Method for generating trained model, device for generating training data, device for generating trained model, and device for predicting risk of disease onset

Abstract

To predict a disease development risk that a subject potentially has.SOLUTION: A teacher data generation device 60 generates teacher data TR for generating a learned model for inferring a disease development risk on the basis of biological information acquired from a subject. A reception unit 61 receives first data D1 including biological information acquired from the subject 31 and a result that a disease has developed, determined on the basis of the biological information, and receives second data D2 including biological information acquired from the subject 31 and a result that a disease has not developed, determined on the basis of the biological information. A processing unit 62 specifies a first time point that the biological information included in the first data D1 is acquired, specifies a second time point that the biological information included in the second data D2 is acquired, and if a time interval between the first time point and the second time point is smaller than a predetermined value, generates the teacher data TR by imparting, to the second data, a weighting index number that can take a plurality of values according to the time interval and a first teacher label indicating that a disease has developed.SELECTED DRAWING: Figure 3

Description

[Technical field]

[0001] The present invention relates to a method and an apparatus for generating a trained model for inferring a disease development risk of a subject based on biological information acquired from the subject. The present invention also relates to an apparatus for generating training data used to generate the trained model. The present invention also relates to an apparatus for predicting a disease development risk of a subject using the trained model. [Background technology]

[0002] Patent Document 1 discloses a Holter electrocardiograph that acquires an electrocardiogram, which is an example of a biological parameter of a subject. Data corresponding to the measured waveform of the electrocardiogram is received by a remotely placed information processing device. The information processing device is configured to analyze the data to detect abnormalities in the measured waveform and identify the onset of a disease such as heart failure in the subject. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2014-054391 A Summary of the Invention [Problem to be solved by the invention]

[0004] An object of the present invention is to predict a subject's potential risk of developing a disease. [Means for solving the problem]

[0005] A first aspect for achieving the above object is a method for generating a trained model for inferring a disease development risk of a subject based on biological information acquired from the subject, comprising: receiving first data including the biological information acquired from a first subject and a result indicating that a disease has developed based on the biological information; Identifying a first time point at which the biometric information included in the first data was acquired; receiving second data including the biological information acquired from the first subject and a result indicating that the first subject is not suffering from a disease determined based on the biological information; Identifying a second time point at which the biometric information included in the second data was acquired; When the time interval between the first time point and the second time point is less than a first predetermined value, a first teacher label indicating that the disease has developed is assigned to the second data together with a weighting index that can take a plurality of values ​​according to the time interval; The second data to which the first teacher label has been assigned is used as teacher data to perform machine learning of a model.

[0006] A second aspect for achieving the above object is a device for generating teacher data used to generate a trained model that infers a disease development risk of a subject based on biological information acquired from the subject, the device comprising: a receiving unit that receives first data including the biometric information acquired from a first subject and a result indicating that a disease has developed based on the biometric information, and receives second data including the biometric information acquired from the first subject and a result indicating that a disease has not developed based on the biometric information; a processing unit which identifies a first time point when the biometric information included in the first data was acquired, identifies a second time point when the biometric information included in the second data was acquired, and, when a time interval between the first time point and the second time point is less than a first predetermined value, generates teacher data by assigning a first teacher label indicating the onset of the disease to the second data together with a weighting index which may take multiple values ​​according to the time interval; It is equipped with:

[0007] Even if the biometric information is determined to be free of disease, if the time interval between the time the biometric information was acquired and the time the biometric information determined to be free of disease was acquired is not very long, there is a possibility that the biometric information determined to be free of disease contains some factor that is a predictor of disease or some factor that is influenced by disease. According to the above configuration, it is possible to generate training data in which a training label indicating disease is assigned to biometric information determined to be free of disease that satisfies such a time condition.

[0008] Therefore, a trained model generated through machine learning using such training data can output an inference result that there is a risk of developing a disease when inputted biological information that may be judged to be disease-free is present, thereby making it possible to predict a subject's potential risk of developing a disease.

[0009] In addition, a weight index that can take multiple values ​​according to the time interval is assigned to the first teacher label indicating the onset of a disease. This makes it possible to generate teacher data with different degrees of certainty of inferring the risk of developing a disease from biological information determined to be free of a disease. This allows for more precise machine learning, and ultimately more precise prediction of the risk of developing a disease.

[0010] A third aspect for achieving the above object is a method for generating a trained model for inferring a disease development risk of a subject based on biological information acquired from the subject, comprising: receiving first data including the biological information acquired from a first subject and a result indicating that a disease has developed based on the biological information; Identifying a first time point at which the biometric information included in the first data was acquired; receiving second data including the biological information acquired from the first subject and a result indicating that the first subject is not suffering from a disease determined based on the biological information; Identifying a second time point at which the biometric information included in the second data was acquired; When the time interval between the first time point and the second time point is less than a first predetermined value, a first teacher label indicating that the disease has developed is assigned to the second data; When the second time point is earlier than the reference time point by a second predetermined value or more, a second teacher label indicating that the disease has not developed is assigned to the second data; performing machine learning of a model using at least one of the second data to which the first teacher label is assigned and the second data to which the second teacher label is assigned as teacher data; How to generate a trained model.

