple types of cumulative data as part of the input information K2. The cumulative data for heart rate statistics, physical and mental recovery indicators, and exercise responsiveness indicators are examples of "vital information H21 on which cumulative processing has been performed" in the input information K2. The cumulative data for step count statistics, exercise intensity statistics, sleep indicators, and alcohol consumption estimation indicators are examples of "behavioral information J21 on which cumulative processing has been performed" in the input information K2. Hereinafter, cumulative data may be referred to as cumulative indicators.

[0184] The cumulative processing unit 416 may perform cumulative processing on all of the multiple types of indicators, or it may perform cumulative processing on some of them.

[0185] The details of the cumulative processing unit 416 will now be explained. In this case, the indicator for a unit period (for example, one day) is referred to as unit data. That is, the unit data is the heart rate statistical indicator, step count statistical indicator, exercise intensity statistical indicator, sleep indicator, alcohol consumption estimation indicator, physical and mental recovery indicator, or exercise responsiveness indicator for the unit period. The heart rate statistical indicator, physical and mental recovery indicator, and exercise responsiveness indicator are each vital data. The step count statistical indicator, exercise intensity statistical indicator, sleep indicator, and alcohol consumption indicator are each behavioral data. Therefore, the unit data is either vital data or behavioral data.

[0186] The cumulative processing includes classification processing, correction processing, and cumulative processing. Furthermore, multiple unit data points are arranged in a time series during the observation period.

[0187] The cumulative processing unit 416 performs classification execution processing. The classification execution processing involves classifying multiple unit data arranged in a time series into at least a first state group representing a first state of mind and body and a second state group representing a second state of mind and body, thereby obtaining multiple first state data belonging to the first state group and multiple second state data belonging to the second state group.

[0188] Next, the cumulative processing unit 416 performs a correction execution process. The correction execution process involves calculating multiple time-series correction data by adjusting the balance of magnitudes between multiple first state data and multiple second state data.

[0189] Next, the cumulative processing unit 416 performs cumulative execution processing. Cumulative execution processing involves calculating multiple time-series cumulative data by accumulating correction data along the time axis.

[0190] As described above, according to Embodiment 1, the correction execution process is a process that adjusts the balance of magnitude between the first state data and the second state data, which represent different states. Therefore, when the accumulation processing unit 416 accumulates the correction data, it is possible to suppress the divergence of the accumulated data. By performing a correction that suppresses divergence, that is, a correction that adjusts the balance, the accumulated data can be given characteristics that approximate the mechanisms of living organisms that ensure homeostasis. In other words, it is possible to derive accumulated data that reflects the homeostasis of living organisms.

[0191] Furthermore, in Embodiment 1, cumulative data is calculated by accumulating multiple correction data representing the first and second states of the mind and body on a time axis. Therefore, the cumulative data reflects the accumulated states of the mind and body (first and second states) over time. As a result, the cumulative data quantitatively and temporally represents the accumulated states of the mind and body (first and second states) over time.

[0192] For example, living organisms maintain homeostasis by having a biological mechanism that repeats between a first state such as deterioration and a second state such as improvement, thereby attempting to keep the state of mind and body constant. Therefore, if the accumulated biological data diverges to one side of either the first or second state, the correlation with the biological mechanism is lost. In Embodiment 1, therefore, the divergence of the accumulated data is suppressed by a correction execution process, and the correlation between the accumulated data and the biological mechanism is ensured. Furthermore, the accumulated data quantitatively and temporally represents the first state such as deterioration and the second state such as improvement that have accumulated over time.

[0193] Referring to Figures 7 to 11, we will now explain the cumulative processing unit 416 in more detail. First, we will explain the classification execution process. Figure 7 is a graph showing an example of multiple unit data d1 arranged in a time series. Figure 8 is a graph for explaining the classification execution process. In Figures 7 and 8, the vertical axis shows the value of the unit data d1 (arbitrary unit). For ease of understanding, a scale is shown on the vertical axis. q1 to q4 represent positive real numbers. The horizontal axis shows time t (for example, days). The unit of time t is a unit period (for example, 1 day). This is the same in Figures 8 to 11.

[0194] As shown in Figures 7 and 8, the cumulative processing unit 416 classifies multiple unit data d1 arranged in time series during the observation period T1 into a first state group g1 representing the first state of mind and body and a second state group g2 representing the second state of mind and body. In Figures 8 to 10, black circles represent the first state, and white diamonds represent the second state.

[0195] The first state is different from the second state. Specifically, the first state represents a mental and physical state that is incompatible with the second state. For example, the first state may represent a mental and physical state opposite to that of the second state.

[0196] Specifically, the first state data b1 of the first state group g1 has one of two signs: positive or negative. The second state data b2 of the second state group g2 has the other of two signs: positive or negative. Figure 9 is a graph showing an example of the first state data b1 and second state data b2 arranged in a time series. The horizontal axis represents time t (for example, days), and the vertical axis represents the relative value of the unit data d1 (= d1 - c1). For ease of understanding, a scale is shown on the vertical axis. p represents a positive real number. In the example in Figure 9, the first state data b1 is a positive value with a positive sign. The second state data b2 is a negative value with a negative sign.

[0197] More specifically, as shown in Figure 7, the cumulative processing unit 416 calculates first classification criterion information c1 based on multiple unit data d1 during the observation period T1. The first classification criterion information c1 is, for example, the mean, median, mode, or percentile value of the multiple unit data d1 during the observation period T1. In the example in Figure 7, the first classification criterion information c1 shows the median.

[0198] As shown in Figure 8, the cumulative processing unit 416 compares the first classification criterion information c1 with the unit data d1 and classifies the unit data d1 based on the comparison result. As an example, the cumulative processing unit 416 subtracts the first classification criterion information c1 from each unit data d1. Then, as shown in Figure 9, the cumulative processing unit 416 classifies each unit data d1 into either the first state group g1 or the second state group g2 depending on whether each subtraction result (= d1 - c1) is positive or negative. In Figure 9, the cumulative processing unit 416 sets the subtraction result showing a positive value as the first state data b1 and the subtraction result showing a negative value as the second state data b2.

[0199] The cumulative processing unit 416 may also subtract each unit data d1 from the first classification criterion information c1.

[0200] Next, the correction execution process will be explained. As shown in Figure 9, the cumulative processing unit 416 calculates multiple time-series correction data d2 by performing a correction execution process on multiple first state data b1 and multiple second state data b2 during the observation period T1. Figure 10 is a graph showing an example of correction data d2 arranged in a time series. The horizontal axis represents time t (for example, days), and the vertical axis represents the corrected relative value of the unit data. For ease of understanding, a scale is shown on the vertical axis as an example. p represents a positive real number. This is the same as in Figure 9.

[0201] Returning to Figure 9, the correction execution process is a process that adjusts the balance of magnitude between the multiple first state data b1 and the multiple second state data b2. In other words, the correction execution process is a process that adjusts the ratio R (=FS2 / FS1) of the sum of the absolute values ​​of the multiple second state data b2 to the sum of the absolute values ​​of the multiple first state data b1 FS1 during the observation period T1. Preferably, the correction execution process represents a process that makes the ratio R (=FS2 / FS1) during the observation period T1 substantially "1".

[0202] In this preferred example, the correction execution process includes a process for calculating a ratio R and a process for correcting the first state data b1 or the second state data b2 based on the ratio R. That is, the cumulative processing unit 416 calculates the ratio R during the observation period T1. Then, for example, if the total value FS1 is greater than the total value FS2, the cumulative processing unit 416 corrects each second state data b2 during the observation period T1 by multiplying it by the reciprocal of the ratio R (1 / R). Thus, the cumulative processing unit 416 obtains a plurality of corrected second state data b2# as a result of the multiplication. As a result, after the correction execution process, the ratio R becomes "1". As shown in Figure 10, the plurality of first state data b1 and the plurality of corrected second state data b2# constitute a plurality of corrected data d2.

[0203] For example, if the total value FS1 is greater than the total value FS2, the cumulative processing unit 416 may multiply each first state data b1 in the observation period T1 by a ratio R to obtain a plurality of corrected first state data b1# (not shown) as a result of multiplication. In this case, the plurality of corrected first state data b1# and the plurality of second state data b2 constitute a plurality of corrected data d2.

[0204] Next, the cumulative execution process will be explained. As shown in Figure 10, the cumulative processing unit 416 calculates multiple time-series cumulative data d3 by accumulating time-series correction data d2 along the time axis. Figure 11 is a graph showing an example of cumulative data d3 arranged in a time series. The horizontal axis represents time t (e.g., days), and the vertical axis represents the cumulative value of the corrected relative value of the unit data d1. For ease of understanding, a scale is shown on the vertical axis. m represents a positive real number. The correction data d2 to be accumulated is the correction data d2 during the observation period T1. Therefore, the cumulative processing unit 416 calculates cumulative data d3 at each time t (e.g., each day) during the observation period T1. In other words, the cumulative data d3 at each time t (e.g., each day) represents the cumulative value of the correction data d2 from the start of accumulation (e.g., the start date of accumulation) to that time t (e.g., that day).

[0205] As an example, the cumulative processing unit 416 calculates the cumulative sum d3j as cumulative data d3 using equation (6). d3j represents the cumulative sum at time t=j during the observation period T1. d2i represents the corrected data d2 at time t=i. ts represents the start of the cumulative process (for example, the start date of the cumulative process). The cumulative processing unit 416 calculates multiple cumulative sums d3j of time series during the observation period T1 while updating j.

[0206]

[0207] The cumulative processing unit 416 may calculate multiple cumulative data d4 after statistical processing by performing statistical processing on multiple cumulative data d3 during the observation period T1. Statistical processing may include, for example, centering, standardization, or normalization.

[0208] As described above with reference to Figures 7 to 11, according to Embodiment 1, the correction execution process represents a process that substantially sets the ratio R to "1". Therefore, the divergence of the cumulative data d3 can be suppressed more effectively, and cumulative data d3 that better reflects the homeostasis of the living organism can be derived.

[0209] The fact that the ratio R is substantially "1" indicates that it is not only the case that the ratio R is "1", but that the ratio R can reflect constancy in the cumulative data d3. For example, the ratio R may include "0.8 or more and 1.2 or less", or it may include "0.9 or more and 1.1 or less". The closer the ratio R is to "1", the better, but it is most preferable when it is "1".

[0210] Here, the cumulative processing unit 416 may, for example, use the second classification criterion information c2 instead of the first classification criterion information c1 when the unit data d1 is a step count statistical index or an exercise intensity statistical index. The second classification criterion information c2 is set to a value of the unit data d1 that substantially indicates a cessation of physical and mental activity. The cessation of physical and mental activity may include a state of rest for the living body. For example, the second classification criterion information c2 is set to a step count value (zero) that substantially indicates a cessation of physical and mental activity. Alternatively, for example, the second classification criterion information c2 is set to an exercise intensity value (METs = 1) that substantially indicates a cessation of physical and mental activity. When using the second classification criterion information c2, the unit data d1 is classified into active states and rest states, and then corrected and accumulated. Therefore, the cumulative data d3 reflects the comparison results and balance between active states and rest states. As a result, the cumulative data d3 is more suitable for evaluating the mechanism of the living body that repeats activity and rest.

[0211] Specifically, the accumulation processing unit 416 calculates the first state data b1 belonging to the first state group g1 by classifying unit data d1 that are greater than the second classification criterion information c2 from among a plurality of unit data d1. Furthermore, the accumulation processing unit 416 calculates the second state data b2 belonging to the second state group g2 by classifying unit data d1 that are equal to the second classification criterion information c2 from among a plurality of unit data d1. In this case, the accumulation processing unit 416 sets the second state data b2 to a predetermined value. For example, the predetermined value may be a negative value. The predetermined value is determined experimentally and / or empirically.

