Information processing device and method, computer program, and recording medium

Abstract

To improve the generalization performance of a learning model.SOLUTION: An information processing device (1) performs machine learning by using learning data generated from biological sound data including characteristics about one symptom. The information processing device includes acquisition means (12) for acquiring a first characteristic of the learning data by determining whether a frequency component as the characteristics about the one symptom is included in the learning data for every n (n is a natural number of 3 or more) frequency bands defined by dividing a prescribed frequency band, and acquiring a second characteristic of the learning data by determining whether the frequency component as the characteristics about the one symptom is included in the learning data for every m (m is a natural number of 2 or more and is smaller than n) defined by dividing the prescribed frequency band, and generation means (13) for generating a learning model related to the one symptom on the basis of the first characteristic and the second characteristic.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention relates to the technical field of an information processing device and method, a computer program, and a recording medium for performing, for example, identification of input data. [Background technology]

[0002] As this type of device, for example, a device has been proposed that analyzes high-frequency signals transmitted from a transmitting device using a machine learning technique to identify the transmitting device (see Patent Document 1). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2016-158209 A Summary of the Invention [Problem to be solved by the invention]

[0004] Analysis of input data (corresponding to the above-mentioned high-frequency signal) using machine learning techniques is often performed by comparing the features of the input data with a learning model. The accuracy of analysis using machine learning techniques depends on the quality of the learning model. For example, to build a learning model for analyzing body sound data, learning data based on body sound data is used.

[0005] However, since the physical constitution of living bodies is different from one another, even if a body sound is classified into one type, its characteristics may vary widely. On the other hand, there is a limit to the body sound data available at the stage of constructing a learning model. Therefore, if a learning model is constructed using learning data based on limited body sound data, there is a technical problem that the generalization performance of the learning model may be relatively low.

[0006] The present invention has been made in consideration of the above problems, and an object of the present invention is to provide an information processing device and method, a computer program, and a recording medium that can improve the generalization performance of a learning model. [Means for solving the problem]

[0007] In order to solve the above problem, an information processing device according to the present invention is an information processing device that performs machine learning using learning data generated from body sound data including features related to one symptom, and includes an acquisition means that acquires a first characteristic of the learning data by determining whether or not the learning data includes a frequency component as a feature related to the one symptom for each of n frequency bands (where n is a natural number equal to or greater than 3) defined by dividing a predetermined frequency band, and acquires a second characteristic of the learning data by determining whether or not the learning data includes a frequency component as a feature related to the one symptom for each of m frequency bands (where m is a natural number equal to or greater than 2 and smaller than n) defined by dividing the predetermined frequency band, and a generation means that generates a learning model related to the one symptom based on the first characteristic and the second characteristic.

[0008] In order to solve the above-mentioned problems, an information processing method according to the present invention is an information processing method for performing machine learning using learning data generated from body sound data including features related to one symptom, and includes a first acquisition step of acquiring a first characteristic of the learning data by determining whether or not the learning data includes a frequency component as a feature related to the one symptom for each of n frequency bands (where n is a natural number equal to or greater than 3) defined by dividing a predetermined frequency band; a second acquisition step of acquiring a second characteristic of the learning data by determining whether or not the learning data includes a frequency component as a feature related to the one symptom for each of m frequency bands (where m is a natural number equal to or greater than 2 and is smaller than n) defined by dividing the predetermined frequency band; and a generation step of generating a learning model related to the one symptom based on the first characteristic and the second characteristic.

[0009] In order to solve the above problem, the computer program of the present invention causes a computer of an information processing device that performs machine learning using learning data generated from body sound data including features related to one symptom to function as an acquisition means that acquires a first characteristic of the learning data by determining whether or not the learning data includes a frequency component as a feature related to the one symptom for each of n frequency bands (where n is a natural number equal to or greater than 3) defined by dividing a predetermined frequency band, and acquires a second characteristic of the learning data by determining whether or not the learning data includes a frequency component as a feature related to the one symptom for each of m frequency bands (where m is a natural number equal to or greater than 2 and smaller than n) defined by dividing the predetermined frequency band, and a generation means that generates a learning model related to the one symptom based on the first characteristic and the second characteristic.

[0010] In order to solve the above problems, a recording medium according to the present invention has the above-mentioned computer program according to the present invention recorded thereon.

