Evaluation device, evaluation system, evaluation method, and evaluation program for evaluating state of intestine

Abstract

An evaluation device (3) for evaluating the state of an intestine of a subject comprises: an extraction unit (32) for extracting intestinal sounds from acoustic data obtained from the subject; a first feature amount calculation unit (331) for calculating a first feature amount, which is a feature amount of the intestinal sound before stimulation is given to the intestine, from among the extracted intestinal sounds; a second feature amount calculation unit (332) for calculating a second feature amount, which is a feature amount of the intestinal sound after stimulation is given to the intestine, from among the extracted intestinal sounds; and an evaluation unit (34) for evaluating the state of the intestine of the subject on the basis of the first feature amount and the second feature amount. The evaluation unit (34) evaluates the state of the intestine using a machine-learned evaluation model (M2).

Description

Evaluation device, evaluation system, evaluation method, and evaluation program for evaluating intestinal condition

[0001] The present invention relates to an evaluation device, an evaluation system, an evaluation method, and an evaluation program for evaluating the state of the intestines.

[0002] Techniques for evaluating the state of the intestines based on bowel sounds have been proposed (for example, Patent Document 1). Specifically, Patent Document 1 discloses a method for calculating the amount of gas in the intestines based on time domain features such as the duration and frequency of bowel sounds.

[0003] Japanese Patent Application Laid-Open No. 2021-74238

[0004] However, even among subjects with similar intestinal conditions, the duration and frequency of bowel sounds vary from person to person. For example, among subjects with healthy intestinal conditions, there are subjects who have relatively frequent bowel sounds and subjects who have relatively infrequent bowel sounds. Therefore, the conventional technology described in Patent Document 1 has the problem of being unable to accurately evaluate intestinal condition.

[0005] An object of the present invention is to accurately evaluate the state of the intestines.