[0212] Furthermore, the accumulation processing unit 416 may, for example, compare the accumulated data d3 with a threshold when the unit data d1 is a sleep index, and based on the comparison result, replace the accumulated data d3 with the threshold and set it as the new accumulated data d3. For example, in cases where sleep duration cannot be accumulated and sleep debt arises due to sleep deprivation, it may be possible to more appropriately approximate the biological mechanism by setting a limit on the accumulated data d3 based on the premise that sleep duration cannot be accumulated, rather than uniformly accumulating the correction data d2. For example, by setting a threshold as a limit, the characteristics of sleep duration that cannot be accumulated can be represented by the accumulated data d3. In other words, sleep debt is represented by the accumulated data d3.

[0213] As an example, the accumulation processing unit 416 calculates the cumulative sum d3j as cumulative data d3 using equations (7) to (9). d3#j represents the cumulative sum at time t=j. d2i represents the corrected data d2 at time t=i. ts represents the start of accumulation (for example, the start date of accumulation). The accumulation processing unit 416 calculates multiple cumulative sums d3#j of the time series while updating j. This is the same as in equation (6).

[0214]

[0215] The cumulative processing unit 416 determines whether the cumulative sum d3#j exceeds the threshold Th. If the cumulative processing unit 416 determines that the cumulative sum d3#j does not exceed the threshold Th, it calculates the cumulative sum d3j using equation (8). On the other hand, if the cumulative processing unit 416 determines that the cumulative sum d3#j exceeds the threshold Th, it calculates the cumulative sum d3j using equation (9).

[0216]

[0217]

[0218] As explained above with reference to Figures 6 to 11, the heart rate statistical index, step count statistical index, exercise intensity statistical index, sleep index, alcohol consumption estimation index, physical and mental recovery index, exercise responsiveness index, and cumulative data constitute the input information K2. Hereinafter, the heart rate statistical index, step count statistical index, exercise intensity statistical index, sleep index, alcohol consumption estimation index, physical and mental recovery index, exercise responsiveness index, cumulative data, basic physical information Q2, and environmental information R2 may be collectively referred to as "features." The number of features is not particularly limited. Note that "features" may also be referred to as "feature FT."

[0219] Here, we illustrate why the above features can contribute to the estimation of rheumatoid arthritis index information M2 from the perspective of inflammatory response.

[0220] Since inflammatory responses caused by rheumatoid arthritis are accompanied by sympathetic nerve activation, changes in heart rate (e.g., xth percentile value) and heart rate variability indices that capture autonomic nervous system activity (e.g., RMSSD, CVRR, or LF / HF) can contribute to the estimation of rheumatoid arthritis index information M2.

[0221] Inflammatory responses are thought to follow a circadian rhythm, and in particular, in rheumatoid arthritis, the response is heightened at night and during sleep. Therefore, for example, changes in heart rate and heart rate variability during the night (e.g., from 8 PM to 4 AM the following morning), during sleep, and before and after falling asleep can contribute to the estimation of rheumatoid arthritis indicator information M2.

[0222] The systemic inflammatory response observed in rheumatoid arthritis causes fatigue, malaise, and pain. Fatigue, malaise, and pain lead to a decrease in physical activity. Therefore, a decrease in physical activity can contribute to the estimation of rheumatoid arthritis index information M2. A decrease in physical activity is indicated, for example, by a low number of steps taken while awake (e.g., xth percentile value) and a low exercise intensity (e.g., xth percentile value).

[0223] Even a relatively mild inflammatory response, if it persists, can lead to an exacerbation of various symptoms and disease activity. Therefore, for example, changes in heart rate and heart rate variability can contribute to the estimation of rheumatoid arthritis index information M2. Furthermore, for example, changes in heart rate and heart rate variability during the night, during sleep, and before and after falling asleep can contribute to the estimation of rheumatoid arthritis index information M2. In addition, for example, the number of steps taken and exercise intensity while awake can contribute to the estimation of rheumatoid arthritis index information M2. Moreover, for example, features obtained by performing cumulative processing on these features can also contribute to the estimation of rheumatoid arthritis index information M2.

[0224] Next, we will illustrate why the above-mentioned features can contribute to the estimation of rheumatoid arthritis index information M2 from the perspective of joint deformation, pain, and stiffness symptoms.

[0225] Joint symptoms such as pain can cause tossing and turning or "struggle" during sleep. Therefore, features indicating movement during sleep, features indicating the number of temporary awakenings, and features indicating the number of steps taken during temporary awakenings can contribute to the estimation of rheumatoid arthritis index information M2. Features indicating movement include, for example, the number of steps or the intensity of exercise. For example, if the number of steps is greater than "0", or if the intensity of exercise is greater than "1" in METs, it can be determined that the subject has moved. In particular, features indicating movement due to tossing and turning and "struggle" before waking, when the subject is approaching a state of wakefulness, can contribute to the estimation of rheumatoid arthritis index information M2.

[0226] Joint symptoms such as pain and stiffness may particularly worsen around the time of waking, when the viscosity of synovial fluid within the joint capsule is highest. Therefore, for example, features indicating a decrease in the number of steps taken immediately after waking and features indicating a decrease in exercise intensity can contribute to the estimation of rheumatoid arthritis index information M2.

[0227] Excessive exercise on the day of and the day before can be a cause of joint symptoms such as pain. Therefore, for example, extremely high step counts (e.g., xth percentile value) and extremely high exercise intensity (e.g., xth percentile value), as well as low variability in these factors, can contribute to the estimation of rheumatoid arthritis index information M2. On the other hand, insufficient exercise (limited joint movement) on the day of and the day before can also be a cause of joint symptoms such as pain. Therefore, for example, low step counts and low exercise intensity, as well as low variability in these factors, can contribute to the estimation of rheumatoid arthritis index information M2.

[0228] The level of exercise load on a joint is affected not only by the day of the exercise and the day before, but also by the accumulation of load over several days or more. Therefore, for example, features obtained by performing cumulative processing on features related to the number of steps, the intensity of exercise, and the variability of these features can contribute to the estimation of rheumatoid arthritis index information M2.

[0229] Next, we will illustrate why the above features can contribute to the estimation of rheumatoid arthritis index information M2 from the perspective of vicious cycles and virtuous cycles caused by rheumatoid arthritis.

[0230] In rheumatoid arthritis, a vicious cycle of symptom exacerbation is anticipated. For example, joint pain reduces physical activity such as walking, inducing a depressive state. Consequently, the perceived pain intensifies, further reducing activity. As a result, joint symptoms cyclically worsen. For example, joint pain reduces physical activity such as walking, narrowing the range of motion of the joints. Consequently, joint pain becomes more likely, further reducing physical activity such as walking. As a result, joint symptoms cyclically worsen. For example, joint pain makes it difficult to fall asleep, shortening sleep duration. Consequently, daytime physical activity decreases and a depressive state occurs, reducing activity. Consequently, pain symptoms worsen further, and falling asleep becomes even more difficult. Quantifying such cycles can contribute to the estimation of rheumatoid arthritis index information M2. Quantifying these cycles is, for example, a feature obtained by performing cumulative processing on features related to low step count, low exercise intensity, and low variability of these features. Furthermore, the features that quantify the cycle are, for example, features obtained by performing cumulative processing on features related to sleep duration or sleep rhythm. Features related to sleep rhythm are, for example, the deviation from the mean of the mid-sleep time.

[0231] Next, we will illustrate why the above-mentioned features can contribute to the estimation of rheumatoid arthritis index information M2 from the perspective of individual characteristics and lifestyle (including preferences).

[0232] It is known that obesity is associated with elevated levels of inflammatory biomarkers (e.g., CRP). Furthermore, some studies suggest that moderate alcohol consumption may contribute to slowing the progression of rheumatoid arthritis. Therefore, for example, the relative levels of features related to height, weight, BMI (body mass index), alcohol consumption, and alcohol intake may contribute to the estimation of rheumatoid arthritis indicator information M2.

[0233] Next, from an environmental perspective, we will illustrate why the above features can contribute to the estimation of rheumatoid arthritis index information M2.

[0234] Symptoms such as joint pain and stiffness are expected to worsen when living temperatures (air temperature, room temperature) are low, as this increases the viscosity of synovial fluid within the joint capsule. Furthermore, drastic changes in living temperature can place various stresses on the body, potentially disrupting autonomic nervous system function and affecting the level of inflammatory responses. Therefore, for example, seasonal features can contribute to the estimation of rheumatoid arthritis index information M2. Seasonal features include winter or months that fall during seasonal transitions.

[0235] Next, from the perspective of secondary symptoms, we will illustrate why the above-mentioned features can contribute to the estimation of rheumatoid arthritis index information M2.

[0236] Rheumatoid arthritis can cause various accompanying symptoms in addition to joint symptoms. For example, if interstitial pneumonia occurs, it can also negatively affect cardiopulmonary and respiratory function. One method for quantifying the decline in cardiopulmonary and respiratory function uses the change in heart rate at the start of exercise, such as walking, and the change in heart rate at the end of exercise, such as walking (when transitioning to rest). Therefore, given that pneumonia occurs as a secondary disease, chronic lack of exercise due to joint pain and fatigue leads to a decline in cardiopulmonary function, and autonomic nervous system dysfunction occurs due to severe symptoms, features that can express the first specific state or the second specific state can contribute to the estimation of rheumatoid arthritis index information M2. The first specific state is characterized by a high or low increase in heart rate at the start of exercise. Therefore, the features that can express the first specific state are exercise responsiveness information. The second specific state is characterized by a slow recovery rate of heart rate at the end of exercise. Therefore, the feature quantity that can represent the second specific state is information about the mental and physical recovery capacity.

[0237] The above examples illustrate, from various perspectives, why the above-mentioned features can contribute to the estimation of rheumatoid arthritis index information M2. The specific statistical processing that expresses the "height and lowness" and "high and low variability" shown above is implemented by the statistical processing unit 410.

[0238] Returning to Figure 6, the estimation unit 42 uses the learning model TM1 to estimate the subject's rheumatoid arthritis index information M2. Specifically, the estimation unit 42 inputs input information K2 to the learning model TM1 and obtains output information L2 from the learning model TM1. The storage unit 402 stores the output information L2.

[0239] The learning model TM1 takes input information K2 as input and outputs output information L2 according to a machine learning algorithm. The machine learning algorithm is not particularly limited, but examples include linear regression, neural networks, deep neural networks (hereinafter referred to as DNN), decision trees, random forests, gradient boosting, or regularized regression. Regularized regression is, for example, L1 regularized regression or L2 regularized regression. Below, as an example, we will explain the case where the machine learning algorithm of the learning model TM1 is a DNN.

[0240] Figure 12 is a schematic diagram showing an example of a DNN 50. As shown in Figure 12, the DNN 50 includes an input layer 51, a plurality of hidden layers 52, and an output layer 53. Figure 12 shows a fully connected example. The input layer 51 includes a plurality of nodes 511. Each of the hidden layers 52 includes a plurality of nodes 521. The output layer 53 includes at least one node 531. In the example of Figure 7, the output layer 53 includes a plurality of nodes 531.

[0241] Multiple nodes 511 of the input layer 51 are each input to multiple feature quantities FT that constitute the input information K2. Each node 521 of the hidden layer 52 uses the learned weights and biases and the activation function to convert the output of the previous layer into input to the next layer. Each node 531 of the output layer 53 uses the learned weights and biases and the activation function according to the final output format to output the final result (estimated result) based on the output of the previous layer. In other words, the output layer 53 outputs output information L2. Output information L2 includes rheumatoid arthritis index information M21, M22, M23, and M24.