[0011] A fourth aspect for achieving the above object is a device for generating teacher data used to generate a trained model that infers a disease development risk of a subject based on biological information acquired from the subject, the device comprising: a receiving unit that receives first data including the biometric information acquired from a first subject and a result indicating that a disease has developed based on the biometric information, and receives second data including the biometric information acquired from the first subject and a result indicating that a disease has not developed based on the biometric information; a processing unit that generates teacher data by identifying a first time point when the biometric information included in the first data was acquired, identifying a second time point when the biometric information included in the second data was acquired, and assigning a first teacher label to the second data indicating that the disease has developed if a time interval between the first time point and the second time point is less than a first predetermined value, and assigning a second teacher label to the second data indicating that the disease has not developed if the second time point is earlier than a reference time point by a second predetermined value or more; It is equipped with:

[0012] When biometric information determined to be free of disease is acquired sufficiently before the reference time point, the biometric information is highly likely to not contain factors involved in disease onset. According to the configurations of the third and fourth aspects, a second teacher label indicating that the disease is not present is proactively assigned to such biometric information and included in the teacher data. This makes it possible to generate both teacher data to which a teacher label indicating that the disease is present and teacher data determined to be free of disease based on biometric information acquired from the same subject and determined to be free of disease. Therefore, it is possible to improve the effectiveness of machine learning for classifying biometric information determined to be free of disease based on the presence or absence of a disease onset risk.

[0013] A fifth aspect for achieving the above object is a trained model generation device that infers a disease development risk of a subject based on biological information acquired from the subject, comprising: A reception unit that receives teacher data generated by the teacher data generation device according to the second or fourth aspect; A processing unit that performs machine learning of a model using the training data; It is equipped with:

[0014] A sixth aspect of the present invention for achieving the above object is a prediction device for predicting a disease onset risk, comprising: A reception unit that receives biological information acquired from a subject; A processing unit that inputs the biological information into a trained model generated by a trained model generation device according to the fifth aspect, and predicts the subject's risk of developing a disease based on an inference result output from the trained model; An output unit that outputs information corresponding to the disease development risk; It is equipped with: [Brief description of the drawings]

[0015] [Figure 1] 2 illustrates an example of a functional configuration of a prediction device according to an embodiment. [Diagram 2]2 illustrates an example of a functional configuration of a model generating device according to an embodiment. [Diagram 3] 1 illustrates an example of the functional configuration of a teacher data generation device according to an embodiment. [Figure 4] 4 illustrates an example of the configuration of the first data and the second data in FIG. 3. [Diagram 5] 4 shows an example of a flow of processing executed by the processing unit of FIG. 3. [Figure 6] 4 is a diagram for explaining a process executed by the processing unit in FIG. 3. [Figure 7] 6 shows an example of weighting indices used in the processing of FIG. 5. [Figure 8] 6 shows another example of weighting indices used in the process of FIG. 5. [Figure 9] 6 shows another example of weighting indices used in the process of FIG. 5. [Figure 10] 4 is a diagram for explaining a process executed by the processing unit in FIG. 3. [Figure 11] 4 shows another example of the flow of processing executed by the processing unit in FIG. 3. [Figure 12] 3 illustrates an example of a process flow executed by the model evaluation device of FIG. 2. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0016] Exemplary embodiments are described in detail below with reference to the accompanying drawings.

[0017] The expression "at least one of A and B" used in this specification with respect to two entities A and B includes the cases where only A is specified, only B is specified, and both A and B are specified. Each entity A and B may be singular or plural unless otherwise specified.

[0018] The expression "at least one of A, B, and C" used in this specification with respect to three entities A, B, and C means a case where only A is specified, a case where only B is specified, a case where only C is specified, a case where A and B are specified, a case where B and C are specified, a case where A and C are specified, and a case where all of A, B, and C are specified. Each of the entities A, B, and C may be singular or plural unless otherwise specified. The same applies to a case where the number of entities to be described is four or more.

[0019] 1 illustrates an example of a configuration of a prediction device 10 according to an embodiment. The prediction device 10 is configured to predict a risk of a subject 30 developing atrial fibrillation based on an electrocardiogram waveform acquired from the subject 30 via an electrocardiograph 20. The electrocardiogram waveform is an example of a biological signal waveform, and is an example of biological information. Atrial fibrillation is an example of a heart disease. The risk of developing atrial fibrillation is an example of a disease onset risk.

[0020] The prediction device 10 includes a receiving unit 11. The receiving unit 11 is configured as an interface that receives electrocardiogram data EC corresponding to an electrocardiogram waveform of the subject 30 acquired by the electrocardiograph 20. The electrocardiogram data EC corresponds to a change over time in a bioelectric potential acquired within a predetermined measurement time by electrodes attached to the body surface of the subject 30.