[0006] In order to solve the above problems, the present invention includes the following aspects. Item 1. An evaluation device for evaluating a subject's intestinal condition, comprising: an extraction unit that extracts bowel sounds from acoustic data obtained from the subject; a first feature amount calculation unit that calculates, from the extracted bowel sounds, a first feature amount that is a feature amount of the bowel sounds before stimulation of the intestines is applied; a second feature amount calculation unit that calculates, from the extracted bowel sounds, a second feature amount that is a feature amount of the bowel sounds after stimulation of the intestines is applied; and an evaluation unit that evaluates the subject's intestinal condition based on the first feature amount and the second feature amount, wherein the evaluation unit evaluates the intestinal condition using a machine-learned evaluation model. Item 2. The evaluation device according to Item 1, wherein the first feature amount calculation unit calculates a plurality of the first feature amounts, and the second feature amount calculation unit calculates a plurality of the second feature amounts, and the evaluation unit evaluates the intestinal condition based on the plurality of first feature amounts and the plurality of second feature amounts. Item 3. 3. The evaluation device according to claim 1 or 2, wherein the evaluation model extracts bowel sounds from acoustic data obtained from a plurality of subjects, calculates, of the bowel sounds of each subject, a first feature amount that is a feature amount of the bowel sounds of the subject before stimulation of the intestines of the subject, and a second feature amount that is a feature amount of the bowel sounds of the subject after stimulation of the intestines of the subject, generates teacher data by labeling the intestinal condition of each subject to any of the calculated first feature amount and its logarithm, and any of the second feature amount, a value obtained by arithmetic operations on the first feature amount and the second feature amount, and their logarithms, and performs machine learning based on the teacher data, and the evaluation unit inputs to the evaluation model the first feature amount calculated by the first feature amount calculation unit and its logarithm, and any of the second feature amount calculated by the second feature amount calculation unit, a value obtained by arithmetic operations on the first feature amount and the second feature amount, and their logarithms, and evaluates the condition based on an output of the evaluation model in response to the input.Item 4. The evaluation device according to Item 3, wherein the training data is generated by labeling the intestinal condition of each subject to the logarithm of the calculated first feature amount and the logarithm of the ratio between the first feature amount and the second feature amount, and the evaluation unit inputs the logarithm of the first feature amount calculated by the first feature amount calculation unit and the logarithm of the ratio between the first feature amount and the second feature amount calculated by the second feature amount calculation unit to the evaluation model. Item 6. The evaluation device according to Item 1 or 2, wherein the evaluation model extracts bowel sounds from acoustic data obtained from a plurality of subjects with healthy bowel conditions, calculates, of the bowel sounds of each subject, a first feature amount that is a feature amount of the bowel sounds before stimulation of the subject's intestines and a second feature amount that is a feature amount of the bowel sounds after stimulation of the subject's intestines, generates teacher data by labeling either the calculated first feature amount or its logarithm with either the second feature amount, a value obtained by arithmetic operations on the first feature amount and the second feature amount, or their logarithms, and performs machine learning on the teacher data, and the evaluation unit inputs either the first feature amount or its logarithm calculated by the first feature amount calculation unit into the evaluation model, and evaluates the condition by comparing a value output by the evaluation model in response to the input with either the second feature amount calculated by the second feature amount calculation unit, a value obtained by arithmetic operations on the first feature amount and the second feature amount, or their logarithms. Item 6. The evaluation device according to Item 5, wherein the training data is generated by labeling the intestinal condition of each subject with the logarithm of the calculated first feature amount and the logarithm of the ratio between the first feature amount and the second feature amount, and the evaluation unit inputs the logarithm of the first feature amount calculated by the first feature amount calculation unit to the evaluation model, and compares a value output by the evaluation model in response to the input with the logarithm of the ratio between the first feature amount and the second feature amount calculated by the second feature amount calculation unit.Item 7. The evaluation device according to Item 1 or 2, wherein the evaluation model extracts bowel sounds from acoustic data obtained from a plurality of subjects, calculates, of the bowel sounds of each subject, a first feature amount that is a feature amount of the bowel sounds before stimulation of the subject's intestines and a second feature amount that is a feature amount of the bowel sounds after stimulation of the subject's intestines, generates teacher data by labeling the intestinal condition of each subject to either the calculated first feature amount or its logarithm, and any difference between the second feature amount, a value obtained by arithmetic operations between the first feature amount and the second feature amount, and their logarithms, and performs machine learning based on the teacher data, and the evaluation unit inputs to the evaluation model either the first feature amount calculated by the first feature amount calculation unit or its logarithm, and any difference between the second feature amount calculated by the second feature amount calculation unit, a value obtained by arithmetic operations between the first feature amount and the second feature amount, and their logarithms, and evaluates the condition based on the output of the evaluation model in response to the input. The evaluation device according to any one of Items 1 to 7, wherein the feature amounts include feature amounts related to the time domain of the bowel sounds. Item 9. The evaluation device according to Item 8, wherein the feature amounts include the frequency of occurrence of the bowel sounds. Item 10. The evaluation device according to any one of Items 1 to 7, wherein the feature amounts include feature amounts related to the frequency domain of the bowel sounds. Item 11. The evaluation device according to Item 10, wherein the feature amounts include at least one of the peak frequency of the bowel sounds and the spectral bandwidth of the bowel sounds. Item 12. The evaluation device according to any one of Items 1 to 7, wherein the feature amounts include the frequency of occurrence of the bowel sounds, the peak frequency of the bowel sounds, and the spectral bandwidth of the bowel sounds. Item 13. The evaluation device according to any one of Items 1 to 12, wherein the evaluation model is an artificial neural network model. Item 14. The evaluation device according to Item 1, wherein the length of each acoustic data from which the bowel sounds before and after the bowel stimulation are extracted is 1 to 30 minutes. Item 15. Item 15. The evaluation device according to Item 14, wherein the length of each acoustic data from which the bowel sounds before and after the bowel stimulation are extracted is 1 to 10 minutes.Item 16. The evaluation device according to Item 15, wherein the length of each piece of acoustic data from which bowel sounds before and after intestinal stimulation are extracted is 1 to 3 minutes. Item 17. The evaluation device according to any one of Items 1 to 16, wherein the stimulation is stimulation by ingestion of a beverage or solid food. Item 18. An evaluation method for evaluating a subject's intestinal condition, comprising: an extraction step of extracting bowel sounds from acoustic data obtained from the subject; a first feature calculation step of calculating, from the extracted bowel sounds, a first feature that is a feature of the bowel sounds before the intestinal stimulation is applied; a second feature calculation step of calculating, from the extracted bowel sounds, a second feature that is a feature of the bowel sounds after the intestinal stimulation is applied; and an evaluation step of evaluating the intestinal condition of the subject based on the first feature and the second feature, wherein the evaluation step uses a machine-learned evaluation model to evaluate the intestinal condition. Item 19. 1. An evaluation program for evaluating a state of the intestines of a subject, the evaluation program causing a computer to execute the following steps: an extraction step of extracting bowel sounds from acoustic data obtained from the subject; a first feature calculation step of calculating, from the extracted bowel sounds, a first feature which is a feature of the bowel sounds before stimulation of the intestines; a second feature calculation step of calculating, from the extracted bowel sounds, a second feature which is a feature of the bowel sounds after stimulation of the intestines; and an evaluation step of evaluating the state of the intestines of the subject based on the first feature and the second feature, wherein in the evaluation step, the state of the intestines is evaluated using a machine-learned evaluation model.