[0011] The functions and other advantages of the present invention will become apparent from the following detailed description of the preferred embodiments. [Brief description of the drawings]

[0012] [Figure 1] 1 is a block diagram showing a configuration of an information processing device according to an embodiment; [Diagram 2] FIG. 1 is a conceptual diagram illustrating a concept of machine learning according to an embodiment. [Diagram 3] FIG. 11 is a diagram showing a specific example of classification according to the embodiment. [Figure 4] FIG. 2 is a conceptual diagram showing the concept of a learning model according to an embodiment. [Diagram 5] 4 is a flowchart illustrating an operation of the information processing device according to the embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0013] <Information processing device> An embodiment of an information processing device will be described. The information processing device according to the embodiment is an information processing device that performs machine learning using learning data generated from body sound data including features related to one symptom. The body sound may be, for example, a respiratory sound, a heart sound, a shunt sound, or the like. The symptom refers to a state that appears when a function of a living body falls outside a normal range due to a disease or the like and becomes abnormal. For example, a symptom of a whistling sound as an abnormal respiratory sound appears due to bronchial asthma.

[0014] It is not necessary that all of the learning data used in the machine learning is generated from body sound data including features related to one symptom. In other words, a part of the learning data may be generated from body sound data including features related to a normal body, for example.

[0015] The information processing device includes an acquiring means and a generating means. The acquiring means acquires a first characteristic of the learning data by determining whether or not the learning data contains a frequency component as a feature related to one symptom for each of n frequency bands (where n is a natural number equal to or greater than 3) obtained by dividing a predetermined frequency band.

[0016] The obtaining means also obtains a second characteristic of the learning data by determining, for each of m frequency bands (where m is a natural number equal to or greater than 2 and smaller than the n) obtained by dividing the predetermined frequency band, whether or not the learning data includes a frequency component that is a feature related to one symptom.

[0017] That is, the acquiring means classifies the same learning data into n frequency bands to acquire a first characteristic thereof, and also classifies the same learning data into m frequency bands to acquire a second characteristic thereof. The generating means generates a learning model related to one symptom based on the first characteristic and the second characteristic acquired by the acquiring means.

[0018] For example, when generating a learning model for identifying objects that have many common features but are of different types, such as flower varieties, a learning model with relatively high discrimination ability is often generated by classifying the features contained in the training data in relatively fine detail.

[0019] In contrast, in the case of body sounds, even if the disease is the same, the severity of the disease, the part of the body where the disease occurs, the physical characteristics of the body (e.g., height, weight, obesity level, muscle mass, etc.) may vary, and the degree of the symptoms (here, body sounds) that appear may change. For example, the characteristics of the above-mentioned whistle sound may change depending on the severity of bronchial asthma, etc. For this reason, when it comes to body sounds, if the characteristics contained in the learning data are classified relatively finely, the discrimination ability of the learning model may decrease.

[0020] On the other hand, if it were possible to collect a sufficient amount of body sound data corresponding to one symptom without bias, it would be possible to generate a learning model with relatively high discrimination ability even if the features included in the learning data are classified relatively finely. However, it is not realistic to collect a sufficient amount of body sound data corresponding to one symptom without bias.

[0021] In this information processing device, the learning data is classified relatively finely (i.e., into n frequency bands) to obtain a first characteristic, and the same learning data is classified relatively roughly (i.e., into m frequency bands) to obtain a second characteristic. Here, the first characteristic is expected to indicate relatively fine characteristics of one symptom, and the second characteristic is expected to indicate rough characteristics of one symptom.

[0022] In the information processing device, a learning model is generated based on the first characteristic and the second characteristic. The inventor's research has revealed that the learning model generated in this way has good discrimination ability for training data and also has relatively high discrimination ability for unknown data. Therefore, the information processing device can improve the generalization performance of the learning model.

[0023] In one aspect of the information processing device, the generating means may generate a learning model by specifying a weighting factor for each of the n frequency bands based on the first characteristic and specifying a weighting factor for each of the m frequency bands based on the second characteristic. Here, the weighting factor may be larger for a frequency band that includes more frequency components as features related to a symptom.