[0021] The prediction device 10 includes a processing unit 12. The processing unit 12 is configured to perform a process of acquiring the probability that the subject 30 will have atrial fibrillation by inputting the electrocardiogram data EC received by the receiving unit 11 to the trained model 13. The trained model 13 according to this example is an inference algorithm generated through machine learning using a neural network, which will be described later. The trained model 13 is configured to receive the electrocardiogram data EC corresponding to the electrocardiogram of the subject 30 as an input, and to output the probability that the subject 30 will have atrial fibrillation as an inference result. The inference result may be associated with a score (e.g., any of values ​​from 1 to 5) corresponding to the inferred probability.

[0022] The prediction device 10 includes an output unit 14. The processing unit 12 is configured to output prediction data PR corresponding to the probability inferred by the trained model 13 from the output unit 14. That is, the output unit 14 is configured as an interface capable of outputting the prediction data PR.

[0023] The prediction data PR is transmitted to the output device 40. The output device 40 is configured to notify a user of the risk of the subject 30 developing atrial fibrillation predicted by the prediction device 10 based on the prediction data PR. The notification is performed using at least one of a visual notification, an audio notification, and a tactile notification.

[0024] The trained model 13 is generated by a model generation device 50 illustrated in Fig. 2. That is, the model generation device 50 is configured to generate the trained model 13 to be installed in the prediction device 10 described above.

[0025] The model generating device 50 includes a receiving unit 51. The receiving unit 51 is configured as an interface that receives the teacher data TR. The teacher data TR includes an electrocardiogram waveform acquired from a subject, and a teacher label indicating whether the subject is suffering from atrial fibrillation or not. In other words, the teacher data TR teaches what kind of electrocardiogram waveform acquired from the subject can be used to determine whether the subject is suffering from atrial fibrillation (or not).

[0026] The model generating device 50 includes a processing unit 52. The processing unit 52 is configured to generate a trained model 13 by training a neural network using the teacher data TR. As the process of training the neural network, a well-known method related to supervised learning is appropriately used.

[0027] The model generating device 50 includes an output unit 53. The output unit 53 is configured as an interface that outputs the trained model 13 generated by the processing unit 52 in a form that can be implemented in the prediction device 10.

[0028] The teacher data TR is generated by a teacher data generating device 60 illustrated in Fig. 3. That is, the teacher data generating device 60 is configured to generate teacher data TR used in the machine learning executed by the model generating device 50.

[0029] The teacher data generating device 60 includes a receiving unit 61. The receiving unit 61 is configured as an interface capable of receiving the first data D1 and the second data D2. The first data D1 and the second data D2 are generated based on an electrocardiogram waveform acquired from the subject 31 by the electrocardiograph 21. The subject 31 is an example of a first subject. The electrocardiograph 21 may be the same device as the electrocardiograph 20 illustrated in FIG. 1, or may be a different device. The electrocardiogram waveform may be acquired through a medical information database or the like.

[0030] As illustrated in Fig. 4, each of the first data D1 and the second data D2 includes, as information, an acquired electrocardiogram waveform, the time when the electrocardiogram waveform was acquired, and a finding on whether atrial fibrillation is occurring that is assigned to the electrocardiogram waveform. More specifically, the first data D1 includes a determination result that atrial fibrillation is occurring. The second data D2 includes a determination result that atrial fibrillation is not occurring. The finding may be given through visual confirmation by a medical professional, or may be given by an algorithm for automatically determining atrial fibrillation that may be installed in the electrocardiograph 21 or the like.

[0031] 3, the teacher data generation device 60 includes a processing unit 62. The processing unit 62 is configured to identify a time point when an electrocardiogram waveform included in the first data D1 is acquired as a first time point, identify a time point when an electrocardiogram waveform included in the second data D2 is acquired as a second time point, and, when the time interval between the first time point and the second time point is less than a first predetermined value T1, generate teacher data TR by assigning a first teacher label indicating the occurrence of atrial fibrillation to the second data D2 together with a weighting index that can take multiple values ​​according to the time interval.

[0032] The teacher data generating device 60 includes an output unit 63. The processing unit 62 is configured to output the generated teacher data TR from the output unit 63. That is, the output unit 63 is configured as an interface capable of outputting the teacher data TR.

[0033] Fig. 5 shows an example of the flow of processing performed by the processing unit 62. A case in which the processing is performed on the first data D1 and second data D2 illustrated in Fig. 4 will be described with reference to Fig. 6.

[0034] The processing unit 62 acquires the first data D1 through the receiving unit 61 (STEP 11). That is, the processing unit 62 acquires data including an electrocardiogram waveform that has been determined to indicate the occurrence of atrial fibrillation.

[0035] Next, the processing unit 62 specifies the time point at which the electrocardiogram waveform included in the first data D1 was acquired as the first time point (STEP 12). In the example shown in Fig. 6, the time point t4 is specified as the first time point.

[0036] Next, the processing unit 62 acquires the second data D2 through the receiving unit 61 (STEP 13 in FIG. 5). That is, the processing unit 62 acquires data including an electrocardiogram waveform determined to be free of atrial fibrillation.

[0037] Next, the processing unit 62 specifies the time point at which the electrocardiogram waveform included in the second data D2 was acquired as the second time point (STEP 14).