[0242] In the example in Figure 12, the rheumatoid arthritis index information M21, M22, M23, and M24 are CDAI, SDAI, DAS28, and pain assessment scale values, respectively. In this example, the rheumatoid arthritis index information M21, M22, M23, and M24 directly show the CDAI, SDAI, DAS28, and pain assessment scale values ​​numerically.

[0243] In the example shown in Figure 12, the storage unit 402 stores the rheumatoid arthritis index information M21, M22, M23, and M24 as output information L2. The communication unit 401 transmits the output information L2 to the information provision server 47 (Figure 4). The number of nodes 531 is equal to the number of rheumatoid arthritis index information to be output, and may be one or two or more.

[0244] On the other hand, the rheumatoid arthritis index information M21, M22, M23, and M24 may indirectly represent CDAI, SDAI, DAS28, and pain assessment scale values, respectively. In this case, the post-processing unit 43 in Figure 6 performs post-processing on each of the rheumatoid arthritis index information M21, M22, M23, and M24 to output the rheumatoid arthritis index information M31, M32, M33, and M34. These points will be explained with reference to Figures 13 and 14.

[0245] Figure 13 schematically illustrates another example of the DNN 50. Figure 13 shows an example where binary classification is performed. In binary classification, each of the rheumatoid arthritis index information M21, M22, M23, and M24 output by the output layer 53 is, for example, a real number between 0 and 1.

[0246] For the rheumatoid arthritis index information M21, M22, and M23, for example, "0" indicates remission, and "1" indicates that remission is not achieved, i.e., disease activity is present. A value close to "0" for the rheumatoid arthritis index information M21, M22, and M23 indicates a high probability of determining remission. A value close to "1" for the rheumatoid arthritis index information M21, M22, and M23 indicates a high probability of determining that remission is not achieved. For the rheumatoid arthritis index information M24, for example, "0" indicates non-severe pain (no pain, mild pain, or moderate pain), and "1" indicates severe pain. A value close to "0" for the rheumatoid arthritis index information M24 indicates a high probability of determining that the pain is not severe. A value close to "1" for the rheumatoid arthritis index information M24 indicates a high probability of determining that the pain is severe.

[0247] In binary classification, the post-processing unit 43 performs threshold processing on each of the rheumatoid arthritis index information M21, M22, M23, and M24. In other words, the post-processing unit 43 sets "0" or "1" for the rheumatoid arthritis index information M31, M32, M33, and M34 depending on whether the rheumatoid arthritis index information M21, M22, M23, and M24 are below the threshold or above the threshold. In short, binary classification is the process of classifying the output information L2 into either the "0" class or the "1" class.

[0248] For example, if the rheumatoid arthritis index information M21 is less than a threshold (e.g., 0.5), the post-processing unit 43 sets the rheumatoid arthritis index information M31 to "0" to indicate remission from the perspective of CDAI. On the other hand, if the rheumatoid arthritis index information M21 is greater than or equal to a threshold (e.g., 0.5), the post-processing unit 43 sets the rheumatoid arthritis index information M31 to "1" to indicate that remission has not occurred from the perspective of CDAI.

[0249] The threshold is not limited to 0.5 and can be set to any value. The threshold is determined, for example, based on experimental, empirical, and / or medical standards. Furthermore, the thresholds for the rheumatoid arthritis index information M21, M22, M23, and M24 may be the same or different. Although "0" and "1" were used as binary values, any value can be used. The number of nodes 531 is the same as the number of rheumatoid arthritis index information to be output, and may be one or two or more.

[0250] Figure 14 schematically illustrates yet another example of DNN50. Figure 14 shows an example of when multi-class classification is performed. For example, for CDAI, SDAI, and DAS28, the assessment of disease activity in rheumatoid arthritis is classified into four classes. For example, the four classes are "remission," "low," "moderate," and "high." Also, for example, for pain assessment scale values, the assessment of joint pain is classified into five classes. For example, for a 100 mm pain assessment scale (e.g., VAS), the five classes are "0-20 mm," "21-40 mm," "41-60 mm," "61-80 mm," and "81-100 mm." Note that, for example, the five classes may be expressed as percentages, with the total length of the pain assessment scale being 100%.

[0251] The number of classifications (classes) is not particularly limited and can be set arbitrarily. Furthermore, the numerical range for each class is determined, for example, based on experimental, empirical, and / or medical standards.

[0252] As shown in Figure 14, the DNN 50 includes an input layer 51, a plurality of intermediate layers 52, and at least one output layer 53A. In the example in Figure 14, the DNN 50 includes a plurality of output layers 53A.

[0253] Multiple output layers 53A are provided, each corresponding to CDAI, SDAI, DAS28, and pain assessment scale value. Each output layer 53A outputs output information L2. Specifically, each output layer 53A contains the same number of nodes 531A as the number of classifications (number of classes). Note that in Figure 14, for the sake of simplifying the diagram, only the nodes 531A of the output layers 53A corresponding to CDAI and SDAI are shown, and the nodes 531A of the output layers 53A corresponding to DAS28 and pain assessment scale value are omitted.

[0254] Let's focus on one output layer 53A. The four nodes 531A each correspond to one of four classes. That is, the output layer 53A includes a node 531A corresponding to "remission," a node 531A corresponding to "low," a node 531A corresponding to "moderate," and a node 531A corresponding to "high."

[0255] Each node 531A outputs an output value VL. The output value VL is, for example, a real number between 0 and 1 (inclusive). The sum of the output values ​​VL of the four nodes 531A is "1". Therefore, the output value VL indicates the probability of being classified into the corresponding class. In other words, the class corresponding to the node 531A that outputs the largest output value VL has the highest probability of being correct.

[0256] Therefore, the post-processing unit 43 acquires the maximum output value from among the multiple output values ​​VL from the multiple nodes 531A for each output layer 53A. The multiple maximum output values ​​output from each of the multiple output layers 53A correspond to the rheumatoid arthritis index information M21, M22, M23, and M24. In this way, the multiple output values ​​VL (output information L2) from each output layer 53A substantially include rheumatoid arthritis index information.

[0257] Then, the post-processing unit 43 sets the class corresponding to the node 531A that output the maximum output value for each output layer 53A in the rheumatoid arthritis index information M31, M32, M33, and M34. For example, in the output layer 53A related to CDAI, if the output value VL of node 531A corresponding to "remission" is "0.9", the output value VL of node 531A corresponding to "low" is "0.05", the output value VL of node 531A corresponding to "moderate" is "0.03", and the output value VL of node 531A corresponding to "high" is "0.02", the post-processing unit 43 sets "remission" in the rheumatoid arthritis index information M31, corresponding to node 531A that shows the maximum output value.

[0258] The number of output layers 53A is equal to the number of rheumatoid arthritis index information to be output, and may be one or two or more. In addition, the number of nodes 531A in one output layer 53A is equal to the number of classifications (classes).

[0259] In the examples shown in Figures 13 and 14, the storage unit 402 stores the rheumatoid arthritis index information M31, M32, M33, and M34 obtained by the post-processing unit 43 as output information N2. The communication unit 401 transmits the output information N2 to the information provision server 47 (Figure 4).

[0260] Furthermore, as can be seen from Figures 13 and 14, the post-processing is a process that obtains new output information N2 (rheumatoid arthritis index information M31, M32, M33, M34) by classifying the output information L2 (rheumatoid arthritis index information M21, M22, M23, M24) of the learning model TM1 into one of several classes indicating the activity or symptoms of rheumatoid arthritis, based on the numerical values ​​indicated by the output information L2.

[0261] As described above with reference to Figures 1 to 14, according to Embodiment 1, the estimation device 4 can output rheumatoid arthritis index information MX with high estimation accuracy by inputting at least vital information H2 and behavioral information J21 to the learning model TM1.

[0262] In particular, in Embodiment 1, the input information K2 does not include information indicating the results of diagnosis and evaluation by medical professionals such as doctors, nor information indicating the results of blood tests. Therefore, during the utilization stage of the learning model TM1, the estimation device 4 can output rheumatoid arthritis index information MX with high estimation accuracy without using the results of diagnosis and evaluation by medical professionals or blood tests. In other words, the subject is not required to undergo diagnosis and evaluation by medical professionals or blood tests. Therefore, rheumatoid arthritis index information MX can be obtained while reducing the burden on the subject. In this way, rheumatoid arthritis index information MX can be obtained without reducing the subject's QOL (Quality of Life).

[0263] In particular, the rheumatoid arthritis index information MX includes at least one piece of information from among CDAI, SDAI, DAS28, and joint pain. Therefore, it is possible to more accurately grasp the activity or symptoms of rheumatoid arthritis.

[0264] In addition, the estimation device 4 can continuously acquire raw biological state data A2 from a wearable device (biological state detection device 103) and / or a mobile terminal (first terminal 102). Therefore, the calculation and monitoring of rheumatoid arthritis index information MX can be continuously performed. In particular, the raw biological state data A2 is automatically transmitted from the wearable device (biological state detection device 103) and / or the mobile terminal (first terminal 102) to the estimation system 40 (estimation device 4). As a result, the burden on the subject is further reduced. In other words, rheumatoid arthritis index information MX can be obtained while reducing the burden on the subject.

[0265] As an example, Embodiment 1 offers the following benefits: For instance, the severity of rheumatoid arthritis symptoms (disease activity, inflammatory response, joint pain) can be constantly monitored without direct medical examination, interviews, or blood tests by a physician. For instance, disease activity assessment results equivalent to those obtained by a physician or specialist can be obtained. For instance, in cases where patients only visit the hospital once every few months, they may not be able to recall the changes from the previous visit to the current one. In this case, the inability to grasp the changes in disease activity can lead to delays or errors in decisions regarding medication changes, etc. Therefore, estimation by the learning model TM1 can increase the speed and efficiency of decisions regarding medication changes and adjustments, leading to the optimization of medical decisions. For instance, it enables telemedicine in rural areas where specialists or hospitals are scarce. For instance, it allows for immediate detection of worsening conditions and interventions to prevent further deterioration (medication, lifestyle guidance, exercise therapy, etc.). For instance, it can facilitate support from family members, etc. For instance, it allows for constant guidance on avoiding insufficient or excessive exercise. As a result, it contributes to symptom control.

[0266] Next, with reference to Figures 6 and 15, the estimation process for rheumatoid arthritis index information MX, which is performed in step S3 of Figure 5, will be described. Figure 15 is a flowchart of an example of the estimation process. The estimation process is performed by the estimation device 4 in Figure 6. As shown in Figure 15, the estimation process includes steps S31 to S42.

[0267] First, in step S31, the condition processing unit 411 obtains the subject's biological state raw data A2 from the first database 45 (Figure 4). The biological state raw data A2 is stored in the storage unit 402.

[0268] Next, in step S32, the condition processing unit 411 performs processing (condition processing) according to specific conditions on the heart rate data, step count data, and exercise intensity data of the subject's biological state raw data A2. As a result, the condition processing unit 411 outputs extracted heart rate data, extracted step count data, and extracted exercise intensity data. The specific conditions are conditions related to sleep, conditions related to wakefulness, and conditions related to the time of day caused by the sun.