[0007] According to the present invention, the state of the intestines can be evaluated with high accuracy.

[0008] Fig. 1 is a block diagram showing the configuration of an evaluation system according to one embodiment of the present invention; Fig. 2 is a flowchart showing the processing procedure of an evaluation method according to one embodiment of the present invention; Fig. 3 is a flowchart showing more specific processing steps of the step of collecting acoustic data; and Fig. 4 is an example of the waveform of acoustic data obtained from a subject.

[0009] Hereinafter, an embodiment of the present invention will be described with reference to the accompanying drawings. Note that the present invention is not limited to the following embodiment, and various modifications are possible without departing from the spirit of the present invention.

[0010] 1 is a block diagram showing the configuration of an evaluation system 1 according to one embodiment of the present invention. The evaluation system 1 includes a sound collection device 2 and an evaluation device 3.

[0011] The sound collection device 2 is a device for obtaining acoustic data from a subject. When acoustic data is collected at a medical institution, the sound collection device 2 is, for example, a stethoscope, and when acoustic data is collected outside a medical institution, the sound collection device 2 is, for example, a microphone. When the sound collection device 2 is a microphone, the sound collection device 2 is attached to the abdomen of the subject directly or via clothing, and obtains acoustic data including the subject's bowel sounds. Note that there are no particular limitations on the installation location of the sound collection device 2.

[0012] The sound collection device 2 is connected to the evaluation device 3 by wire or wirelessly, and transfers the acquired acoustic data to the evaluation device 3. Note that the acoustic data may be transferred from the sound collection device 2 to the evaluation device 3 via a recording medium or a communication device.

[0013] The evaluation device 3 is a device for evaluating the intestinal condition of a subject. The evaluation device 3 can be configured as a general-purpose computer or a portable computer such as a smartphone or tablet terminal. The hardware configuration of the evaluation device 3 includes a processor such as a CPU or GPU, a main memory device such as a DRAM or SRAM (not shown), and an auxiliary memory device 30 such as an HDD or SSD. The auxiliary memory device 30 stores an evaluation program P, an extraction model M1, an evaluation model M2, etc.

[0014] The auxiliary storage device 30 may be external to the evaluation device 3. The evaluation device 3 may also be provided on a cloud.

