Body condition abnormality detection method, body condition abnormality detection program, and body condition abnormality detection system

Abstract

To provide a body condition abnormality detection method, a body condition abnormality detection program, and a body condition abnormality detection system capable of detecting, with a higher degree of precision, body condition abnormality of a subject on the basis of measurement data on a body surface temperature.SOLUTION: A first identifier 26 is created by performing machine learning with a plurality of first normal time feature amount vectors S1B consisting of a normal time body surface temperature feature amount STT and a normal time air temperature feature amount SKT as inputs. A second identifier 28 is created by performing machine learning with a plurality of second normal time feature amount vectors S2B consisting of the normal time body surface temperature feature amount STT and a normal time humidity feature amount SST as inputs. A first determination is performed by inputting a first detection feature amount vector K1B consisting of a detection body surface temperature feature amount KTT and a detection air temperature feature amount KKT to the first identifier 26. A second determination is performed by inputting a second detection feature amount vector K2B consisting of the detection body surface temperature feature amount KTT and a detection humidity feature amount KST to the second identifier 28. A total value An of a non-normality degree A in continuous n periods T, which is a total value of counter values in the first and second determinations for each period T, is obtained.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to an abnormal physical condition detection method, an abnormal physical condition detection program, and an abnormal physical condition detection system for detecting an abnormal physical condition in a subject. [Background technology]

[0002] In nursing care facilities, it is necessary to quickly detect abnormalities in the health of elderly people. Body temperature is an important indicator for detecting abnormalities in health, but measuring body temperature is restrictive for the subject, making it unsuitable for monitoring, which requires frequent measurements.

[0003] In recent years, a technology has been proposed that uses a non-contact infrared thermometer to observe the body surface temperature of a subject and detect abnormalities based on the measured body surface temperature data. However, since body surface temperature is constantly fluctuating due to environmental changes, it is difficult to determine the range of body surface temperature within which a person's physical condition can be considered normal. Therefore, the challenge is how to accurately detect abnormalities in the subject's physical condition from the measured body surface temperature data.

[0004] Conventionally, as disclosed in Patent Document 1, for example, there has been a monitoring method that detects information on the subject's body surface temperature, sets the body surface temperature at a predetermined reference time as a reference value, and outputs an alert display when the difference between the body surface temperature detected thereafter and the reference value exceeds a predetermined threshold. This monitoring method focuses on the range of change in body surface temperature, rather than the absolute value, making it less susceptible to environmental fluctuations. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Patent Publication No. 2021-96760 Summary of the Invention [Problem to be solved by the invention]

[0006] The monitoring method of Patent Document 1 is considered to improve detection accuracy to a certain extent compared to detecting abnormal physical conditions directly from body surface temperature, but the effect is limited.

[0007] The present invention has been made in consideration of the above-mentioned background technology, and aims to provide an abnormal physical condition detection method, an abnormal physical condition detection program, and an abnormal physical condition detection system that can detect abnormal physical condition of a subject with greater accuracy based on measurement data of body surface temperature. [Means for solving the problem]

[0008] The present invention relates to a method for detecting an abnormality in a physical condition of a subject, the method being executed by a computer system, a classifier preparation step of collecting a plurality of combinations of information on normal-state body surface temperature features indicating characteristics of time-varying changes in body surface temperature when the subject is normal during a certain period T, information on normal-state air temperature features indicating characteristics of time-varying changes in air temperature in the living space of the subject during the period T, and information on normal-state humidity features indicating characteristics of time-varying changes in humidity in the living space of the subject during the period T, and creating a first classifier by performing machine learning using as input a plurality of first normal-state feature vectors consisting of combinations of the normal-state body surface temperature features and the normal-state air temperature features for contemporaneous periods T, and creating a second classifier by performing machine learning using as input a plurality of second normal-state feature vectors consisting of combinations of the normal-state body surface temperature features and the normal-state humidity features for contemporaneous periods T, or acquiring the created first and second classifiers; a subject data acquisition step of acquiring measurement data of the subject's body surface temperature, the air temperature of the living space, and the humidity of the living space; a detected feature extraction step of analyzing the measurement data acquired in the subject data acquisition step for each successive period T of a certain length, and extracting a detected body surface temperature feature that indicates a characteristic of a time change in the body surface temperature of the subject, extracting a detected air temperature feature that indicates a characteristic of a time change in the air temperature of the living space of the subject, and extracting a detected humidity feature that indicates a characteristic of a time change in the humidity of the living space of the subject; a tentative determination step of making a first determination as to whether the first detected feature vector "belongs to the normal group" or "does not belong to the normal group" by inputting a first detected feature vector consisting of the detected body surface temperature feature and the detected air temperature feature for each period T to the first classifier, and making a second determination as to whether the second detected feature vector "belongs to the normal group" or "does not belong to the normal group" by inputting a second detected feature vector consisting of the detected body surface temperature feature and the detected humidity feature for each period T to the second classifier; an abnormality degree calculation step of counting, for each period T, zero when the first judgment judges that the test result "belongs to the normal group" and +1 when the first judgment judges that the test result "does not belong to the normal group", and counting zero when the second judgment judges that the test result "belongs to the normal group" and +1 when the second judgment judges that the test result "does not belong to the normal group", and determining the sum of the counts of the first and second judgments for each period T as the abnormality degree A for that period T; This method for detecting abnormal physical condition includes a main judgment step of calculating a sum An of the degree of abnormality A for each of n consecutive periods T (wherein the constant n is a natural number greater than or equal to 2) and judging that "an abnormal physical condition of the subject has been detected" if the sum An is greater than or equal to a reference value Ar.

[0009] In the abnormality calculation step, a total value An(k) of the degrees of abnormality A for each of n periods T(k) to T(k+n-1), a total value An(k+1), ... are calculated in sequence, and in the main judgment step, a judgment is made as to whether or not "an abnormality in the subject's physical condition has been detected" for each of the total values ​​An(k), An(k+1), ....