[0269] Next, in step S33, the statistical processing unit 410 performs statistical processing and differential processing on each of the extracted heart rate data, extracted step count data, and extracted exercise intensity data. In addition, the statistical processing unit 410 performs ratio processing on each of the extracted step count data and extracted exercise intensity data. As a result, the statistical processing unit 410 outputs a heart rate statistical index based on the extracted heart rate data, a step count statistical index based on the extracted step count data, and an exercise intensity statistical index based on the extracted exercise intensity data. These indices are stored in the storage unit 402 as input information K2. Note that differential processing and ratio processing can be considered as statistical processing in a broad sense.

[0270] Next, in step S34, the sleep processing unit 412 calculates a sleep index based on the sleep data of the subject's biological state raw data A2 (an example of sleep processing). The sleep index is stored in the storage unit 402 as input information K2.

[0271] Next, in step S35, the alcohol consumption estimation unit 413 calculates an alcohol consumption estimation index based on the heart rate data and step count data of the subject's biological state raw data A2 (an example of alcohol consumption estimation processing). The alcohol consumption estimation index is stored in the storage unit 402 as input information K2.

[0272] Next, in step S36, the mental and physical recovery ability estimation unit 414 calculates a mental and physical recovery index based on the subject's biological state raw data A2, which includes heart rate data and behavioral raw data (e.g., exercise intensity data) (an example of mental and physical recovery ability estimation processing). The mental and physical recovery index is stored in the storage unit 402 as input information K2.

[0273] Next, in step S37, the exercise responsiveness estimation unit 415 calculates an exercise responsiveness index based on the subject's biological state raw data A2, which includes heart rate data and behavioral raw data (e.g., exercise intensity data) (an example of exercise responsiveness estimation processing). The exercise responsiveness index is stored in the storage unit 402 as input information K2.

[0274] Next, in step S38, the accumulation processing unit 416 performs accumulation processing on the heart rate statistical index, step count statistical index, exercise intensity statistical index, sleep index, physical and mental recovery index, and exercise responsiveness index, respectively. As a result, the accumulation processing unit 416 outputs accumulated data (cumulative index) for the heart rate statistical index, step count statistical index, exercise intensity statistical index, sleep index, physical and mental recovery index, and exercise responsiveness index, respectively. The accumulated data is stored in the storage unit 402 as input information K2. The accumulation processing includes classification execution processing, correction execution processing, and accumulation execution processing.

[0275] Next, in step S39, the estimation unit 42 inputs input information K2, which includes the feature quantities generated in steps S31 to S37, to the learning model TM1. As a result, the learning model TM1 outputs output information L2. In this case, it is preferable that the input information K2 includes basic physical information Q2 and environmental information R2.

[0276] Next, in step S40, the estimation unit 42 acquires output information L2 from the learning model TM1. The output information L2 includes the subject's rheumatoid arthritis index information M2. The output information L2 is stored in the storage unit 402.

[0277] Next, in step S41, the estimation unit 42 determines whether or not post-processing is required for the rheumatoid arthritis index information M2.

[0278] If it is determined in step S41 that no post-processing is required (NO), the estimation process is completed and the system returns to the main routine shown in Figure 5. If no post-processing is required, it is because the rheumatoid arthritis index information M2 directly indicates the rheumatoid arthritis index numerically (for example, Figure 12).

[0279] On the other hand, if it is determined in step S41 that post-processing is necessary (YES), the process proceeds to step S42. Post-processing is necessary when the rheumatoid arthritis index information M2 indirectly indicates rheumatoid arthritis indicators (for example, Figures 13 and 14).

[0280] Next, in step S42, the post-processing unit 43 performs post-processing on the rheumatoid arthritis index information M2 and outputs the rheumatoid arthritis index information M3. The rheumatoid arthritis index information M3 is stored in the storage unit 402 as output information N2. Then the estimation process is completed and the system returns to the main routine shown in Figure 5.

[0281] As described above with reference to Figure 15, according to Embodiment 1, the estimation device 4 can obtain highly accurate output information L2 (rheumatoid arthritis index information M2) from the learning model TM1 by inputting input information K2 to the learning model TM1. Furthermore, the estimation device 4 can obtain output information N2 (rheumatoid arthritis index information M3) by performing post-processing on the output information L2.

[0282] In this case, the input information K2 does not include the results of diagnosis and evaluation by medical professionals or the results of blood tests. Therefore, according to Embodiment 1, by using the learning model TM1, it is possible to estimate the rheumatoid arthritis index information MX with high estimation accuracy while improving the quality of life of the subject.

[0283] Next, the generation stage of the learning model TM1 will be described with reference to Figures 1, 16, and 17. Figure 16 is a block diagram showing an example configuration of the learning device 3 according to Embodiment 1. As shown in Figure 16, the learning device 3 comprises a processing unit 31, a communication unit 34, and a storage unit 35. The learning device 3 may also include an input unit 32 and an output unit 33. The hardware configuration of the processing unit 31, communication unit 34, storage unit 35, input unit 32, and output unit 33 is the same as the hardware configuration of the processing unit 400, communication unit 401, storage unit 402, input unit 403, and output unit 404 of the estimation device 4 in Figure 6.

[0284] The storage unit 35 stores data and computer programs. The processing unit 31 includes a learning data acquisition unit 310 and a learning unit 311. For example, the processing unit 31 functions as the learning data acquisition unit 310 and the learning unit 311 by executing the computer programs stored in the storage unit 35.

[0285] Figure 17 is a flowchart showing an example of a learning method using the learning device 3. The learning method is an example of the "learning model generation method" of this disclosure. As shown in Figure 17, the learning method includes steps S101 to S108.

[0286] First, in step S101, the learning data acquisition unit 310 acquires multiple learning datasets F1 from the learning data creation device 2 (Figure 1). The storage unit 35 stores the multiple learning datasets F1. A portion of the multiple learning datasets F1 is training data, another portion is evaluation data, and yet another portion is test data.

[0287] Next, in step S102, the learning unit 311 prepares the pre-training learning model TM1. In the pre-training learning model TM1, various parameters are set to their initial values.

[0288] Next, in step S103, the learning unit 311 obtains one learning dataset F1 from the multiple learning datasets F1 stored in the storage unit 35. In this case, the learning dataset F1 is the training data.

[0289] Next, in step S104, the learning unit 311 inputs the feature information G1 contained in the learning dataset F1 to the learning model TM1 before (or during) learning. As a result, output information is output from the learning model TM1 as an estimation result according to the machine learning algorithm.

[0290] Next, in step S105, the learning unit 311 compares the correct labels B1 included in the learning dataset F1 with the output information output as the estimation result in step S104, and performs machine learning by adjusting various parameters according to the machine learning algorithm. In this way, the learning unit 311 trains the learning model TM1 on the correlation between the feature information G1 and the correct labels B1.

[0291] Machine learning algorithms are not particularly limited, but examples include linear regression, neural networks, deep neural networks (DNNs), decision trees, random forests, gradient boosting, or regularized regression.

[0292] Next, in step S106, the learning unit 311 determines whether the learning termination condition has been met. The learning termination condition is, for example, that the evaluation value of the loss function based on the correct label B1 and the output information output as an estimation result has reached the target value. Alternatively, the learning termination condition is, for example, that the number of learning iterations (epochs) has reached the target number.

[0293] If it is determined in step S106 that the learning termination condition is not met (NO), the process proceeds to step S103. Steps S103 to S105 are repeated until the learning termination condition is met.

[0294] On the other hand, if it is determined in step S106 that the learning termination condition has been met (YES), the process proceeds to step S107.

[0295] Next, in step S107, the learning unit 311 adjusts the hyperparameters of the learning model TM1 based on the learning dataset F1 as evaluation data and the input values ​​input from the input unit 32 by the machine learning engineer.

[0296] Next, in step S108, the learning unit 311 evaluates the estimation accuracy of the learning model TM1 using the learning dataset F1 as test data. Then the learning method is completed.

[0297] As described above with reference to Figure 17, according to Embodiment 1, the learning device 3 generates a learning model TM1 that outputs output information L2 when input information K2 is input by performing learning using the learning dataset F1 (step S105). That is, the learning device 3 generates a trained learning model TM1 having various trained parameters by repeating learning using a plurality of learning datasets F1. The storage unit 35 stores the trained learning model TM1.

[0298] The input information K2 includes the subject's vital information H2 and the subject's behavioral information J2. The output information L2 includes the subject's rheumatoid arthritis index information M2. The vital information H2 and behavioral information J2 do not include the results of diagnosis and evaluation by healthcare professionals or the results of blood tests. Therefore, according to Embodiment 1, a learning model TM1 can be generated that can estimate the rheumatoid arthritis index information M2 while improving the subject's quality of life.

[0299] Next, the generation stage of the learning dataset F1 will be described with reference to Figures 1, 18, and 19. Figure 18 is a block diagram showing an example configuration of the learning data creation device 2 according to Embodiment 1. As shown in Figure 18, the learning data creation device 2 comprises a processing unit 21, a communication unit 24, and a storage unit 25. The learning data creation device 2 may also include an input unit 22 and an output unit 23. The hardware configuration of the processing unit 21, communication unit 24, storage unit 25, input unit 22, and output unit 23 is the same as the hardware configuration of the processing unit 400, communication unit 401, storage unit 402, input unit 403, and output unit 404 of the estimation device 4 in Figure 6.

[0300] The memory unit 25 stores data and computer programs. The processing unit 21 preferably includes a statistical processing unit 210, a conditional processing unit 211, a sleep processing unit 212, an accumulation processing unit 216, and a correct label creation unit 217. The processing unit 21 may further include at least one of the following: an alcohol consumption estimation unit 213, a mental and physical recovery ability estimation unit 214, and an exercise responsiveness estimation unit 215. For example, the processing unit 400 functions as the statistical processing unit 210, the conditional processing unit 211, the sleep processing unit 212, the accumulation processing unit 216, the correct label creation unit 217, the alcohol consumption estimation unit 213, the mental and physical recovery ability estimation unit 214, and the exercise responsiveness estimation unit 215 by executing a computer program stored in the memory unit 402.

[0301] The processing of the statistical processing unit 210, the condition processing unit 211, the sleep processing unit 212, the accumulation processing unit 216, the alcohol consumption estimation unit 213, the mental and physical recovery ability estimation unit 214, and the exercise responsiveness estimation unit 215 are the same as the processing of the statistical processing unit 410, the condition processing unit 411, the sleep processing unit 412, the accumulation processing unit 416, the alcohol consumption estimation unit 413, the mental and physical recovery ability estimation unit 414, and the exercise responsiveness estimation unit 415 of the estimation device 4 (Figure 6).

[0302] For example, in the descriptions of the statistical processing unit 410, condition processing unit 411, sleep processing unit 412, accumulation processing unit 416, alcohol consumption estimation unit 413, mental and physical recovery ability estimation unit 414, and exercise responsiveness estimation unit 415 of the estimation device 4, the subject can be read as the learning subject, the biological state raw data A2 as the biological state raw data A1, the input information K2 as the feature information G1, the output information L2 as the correct label B1, the basic physical information Q2 as the basic physical information Q1, the environmental information R2 as the environmental information R1, and the rheumatoid arthritis index information M2 as the rheumatoid arthritis index information M1, thereby substituting the descriptions of the statistical processing unit 210, condition processing unit 211, sleep processing unit 212, accumulation processing unit 216, alcohol consumption estimation unit 213, mental and physical recovery ability estimation unit 214, and exercise responsiveness estimation unit 215 of the learning data creation device 2.

[0303] Figure 19 is a flowchart showing an example of a method for creating learning data using the learning data creation device 2. As shown in Figure 19, the learning data creation method includes steps S201 to S209.