[0024] In another aspect of the information processing device, the body sound data may include data acquired from a subject exhibiting a certain symptom and data artificially created to imitate the certain symptom. With this configuration, the number of learning data can be increased relatively easily, which is very advantageous in practice.

[0025] <Information processing method> An embodiment of an information processing method will be described. The information processing method according to the embodiment is an information processing method for performing machine learning using learning data generated from body sound data including features related to one symptom.

[0026] The information processing method includes a first acquisition step of acquiring a first characteristic of the learning data by determining whether or not the learning data contains a frequency component that is a feature related to one symptom, for each of n frequency bands (where n is a natural number equal to or greater than 3) obtained by dividing a predetermined frequency band.

[0027] The information processing method includes a second acquisition step of determining, for each of m frequency bands (where m is a natural number equal to or greater than 2 and smaller than the n) into which the predetermined frequency band is divided, whether or not the learning data contains a frequency component as a feature related to one symptom, thereby acquiring a second characteristic of the learning data.

[0028] The information processing method includes a generating step of generating a learning model related to one symptom based on the first characteristic and the second characteristic. According to the information processing method, it is possible to improve the generalization performance of the learning model, similarly to the information processing device according to the above-described embodiment.

[0029] <Computer Program> An embodiment of a computer program will be described. The computer program according to the embodiment causes a computer of an information processing device that performs machine learning using learning data generated from body sound data including features related to one symptom to function as an acquisition means for determining whether or not the learning data includes a frequency component as a feature related to one symptom for each of n (n is a natural number equal to or greater than 3) frequency bands defined by dividing a predetermined frequency band, thereby acquiring a first characteristic of the learning data, and for determining whether or not the learning data includes a frequency component as a feature related to one symptom for each of m (m is a natural number equal to or greater than 2 and smaller than n) frequency bands defined by dividing the predetermined frequency band, thereby acquiring a second characteristic of the learning data, and a generation means for generating a learning model related to one symptom based on the first characteristic and the second characteristic.

[0030] According to the computer program, the computer program is read from a recording medium such as a RAM (Random Access Memory), a CD-ROM (Compact Disc Read Only Memory), or a DVD-ROM (DVD Read Only Memory) that stores the computer program, into a computer constituting an information processing device, and executed, or the computer program is downloaded via a communication means and then executed, whereby the information processing device according to the above-mentioned embodiment can be realized relatively easily. As a result, the generalization performance of the learning model can be improved, similar to the information processing device according to the above-mentioned embodiment.

[0031] Incidentally, a recording medium, such as a CD-ROM or a DVD-ROM, on which the computer program is recorded corresponds to an embodiment of the recording medium. EXAMPLES

[0032] An embodiment of an information processing device will be described with reference to Figs. 1 to 5. In Fig. 1, an information processing device 1 according to the embodiment is configured to include a machine learning unit 10 and a classification unit 20. The machine learning unit 10 is configured to generate a learning model from learning data as input data. In order to generate the learning model, the machine learning unit 10 has a feature extraction unit 11, an evaluation calculation unit 12, and a learning unit 13. The classification unit 20 is configured to classify the input data using the learning model generated by the machine learning unit 10.

[0033] The machine learning unit 10 will be described with reference to Fig. 2 in addition to Fig. 1. The learning data (i.e., input data) input to the machine learning unit 10 may include data acquired from a subject exhibiting a certain symptom (in other words, clinical data) and data artificially created to imitate the certain symptom. The learning data may include data acquired from a subject not exhibiting the certain symptom.

[0034] The data included in the learning data, which is obtained from a subject exhibiting a certain symptom, and the data artificially created to imitate a certain symptom, may each be given a label indicating that it is a certain symptom. The data included in the learning data, which is obtained from a subject not exhibiting a certain symptom, may be given a label indicating, for example, that it is normal. Note that, since various existing aspects can be applied to the method of generating the learning data, detailed explanation thereof will be omitted.

[0035] The feature extraction unit 11 of the machine learning unit 10 extracts frequency components as features from the learning data. The evaluation calculation unit 12 refers to the feature extraction results by the feature extraction unit 11 and determines whether or not the learning data includes a frequency component as a feature related to one symptom for each of n frequency bands included in the detailed classification (see FIG. 2). The evaluation calculation unit 12 obtains a first evaluation (corresponding to "Evaluation 1" in FIG. 2) indicating the determination result as a first characteristic of the learning data. Note that the above "n" is a natural number equal to or greater than 3.