[0010] The first classifier is a classifier in which information on one or more normal clusters generated by applying a predetermined clustering algorithm to a plurality of the first normal feature vectors is registered, and the second classifier is a classifier in which information on one or more normal clusters generated by applying a predetermined clustering algorithm to a plurality of the second normal feature vectors is registered, and in the first determination of the provisional determination step, the determination of whether the object “belongs to a normal group” or “does not belong to a normal group” is made by using a Mahalanobis function between a cluster centroid in the one or more normal clusters registered in the first classifier and the first detection feature vector. The normal cluster having the closest distance is extracted, and if the Mahalanobis distance of the extracted normal cluster to the cluster centroid is equal to or greater than a first threshold Dr1, it is determined that the cluster does not belong to the normal group. In the second determination of the provisional determination step, the determination of whether the cluster belongs to the normal group or not is performed by extracting the normal cluster having the closest Mahalanobis distance between the cluster centroid and the second detection feature vector from among one or more normal clusters registered in the second classifier, and if the Mahalanobis distance of the extracted normal cluster to the cluster centroid is equal to or greater than a second threshold Dr2, it is determined that the cluster does not belong to the normal group. In this case, it is preferable that information on normal clusters generated by applying the k-means method is registered in the first and second classifiers.

[0011] Furthermore, in the classifier preparation step, a plurality of combinations of information on the normal state body surface temperature feature amount and information on normal state environment information feature amount indicating characteristics of time change of environmental information such as air pressure other than the air temperature and humidity in the living space of the subject during a period T that is the same as the normal state body surface temperature feature amount are collected, and a third classifier is created by performing machine learning using as input a plurality of third normal state feature amount vectors each consisting of combinations of the normal state body surface temperature feature amount and the normal state environment information feature amount during the same period T, or the created third classifier is acquired, and in the subject data acquisition step, measurement data of environmental information of the living space of the subject is acquired, and in the detected feature amount extraction step, the measured data of the environmental information acquired in the subject data acquisition step is used as the detected body surface temperature feature amount. The feature extraction step performs an analysis for each period T when the feature is extracted, and performs a process of extracting detected environmental information feature values ​​that indicate characteristics of time-dependent changes in environmental information of the subject's living space. In the provisional judgment step, a third detected feature vector consisting of the detected body surface temperature feature value and the detected environmental information feature value for each period T is input to the third classifier, and a third judgment is made to determine whether the third detected feature vector "belongs to the normal group" or "does not belong to the normal group". In the abnormality calculation step, if the third judgment determines that the feature vector "belongs to the normal group", zero is counted, and if the third judgment determines that the feature vector "does not belong to the normal group", +1 is counted. The total value of the counts of the first, second and third judgments for each period T can be set as the abnormality A for that period T.

[0012] In this case, the third classifier is configured to register information on one or more normal clusters generated by applying a predetermined clustering algorithm to the third normal feature vectors, and in the third judgment in the provisional judgment step, the judgment of "belonging to the normal group" or "not belonging to the normal group" is configured to extract the normal cluster having the shortest Mahalanobis distance between the cluster centroid and the third detection feature vector from the one or more normal clusters registered in the third classifier, and judge "not belonging to the normal group" when the Mahalanobis distance of the extracted normal cluster to the cluster centroid is equal to or greater than a third threshold Dr3. Also, it is preferable that the third classifier is configured to register information on the normal cluster generated by applying the k-means method.

[0013] The method may further include a classifier update step for performing a process of updating each of the classifiers based on the measurement data of the subject on which the main judgment step is based when the main judgment step does not judge that "an abnormality in the physical condition of the subject has been detected." In addition, it is preferable that each of the feature amounts for each period T is a value obtained by subtracting the minimum value from the maximum value in the measurement data for the period T.

[0014] The present invention also provides an abnormal physical condition detection program, which is made up of programs for executing each step, for causing a computer system to execute the abnormal physical condition detection method described above.

[0015] The present invention also provides a system for detecting abnormalities in the physical condition of a subject, the system comprising a computer, a classifier preparation unit that collects a plurality of combinations of information on normal-state body surface temperature features indicating characteristics of a time change in body surface temperature when the subject is normal during a certain period T, information on normal-state air temperature features indicating characteristics of a time change in air temperature in the living space of the subject during the period T, and information on normal-state humidity features indicating characteristics of a time change in humidity in the living space of the subject during the period T, and creates a first classifier by performing machine learning using a plurality of first normal-state feature vectors consisting of the normal-state body surface temperature features and the normal-state air temperature features during the same period T as an input, and creates a second classifier by performing machine learning using a plurality of second normal-state feature vectors consisting of the normal-state body surface temperature features and the normal-state humidity features during the same period T as an input, or acquires the created first and second classifiers; a subject data acquisition unit that acquires measurement data of the subject's body surface temperature, measurement data of the air temperature of the living space, and measurement data of the humidity of the living space; a detected feature extraction unit that performs a process of analyzing the measurement data acquired in the subject data acquisition step for each successive period T of a certain length, and extracting a detected body surface temperature feature that indicates a feature of a time change in the body surface temperature of the subject, a process of extracting a detected air temperature feature that indicates a feature of a time change in the air temperature of the living space of the subject, and a process of extracting a detected humidity feature that indicates a feature of a time change in the humidity of the living space of the subject; a tentative determination unit that performs a first determination as to whether the first detected feature vector "belongs to the normal group" or "does not belong to the normal group" by inputting a first detected feature vector consisting of the detected body surface temperature feature and the detected air temperature feature for each time period T to the first classifier, and performs a second determination as to whether the second detected feature vector "belongs to the normal group" or "does not belong to the normal group" by inputting a second detected feature vector consisting of the detected body surface temperature feature and the detected humidity feature for each time period T to the second classifier; an abnormality degree calculation unit that, for each period T, counts zero when the first judgment determines that the test result "belongs to the normal group" and counts +1 when the first judgment determines that the test result "does not belong to the normal group", and counts zero when the second judgment determines that the test result "belongs to the normal group" and counts +1 when the second judgment determines that the test result "does not belong to the normal group", and determines the sum of the counts of the first and second judgments for each period T as the abnormality degree A for that period T; This is an abnormal physical condition detection system that includes a main judgment unit that calculates a sum An of the degree of abnormality A for each of n consecutive periods T (where the constant n is a natural number greater than or equal to 2) and judges that "an abnormal physical condition of the subject has been detected" if the sum An is greater than or equal to a reference value Ar.