[0304] First, in step S201, the processing unit 21 acquires the raw biological state data A1 and the correct answer information Z1 of the training subject. The raw biological state data A1 and the correct answer information Z1 are stored in the storage unit 25. Preferably, the processing unit 21 acquires basic physical information Q1 and environmental information R1, and stores the basic physical information Q1 and environmental information R1 as feature information G1 in the storage unit 25.

[0305] Next, in step S202, the condition processing unit 211 performs processing (condition processing) according to specific conditions on the heart rate data, step count data, and exercise intensity data of the learning subject's biological state raw data A1. As a result, the condition processing unit 211 outputs extracted heart rate data, extracted step count data, and extracted exercise intensity data. The specific conditions are conditions related to sleep, conditions related to wakefulness, and conditions related to the time of day caused by the sun.

[0306] Next, in step S203, the statistical processing unit 210 performs statistical processing and differential processing on each of the extracted heart rate data, extracted step count data, and extracted exercise intensity data. In addition, the statistical processing unit 210 performs ratio processing on each of the extracted step count data and extracted exercise intensity data. As a result, the statistical processing unit 210 outputs a heart rate statistical index based on the extracted heart rate data, a step count statistical index based on the extracted step count data, and an exercise intensity statistical index based on the extracted exercise intensity data. These indices are stored in the storage unit 25 as feature information G1. Note that differential processing and ratio processing can be considered as statistical processing in a broad sense.

[0307] Next, in step S204, the sleep processing unit 212 calculates a sleep index based on the sleep data of the learning subject's biological state raw data A1 (an example of sleep processing). The sleep index is stored in the storage unit 25 as feature information G1.

[0308] Next, in step S205, the alcohol consumption estimation unit 213 calculates an alcohol consumption estimation index based on the heart rate data and step count data of the learning subject's biological state raw data A1 (an example of alcohol consumption estimation processing). The alcohol consumption estimation index is stored in the storage unit 25 as feature information G1.

[0309] Next, in step S206, the mental and physical recovery ability estimation unit 214 calculates a mental and physical recovery index based on the heart rate data and behavioral raw data (e.g., exercise intensity data) of the learning subject's biological state raw data A1 (an example of mental and physical recovery ability estimation processing). The mental and physical recovery index is stored in the storage unit 25 as feature information G1.

[0310] Next, in step S207, the exercise responsiveness estimation unit 215 calculates an exercise responsiveness index based on the heart rate data and behavioral raw data (e.g., exercise intensity data) of the learning subject's biological state raw data A1 (an example of exercise responsiveness estimation processing). The exercise responsiveness index is stored in the storage unit 25 as feature information G1.

[0311] Next, in step S208, the accumulation processing unit 216 performs accumulation processing on the heart rate statistical index, step count statistical index, exercise intensity statistical index, sleep index, physical and mental recovery index, and exercise responsiveness index, respectively. As a result, the accumulation processing unit 216 outputs accumulated data (cumulative index) for the heart rate statistical index, step count statistical index, exercise intensity statistical index, sleep index, physical and mental recovery index, and exercise responsiveness index, respectively. The accumulated data is stored in the storage unit 25 as feature information G1. The accumulation processing includes classification execution processing, correction execution processing, and accumulation execution processing.

[0312] Next, in step S209, the correct label creation unit 217 creates a correct label B1 based on the correct information Z1 and associates it with the feature information G1 (correct label creation process). The correct label B1 is stored in the storage unit 25. As an example, the correct information Z1 is a numerical value that directly represents a rheumatoid arthritis index. For example, the correct information Z1 is a numerical value that directly represents CDAI, a numerical value that directly represents SDAI, a numerical value that directly represents DAS28, and / or a numerical value that directly represents a pain assessment scale value.

[0313] As a first example, when the estimation device 4 estimates a numerical value that directly represents the rheumatoid arthritis index (for example, in the case of Figure 12), the correct label creation unit 217 associates the correct information Z1 with the correct label B1 and the feature information G1.

[0314] As a second example, when the estimation device 4 performs binary classification (for example, in the case of Figure 13), the correct label creation unit 217 sets the correct label B1 to, for example, "0" or "1" depending on whether the correct information Z1 (a numerical value that directly indicates the rheumatoid arthritis index) is below a threshold or above a threshold. The threshold is determined, for example, based on experimental, empirical, and / or medical standards. For example, if the correct information Z1 is CDAI, SDAI, or DAS28, then in the correct label B1, "0" indicates remission and "1" indicates non-remission. For example, if the correct information Z1 is a pain assessment scale value, then in the correct label B1, "0" indicates mild pain and "1" indicates severe pain.

[0315] As a third example, when the estimation device 4 performs multi-class classification (for example, in the case of Figure 14), the correct label creation unit 217 sets an evaluation corresponding to a class for the correct label B1, depending on whether the correct information Z1 (a numerical value directly indicating the rheumatoid arthritis index) falls within the numerical range of any of the multiple classes. Depending on the class to which the correct information Z1 indicating CDAI etc. belongs, the correct label creation unit 217 sets, for example, "remission," "low," "moderate," or "high" for the correct label B1. If the correct information Z1 is a pain assessment scale value, for example, the pain assessment scale is divided into five equal parts to create five classes. The numerical range for each class is determined, for example, based on experimental, empirical, and / or medical standards.

[0316] In the second and third examples, the correct answer information Z1 after binary classification or multi-class classification may be input via the input unit 22. In this case, the correct label creation unit 217 associates the correct answer information Z1 with the feature information G1 as the correct label B1, similar to the first example.

[0317] Once step S209 is completed, the training data creation method is finished. Steps S201 to S209 are repeatedly executed to create multiple training datasets F1.

[0318] As described above with reference to Figure 19, according to Embodiment 1, the learning data creation device 2 can create a learning dataset F1 suitable for a learning model TM1 that can estimate rheumatoid arthritis index information M2 while improving the quality of life of the subject.

[0319] Note that the training dataset F1 of Embodiment 1 does not include the evaluation report information V1 of Embodiment 2. Also, the input information K2 of Embodiment 1 does not include the self-report information V2 of Embodiment 2.

[0320] (Embodiment 2) An estimation device 4, a learning device 3, and a learning data creation device 2 according to Embodiment 2 of the present disclosure will be described with reference to Figures 4 to 6 and Figures 15 to 21. Embodiment 2 differs from Embodiment 1 mainly in that the input information K20 includes the subject's self-reported information V2. The differences between Embodiment 2 and Embodiment 1 will be mainly described below.

[0321] Figure 20 shows an example of a training dataset F10 according to Embodiment 2. As shown in Figure 20, the training dataset F10 includes feature information G10 and ground truth labels B1 of the training subject. Feature information G10 are explanatory variables, and ground truth labels B1 are the target variables. Feature information G10 includes at least evaluation report information V1, vital information H1, and behavioral information J1 of the training subject.

[0322] Evaluation report information V1 includes a report on the learning subject's assessment of rheumatoid arthritis. Typically, evaluation report information V1 includes a report on an objective assessment of the learning subject's rheumatoid arthritis by a healthcare professional. The healthcare professional is, for example, a physician or nurse. Alternatively, evaluation report information V1 may include a report on the learning subject's subjective assessment of their own rheumatoid arthritis instead of an objective assessment.

[0323] Specifically, the evaluation report information V1 includes at least one of the following: information indicating the medical professional's evaluation of the learning subject's overall condition (hereinafter, overall condition information); information indicating the medical professional's evaluation of the learning subject's overall joint pain (hereinafter, overall pain information); information indicating the number of tender joints of the learning subject (hereinafter, tender joint count information); and information indicating the number of swollen joints of the learning subject (hereinafter, swollen joint count information). In Embodiment 2, the evaluation report information V1 includes the medical professional's overall condition information of the learning subject, the medical professional's overall pain information of the learning subject, the medical professional's tender joint count information of the learning subject, and the medical professional's swollen joint count information of the learning subject.

[0324] Furthermore, in Evaluation Report Information V1, "General condition information of the learning subject by healthcare professionals" may be replaced with "General condition information of the learning subject themselves." In Evaluation Report Information V1, "General pain information of the learning subject by healthcare professionals" may be replaced with "General pain information of the learning subject themselves." In Evaluation Report Information V1, "Information on the number of tender joints of the learning subject by healthcare professionals" may be replaced with "Information on the number of tender joints of the learning subject themselves." In Evaluation Report Information V1, "Information on the number of swollen joints of the learning subject by healthcare professionals" may be replaced with "Information on the number of swollen joints of the learning subject themselves."

[0325] The training dataset F10 may include at least one of the following: basic physical information Q1 and environmental information R1. The details of vital information H1, behavioral information J1, basic physical information Q1, and environmental information R1 are the same as those of vital information H1, behavioral information J1, basic physical information Q1, and environmental information R1 in Figure 2, respectively.

[0326] Furthermore, the rheumatoid arthritis index information M1 of the correct label B1 includes at least one piece of information from among CDAI, SDAI, and DAS28. In particular, in Embodiment 2, the rheumatoid arthritis index information M1 does not include information about joint pain. This is because the feature information G10 includes the evaluation report information V1, so there is no need to estimate it.

[0327] Figure 21 shows an example of input information K20 and output information L2 according to Embodiment 2. As shown in Figure 20, input information K20 includes at least the subject's self-reported information V2, vital information H2, and behavioral information J2. Input information K20 is an explanatory variable, and output information L2 is the dependent variable. Self-reported information V2 may be referred to as a "feature."

[0328] Self-reported information V2 includes at least one of the following: information indicating the subject's own assessment of their overall physical condition (hereinafter, overall physical condition information); information indicating the subject's own assessment of their overall joint pain (hereinafter, overall pain information); information indicating the number of tender joints as assessed by the subject (hereinafter, number of tender joints information); and information indicating the number of swollen joints as assessed by the subject (hereinafter, number of swollen joints information). In Embodiment 2, self-reported information V2 includes the subject's own overall physical condition information, the subject's own overall pain information, the number of tender joints as assessed by the subject, and the number of swollen joints as assessed by the subject.

[0329] The input information K20 may include at least one of the following: basic physical information Q2 and environmental information R2. The details of vital information H2, behavioral information J2, basic physical information Q2, and environmental information R2 are the same as those of vital information H2, behavioral information J2, basic physical information Q2, and environmental information R2 in Figure 3, respectively.

[0330] Furthermore, the rheumatoid arthritis index information M2 of the output information L2 includes at least one of the following: information regarding CDAI, information regarding SDAI, and information regarding DAS28. In particular, in Embodiment 2, the rheumatoid arthritis index information M2 does not include information regarding joint pain. This is because the input information K20 includes self-reported information V2, so there is no need to estimate it.

[0331] In Figures 20 and 21, general condition information and general pain information are indicated by symptom assessment scale values. Typically, general condition information and general pain information are indicated by VAS values. For example, the general condition information in assessment report information V1 is indicated by the physician's VAS value (see formula (1)). For example, the general condition information in self-reported information V2 is indicated by the subject's VAS value (see formula (1)). Furthermore, the definitions of the number of tender joints and the number of swollen joints are the same as the number of tender joints and the number of swollen joints in formula (1) that shows CDAI, respectively.

[0332] As described above with reference to Figures 20 and 21, in Embodiment 2, the training dataset F10 is used instead of the training dataset F1 of Embodiment 1, and the input information K20 is used instead of the input information K2 of Embodiment 1.

[0333] In particular, in Embodiment 2, the learning dataset F10 includes evaluation report information V1 related to rheumatoid arthritis. Therefore, the learning model TM2 constructed by learning using the learning dataset F10 can output the subject's rheumatoid arthritis index information M2 with even higher estimation accuracy when the subject's input information K20 (self-report information V2, vital information H2, and behavioral information J2) is input. In other words, the estimation device 4 can estimate the rheumatoid arthritis index information M2 by using the learning model TM2.