[0036] The evaluation calculation unit 12 refers to the feature extraction result by the feature extraction unit 11 for the same learning data, and judges whether or not the learning data contains a frequency component as a feature related to one symptom for each of m frequency bands included in the major classification (see FIG. 2). The evaluation calculation unit 12 obtains a second evaluation (corresponding to "Evaluation 2" in FIG. 2) indicating the judgment result as a second characteristic of the learning data. Note that the above "m" is a natural number equal to or greater than 2 and smaller than the above-mentioned "n".

[0037] As an example of the detailed classification described above, as shown in Fig. 3(a), a classification method is given in which the frequency band from 0 Hz to 4000 Hz is divided into 13 frequency bands. As an example of the major classification described above, as shown in Fig. 3(b), a classification method is given in which the frequency band from 0 Hz to 4000 Hz is divided into three frequency bands.

[0038] The learning unit 13 of the machine learning unit 10 generates a learning model for one symptom based on the first characteristic and the second characteristic acquired by the evaluation calculation unit 12. Specifically, for example, the learning unit 13 identifies a weighting factor for each of the n frequency bands included in the detailed classification based on the first characteristic. The learning unit 13 identifies a weighting factor for each of the m frequency bands included in the large classification based on the second characteristic. Then, the learning unit 13 generates the learning model based on the weighting factor for each of the n frequency bands included in the detailed classification and the weighting factor for each of the m frequency bands included in the large classification.

[0039] Here, a specific example of the learning model will be described with reference to Fig. 4. The vertical axis of the graph shown in Fig. 4 corresponds to, for example, the amount of feature, weight, frequency, etc. Here, for convenience, the vertical axis of the graph shown in Fig. 4 is assumed to be weight. Fig. 4(a) is an example of a learning model generated based only on the above-mentioned first characteristic. In the learning model shown in Fig. 4(a), the weight of the frequency band surrounded by the dotted circle C1 is relatively small.

[0040] In the learning model shown in Fig. 4(a), the weight of the frequency band surrounded by the dotted circle C1 is relatively small for the following two reasons: (i) there is little or no frequency component characteristic of a symptom, and (ii) there is a bias in the learning data.

[0041] Considering that the degree of symptoms (e.g., body sounds) that appear due to one disease varies depending on the severity of the disease, the weight of the frequency band surrounded by the dotted circle C1 in the learning model shown in Fig. 4(a) is relatively small, it is highly likely that the learning data was biased. In such a learning model shown in Fig. 4(a), the ability to identify unknown data may be relatively low.

[0042] In contrast, Fig. 4(b) is an example of a learning model generated based on the above-mentioned first and second characteristics. In the learning model shown in Fig. 4(b), the weight of the frequency band surrounded by the dotted circle C2 (i.e., corresponding to the frequency band surrounded by the dotted circle C1 in Fig. 4(a)) is relatively large. In other words, by generating a learning model based on the first and second characteristics, even if the learning data is biased, it is possible to generate a learning model in which the influence of the bias in the learning data is mitigated.

[0043] Therefore, a learning model generated based on the first characteristic and the second characteristic as shown in FIG. 4(b) can improve the discrimination ability for unknown data compared to a learning model generated based only on the first characteristic as shown in FIG. 4(a).

[0044] Next, the operation of the machine learning unit 10 of the information processing device 1 will be described with reference to the flowchart of Fig. 5. In Fig. 5, the evaluation calculation unit 12 of the machine learning unit 10 refers to the feature extraction result by the feature extraction unit 11, and determines whether or not the learning data includes a frequency component as a feature related to one symptom for each of n frequency bands included in the detailed classification. Then, the evaluation calculation unit 12 calculates a first evaluation corresponding to the determination result (step S101). The evaluation calculation unit 12 acquires the first evaluation as a first characteristic of the learning data.

[0045] In parallel with or before or after the process of step S101, the evaluation calculation unit 12 refers to the feature extraction result by the feature extraction unit 11 for the same learning data, and judges whether or not the learning data contains a frequency component as a feature related to one symptom for each of the m frequency bands included in the major classification. Then, the evaluation calculation unit 12 calculates a second evaluation corresponding to the judgment result (step S102). The evaluation calculation unit 12 acquires the second evaluation as a second characteristic of the learning data. Thereafter, the learning unit 13 of the machine learning unit 10 generates a learning model related to one symptom based on the first characteristic and the second characteristic acquired by the evaluation calculation unit 12 (step S103).