[0016] Furthermore, the classifier preparation unit collects a plurality of combinations of information on the normal state body surface temperature feature amount and information on normal state environment information feature amount indicating characteristics of time change of environmental information such as air pressure other than the air temperature and the humidity in the living space of the subject, during a period T that is the same as the normal state body surface temperature feature amount, and performs machine learning using as input a plurality of third normal state feature amount vectors consisting of the normal state body surface temperature feature amount and the normal state environment information feature amount during the same period T, to create a third classifier, or to acquire the created third classifier, the subject data acquisition unit acquires measurement data of environmental information of the living space of the subject, and the detected feature amount extraction unit extracts the detected body surface temperature feature amount from the measurement data of the environmental information acquired in the subject data acquisition step. The subject is analyzed for each period T when the subject is examined, and a process is performed to extract detected environmental information features indicating characteristics of time-dependent changes in environmental information of the subject's living space. The tentative judgment unit inputs a third detected feature vector consisting of the detected body surface temperature feature and the detected environmental information feature for each period T to the third classifier, thereby making a third judgment to judge whether the third detected feature vector "belongs to the normal group" or "does not belong to the normal group". The non-normality calculation unit counts zero if the third judgment judges that the third detected feature vector "belongs to the normal group" and counts +1 if the third judgment judges that the third detected feature vector "does not belong to the normal group". The total value of the counts of the first, second and third judgments for each period T can be set as the non-normality A for the period T.

[0017] The configuration may include a classifier update unit that performs a process of updating each of the classifiers based on the measurement data of the subject on which the main judgment unit has based its judgment that "an abnormality in the subject's physical condition has been detected" when the main judgment unit has not made that judgment. Effect of the Invention

[0018] The abnormal physical condition detection method, abnormal physical condition detection program, and abnormal physical condition detection system of the present invention are configured to detect abnormal physical condition based on measurement data such as body surface temperature and air temperature that can be measured without contact, so that the subject can be gently watched over without feeling restrained or uncomfortable. In addition, multiple classifiers created by machine learning that input feature values ​​based on body surface temperature and surrounding environmental information (air temperature, humidity, etc.) when the subject is in normal physical condition are prepared, and a unique method is used to make judgments, so that abnormal physical condition of the subject can be detected objectively and with high accuracy. [Brief description of the drawings]

[0019] [Figure 1] FIG. 1A is a system configuration diagram showing one embodiment of an abnormal health detection system of the present invention, and FIG. 1B is a flowchart showing an abnormal health detection method executed by this abnormal health detection system (one embodiment of the abnormal health detection method of the present invention). [Diagram 2] FIG. 1(b) shows schematic diagrams (a) to (d) illustrating the contents of the four types of normal feature quantities collected in the classifier preparation step, and (e) to (f) illustrating the contents of three types of classifiers created based on the four types of normal feature quantities. [Diagram 3] 1(b) is a time chart showing four types of measurement data acquired in a subject data acquiring step and four types of detection feature amounts extracted in a detection feature amount extracting step in FIG. [Figure 4] 2 is a table showing a specific example of the processing contents of a tentative judgment step, a degree of abnormality calculation step, and a main judgment step in FIG. [Diagram 5]11 is a table showing the results of a verification experiment conducted to confirm the effectiveness of the abnormal physical condition detection system and abnormal physical condition detection method of this embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0020] <<<One embodiment of the present invention>>> An embodiment of the abnormal physical condition detection method, abnormal physical condition detection program, and abnormal physical condition detection system of the present invention will be described below with reference to the drawings. The abnormal physical condition detection system 10 of this embodiment is provided in a computer, and as shown in Fig. 1(a), is composed of a classifier preparation unit 12, a subject data acquisition unit 14, a detected feature amount extraction unit 16, a tentative judgment unit 18, an abnormality degree calculation unit 20, a main judgment unit 22, and a classifier update unit 24, and judges whether or not "an abnormal physical condition of subject M has been detected" based on multiple measurement data acquired by the subject data acquisition unit 14.

[0021] 1(b), the abnormal physical condition detection method of this embodiment is composed of a classifier preparation step S11, a subject data acquisition step S12, a detected feature extraction step S13, a tentative judgment step S14, a degree of abnormality calculation step S15, a main judgment step S16, and a classifier update step S17, and is executed by the above-mentioned classifier preparation unit 12, subject data acquisition unit 14, detected feature extraction unit 16, tentative judgment unit 18, degree of abnormality calculation unit 20, main judgment unit 22, and classifier update unit 24, respectively. Moreover, the abnormal physical condition detection program of this embodiment is composed of execution programs for each of the steps S11 to S17, and is installed in a computer (abnormal physical condition detection system 10).

[0022] The main feature of this embodiment is that measurement data of the body surface temperature of the subject M, measurement data of the temperature and humidity of the living space in which the subject M lives, and atmospheric pressure as other environmental information are acquired, and these are comprehensively analyzed to detect abnormalities in the physical condition of the subject M. The contents of each of steps S11 to S17 shown in Fig. 1(b) will be described in detail below.

[0023] <<Description of each step of the abnormal physical condition detection method of this embodiment>> <Classifier preparation step S11> In the classifier preparation step S11, multiple sets of information on four types of normal state features are collected, and three types of classifiers are created using the collected information. The four types of normal state features are normal state body surface temperature feature STT, normal state air temperature feature SKT, normal state humidity feature SST, and normal state air pressure feature SAT. The three types of classifiers are a first classifier 26, a second classifier 28, and a third classifier 30.

[0024] The normal body surface temperature feature STT is numerical information indicating the feature of the time change in body surface temperature when the subject M is normal, during a certain period of time T. For example, as shown in Fig. 2(a), the feature STT can be obtained by subtracting the maximum value from the minimum value in the body surface temperature measurement data STD during the period T.

[0025] The normal temperature feature value SKT is numerical information that indicates the characteristics of the time change in temperature in the living space of the subject M during the above-mentioned period T. For example, as shown in FIG. 2(b), the feature value SKT can be obtained by subtracting the minimum value from the maximum value in the temperature measurement data SKD during the period T.

[0026] The normal humidity characteristic value SST is numerical information indicating the characteristics of the time change in humidity in the living space of the subject M during the above-mentioned period T. For example, as shown in FIG. 2(c), the characteristic value SST can be obtained by subtracting the maximum value from the minimum value in the humidity measurement data SSD during the period T.