[0334] Refer to Figure 4. The first terminal 102 receives self-report information V2 from the subject. The first terminal 102 sends the self-report information V2 to the cloud server 101. The cloud server 101 sends the self-report information V2 to the relay server 44. Alternatively, the first terminal 102 may also send the self-report information V2 to the relay server 44.

[0335] Refer to Figure 5. The estimation method performed by the estimation system 40 according to Embodiment 2 includes, as an example, steps S1 to S5. However, in step S1, the relay server 44 receives self-report information V2 in addition to the raw biological state data A2, the subject's basic physical information Q2, and the subject's environmental information R2. In step S2, the first database 45 stores the self-report information V2 in addition to the raw biological state data A2, the basic physical information Q2, and the environmental information R2. Furthermore, in step S3, the estimation device 4 estimates the subject's rheumatoid arthritis index information MX using the self-report information V2, the raw biological state data A2, the basic physical information Q2, the environmental information R2, and the learning model TM2.

[0336] Refer to Figure 6. In the estimation device 4 according to Embodiment 2, instead of the learning model TM1, a learning model TM2 constructed by learning using the learning dataset F10 is stored in the storage unit 402. The estimation unit 42 receives input information K20 from the learning model TM2 and obtains output information L2 from the learning model TM2. The storage unit 402 stores the output information L2. The learning model TM2 receives input information K20 and outputs output information L2 according to the machine learning algorithm. Typically, the learning model TM2 is a trained model. Alternatively, the learning model TM2 may be a computer program.

[0337] During the utilization phase of the learning model TM2, including self-reported information V2 in the input information K20 allows for more accurate output of the subject's rheumatoid arthritis index information M2. Self-reported information V2 is information about rheumatoid arthritis symptoms that the subject is aware of, or information about accompanying symptoms that the subject is aware of. Accompanying symptoms include, for example, symptoms indirectly caused by inflammation or pain, or secondary effects such as depressed mood or impairment of daily life.

[0338] Refer to Figure 15. The estimation process performed by the estimation device 4 according to Embodiment 2 includes, as an example, steps S31 to S42. However, in step S39, the estimation unit 42 inputs input information K20, which includes the feature quantities generated in steps S31 to S37 and self-reported information V2, to the learning model TM2. As a result, the learning model TM2 outputs output information L2.

[0339] Refer to Figures 16 and 17. The learning device 3 according to Embodiment 2 generates a learning model TM2 by repeatedly performing learning using a plurality of learning datasets F10 according to a machine learning algorithm. The learning device 3 performs supervised learning. The learning method performed by the learning device 3 according to Embodiment 2 includes, as an example, steps S101 to S108. The learning method is an example of the "learning model generation method" of this disclosure. However, in step S101, the learning data acquisition unit 310 acquires a plurality of learning datasets F10 from the learning data creation device 2 (Figure 1). In step S102, the learning unit 311 prepares the learning model TM2 before learning. Hereafter, in steps S54 to S58 according to Embodiment 2, the learning dataset F10 is used instead of the learning dataset F1 of Embodiment 1, and the learning model TM2 before learning (or during learning) is used instead of the learning model TM1 before learning (or during learning), thereby generating a learned learning model TM2.

[0340] In the generation stage of the learning model TM2, including evaluation report information V1 in the learning dataset F10 allows for the generation of a learning model TM2 with higher estimation accuracy. The evaluation report information V1 is, for example, objective information from healthcare professionals regarding the subject's rheumatoid arthritis symptoms, or objective information from healthcare professionals regarding the subject's accompanying symptoms.

[0341] Refer to Figures 18 and 19. The learning data creation device 2 according to Embodiment 2 creates a learning dataset F10 based on at least the raw biological state data A1 of the learning subject, evaluation report information V1, and correct answer information Z1.

[0342] The learning data creation method executed by the learning data creation device 2 according to Embodiment 2 includes, as an example, steps S201 to S209. However, in step S201, the processing unit 21 acquires the biological state raw data A1 of the learning subject, basic physical information Q1, environmental information R1, evaluation report information V1, and correct answer information Z1. The basic physical information Q1, environmental information R1, and evaluation report information V1 are stored in the storage unit 25 as feature information G10. In the following description of steps S202 to S209 according to Embodiment 2, feature information G1 will be read as feature information G10. As a result, the learning data creation device 2 creates a learning dataset F10.

[0343] Here, the evaluation report information V1 and the self-report information V2 may include at least one of the following score information: score information based on a questionnaire regarding the degree of impairment to daily life, score information based on questions regarding depressive symptoms, score information based on questions regarding sleep disorders and sleep quality, score information based on questions regarding fatigue and lethargy, and score information based on questions regarding QOL. The score information of the evaluation report information V1 is based on the responses of the learning subject, and the score information of the self-report information V2 is based on the responses of the subject.

[0344] (Embodiment 3) Embodiment 3 of this disclosure differs from Embodiment 2 in that the input information K200 includes self-report information V2 but does not include vital information H2 and behavioral information J2. The differences between Embodiment 3 and Embodiment 2 will be mainly described below.

[0345] The training dataset F100 according to Embodiment 3 includes feature information G100 and ground truth labels B1 of the training subject. The feature information G100 is an explanatory variable, and the ground truth labels B1 is the target variable. The feature information G100 includes at least evaluation report information V1 of the training subject. However, the feature information G100 does not include vital information H1, H11 and behavioral information J1, J11.

[0346] The input information K200 according to Embodiment 3 includes at least the subject's self-reported information V2. However, the input information K200 does not include vital information H2, H21 and behavioral information J2, J21. The input information K200 is an explanatory variable, and the output information L2 is the dependent variable.

[0347] In Embodiment 3, the training dataset F100 is used instead of the training dataset F10 in Embodiment 2, and the input information K200 is used instead of the input information K20 in Embodiment 2.

[0348] In Embodiment 3, the learning model TM3, constructed by learning using a learning dataset F100 containing evaluation report information V1 related to rheumatoid arthritis, can output the subject's rheumatoid arthritis index information M2 with high estimation accuracy when the subject's input information K20 (self-reported information V2) is input. This is because the rheumatoid arthritis index information M2 is based on the subject's own self-report (self-reported information V2).

[0349] The estimation unit 42 according to Embodiment 3 receives input information K200 from the learning model TM3 and obtains output information L2 from the learning model TM3. The learning model TM3 receives input information K200 and outputs output information L2 according to the machine learning algorithm.

[0350] The learning device 3 according to Embodiment 3 generates a learning model TM3 by repeatedly performing learning using a plurality of learning datasets F100.

[0351] The learning data creation device 2 according to Embodiment 3 creates a learning dataset F100 based on the evaluation report information V1 and the correct answer information Z1.

[0352] Next, the present disclosure will be specifically described based on examples, but the present disclosure is not limited to the following examples. For example, the examples do not indicate that all features described in the examples are essential, but merely provide preferred examples or illustrations. For example, the estimated rheumatoid arthritis index information is merely an example. For example, the machine learning algorithm in the examples is merely an example.

[0353] In the embodiments of this disclosure, the estimation accuracy (classification ability) of the learning model and each feature were evaluated. In particular, the contribution of each feature to the estimation result of the learning model was evaluated by the average absolute value M of the SHAP (SHapley Additive exPlanation) value. The average absolute value M of the SHAP value has the characteristic that the value itself becomes low when multiple features contribute to the prediction, that is, when the contribution of a feature is distributed among many features. Therefore, the magnitude of the average M did not indicate the absolute contribution to the estimation result. For this reason, even features with a small average M only indicate a relatively low contribution, and it was inferred that they still contribute to the estimation result and improve the estimation accuracy.

[0354] Furthermore, in Figures 23 and 24, which show the SHAP values, the feature quantities on the horizontal axis of the graph are abbreviated using the alphabets defined in Tables 3 to 8.

[0355] (Examples 1 to 4) The learning model according to Examples 1 to 4 of this disclosure was the learning model TM2 according to Embodiment 2. The learning model TM2 was a model for binary classification. The machine learning algorithm was gradient boosting. The feature information G10 of the learning dataset F10 used to train the learning model TM2 consisted of 1022 types of data (various statistical indicators, various cumulative indicators) from the multiple types of data (vital information H1, behavioral information J1) and multiple types of cumulative data for the multiple types of data shown in Embodiment 1, evaluation report information V1 (overall condition information, overall pain information, number of tender joints information, number of swollen joints information) shown in Embodiment 2, basic physical information Q1, and environmental information R1.

[0356] The correct label B1 for the training dataset F10 was "0" indicating remission and "1" indicating non-remission. In creating the correct label B1 for Example 1, a value of 3.33 or less was set to "0" indicating remission, and a value greater than 3.33 was set to "1" indicating non-remission. In creating the correct label B1 for Example 2, a value of 2.8 or less was set to "0" indicating remission, and a value greater than 2.8 was set to "1" indicating non-remission. In creating the correct label B1 for Example 3, a value of less than 2.3 was set to "0" indicating remission, and a value of 2.3 or greater was set to "1" indicating non-remission. In creating the correct label B1 for Example 4, a DAS28_ESR of less than 2.6 was assigned a value of "0" indicating remission, and a DAS28_ESR of 2.6 or greater was assigned a value of "1" indicating non-remission.

[0357] The input information K20 to the learning model TM2 according to Examples 1 to 4 consisted of 1022 types of data (various statistical indicators, various cumulative indicators) from the multiple types of data (vital information H2, behavioral information J2) and multiple types of cumulative data for the multiple types of data shown in Embodiment 1, as well as self-report information V2 (general condition information, general pain information, number of tender joints information, number of swollen joints information) shown in Embodiment 2, basic physical information Q2, and environmental information R2.

[0358] In Example 1, the output information L2 of the learning model TM2 was rheumatoid arthritis index information M2 related to SDAI. In Example 2, the output information L2 of the learning model TM2 was rheumatoid arthritis index information M2 related to CDAI. In Example 3, the output information L2 of the learning model TM2 was rheumatoid arthritis index information M2 related to DAS28_CRP. In Example 4, the output information L2 of the learning model TM2 was rheumatoid arthritis index information M2 related to DAS28_ESR.

[0359] Figure 22 is a graph showing the ROC (Receiver Operating Characteristic) curve 500 calculated for the learning model TM2 according to Example 1. The horizontal axis represents "1 - specificity," and the vertical axis represents sensitivity. The AUC (Area Under the Curve) was 0.936. AUC represents the area under the ROC curve 500. The closer the AUC is to 1, the higher the estimation accuracy (classification ability) of the learning model TM2. Therefore, it was confirmed that the estimation accuracy of the learning model TM2 according to Example 1 is high.

[0360] Figure 23 is a graph showing the SHAP values ​​calculated for the learning model TM2 according to Example 1. Figure 23 shows the top 20 features. The horizontal axis represents each feature, and the vertical axis represents the average of the absolute values ​​of the SHAP values ​​for each feature, M(|SHAP|).