[0038] Next, the processing unit 62 determines whether the time interval between the identified second time point and the first time point is less than a first predetermined value T1 (STEP 15). An example of the first predetermined value T1 is 30 days.

[0039] If the time interval is less than the first predetermined value T1 (YES in STEP 15), the processing unit 62 assigns a first teacher label indicating that atrial fibrillation is occurring to the second data D2 used to identify the second time point in STEP 14 (STEP 16).

[0040] At this time, the processing unit 62 assigns a weighting index to the second data D2 together with the first teacher label. The weighting index represents the degree to which the instruction of the first teacher label (i.e., the occurrence of atrial fibrillation) is taken into consideration when machine learning based on the teacher data TR is performed.

[0041] As illustrated in Fig. 7, the weighting index is set to take a plurality of values ​​according to the time interval from the reference time t0 at which the electrocardiogram waveform included in the first data D1 was acquired. In the example of Fig. 7, the weighting index takes a value of 1 at the reference time t0. As the time interval from the reference time t0 becomes longer, the weighting index value decreases in a normal distribution manner. When the time interval from the reference time t0 reaches a first predetermined value T1, the weighting index value takes a value of 0. The normal distribution decrease is an example of a nonlinear decrease.

[0042] As long as the time interval from the reference time t0 can take multiple values ​​within a range less than the first predetermined value T1, the manner in which the weighting index changes can be determined appropriately. For example, as shown in Fig. 8, the value of the weighting index may decrease linearly as the time interval from the reference time t0 increases. Alternatively, as shown in Fig. 9, the value of the weighting index may decrease stepwise as the time interval from the reference time t0 increases.

[0043] Next, the processing unit 62 judges whether the second time point has been specified for all the second data D2 (STEP 17 in FIG. 5). If it is judged that the second time point has not been specified for all the second data D2 (NO in STEP 17), the processing returns to STEP 14, and the second time point is specified for another second data D2.

[0044] If it is determined that the time interval between the identified second point in time and the first point in time is not less than the first predetermined value T1 (NO in STEP 15), the process proceeds to STEP 17.

[0045] In the example shown in Fig. 6, time points t1, t2, t3, t5, and t6 are identified as the second time points. Among them, the time interval between time points t3 and t4 is less than the first predetermined value T1, so the processing unit 62 assigns a first teacher label indicating the occurrence of atrial fibrillation to the second data D2 based on the electrocardiogram waveform acquired at time point t3. The same process is performed on the second data D2 based on the electrocardiogram waveform acquired at time point t5.

[0046] Time t3 is longer than time t5 from time t4 when the electrocardiogram waveform included in the first data D1 is acquired. Therefore, when the weighting indexes shown in Fig. 7 or 8 are used, the weighting index value assigned to the second data D2 based on the electrocardiogram waveform acquired at time t3 is smaller than the weighting index value assigned to the second data D2 based on the electrocardiogram waveform acquired at time t5. When the weighting indexes shown in Fig. 9 are used, the weighting index value is not always smaller even if the time interval from time t4 is longer.

[0047] In the example shown in FIG. 6, since the time intervals between each of time points t1, t2, and t6 and time point t4 are not less than the first predetermined value T1, the processing unit 62 does not assign a first teacher label to the second data D2 based on the electrocardiogram waveforms acquired at each time point.

[0048] Next, the processing unit 62 outputs the second data D2 to which the first teacher label has been assigned along with the weighting index from the output unit 63 as teacher data TR (STEP 18 in FIG. 5).

[0049] The teacher data TR generated as described above is used for machine learning of a neural network executed by a model generating device 50, as described with reference to Fig. 2. The trained model 13 generated through the machine learning is implemented in the prediction device 10 illustrated in Fig. 1.

[0050] Even if an electrocardiogram waveform is given a finding of sinus rhythm, if the time interval between the time when the electrocardiogram waveform was acquired and the time when an electrocardiogram waveform determined to be in atrial fibrillation was acquired is not very long, there is a possibility that some factor that is a sign of atrial fibrillation or some factor affected by atrial fibrillation is included in the sinus rhythm electrocardiogram waveform. With the above-mentioned configuration, it is possible to generate training data in which a training label indicating the occurrence of atrial fibrillation is assigned to an electrocardiogram waveform in sinus rhythm that satisfies such a time condition.

[0051] Therefore, a trained model generated through machine learning using such training data can output an inference result that there is a risk of atrial fibrillation when an electrocardiogram waveform that can be determined to be sinus rhythm is input, thereby making it possible to predict the subject's potential risk of developing atrial fibrillation.

[0052] In addition, a weighting index that can take multiple values ​​depending on the time interval is assigned to the first teacher label indicating the occurrence of atrial fibrillation. This makes it possible to generate teacher data with different degrees of certainty for inferring the risk of atrial fibrillation from an electrocardiogram waveform showing sinus rhythm. This allows for more precise machine learning, and ultimately makes it possible to more precisely predict the risk of atrial fibrillation.