[0046] Incidentally, the "evaluation calculation unit 12" and the "learning unit 13" in the examples correspond to examples of the "acquisition means" and the "generation means" in the above-mentioned embodiment, respectively.

[0047] The identification unit 20 of the information processing device 1 identifies unknown body sound data as input data using the learning model generated as described above. Specifically, the identification unit 20 first separates the frequency components included in the body sound data into n frequency bands in a predetermined frequency band to obtain a first analysis result (n is a natural number of 3 or more). The identification unit 20 also separates the frequency components included in the body sound data into m frequency bands in a predetermined frequency band to obtain a second analysis result (m is a natural number of 2 or more and smaller than the above n).

[0048] The identification unit 20 determines whether or not the body sound data includes a feature related to the certain symptom, based on the first analysis result, the second analysis result, and the learning model related to the certain symptom generated by the machine learning unit 10. The identification unit 20 outputs the determination result to, for example, a display device (not shown).

[0049] The present invention is not limited to the above-described embodiments, but may be modified as appropriate within the scope of the claims and the gist or concept of the invention as can be read from the entire specification, and information processing devices and methods, computer programs, and recording media that involve such modifications are also included in the technical scope of the present invention. [Explanation of symbols]

[0050] 1...information processing device, 10...machine learning unit, 11...feature extraction unit, 12...evaluation calculation unit, 13...learning unit, 20...recognition unit

Claims

1. An information processing device that performs machine learning using learning data generated from body sound data including features related to a symptom, an acquiring means for determining whether or not the learning data includes a frequency component as a feature related to the one symptom for each of n frequency bands (n is a natural number equal to or greater than 3) defined by dividing a predetermined frequency band, thereby acquiring a first characteristic of the learning data, and for determining whether or not the learning data includes a frequency component as a feature related to the one symptom for each of m frequency bands (m is a natural number equal to or greater than 2 and smaller than n) defined by dividing the predetermined frequency band, thereby acquiring a second characteristic of the learning data; A generation means for generating a learning model related to the one symptom based on the first characteristic and the second characteristic; An information processing device comprising:

2. The information processing device according to claim 1, characterized in that the generation means specifies a weighting coefficient for each of the n frequency bands based on the first characteristic, and specifies a weighting coefficient for each of the m frequency bands based on the second characteristic, thereby generating the learning model.

3. 3. The information processing apparatus according to claim 1, wherein the body sound data includes data acquired from a subject exhibiting the one symptom and data artificially created to imitate the one symptom.

4. An information processing method for performing machine learning using learning data generated from body sound data including features related to a symptom, comprising: a first acquisition step of determining whether or not the learning data includes a frequency component as a feature related to the one symptom for each of n frequency bands (n being a natural number equal to or greater than 3) defined by dividing a predetermined frequency band, thereby acquiring a first characteristic of the learning data; a second acquisition step of determining whether or not the learning data includes a frequency component as a feature related to the one symptom for each of m frequency bands (where m is a natural number equal to or greater than 2 and smaller than n) defined by dividing the predetermined frequency band, thereby acquiring a second characteristic of the learning data; A generation step of generating a learning model related to the one symptom based on the first characteristic and the second characteristic; 13. An information processing method comprising:

5. A computer of an information processing device that performs machine learning using learning data generated from body sound data including features related to one symptom, an acquiring means for determining whether or not the learning data includes a frequency component as a feature related to the one symptom for each of n frequency bands (n is a natural number equal to or greater than 3) defined by dividing a predetermined frequency band, thereby acquiring a first characteristic of the learning data, and for determining whether or not the learning data includes a frequency component as a feature related to the one symptom for each of m frequency bands (m is a natural number equal to or greater than 2 and smaller than n) defined by dividing the predetermined frequency band, thereby acquiring a second characteristic of the learning data; A generation means for generating a learning model related to the one symptom based on the first characteristic and the second characteristic; A computer program characterized by causing the computer to function as follows.

6. A recording medium on which the computer program according to claim 5 is recorded.

Images