[0361] As shown in Figure 23, it was confirmed that self-reported information V2, such as the feature "Pt_G_VAS," had a relatively high contribution. In other words, the top three features were self-reported information V2. In addition, it was confirmed that vital information H2, such as the feature "lAwake_tMax_HR," and behavioral information J2, such as the features "Awake_dtMin_METs" and "Sleep_RMSSD_STEP," had a relatively high contribution. Furthermore, it was confirmed that cumulative data such as the features "AdjCumOrg_Awake_Med_STEP" and "AdjCumOrg_Sleep_dtMin_METs" had a relatively high contribution. It was also confirmed that information with added conditions related to sleep, such as the features "Sleep_tLiftR_METs" and "Sleep_Slope_HR," had a relatively high contribution. It was confirmed that information with conditions related to arousal, such as the features "Awake_dtMin_METs" and "lAwake_tMax_HR," had a relatively high contribution. It was also confirmed that information with conditions related to the time of day caused by sunlight, such as the feature "Night_M_HR," had a relatively high contribution. It was also confirmed that information related to alcohol consumption, such as the feature "AdjCumOrg_score3," had a relatively high contribution. In Figure 23, on the horizontal axis, "minute_count" represents the wearing time of the wearable device (biometric state detection device 103).

[0362] Table 1 shows the AUC obtained from the ROC curves calculated for the learning models TM2 of Examples 1 to 4. From the AUC of Examples 2 to 4, not just Example 1, it was confirmed that the estimation accuracy of the learning models TM2 related to Examples 2 to 4 is high.

[0363]

[0364] Although Examples 1 to 4 demonstrate binary classification, it was hypothesized that a similarly high-accuracy learning model TM2 could be generated when directly estimating rheumatoid arthritis indices numerically, or when performing multi-class classification, by using the same learning dataset F10.

[0365] Furthermore, in the graph in Figure 23, the relative contributions of features related to self-reported information V2 were among the highest. Therefore, it could be inferred that the learning model TM3 according to Embodiment 3, which does not use vital information H1, H11, H2, H21 and behavioral information J1, J11, J2, J22, has good estimation accuracy, similar to the learning models TM2 of Embodiments 1 to 4.

[0366] (Examples 5 to 8) The learning models in Examples 5 to 8 of this disclosure were the learning model TM1 according to Embodiment 1. The learning model TM1 was a model for performing binary classification. The machine learning algorithm was gradient boosting. The feature information G1 of the learning dataset F1 used to train the learning model TM1 consisted of 1022 types of data (various statistical indicators, various cumulative indicators) from the multiple types of data (vital information H1, behavioral information J1) and multiple types of cumulative data for the multiple types of data shown in Embodiment 1, as well as basic physical information Q1 and environmental information R1. The correct labels B1 were the same as in Examples 1 to 4.

[0367] The input information K2 to the learning model TM1 in Examples 5 to 8 consisted of 1022 types of data (various statistical indicators, various cumulative indicators) from the multiple types of data (vital information H2, behavioral information J2) and multiple types of cumulative data for the multiple types of data shown in Embodiment 1, as well as basic physical information Q2 and environmental information R2. The output information L2 was the same as in Examples 1 to 4.

[0368] Table 2 shows the AUC obtained from the ROC curves calculated for the learning models TM1 of Examples 5 to 8. From the AUC of Examples 5 to 8, it was confirmed that the estimation accuracy of the learning models TM1 related to Examples 5 to 8 is high.

[0369]

[0370] Although Examples 5 to 8 demonstrated binary classification, it was hypothesized that a similarly high-accuracy learning model TM1 could be generated when directly estimating rheumatoid arthritis indices numerically, or in the case of multi-class classification, by using the same learning dataset F1.

[0371] Furthermore, the difference between Examples 5-8 and Examples 1-4 was that Examples 5-8 did not use the evaluation report information V1 and self-reported information V2 from Examples 1-4. Therefore, it could be inferred that the state in which the features related to self-reported information V2 (Pt_G_VAS, Pt_P_VAS, PtSJCount, PtTJCount) are excluded from the features on the horizontal axis of the graph in Figure 23 is similar to the contribution of the features to the estimation results of the learning model TM1 in Examples 5-8.

[0372] (Example 9) The learning model according to Example 9 of this disclosure was the learning model TM1 according to Embodiment 1. The learning model TM1 was a model that directly estimated joint pain scale values. The machine learning algorithm was gradient boosting. The feature information G1 of the training dataset F1 used to train the learning model TM1 was the same as in Examples 5 to 8. The correct label B1 was the joint pain scale value (VAS value). The input information K2 to the learning model TM1 according to Example 9 was the same as in Examples 5 to 8. The output information L2 was the joint pain scale value (VAS value). The joint pain scale value was expressed as a percentage with the total length of the joint pain scale set to 100%.

[0373] The mean absolute error (MAE) of the output information L2 was 15.203%. This confirmed that the learning model TM1 according to Example 9 has high estimation accuracy.

[0374] Figure 24 is a graph showing the SHAP values ​​calculated for the learning model TM1 according to Example 9. Figure 24 shows the top 30 features. The horizontal axis represents each feature, and the vertical axis represents the average of the absolute values ​​of the SHAP values ​​for each feature, M(|SHAP|).

[0375] As shown in Figure 24, vital information H2 such as the feature "AdjCumOrg_Sleep_Slope_HR", and behavioral information J2 such as the features "CENTERED_lSleep_tLiftR_STEP" and "CENTERED_Night_tMin_METs" were found to have a relatively high contribution. Cumulative data such as the features "AdjCumOrg_Awake_Low_METs" and "AdjCumOrg_Sleep_Slope_HR" were found to have a relatively high contribution. Furthermore, information with added conditions related to sleep, such as the feature "CENTERED_lSleep_tLiftR_STEP", was found to have a relatively high contribution. Information with added conditions related to arousal, such as the feature "AdjCumOrg_Awake_High_STEP", was found to have a relatively high contribution. We confirmed that information with conditions related to the time of day caused by the sun, such as the feature "CENTERED_Night_tMin_METs," has a relatively high contribution. We also confirmed that information on the difference between day and night, such as the feature "CENTERED_Diff_tMin_METs," has a relatively high contribution. Furthermore, we confirmed that environmental information R2, such as the feature "month," has a relatively high contribution.

[0376] Although Example 9 demonstrated the case of directly estimating rheumatoid arthritis indices numerically, it was inferred that a similarly high-accuracy learning model TM1 could be generated in the cases of binary classification and multi-class classification using the same learning dataset F1.

[0377] (Explanation of abbreviations on the horizontal axis of SHAP values) Table 3 shows abbreviations for self-reported information V2 and basic physical information Q2.

[0378]

[0379] Table 4 shows the abbreviations for cumulative processing and centering processing applied to each indicator. Centering processing is the process of subtracting the average value of an individual's indicators over the observation period from the individual's indicators over each unit period.

[0380]

[0381] Table 5 shows abbreviated information regarding sleep, wakefulness, and the time of day caused by the sun.

[0382]

[0383] Table 6 shows the processing applied to heart rate, steps, and exercise intensity.

[0384]

[0385] Table 7 shows abbreviations for heart rate, steps, and exercise intensity.

[0386]

[0387] Table 8 shows other abbreviations.

[0388]

[0389] Preferred embodiments and modifications of the present disclosure have been described in detail above with reference to the attached drawings, but the technical scope of the present disclosure is not limited to such examples. It is clear to any person with ordinary skill in the art of the present disclosure that various modifications or alterations can be conceived within the scope of the technical idea set forth in the claims, and these too are understood to fall within the technical scope of the present disclosure.

[0390] The apparatus or system described herein may be implemented as a single apparatus, or it may be implemented by a plurality of apparatuses (e.g., a cloud server) that are partially or entirely connected by a network. For example, some or all of the statistical processing unit 410, condition processing unit 411, sleep processing unit 412, alcohol consumption estimation unit 413, mental and physical recovery ability estimation unit 414, motor responsiveness estimation unit 415, accumulation processing unit 416, estimation unit 42, and post-processing unit 43 in Figure 6 may be implemented by the same computer or server. For example, the statistical processing unit 410, condition processing unit 411, sleep processing unit 412, alcohol consumption estimation unit 413, mental and physical recovery ability estimation unit 414, motor responsiveness estimation unit 415, accumulation processing unit 416, estimation unit 42, and post-processing unit 43 in Figure 6 may each be implemented by separate computers or servers. These points also apply to the statistical processing unit 210, condition processing unit 211, sleep processing unit 212, alcohol consumption estimation unit 213, mental and physical recovery ability estimation unit 214, motor responsiveness estimation unit 215, cumulative processing unit 216, and correct label creation unit 217 in Figure 18. For example, the relay server 44 or estimation device 4 in Figure 4 may also have the functions of an information provision server 47. For example, the first database 45 and the second database 46 may be implemented on a single computer or server. For example, the estimation unit 42 may have the functions of a relay server 44.

[0391] The series of processes performed by the apparatus described herein may be implemented using software, hardware, or a combination of software and hardware. Computer programs for implementing each function of the processing units 21, 31, and 400 can be created and implemented on a PC or the like. Furthermore, a computer-readable storage medium containing such a computer program can also be provided. Examples of storage media include magnetic disks, optical disks, magneto-optical disks, and flash memory. Alternatively, the above-mentioned computer program may be distributed without using a storage medium, for example, via a network.

[0392] Furthermore, the processes described using flowcharts in this specification do not necessarily have to be executed in the order shown. Some processing steps may be executed in parallel. Additional processing steps may be adopted, and some processing steps may be omitted.

[0393] In Figures 15, 17, and 19, the processing units 400, 31, and 21 execute the computer programs stored in the memory units 402, 35, and 25, thereby executing each step included in the estimation process (estimation method), learning method, and learning data creation method. In other words, the computer program causes the processing units 400, 31, and 21 to execute each step included in the estimation process (estimation method), learning method, and learning data creation method. The processing units 400, 31, and 21 correspond to an example of the "computer" in this disclosure. To put it another way, the computer program product realizes each step included in the estimation process (estimation method), learning method, and learning data creation method when the computer program is executed by the processing units 400, 31, and 21.

[0394] Furthermore, the estimation device 4 in Figure 6 may not include all or part of, for example, the sleep processing unit 412, the alcohol consumption estimation unit 413, the mental and physical recovery ability estimation unit 414, the motor responsiveness estimation unit 415, and the accumulation processing unit 416. The estimation process in Figure 15 may not include all or part of, for example, steps S34 to S38. Also, the estimation process in Figure 15 may not include, for example, steps S41 and S42. The learning data creation device 2 in Figure 18 may not include all or part of, for example, the sleep processing unit 212, the alcohol consumption estimation unit 213, the mental and physical recovery ability estimation unit 214, the motor responsiveness estimation unit 215, and the accumulation processing unit 216. The learning data creation method in Figure 19 may not include all or part of, for example, steps S304 to S308.

[0395] Furthermore, the learning device 3 may generate a new learning model (distilled model) by performing learning using the input information K2 input to the learning model TM1 and the output information L2 or output information N2 output by the learning model TM1 as a learning dataset. The input information K2 and output information L2 and N2 are information about the subject for the learning model TM1 at the utilization stage. However, in the generation of the distilled model, the input information K2 and output information L2 and N2 constituting the learning dataset correspond to information about the learning subject for the learning device 3. Similarly, the learning device 3 may generate a distilled model by using the input information K20 and output information L2 and N2 from Embodiment 2, or the input information K200 and output information L2 and N2 from Embodiment 3, as a learning dataset.

[0396] Furthermore, in embodiments 1 and 2, the learning datasets F1 and F10 may include at least one of the vital information H1 and behavioral information J1 of the learning subject, and the rheumatoid arthritis index information of the learning subject. The input information K2 and K20 may include at least one of the vital information H2 and behavioral information J2 of the subject.

[0397] Furthermore, the effects described herein are merely descriptive or illustrative and not limiting. In other words, the technology relating to this disclosure may produce other effects that are obvious to those skilled in the art from the description herein, in addition to or in lieu of the effects described herein.