[0053] As illustrated in FIG. 3, the receiving unit 61 of the teacher data generating device 60 may be configured as an interface capable of acquiring the third data D3 as well. The third data D3 is generated based on an electrocardiogram waveform acquired from the subject 32 by the electrocardiograph 21. The subject 32 is different from the subject 31. For example, the subject 31 may be a patient with a history of atrial fibrillation, whereas the subject 32 may be a patient with no history of atrial fibrillation. The subject 32 is an example of a second subject. The electrocardiograph for acquiring the electrocardiogram waveform of the subject 32 does not necessarily need to be the same device as the electrocardiograph 21.

[0054] The third data D3 has the same configuration as the second data D2 illustrated in Fig. 4. That is, the third data D3 includes an electrocardiogram waveform acquired by the electrocardiograph 21, the time point when the electrocardiogram waveform was acquired, and a determination made by a medical professional based on the electrocardiogram waveform that atrial fibrillation is not occurring.

[0055] In this case, the processing unit 62 of the teacher data generating device 60 is configured to identify the time when the electrocardiogram waveform included in the third data D3 was acquired as the third time point, and if the third time point is earlier than the reference time point by a second predetermined value T2 or more, generate teacher data TR by assigning a second teacher label to the third data D3 indicating that atrial fibrillation is not occurring.

[0056] The flow of the process executed by the processing unit 62 on the third data D3 illustrated in FIG. 10 will be described with reference to FIG.

[0057] When it is determined that the second time points have been identified for all of the second data D2 (YES in STEP 17 in FIG. 5), the processing unit 62 acquires the third data D3 through the receiving unit 61 (STEP 21).

[0058] Next, the processing unit 62 specifies the time point at which the electrocardiogram waveform included in the third data D3 was acquired as the third time point (STEP 22).

[0059] Next, the processing unit 62 determines whether the identified third time point is earlier than the reference time point t0 by a second predetermined value T2 or more (STEP 23). Examples of the reference time point t0 include the time point at which this process is executed and the time point at which the latest electrocardiogram waveform is acquired from the subject 32. An example of the second predetermined value T2 is one year.

[0060] If it is determined that the identified third point in time is earlier than the reference point in time t0 by a second predetermined value T2 or more (YES in STEP 23), the processing unit 62 assigns a second teaching label indicating that atrial fibrillation is not occurring to the third data D3 obtained in STEP 21 (STEP 24).

[0061] Next, the processing unit 62 determines whether the third time point has been specified for all the third data D3 (STEP 25). If it is determined that the third time point has not been specified for all the third data D3 (NO in STEP 25), the process returns to STEP 22, and the third time point is specified for another third data D3.

[0062] If it is determined that the identified third time point is not earlier than the reference time point t0 by the second predetermined value T2 or more (NO in STEP 23), the process proceeds to STEP 25.

[0063] 10, time points t7, t8, t9, and t10 are identified as the third time points. Among these, time points t7 and t8 are earlier than the reference time point t0 by the second predetermined value T2 or more, so the processing unit 62 assigns a second teacher label indicating that atrial fibrillation is not occurring to the third data D3 based on the electrocardiogram waveforms acquired at each of time points t7 and t8.

[0064] Then, the processing unit 62 outputs the teacher data TR including the second data D2 to which the first teacher label has been assigned along with the weighting index and the third data D3 to which the second teacher label has been assigned from the output unit 63 (STEP 18 in FIG. 5).

[0065] An electrocardiogram waveform showing sinus rhythm acquired sufficiently before the reference time t0 is highly likely to not include factors related to atrial fibrillation. In other words, it is preferable that the second predetermined value T2 is set to a length that can ignore the influence of atrial fibrillation even if the subject has atrial fibrillation. According to the above process, the second teacher label indicating that atrial fibrillation is not occurring is proactively assigned to such an electrocardiogram waveform and included in the teacher data TR, thereby improving the effect of machine learning for classifying an electrocardiogram waveform showing sinus rhythm as to whether or not there is a risk of atrial fibrillation occurring.

[0066] Fig. 11 shows another example of a process flow that can be executed by the processing unit 62 of the teacher data generation device 60. Processing elements that are substantially the same as those described with reference to Fig. 5 are given the same reference symbols, and repeated description will be omitted.

[0067] If it is determined that the time interval between the second time point identified in STEP 14 and the first time point identified in STEP 12 is less than the first predetermined value T1 (YES in STEP 15), the processing unit 62 assigns a first teacher label indicating the occurrence of atrial fibrillation to the second data D2 used to identify the second time point in STEP 14 (STEP 36). At this time, the assignment of a weighting index is not essential.

[0068] Next, the processing unit 62 judges whether the second time point has been specified for all the second data D2 (STEP 37). If it is judged that the second time point has not been specified for all the second data D2 (NO in STEP 37), the processing returns to STEP 14, and the second time point is specified for another second data D2.

[0069] If it is determined that the time interval is not less than the first predetermined value T1 (NO in STEP 15), the processing unit 62 determines whether the specified second time point is earlier than the reference time point t0 by at least the second predetermined value T2 (STEP 43). Examples of the reference time point t0 include the time point at which this process is executed, the time point at which the latest electrocardiogram waveform is acquired from the subject 31, and the time point at which an electrocardiogram waveform determined to be causing atrial fibrillation is acquired from the subject 31. The second predetermined value T2 may be the same as or different from the first predetermined value T1.