[0027] The normal atmospheric pressure feature SAT is numerical information indicating the characteristics of the time change in the atmospheric pressure in the living space of the subject M during the above-mentioned period T. For example, as shown in Fig. 2(d), the feature SAT can be determined by subtracting the maximum value from the minimum value in the atmospheric pressure measurement data SAD during the period T. In addition, in Figures 2(a) to (d), the waveforms of the measurement data STD, SKD, SSD, and SAD are shown on the same diagram, but in reality they each have a different waveform.

[0028] The first classifier 26 is created by performing machine learning using a plurality of first normal feature vectors S1B, each of which is a combination of normal body surface temperature feature STT and normal air temperature feature SKT for the same period T. Specifically, a predetermined clustering algorithm is applied to the plurality of first normal feature vectors S1B to generate one or more normal clusters, and the information on the generated clusters is registered. For example, in the case of the first classifier 26 shown in FIG. 2(e), the vertices of the first normal feature vector S1B are represented by plots of "◯", which are grouped into three normal clusters CL1a to CL1c. In addition, the cluster centers J1a to J1c of the normal clusters CL1a to CL1c are represented by plots of "□", and the lines of the first thresholds Dr1a to Dr1c (thresholds defined by the Mahalanobis distance from the cluster centers) that are the boundaries as to whether or not each cluster belongs to a normal cluster are represented by dashed and dotted lines.

[0029] The second classifier 28 is created by performing machine learning using a plurality of second normal feature vectors S2B, each of which is a combination of normal body surface temperature feature STT and normal humidity feature SST for the same period T. Specifically, a predetermined clustering algorithm is applied to the plurality of second normal feature vectors S2B to generate one or more normal clusters, and the information on the generated clusters is registered. For example, in the case of the second classifier 28 shown in FIG. 2(f), the vertices of the second normal feature vector S2B are represented by plots of "◯", which are grouped into three normal clusters CL2a to CL2c. In addition, the cluster centers J2a to J2c of the normal clusters CL2a to CL2c are represented by plots of "□", and the lines of the second thresholds Dr2a to Dr2c (thresholds defined by the Mahalanobis distance from the cluster centers) that are the boundaries as to whether or not each cluster belongs to a normal cluster are represented by dashed lines.

[0030] The third classifier 30 is created by performing machine learning using multiple third normal feature vectors S3B, which are combinations of normal body surface temperature feature STT and normal air pressure feature SAT for the same period T, as input. Specifically, one or more normal clusters are generated by applying a predetermined clustering algorithm to the multiple third normal feature vectors S3B, and the information on the generated clusters is registered. For example, in the case of the third classifier 30 shown in FIG. 2(g), the vertices of the third normal feature vector S3B are represented by plots of "◯", which are grouped into three normal clusters CL3a to CL3c. In addition, the cluster centers J3a to J3c of the normal clusters CL3a to CL3c are represented by plots of "□", and the lines of the third thresholds Dr3a to Dr3c (thresholds defined by the Mahalanobis distance from the cluster centers) that are the boundaries of whether or not each cluster belongs to a normal cluster are represented by dashed and dotted lines.

[0031] In addition, in Figures 2(e) to (g), the normal state feature vectors S1B, S2B, and S3B are shown to be distributed in the same way, and three normal clusters of the same shape are generated, but in reality, each distribution is different, and normal clusters of different shapes are generated.

[0032] It is preferable to use the k-means method as the clustering algorithm used when creating the classifiers 26, 28, and 30, but if the conditions are met, it is also possible to use the EM method, DBSCAN method, or the like.

[0033] In addition, in the above description of the classifier preparation step S11, the three classifiers 26, 28, and 30 are "created," but the previously created classifiers 26, 28, and 30 may be "acquired." "Acquire" refers to retrieving data stored in another device or storage medium, receiving data transmitted from another device, and the like, and also includes reading out data that has already been acquired and stored and making it usable.

[0034] <Subject data acquisition step S12, detection feature extraction step S13> In the subject data acquisition step S12, measurement data is acquired while monitoring the daily life of the subject M. Specifically, as shown in Fig. 3, measurement data KTD of the body surface temperature of the subject M, measurement data KKD of the air temperature of the living space, measurement data KSD of the humidity of the living space, and measurement data KAD of the air pressure of the living space are acquired.

[0035] In the detection feature extraction step S13, the measurement data KTD, KKD, KSD, and KAD acquired in the subject data acquisition step S12 are analyzed in sequence for each period T of a constant length, and four types of detection features are extracted. Specifically, the following processes are performed in sequence: a process of extracting a detected body surface temperature feature KTT(k) showing the characteristics of the time change of the body surface temperature in the period T(k) [k=a natural number of 1, 2, 3, . . .], a process of extracting a detected air temperature feature KKT(k) showing the characteristics of the time change of the air temperature in the period T(k), a process of extracting a detected humidity feature KST(k) showing the characteristics of the time change of the humidity in the period T(k), and a process of extracting a detected air pressure feature KAD(k) showing the characteristics of the time change of the air pressure in the period T(k). Note that each detection feature is a value obtained by subtracting the minimum value from the maximum value of the measurement data in the period T, similar to each normal state feature in the classifier preparation step S11.

[0036] <Temporary Judgment Step S14> In the provisional determination step S14, first to third determinations are made for the four types of detected feature amounts extracted in the detected feature amount extraction step S13.

[0037] The first determination is a process of inputting a first detection feature vector K1B(k) consisting of a detected body surface temperature feature KTT(k) and a detected air temperature feature KKT(k) for a period T(k) into a first classifier 26 to determine whether the first detection feature vector K1B(k) “belongs to the normal group” or “does not belong to the normal group”.

[0038] The second judgment is a process of inputting a second detection feature vector K2B(k) consisting of the detected body surface temperature feature KTT(k) and the detected humidity feature KST(k) for the period T(k) into a second classifier 28 to judge whether the second detection feature vector K2B(k) “belongs to the normal group” or “does not belong to the normal group”.

[0039] The third determination is a process of determining whether the second detection feature vector K2B(k) “belongs to the normal group” or “does not belong to the normal group” by inputting the second detection feature vector K2B(k) consisting of the detected body surface temperature feature KTT(k) and the detected air pressure feature KAT(k) for the period T(k) into a second classifier 28.