[0398] Furthermore, the following configurations also fall within the technical scope of this disclosure.

[0399] (Item 1) An estimation device for estimating rheumatoid arthritis index information of a subject, comprising: an estimation unit that inputs input information to a learning model constructed by learning using a learning dataset and obtains output information from the learning model; and a storage unit that stores the output information, wherein the learning dataset includes vital information of the learning subject, behavioral information of the learning subject, and rheumatoid arthritis index information of the learning subject; the input information includes vital information of the subject and behavioral information of the subject; and the output information includes the rheumatoid arthritis index information of the subject.

[0400] (Item 2) The estimation device according to Item 1, wherein the learning dataset further includes evaluation report information including an evaluation report of the learning subject regarding rheumatoid arthritis, and the input information further includes the subject's self-reported information regarding rheumatoid arthritis.

[0401] (Item 3) The estimation device according to Item 2, wherein the rheumatoid arthritis index information includes at least one of the following: information regarding CDAI, information regarding SDAI, and information regarding DAS28.

[0402] (Item 4) The estimation device according to Item 1, wherein the rheumatoid arthritis index information includes at least one of the following: information on CDAI, information on SDAI, information on DAS28, and information on joint pain.

[0403] (Item 5) The estimation device according to any one of Items 1 to 4, wherein the vital information includes information about heart rate, and the behavioral information includes at least one of the following: information about the number of steps and information about exercise intensity.

[0404] (Item 6) The estimation device according to any one of Items 1 to 5, wherein the vital information further includes at least one of the following: information on physical and mental recovery ability and information on exercise responsiveness, the information on physical and mental recovery ability includes information indicating the degree of physical and mental recovery ability from the end of exercise, and the information on exercise responsiveness is information indicating the degree of physical and mental response to exercise.

[0405] (Item 7) The estimation device according to any one of Items 1 to 6, wherein the behavioral information further includes information relating to sleep.

[0406] (Item 8) The estimation device according to any one of Items 1 to 7, wherein the behavioral information further includes information relating to drinking.

[0407] (Item 9) The estimation device according to any one of Items 1 to 8, wherein at least one of the vital information and the behavioral information includes vital information or behavioral information during a period indicated by specific conditions, and the specific conditions include at least one of the conditions relating to sleep, conditions relating to wakefulness, and conditions relating to the time of day caused by the sun.

[0408] (Item 10) The estimation device according to Item 9, further comprising: a condition processing unit that performs a process for extracting vital raw data from the subject's vital raw data for a period indicated by the specific conditions, and / or a process for extracting behavioral raw data from the subject's behavioral raw data for a period indicated by the specific conditions; and a statistical processing unit that generates vital information by performing statistical processing on the extracted vital raw data, and / or generates behavioral information by performing statistical processing on the extracted behavioral raw data.

[0409] (Item 11) The estimation device according to any one of Items 1 to 10, wherein at least one of the vital information and the behavioral information includes vital information or behavioral information on which cumulative processing has been performed, the cumulative processing indicates a process of accumulating the vital information or behavioral information while avoiding the value diverging, and the device further comprises a cumulative processing unit that performs the cumulative processing.

[0410] (Item 12) The estimation device according to Item 11, wherein the cumulative processing includes a classification execution process, a correction execution process, and a cumulative execution process, wherein the classification execution process classifies a plurality of vital data or behavioral data arranged in a time series into at least a first state group indicating a first state of mind and body and a second state group indicating a second state of mind and body, thereby obtaining a plurality of first state data belonging to the first state group and a plurality of second state data belonging to the second state group, the correction execution process calculates a plurality of correction data in a time series by adjusting the balance of magnitude between the plurality of first state data and the plurality of second state data, and the cumulative execution process calculates a plurality of cumulative data in a time series by accumulating the correction data along the time axis, and the vital information or behavioral information on which the cumulative processing has been performed is composed of the plurality of cumulative data.

[0411] (Item 13) The estimation device according to any one of Items 1 to 12, wherein the learning dataset further includes basic information about the body of the learning subject, the input information further includes basic information about the body of the subject, and the basic information includes at least one of BMI information and body shape information.

[0412] (Item 14) The estimation device according to any one of Items 1 to 13, wherein the training dataset further includes information about the environment of the training subject, the input information further includes information about the environment of the subject, and the environment includes seasonal information.

[0413] (Item 15) A learning model constructed by learning using a learning dataset, which causes a computer to function in order to estimate rheumatoid arthritis index information of a subject, wherein the learning dataset includes vital information of the learning subject, behavioral information of the learning subject, and rheumatoid arthritis index information of the learning subject, and the computer is configured to take input information as input and output information as output, wherein the input information includes vital information of the subject and behavioral information of the subject, and the output information includes the rheumatoid arthritis index information of the subject.

[0414] (Item 16) An estimation method for estimating rheumatoid arthritis index information of a subject, comprising the steps of: inputting input information into a learning model constructed by learning using a learning dataset; and obtaining output information from the learning model into which the input information has been input, wherein the learning dataset includes vital information of the learning subject, behavioral information of the learning subject, and rheumatoid arthritis index information of the learning subject; the input information includes vital information of the subject and behavioral information of the subject; and the output information includes the rheumatoid arthritis index information of the subject.

[0415] (Item 17) A computer program that causes a computer to execute the estimation method described in Item 16.

[0416] (Item 18) A method for generating a learning model, comprising the steps of: obtaining a learning dataset; and generating a learning model that outputs output information when input information is input by performing learning using the learning dataset, wherein the learning dataset includes vital information of a learning subject, behavioral information of the learning subject, and rheumatoid arthritis index information of the learning subject; the input information includes vital information of the subject and behavioral information of the subject; and the output information includes the rheumatoid arthritis index information of the subject.

[0417] (Item 19) A computer program that causes a computer to execute the learning model generation method described in Item 18.

[0418] This disclosure provides an estimation device, a learning model, an estimation method, a learning model generation method, and a computer program, and has industrial applicability.

[0419] 2 Learning data creation device, 3 Learning device, 4 Estimation device, 42 Estimation unit, 402 Memory unit, 410 Statistical processing unit, 411 Condition processing unit, 412 Sleep processing unit, 413 Alcohol consumption estimation unit, 414 Mental and physical recovery ability estimation unit, 415 Motor responsiveness estimation unit, 416 Cumulative processing unit

Claims

1. An estimation device for estimating rheumatoid arthritis index information of a subject, comprising: an estimation unit that inputs input information to a learning model constructed by learning using a learning dataset and obtains output information from the learning model; and a storage unit that stores the output information, wherein the learning dataset includes vital information of the learning subject, behavioral information of the learning subject, and rheumatoid arthritis index information of the learning subject; the input information includes vital information of the subject and behavioral information of the subject; and the output information includes the rheumatoid arthritis index information of the subject.

2. The estimation device according to claim 1, wherein the learning dataset further includes evaluation report information including an evaluation report of the learning subject regarding rheumatoid arthritis, and the input information further includes self-reported information of the subject regarding rheumatoid arthritis.

3. The estimation device according to claim 2, wherein the rheumatoid arthritis index information includes at least one of the following: information regarding CDAI, information regarding SDAI, and information regarding DAS28.

4. The estimation device according to claim 1, wherein the rheumatoid arthritis index information includes at least one of the following: information regarding CDAI, information regarding SDAI, information regarding DAS28, and information regarding joint pain.

5. The estimation device according to claim 1 or 2, wherein the vital information includes information relating to heart rate, and the behavioral information includes at least one of the following: information relating to the number of steps and information relating to exercise intensity.

6. The estimation device according to claim 1 or 2, wherein the vital information further includes at least one of the following: information on physical and mental recovery ability and information on exercise responsiveness, the information on physical and mental recovery ability includes information indicating the degree of physical and mental recovery ability from the end of exercise, and the information on exercise responsiveness is information indicating the degree of physical and mental response to exercise.

7. The estimation device according to claim 1 or 2, wherein the behavioral information further includes information relating to sleep.

8. The estimation device according to claim 1 or 2, wherein the behavioral information further includes information relating to alcohol consumption.

9. The estimation device according to claim 1 or 2, wherein at least one of the vital information and the behavioral information includes vital information or behavioral information during a period indicated by specific conditions, and the specific conditions include at least one of conditions relating to sleep, conditions relating to wakefulness, and conditions relating to the time of day caused by the sun.

10. The estimation device according to claim 9, further comprising: a condition processing unit that performs a process for extracting vital raw data from the subject's vital raw data for a period indicated by the specific conditions, and / or a process for extracting behavioral raw data from the subject's behavioral raw data for a period indicated by the specific conditions; and a statistical processing unit that generates vital information by performing statistical processing on the extracted vital raw data, and / or generates behavioral information by performing statistical processing on the extracted behavioral raw data.

11. The estimation device according to claim 1 or 2, wherein at least one of the vital information and the behavioral information includes vital information or behavioral information on which cumulative processing has been performed, the cumulative processing represents a process of accumulating the vital information or behavioral information while avoiding the value diverging, and the device further comprises a cumulative processing unit that performs the cumulative processing.

12. The estimation device according to claim 11, wherein the cumulative processing includes a classification execution process, a correction execution process, and a cumulative execution process, wherein the classification execution process involves classifying a plurality of vital data or behavioral data arranged in a time series into at least a first state group indicating a first state of mind and body and a second state group indicating a second state of mind and body, thereby obtaining a plurality of first state data belonging to the first state group and a plurality of second state data belonging to the second state group, the correction execution process involves adjusting the balance of magnitude between the plurality of first state data and the plurality of second state data to calculate a plurality of correction data in a time series, the cumulative execution process involves accumulating the correction data along a time axis to calculate a plurality of cumulative data in a time series, and the vital information or behavioral information on which the cumulative processing has been performed is composed of the plurality of cumulative data.

13. The estimation device according to claim 1 or 2, wherein the learning dataset further includes basic information relating to the body of the learning subject, the input information further includes basic information relating to the body of the subject, and the basic information includes at least one of BMI information and body shape information.

14. The estimation apparatus according to claim 1 or 2, wherein the training dataset further includes information about the environment of the training subject, the input information further includes information about the environment of the subject, and the environment includes seasonal information.

15. A learning model constructed by learning using a training dataset, which causes a computer to function in order to estimate rheumatoid arthritis index information of a subject, wherein the training dataset includes vital information of the training subject, behavioral information of the training subject, and rheumatoid arthritis index information of the training subject, and the computer is configured to take input information as input and output information as output, wherein the input information includes the vital information of the subject and behavioral information of the subject, and the output information includes the rheumatoid arthritis index information of the subject.

16. An estimation method for estimating rheumatoid arthritis index information of a subject, comprising the steps of: inputting input information into a learning model constructed by learning using a learning dataset; and obtaining output information from the learning model into which the input information has been input, wherein the learning dataset includes vital information of the learning subject, behavioral information of the learning subject, and rheumatoid arthritis index information of the learning subject; the input information includes vital information of the subject and behavioral information of the subject; and the output information includes the rheumatoid arthritis index information of the subject.

17. A computer program that causes a computer to execute the estimation method described in claim 16.

18. A method for generating a learning model, comprising the steps of: obtaining a learning dataset; and generating a learning model that outputs output information when input information is input by performing learning using the learning dataset, wherein the learning dataset includes vital information of a learning subject, behavioral information of the learning subject, and rheumatoid arthritis index information of the learning subject; the input information includes vital information of the subject and behavioral information of the subject; and the output information includes the rheumatoid arthritis index information of the subject.

19. A computer program that causes a computer to execute the learning model generation method described in claim 18.

Images