[0070] If it is determined that the identified second point in time is earlier than the reference point in time t0 by a second predetermined value T2 or more (YES in STEP 43), the processing unit 62 assigns a second teaching label indicating that atrial fibrillation is not occurring to the second data D2 obtained in STEP 13 (STEP 44).

[0071] Thereafter, the processing unit 62 judges whether the second time point has been specified for all the second data D2 (STEP 37). If it is judged that the second time point has not been specified for all the second data D2 (NO in STEP 37), the processing returns to STEP 14, and the second time point is specified for another second data D2.

[0072] If it is determined that the identified second point in time is not earlier than the reference point in time t0 by the second predetermined value T2 or more (NO in STEP 43), the process also returns to STEP 14, and the second point in time is identified for another second data D2.

[0073] In the example shown in Figure 6, since time points t1 and t2 are earlier than the reference time point t0 by more than the second predetermined value T2, the processing unit 62 assigns a second teacher label indicating that atrial fibrillation is not occurring to the second data D2 based on the electrocardiogram waveforms obtained at each of time points t1 and t2.

[0074] Thereafter, the processing unit 62 outputs the teacher data TR including the second data D2 to which the first teacher label has been assigned and the second data D2 to which the second teacher label has been assigned from the output unit 63 (STEP 38 in FIG. 5).

[0075] According to the process of this example, it is possible to generate both teacher data with a teacher label indicating the occurrence of atrial fibrillation and teacher data determined not to be occurring, based on electrocardiogram waveforms showing sinus rhythm obtained from the same subject. This improves the effectiveness of machine learning for classifying electrocardiogram waveforms showing sinus rhythm as to the presence or absence of a risk of atrial fibrillation.

[0076] The teacher data TR generated by the teacher data generation device 60 described above can be used not only during initial learning to generate the trained model 13, but also during re-learning of the trained model 13. This is performed in the same manner as during initial learning by the model generation device 50. Re-learning of the trained model 13 is performed at an appropriate timing. Examples of appropriate timing include when a predetermined period has passed since the previous learning, when a predetermined number of teacher data TR have been accumulated, etc.

[0077] Each of the processing unit 62 of the teacher data generating device 60, the processing unit 52 of the model generating device 50, and the processing unit 12 of the prediction device 10 having the above-mentioned functions can be realized by a general-purpose microprocessor operating in cooperation with a general-purpose memory. Examples of the general-purpose microprocessor can be a CPU, an MPU, or a GPU. Examples of the general-purpose memory can be a ROM or a RAM. In this case, the ROM can store a computer program that executes the above-mentioned processing. The ROM is an example of a non-transient computer-readable medium that stores a computer program. The general-purpose microprocessor specifies at least a part of the program stored on the ROM, deploys it on the RAM, and executes the above-mentioned processing in cooperation with the RAM. The computer program may be pre-installed in the general-purpose memory, or may be downloaded from an external server via a communication network and then installed in the general-purpose memory. In this case, the external server is an example of a non-transient computer-readable medium that stores a computer program.

[0078] The processing unit 62 of the teacher data generating device 60, the processing unit 52 of the model generating device 50, and the processing unit 12 of the prediction device 10 having the above-mentioned functions may each be realized by a dedicated integrated circuit capable of executing the above-mentioned computer program, such as a microcontroller, an ASIC, or an FPGA. In this case, the above-mentioned computer program is pre-installed in a storage element included in the dedicated integrated circuit. The storage element is an example of a computer-readable medium storing a computer program. The processing unit 62 of the teacher data generating device 60, the processing unit 52 of the model generating device 50, and the processing unit 12 of the prediction device 10 having the above-mentioned functions may also be realized by a combination of a general-purpose microprocessor and a dedicated integrated circuit.

[0079] 2, the trained model 13 generated by the model generation device 50 can be subjected to evaluation by the model evaluation device 70 using the evaluation data EV. FIG 12 illustrates a flow of evaluation processing executed by the model evaluation device 70.

[0080] First, a trained model 13 is prepared (STEP 51). This trained model 13 may be one that has undergone initial training, or one that has undergone re-training.

[0081] Next, the model evaluation device 70 inputs the evaluation data EV to the trained model 13 (STEP 52). The evaluation data EV includes, as information, an electrocardiogram waveform and a finding on whether atrial fibrillation has occurred that is assigned to the electrocardiogram waveform. The evaluation data EV is selected so as to be different from any of the first data D1, the second data D2, and the third data D3 described above. The electrocardiogram waveform included in the evaluation data EV may be actually obtained from the subject, or may be obtained through a medical information database or the like.

[0082] Next, the model evaluation device 70 judges whether the inference performance of the trained model 13 has improved based on the output of the trained model 13 for the input evaluation data EV (STEP 53). Specifically, it is judged whether the accuracy rate of the inference result of whether atrial fibrillation has occurred output from the trained model 13 for the input of the same evaluation data EV, where it is known whether atrial fibrillation has occurred, has improved.