[0040] Here, the concept of the first determination will be explained so as to be visually understandable using the left and right diagrams in the margin of the table in FIG. 4. First, when the first detection feature vector K1B(k) is input to the first classification 26, it is assumed that the vertex "△" of the first detection feature vector K1B(k) is plotted at the position shown in the left diagram. In this case, "△" has the shortest Mahalanobis distance to the cluster center J1b of the normal cluster CL1b among the normal clusters CL1a to CL1c. And, since the Mahalanobis distance to the cluster center J1b is less than the first threshold Dr1b (since the position of "△" is inside the line of the first threshold Dr1b), the first determination for this first detection feature vector K1B(k) is "belongs to the normal group".

[0041] Next, when the first detected feature vector K1B(k) is input to the first classification 26, the vertex "△" of the first detected feature vector K1B(k) is plotted at the position shown in the right diagram. In this case, the "△" has the shortest Mahalanobis distance to the cluster center J1a of the normal cluster CL1a among the normal clusters CL1a to CL1c. Since the Mahalanobis distance to the cluster center J1a is equal to or greater than the first threshold Dr1a (since the position of the "△" is outside the line of the first threshold Dr1a), the first judgment for this first detected feature vector K1B(k) is "not belonging to the normal group."

[0042] In the first judgment in the table of FIG. 4 (specific example), for the first detected feature vector K1B(k) for each period, periods T(1), T(3), T(5), T(7) to T(9) are judged to “belong to the normal group,” and periods T(2), T(4), and T(6) are judged to “not belong to the normal group.”

[0043] The concept of the second judgment is the same as that of the first judgment. That is, the second detection feature vector K2B(k) is input to the second classifier 28, one normal cluster CL2 closest to the second detection feature vector K2B(k) is extracted, and when the Mahalanobis distance between the extracted normal cluster CL2 and the cluster centroid J2 is less than the second threshold Dr2, it is judged to be "belonging to the normal group", and when it is equal to or greater than the second threshold Dr2, it is judged to be "not belonging to the normal group". In the second judgment in the table (specific example) of FIG. 4, for the second detection feature vector K2B(k) of each period, the periods T(1) to T(3), T(6) to T(9) are judged to be "belonging to the normal group", and the periods T(4), T(5) are judged to be "not belonging to the normal group".

[0044] The concept of the third judgment is the same as that of the first judgment. That is, the third detection feature vector K3B(k) is input to the third classifier 30, one normal cluster CL3 closest to the third detection feature vector K3B(k) is extracted, and when the Mahalanobis distance between the extracted normal cluster CL3 and the cluster centroid J3 is less than the third threshold Dr3, it is judged to "belong to the normal group", and when it is equal to or greater than the third threshold Dr3, it is judged to "not belong to the normal group". In the third judgment in the table (specific example) of FIG. 4, the third detection feature vector K3B(k) of each period is judged to "belong to the normal group" except for the period T(6), and the period T(6) is judged to "not belong to the normal group".

[0045] <Abnormality degree calculation step S15> In the tentative judgment step S14, for each period T, if the first judgment judges that the subject "belongs to the normal group", zero is counted, and if the first judgment judges that the subject "does not belong to the normal group", +1 is counted. Similarly, if the second judgment judges that the subject "belongs to the normal group", zero is counted, and if the second judgment judges that the subject "does not belong to the normal group", +1 is counted. Similarly, if the third judgment judges that the subject "belongs to the normal group", zero is counted, and if the third judgment judges that the subject "does not belong to the normal group", +1 is counted. Then, the sum of the counts of the first to third judgments for each period T is calculated as the abnormality degree A for the period T. For example, in the case of period T(1) in the table (specific example) of FIG. 4, the counts for the first, second, and third judgments are "0,0,0", and the abnormality degree A is calculated as zero. Also, in the case of period T(4), the counts for the first, second, and third judgments are "1,1,0", and the abnormality degree A is calculated as 2.

[0046] <Main Judgment Step S16> In the main judgment step S16, for n consecutive periods T [constant n is a natural number equal to or greater than 2], a total value An of the degree of abnormality A for each period T is calculated, and if the total value An is equal to or greater than the reference value Ar, it is judged that "an abnormality in the subject's physical condition has been detected." In the table (specific example) of FIG. 4, n=5 and Ar=5. For example, the total value An for the five periods T(1) to T(5) is 4. Therefore, since the total value An is less than the reference value Ar, it is not judged that "an abnormality in the subject's physical condition has been detected." Moreover, the total value An for the next five periods T(2) to T(6) is 6. Therefore, since the total value An is equal to or greater than the reference value Ar, it is judged that "an abnormality in the subject's physical condition has been detected."

[0047] In this way, in the provisional judgment step S14, the first to third judgments are made sequentially for each period T, and in the abnormality calculation step S15, the total value An(k) of the degree of abnormality A for each of the n periods T(k) to T(k+n-1), the total value An(k+1), ... are calculated sequentially, and in the main judgment step S16, a judgment is made as to whether or not "an abnormality in the subject's physical condition has been detected" for each of the total values ​​An(k), An(k+1), ..., thereby making it possible to continuously monitor the physical condition of the subject M.

[0048] <Classifier update step S17> In the classifier update step S17, if it is not determined in the main judgment step S16 that "an abnormality in the physical condition of the subject M has been detected," a process of updating each of the classifiers 26, 28, and 30 is performed based on the measurement data of the subject M on which the judgment was based.

[0049] For example, in the table of FIG. 4 (specific example), the main judgment for the period T(1) to T(5) did not result in a judgment that "an abnormality in the subject's physical condition was detected." Then, when making the main judgment from this point onwards, the measurement data on which the period T(1) was based is not taken into consideration. Therefore, at the time when the main judgment for the period T(1) to T(5) is made, the measurement data KTD, KKD, KSD, and KAD of the body surface temperature, air temperature, humidity, and air pressure for the period T(1) can be regarded as "new normal measurement data," and machine learning incorporating these data (machine learning in the classifier preparation step S12) is performed to update each of the classifiers 26, 28, and 30.

[0050] On the other hand, the main judgment for the period T(2) to T(6) was that “an abnormality in the subject's physical condition was detected.” Therefore, the measurement data KTD, KKD, KSD, and KAD of the body surface temperature, air temperature, humidity, and air pressure for the periods T(2) to T(6) are measurement data when the subject M is abnormal, and therefore these are not included in the classifiers 26, 28, and 30.

[0051] By repeating this classifier update step S17 several times, it is possible to significantly improve the classification accuracy of the classifiers 26, 28, and 30. Note that the classifier update step S17 may be performed every time "new normal measurement data" is generated after the main judgment step S16 is performed, or may be performed all at once when a certain amount of "new normal measurement data" has been accumulated (for example, every 1 to 3 days).

[0052] <<Verification experiment of the abnormal physical condition detection method of this embodiment>> The inventor conducted a verification experiment to confirm the effectiveness of the above-mentioned method for detecting abnormal physical condition. First, the verification experiment method will be described. Subject M was a healthy man in his twenties, and the living space of subject M was indoors, without sunlight or wind. Subject M's body surface temperature was measured with an infrared thermography camera, and the temperature, humidity, and air pressure of the living space were measured with dedicated sensors, and each measurement data was input into a computer at a predetermined sampling period.

[0053] When creating the classifier in the classifier preparation step S11, the normal measurement data STD, SKT, SSD, SAD of the subject M was data measured when the subject M was at rest for one hour or more. In addition, clustering of the normal state feature vector was performed using the k-means method. The measurement data KTD, KKT, KSD, KAD of the subject M obtained in the subject data acquisition step S12 was of two types: data when the subject M alternated between running and resting [hypothetical abnormal state], and data when the subject M was at rest for one hour or more [hypothetical normal state].

[0054] The length of the period T related to each step was analyzed in two ways: uniformly set to 15 minutes, and uniformly set to 30 minutes. In addition, the feature quantities related to each step (normal feature quantities STT, SKT, SST, SAT, detection feature quantities KTT, KKT, KST, KAT) were analyzed in two ways: the value α obtained by subtracting the minimum value from the maximum value of the measurement data during the period T, and the difference β between the initial value and the final value of the measurement data during the period T, as shown in the margin of the table in Figure 5.

[0055] Next, the results of the verification experiment will be explained based on the table in Figure 5. First, the results when each feature amount is α will be explained. The anomaly detection rate during virtual anomalies was 100% when period T was both 15 minutes and 30 minutes, and good results were obtained. Note that in case 1 and case 2, the distribution of time between running and resting was changed to change the degree of anomaly, but since the anomaly detection rate in both cases was 100%, it can be said that anomalies were detected with high accuracy regardless of the degree of anomaly. In addition, the anomaly detection rate during virtual normality was 0% when period T was both 15 minutes and 30 minutes, and good results were obtained.

[0056] Next, the results when each feature amount was set to β will be explained. In case 2, the anomaly detection rate during virtual anomalies was 100% when period T was both 15 minutes and 30 minutes, and good results were obtained. However, in case 1, it was 90% when period T was 15 minutes and 80% when it was 30 minutes. This shows that when the feature amount of the measurement data is set to β, the accuracy of anomaly detection decreases slightly depending on the level of anomaly. Additionally, the anomaly detection rate during virtual normalities was 0% when period T was both 15 minutes and 30 minutes, and good results were obtained.

[0057] In this verification experiment, the accuracy of anomaly detection decreased slightly when the feature amount was β. Although the detailed cause is under analysis, it is considered that, for example, if the above-mentioned classifier update step S17 is repeated several times to improve the classification accuracy of the classifiers 26, 28, and 30, the anomaly detection rate of the hypothetical anomaly case 1 may become 100%.

[0058] <<Summary of this embodiment>> As described above, the abnormal physical condition detection method, abnormal physical condition detection program, and abnormal physical condition detection system 10 of this embodiment are configured to detect abnormal physical condition based on measurement data such as body surface temperature and air temperature that can be measured without contact, so that it is possible to gently watch over the subject M without making the subject M feel restrained or uncomfortable. In addition, classifiers 26, 28, and 30 created by machine learning that input feature amounts (air temperature, humidity, and air pressure) based on body surface temperature when the physical condition is normal and surrounding environmental information indicating environmental fluctuations are prepared, and a unique method is used to make judgments, so that abnormal physical condition of the subject M can be detected objectively and with high accuracy.

[0059] <<<Other embodiments and modifications>>> The abnormal physical condition detection method, abnormal physical condition detection program, and abnormal physical condition detection system of the present invention are not limited to the above embodiment. For example, the above embodiment uses three measurement data of temperature, humidity, and air pressure as environmental information, but if a desired abnormality detection accuracy is obtained, two measurement data of temperature and humidity may be used for analysis. Conversely, it is also possible to add or replace features such as radiant heat, which is environmental information that may affect body surface temperature, and illuminance and sound, which are environmental information that may affect physical condition, in addition to the environmental information of temperature, humidity, and air pressure, and perform analysis. In addition, the above embodiment is configured to perform the classifier update step S16, but if the classification accuracy of each classifier is sufficient, the classifier update step S16 may be omitted. [Explanation of symbols]

[0060] 10. Abnormal Health Detection System 12 Classifier preparation section 14 Subject data acquisition section 16 Detection feature extraction unit 18 Provisional Judgment Section 20 Abnormality calculation unit 22 Main judgment part 24 Classifier update unit 26 First Classifier 28 Second Classifier 30 Third Classifier M Subject S11 Classifier preparation step S12 Subject data acquisition step S13 Detection feature extraction step S14 Temporary judgment step S15: Calculation step of abnormality S16 Main decision step S17 Classifier update step T,T(k) period A. Degree of abnormality for each period An the sum of the abnormalities over n periods Ar: Criteria for main judgment STD Body surface temperature measurement data when the subject is normal SKD Subject's normal living space temperature measurement data SSD Humidity measurement data of the living space when the subject is normal Air pressure measurements of the living space of SAD subjects when they were normal STT Normal body surface temperature feature SKT Normal temperature feature SST Normal Humidity Feature SAT Normal pressure feature S1B First normal feature vector S2B Second normal feature vector S3B Third normal feature vector KTD Subject's body surface temperature measurement data KKD Temperature measurement data of the subject's living space Humidity measurement data for the living space of KSD subjects KAD Measurement data of the air pressure in the subject's living space KTT, KTT(k) Detected body surface temperature feature KKT,KTT(k) Detected temperature feature KST, KST(k) Detected humidity feature KAT, KAT(k) Detected air pressure feature K1B,K1B(k) First detection feature vector K2B,K2B(k) Second detection feature vector K3B,K3B(k) Third detection feature vector CL1, CL1a to CL1c Normal cluster (first classifier) CL2, CL2a to CL2c Normal cluster (second classifier) CL3, CL3a to CL3c Normal cluster (third classifier) J1, J1a to J1c Cluster centroids (first classifier) J2, J2a to J2c Cluster centroids (second classifier) J3, J3a to J3c Cluster centroids (third classifier) Dr1, Dr1a to Dr1c First threshold (first classifier) Dr2, Dr2a to Dr2c Second threshold (second classifier) Dr3, Dr3a to Dr3c Third threshold (third classifier)