[0083] If it is determined that the inference performance of the trained model 13 has improved (YES in STEP 53), the latest machine learning performed on the trained model 13 is deemed valid (STEP 54). If the trained model prepared in STEP 51 has undergone initial training, the inference performance has in principle improved, and therefore the initial training is confirmed. If the trained model prepared in STEP 51 has undergone re-training, it is determined whether the inference performance has improved based on the output in response to re-input of the same evaluation data EV.

[0084] If it is determined that the inference performance of the trained model 13 has not improved (NO in STEP 53), the latest machine learning performed on the trained model 13 is invalidated (STEP 55).

[0085] With this configuration, it is possible to prevent a situation in which the risk prediction performance of the prediction device 10 in which the trained model 13 is implemented is degraded due to overlearning or the like.

[0086] The process executed by the model evaluation device 70 may be realized by a general-purpose microprocessor operating in cooperation with a general-purpose memory. Examples of the general-purpose microprocessor include a CPU, an MPU, and a GPU. Examples of the general-purpose memory include a ROM and a RAM. In this case, a computer program for executing the above-mentioned process may be stored in the ROM. The ROM is an example of a non-transient computer-readable medium storing the computer program. The general-purpose microprocessor specifies at least a part of the program stored on the ROM, deploys it on the RAM, and executes the above-mentioned process in cooperation with the RAM. The computer program may be pre-installed in the general-purpose memory, or may be downloaded from an external server via a communication network and then installed in the general-purpose memory. In this case, the external server is an example of a non-transient computer-readable medium storing the computer program.

[0087] The processing performed by the model evaluation device 70 may be realized by a dedicated integrated circuit capable of executing the above-mentioned computer program, such as a microcontroller, an ASIC, or an FPGA. In this case, the above-mentioned computer program is pre-installed in a storage element included in the dedicated integrated circuit. The storage element is an example of a computer-readable medium storing a computer program. The processing performed by the model evaluation device 70 may also be realized by a combination of a general-purpose microprocessor and a dedicated integrated circuit.

[0088] The above-described embodiment is merely an example for facilitating understanding of the present invention. The configurations according to the above-described embodiment may be appropriately modified or improved without departing from the spirit and scope of the present invention.

[0089] In the above embodiment, the determination as to whether the time interval between the first time point and the second time point is less than the first predetermined value T1 (STEP 15 in FIG. 5) is performed for both before and after the first time point, as illustrated in FIG. 6. However, it may be determined whether the time interval between the second time point and either one of before or after the first time point is less than the first predetermined value T1.

[0090] 5, it is not necessary that both the second data D2 to which the first teacher label is assigned together with the weighting index and the third data D3 to which the second teacher label is assigned are always output as teacher data TR. The processing unit 62 of the teacher data generation device 60 can output only one of the second data D2 and the third data D3 as teacher data TR from the output unit 63 as necessary.

[0091] 11, it is not necessary that both the second data D2 to which the first teacher label is assigned and the second data D2 to which the second teacher label is assigned are always output as teacher data TR. The processing unit 62 of the teacher data generation device 60 can output only one of the second data D2 to which the first teacher label is assigned and the second data D2 to which the second teacher label is assigned from the output unit 63 as teacher data TR, as necessary.

[0092] In the above embodiment, the model generation device 50 generates the trained model 13 through machine learning using a neural network. However, the trained model 13 may be generated through other machine learning algorithms. Examples of other machine learning algorithms include decision trees, random forests, and support vector machines.

[0093] In the above embodiment, the prediction device 10 is provided as a device independent of the electrocardiograph 20. The prediction device 10 may be disposed in a location where it can be connected to the electrocardiograph 20 via wired communication or short-distance wireless communication, or may be disposed in a location where it can communicate via a local area network or a wide area network. The communication via the local area network or the wide area network may be wired or wireless.

[0094] However, the prediction device 10 may be built into the electrocardiograph 20. In this case, the output device 40 illustrated in FIG. 1 may be a part of the electrocardiograph 20 or may be a device independent of the electrocardiograph 20. Alternatively, the prediction device 10 may be built into another measuring device capable of acquiring an electrocardiogram. Examples of such devices include a vital sign monitor and a defibrillator.

[0095] When the teacher data generation device 60 and the model generation device 50 are provided as devices independent of each other, the output unit 63 of the teacher data generation device 60 and the reception unit 51 of the model generation device 50 may be connected to allow wired or wireless communication. That is, the output unit 63 and the reception unit 51 may be a physical communication interface.

[0096] The teacher data generating device 60 and the model generating device 50 may be functional entities realized within the same device. In this case, at least a part of the function of the processing unit 62 of the teacher data generating device 60 may be realized by the processing unit 52 of the model generating device 50. Furthermore, the output unit 63 and the receiving unit 51 may be a logical interface.