Claims

1. A method for detecting an abnormality in a physical condition of a subject, the method being executed by a computer system, comprising: a classifier preparation step of collecting a plurality of combinations of information on normal-state body surface temperature features indicating characteristics of a time change in body surface temperature of a subject when normal during a certain period T, information on normal-state air temperature features indicating characteristics of a time change in air temperature of the living space of the subject during the period T, and information on normal-state humidity features indicating characteristics of a time change in humidity of the living space of the subject during the period T, and creating a first classifier by performing machine learning using as input a plurality of first normal-state feature vectors consisting of combinations of the normal-state body surface temperature features and the normal-state air temperature features for contemporaneous periods T, and creating a second classifier by performing machine learning using as input a plurality of second normal-state feature vectors consisting of combinations of the normal-state body surface temperature features and the normal-state humidity features for contemporaneous periods T, or acquiring the created first and second classifiers; a subject data acquisition step of acquiring measurement data of the subject's body surface temperature, the air temperature of the living space, and the humidity of the living space; a detected feature extraction step of analyzing the measurement data acquired in the subject data acquisition step for each successive period T of a certain length, and extracting a detected body surface temperature feature that indicates a feature of a time change in the body surface temperature of the subject, a detected air temperature feature that indicates a feature of a time change in the air temperature of the living space of the subject, and a detected humidity feature that indicates a feature of a time change in the humidity of the living space of the subject; a tentative determination step of making a first determination as to whether the first detected feature vector "belongs to the normal group" or "does not belong to the normal group" by inputting a first detected feature vector consisting of the detected body surface temperature feature and the detected air temperature feature for each time period T to the first classifier, and making a second determination as to whether the second detected feature vector "belongs to the normal group" or "does not belong to the normal group" by inputting a second detected feature vector consisting of the detected body surface temperature feature and the detected humidity feature for each time period T to the second classifier; an abnormality degree calculation step of counting, for each period T, zero when the first judgment judges that the test result "belongs to the normal group" and +1 when the first judgment judges that the test result "does not belong to the normal group", and counting zero when the second judgment judges that the test result "belongs to the normal group" and +1 when the second judgment judges that the test result "does not belong to the normal group", and determining the sum of the counts of the first and second judgments for each period T as the abnormality degree A for that period T; A method for detecting abnormal physical condition, comprising a main judgment step of calculating a sum An of the degree of abnormality A for each of n consecutive periods T (where the constant n is a natural number greater than or equal to 2), and judging that "an abnormal physical condition of the subject has been detected" if the sum An is greater than or equal to a reference value Ar.

2. In the abnormality degree calculation step, a sum An(k) of the abnormality degrees A of the n periods T(k) to T(k+n-1) and a sum An(k+1) of the abnormality degrees A of the n periods T(k+1) to T(k+n) are sequentially calculated; 2. The method for detecting an abnormal physical condition according to claim 1, wherein in the main judgment step, it is judged whether or not "an abnormal physical condition of the subject has been detected" for each of the sums An(k) and An(k+1).

3. the first classifier is a classifier in which information of one or more normal clusters generated by applying a predetermined clustering algorithm to a plurality of the first normal feature vectors is registered; the second classifier is a classifier in which information of one or more normal clusters generated by applying a predetermined clustering algorithm to the plurality of second normal feature vectors is registered; In the first determination of the provisional determination step, the determination of whether the cluster "belongs to the normal group" or "does not belong to the normal group" is performed by extracting a normal cluster having a closest Mahalanobis distance between a cluster centroid and the first detected feature vector from one or more of the normal clusters registered in the first classifier, and determining that the cluster "does not belong to the normal group" when the Mahalanobis distance between the cluster centroid of the extracted normal cluster is equal to or greater than a first threshold Dr1; 2. The method for detecting abnormal physical condition according to claim 1, wherein in the second judgment in the provisional judgment step, the judgment of whether the normal cluster "belongs to the normal group" or "does not belong to the normal group" is made by extracting the normal cluster having the shortest Mahalanobis distance between the cluster centroid and the second detection feature vector from the one or more normal clusters registered in the second classifier, and judging the normal cluster to "does not belong to the normal group" if the Mahalanobis distance of the extracted normal cluster to the cluster centroid is equal to or greater than a second threshold value Dr2.

4. The method for detecting an abnormal physical condition according to claim 3, wherein the first and second classifiers are registered with information on normal clusters generated by applying a k-means method.

5. In the classifier preparation step, a plurality of combinations of information on the normal state body surface temperature feature amount and information on normal state environment information feature amount indicating characteristics of time changes in environmental information other than the air temperature and the humidity in the living space of the subject during a period T that is the same as the normal state body surface temperature feature amount are collected, and a third classifier is created by performing machine learning using as input a plurality of third normal state feature amount vectors each consisting of combinations of the normal state body surface temperature feature amount and the normal state environment information feature amount during the same period T, or the created third classifier is acquired; In the subject data acquisition step, measurement data of environmental information of the living space of the subject is acquired, In the detected feature extraction step, the measurement data of the environmental information acquired in the subject data acquisition step is analyzed for each period T when the detected body surface temperature feature is extracted, and a process is performed to extract detected environmental information feature indicating a characteristic of a time change in the environmental information of the living space of the subject, In the tentative determination step, a third detection feature vector including the detected body surface temperature feature and the detected environment information feature for each period T is input to the third classifier to perform a third determination as to whether the third detection feature vector “belongs to a normal group” or “does not belong to a normal group”; 2. The method for detecting abnormal physical condition as described in claim 1, wherein in the abnormality calculation step, if the third judgment judges that the subject "belongs to the normal group", zero is counted, and if the third judgment judges that the subject "does not belong to the normal group", +1 is counted, and the total value of the counts of the first, second and third judgments for each period T is determined to be the abnormality A for that period T.