[0097] In the above embodiment, the prediction device 10 predicts the risk of the subject 30 developing atrial fibrillation. However, the risk of developing other cardiac diseases may also be predicted. Examples of other cardiac diseases include atrial premature contractions, paroxysmal supraventricular tachycardia, ventricular premature contractions, ventricular tachycardia, ventricular fibrillation, and myocardial infarction. Depending on the cardiac disease to be predicted, an appropriate electrocardiogram waveform that characterizes the cardiac disease is used to generate the training data TR.

[0098] The risk of developing a disease other than heart disease may be predicted. Examples of other diseases include epileptic seizures, hypotension, hypertension, and apnea syndrome. When the risk of developing an epileptic seizure is predicted, the teacher data TR may be generated based on an electroencephalogram waveform acquired from the subject. When the risk of developing hypotension or hypertension is predicted, the teacher data TR may be generated based on a blood pressure waveform acquired from the subject. The blood pressure waveform corresponds to a time-dependent change in the subject's blood pressure measured invasively. When the risk of developing apnea syndrome is predicted, the teacher data TR may be generated based on a respiratory waveform acquired from the subject. The respiratory waveform corresponds to a time-dependent change in the carbon dioxide concentration, etc., associated with the subject's breathing. Each of the electroencephalogram waveform, the blood pressure waveform, and the respiratory waveform is an example of biological information. In either case, an appropriate measuring device is used instead of an electrocardiograph.

[0099] The biological information used to generate the teacher data TR does not necessarily have to be a biological signal waveform. The teacher data TR may be generated based on biological parameter values ​​acquired from a subject. Examples of the biological parameter values ​​include a systolic peak time and a diastolic time in blood pressure, a cardiac output, an exhalation time, an inhalation time, and a respiration time in respiration. [Explanation of symbols]

[0100] 10: prediction device, 11: reception unit, 12: processing unit, 13: trained model, 30, 31, 32: subject, 50: model generation device, 51: reception unit, 52: processing unit, 60: teacher data generation device, 61: reception unit, 62: processing unit, D1: first data, D2: second data, D3: third data, T1: first predetermined value, T2: second predetermined value, TR: teacher data

Claims

1. A method for generating a trained model for inferring a disease development risk of a subject based on biological information acquired from the subject, comprising: receiving first data including the biological information acquired from a first subject and a result indicating that a disease has developed based on the biological information; Identifying a first time point at which the biometric information included in the first data was acquired; receiving second data including the biological information acquired from the first subject and a result indicating that the first subject is not suffering from a disease determined based on the biological information; Identifying a second time point at which the biometric information included in the second data was acquired; When the time interval between the first time point and the second time point is less than a first predetermined value, a first teacher label indicating that the disease has developed is assigned to the second data together with a weighting index that can take a plurality of values ​​according to the time interval; performing machine learning of a model using the second data to which the first teacher label has been assigned as teacher data; How to generate a trained model.

2. the value of the weighting index decreases linearly, non-linearly, or in a stepwise manner as the time interval from the first time point increases; A method for generating a trained model according to claim 1.

3. receiving third data including the biological information obtained from a second subject different from the first subject and a result indicating that the second subject is not suffering from a disease determined based on the biological information; Identifying a third time point at which the biometric information included in the third data was acquired; When the third time point is earlier than a reference time point established for the second subject by a second predetermined value or more, a second teacher label indicating that the disease has not developed is assigned to the third data; performing machine learning of the model using the third data to which the second teacher label is assigned as teacher data; A method for generating a trained model according to claim 1 or 2.

4. inputting evaluation data including the biological information and a result of whether or not a disease has developed based on the biological information into the model; Determine whether the inference performance of the model is improved compared to before the machine learning was performed; If it is determined that the inference performance has improved, enabling the machine learning. A method for generating a trained model according to any one of claims 1 to 3.

5. The biological information is a biological signal waveform. A method for generating a trained model according to any one of claims 1 to 4.

6. The disease is a heart disease. A method for generating a trained model according to any one of claims 1 to 5.

7. A device for generating teacher data used to generate a trained model that infers a disease development risk of a subject based on biological information acquired from the subject, comprising: a receiving unit that receives first data including the biometric information acquired from a first subject and a result indicating that a disease has developed based on the biometric information, and receives second data including the biometric information acquired from the first subject and a result indicating that the disease has not developed based on the biometric information; a processing unit which identifies a first time point when the biometric information included in the first data was acquired, identifies a second time point when the biometric information included in the second data was acquired, and, when a time interval between the first time point and the second time point is less than a first predetermined value, generates teacher data by assigning a first teacher label indicating the onset of the disease to the second data together with a weighting index which may take multiple values ​​according to the time interval; Equipped with A device for generating teacher data.

8. A trained model generating device that infers a disease development risk of a subject based on biological information acquired from the subject, comprising: A reception unit that receives the teacher data generated by the teacher data generation device according to claim 7; A processing unit that performs machine learning of a model using the training data; Equipped with A device for generating trained models.

9. a reception unit that receives biological information acquired from a subject; A processing unit that inputs the biological information into a trained model generated by the trained model generation device according to claim 8 and predicts the subject's risk of developing a disease based on an inference result output from the trained model; An output unit that outputs information corresponding to the disease development risk; Equipped with A device for predicting the risk of developing disease.

Images