6. The method for detecting abnormal physical condition according to claim 5, wherein the environmental information of the living space of the subject is atmospheric pressure.

7. the third classifier is a classifier in which information of one or more normal clusters generated by applying a predetermined clustering algorithm to the plurality of third normal feature vectors is registered; The method for detecting abnormal physical condition according to claim 6, wherein in the third judgment in the provisional judgment step, the judgment of whether the normal cluster "belongs to the normal group" or "does not belong to the normal group" is made by extracting the normal cluster having the shortest Mahalanobis distance between its cluster centroid and the third detection feature vector from among the one or more normal clusters registered in the third classifier, and judging the normal cluster "does not belong to the normal group" if the Mahalanobis distance between the cluster centroid of the extracted normal cluster is equal to or greater than a third threshold Dr3.

8. The method for detecting abnormal physical condition according to claim 7 , wherein the third classifier registers information of the normal cluster generated by applying a k-means method.

9. The method for detecting abnormal physical condition as described in claim 1, further comprising a classifier update step for performing a process of updating each of the classifiers based on the measurement data of the subject on which the judgment is based if the main judgment step does not judge that "an abnormal physical condition of the subject has been detected."

10. 10. The method for detecting an abnormality in physical condition according to claim 1, wherein each of the characteristic amounts for each period T is a value obtained by subtracting a minimum value from a maximum value in the measurement data for the period T.

11. 10. An abnormal physical condition detection program comprising programs for executing each step, for causing a computer system to execute the abnormal physical condition detection method according to any one of claims 1 to 9.

12. A physical condition abnormality detection system comprising a computer for detecting an abnormality in the physical condition of a subject, a classifier preparation unit that collects a plurality of combinations of information on normal-state body surface temperature features indicating characteristics of a time change in body surface temperature of a subject when normal during a certain period T, information on normal-state air temperature features indicating characteristics of a time change in air temperature of the living space of the subject during the period T, and information on normal-state humidity features indicating characteristics of a time change in humidity of the living space of the subject during the period T, and creates a first classifier by performing machine learning using a plurality of first normal-state feature vectors made up of the normal-state body surface temperature features and the normal-state air temperature features during the same period T as an input, and creates a second classifier by performing machine learning using a plurality of second normal-state feature vectors made up of the normal-state body surface temperature features and the normal-state humidity features during the same period T as an input, or acquires the created first and second classifiers; a subject data acquisition unit that acquires measurement data of the subject's body surface temperature, measurement data of the air temperature of the living space, and measurement data of the humidity of the living space; a detected feature extraction unit that performs a process of analyzing the measurement data acquired in the subject data acquisition step for each successive period T of a certain length, and extracting a detected body surface temperature feature that indicates a feature of a time change in the body surface temperature of the subject, a process of extracting a detected air temperature feature that indicates a feature of a time change in the air temperature of the living space of the subject, and a process of extracting a detected humidity feature that indicates a feature of a time change in the humidity of the living space of the subject; a tentative determination unit that performs a first determination as to whether the first detected feature vector "belongs to the normal group" or "does not belong to the normal group" by inputting a first detected feature vector consisting of the detected body surface temperature feature and the detected air temperature feature for each time period T to the first classifier, and performs a second determination as to whether the second detected feature vector "belongs to the normal group" or "does not belong to the normal group" by inputting a second detected feature vector consisting of the detected body surface temperature feature and the detected humidity feature for each time period T to the second classifier; an abnormality degree calculation unit that, for each period T, counts zero when the first judgment determines that the target “belongs to the normal group” and counts +1 when the second judgment determines that the target “does not belong to the normal group”, and counts zero when the second judgment determines that the target “belongs to the normal group” and counts +1 when the second judgment determines that the target “does not belong to the normal group”, and determines the sum of the counts of the first and second judgments for each period T as the abnormality degree A for that period T; The abnormal physical condition detection system is characterized by having a main judgment unit that calculates a total value An of the degree of abnormality A for each of n consecutive periods T (where the constant n is a natural number greater than or equal to 2) and judges that "an abnormal physical condition of the subject has been detected" if the total value An is greater than or equal to a reference value Ar.

13. the classifier preparation unit collects a plurality of combinations of information on the normal state body surface temperature feature amount and information on normal state environment information feature amount indicating characteristics of time changes in environmental information other than the air temperature and the humidity in the living space of the subject during a period T that is the same as the normal state body surface temperature feature amount, and performs machine learning using as input a plurality of third normal state feature amount vectors consisting of the normal state body surface temperature feature amount and the normal state environment information feature amount during the same period T to create a third classifier, or acquires the created third classifier; the subject data acquisition unit acquires measurement data of environmental information of the living space of the subject; the detected feature amount extraction unit performs a process of analyzing the measurement data of the environmental information acquired in the subject data acquisition step for each period T when the detected body surface temperature feature amount is extracted, and extracting a detected environmental information feature amount indicating a characteristic of a time change in the environmental information of the living space of the subject; the tentative determination unit inputs a third detection feature vector, which is composed of the detected body surface temperature feature and the detected environment information feature for each period T, to the third classifier, thereby making a third determination as to whether the third detection feature vector “belongs to a normal group” or “does not belong to a normal group”; The abnormality detection system of claim 12, wherein the abnormality calculation unit counts zero if the third judgment determines that the subject "belongs to the normal group" and counts +1 if the third judgment determines that the subject "does not belong to the normal group," and the sum of the counts of the first, second and third judgments for each period T is the abnormality A for that period T.

14. The abnormal physical condition detection system of claim 12 or 13, further comprising a classifier update unit which, when the main judgment unit does not judge that "an abnormal physical condition of the subject has been detected", performs a process of updating each of the classifiers based on the measurement data of the subject that was the basis for the judgment.

Images