[0015] The evaluation device 3 includes, as functional blocks, an acoustic data acquisition unit 31, an extraction unit 32, a feature calculation unit 33, and an evaluation unit 34. Furthermore, the feature calculation unit 33 includes a first feature calculation unit 331 and a second feature calculation unit 332. These functional blocks are realized by the processor of the evaluation device 3 reading an evaluation program P into a main storage device and executing it. The evaluation program P may be downloaded to the evaluation device 3 via a communication network such as the Internet, or may be recorded on a computer-readable non-transitory recording medium such as a CD-ROM or an SD card and installed into the evaluation device 3 via the storage medium. The functions of each functional block will be described later.

[0016] (Processing Procedure of Evaluation Method) The functions of the evaluation system 1 will be described with reference to Fig. 2. Fig. 2 is a flowchart showing the processing procedure of the evaluation method according to this embodiment. The evaluation method has steps S1 to S5, with step S1 being executed by the sound collection device 2 and steps S2 to S5 being executed by the above-mentioned respective functional blocks of the evaluation device 3. That is, the evaluation program P causes the evaluation device 3 to execute steps S2 to S5.

[0017] In step S1, the sound collection device 2 collects acoustic data including bowel sounds of the subject. In this embodiment, a stimulus is applied to the intestines of the subject, and the sound collection device 2 collects acoustic data before and after the stimulus is applied to the intestines.

[0018] FIG. 3 is a flowchart showing more specific processing steps of step S1.

[0019] In step S11, sound collection is started by the sound collection device 2. After a predetermined time (for example, 5 minutes) has elapsed (YES in step S12), in step S13, the subject is made to ingest a beverage to stimulate the subject's intestines.

[0020] The beverage is not particularly limited as long as it can stimulate the intestines, and examples thereof include carbonated water and coffee. Furthermore, the method of stimulating the intestines is not limited to ingestion of a beverage. Examples of methods of stimulating the intestines include ingestion of drugs such as laxatives, ingestion of solid foods such as food, exercises that stimulate the intestines such as twisting the abdomen or applying abdominal pressure, pressing the abdomen, stimulating acupressure points related to the intestines, waking up and going to bed, etc.

[0021] After a predetermined time (for example, 10 minutes) has elapsed since the intestines were stimulated (YES in step S14), sound collection by the sound collection device 2 is terminated in step S15.

[0022] As described above, the sound collection device 2 collects acoustic data before and after stimulating the intestines of the subject.

[0023] In step S2 shown in FIG. 2, the acoustic data acquisition unit 31 of the evaluation device 3 acquires acoustic data obtained from the subject by the sound collection device 2.

[0024] In step S3 (extraction step), the extraction unit 32 extracts bowel sounds from the acoustic data acquired in step S2. The acoustic data contains noise such as breathing sounds and body movement sounds in addition to bowel sounds, but bowel sounds can be extracted using known techniques. In this embodiment, the extraction unit 32 extracts bowel sounds using the technique described in Japanese Patent No. 7197922.

[0025] Specifically, the extraction unit 32 detects multiple segments from the acoustic data, extracts frequency-related features (e.g., linear prediction cepstral (LPC) coefficients, MFCC) from each segment, and inputs the extracted features to the extraction model M1. The extraction model M1 is a neural network model that has undergone machine learning to determine the relationship between bowel sounds in the acoustic data and the features, and outputs a prediction score indicating the likelihood that each segment contains a bowel sound based on the input features. The extraction unit 32 extracts segments with prediction scores greater than a predetermined threshold as bowel sounds. A bowel sound is a continuous sound, and its length (time) is indefinite.

[0026] 4 shows an example of the waveform of acoustic data obtained from a subject. The dashed lines indicate segments, and the extractor 32 extracts segments that are likely to contain bowel sounds.

[0027] Note that the frequency of normal bowel sounds is mainly between approximately 100 and 500 Hz. Therefore, it is preferable that the extraction unit 32 extracts bowel sounds after applying a band-pass filter to the acoustic data acquired in step S2 that passes a frequency band including 100 to 500 Hz (for example, 80 to 1000 Hz). Alternatively, the extraction unit 32 may extract bowel sounds whose signal level is equal to or greater than a threshold.

[0028] In step S4 (first feature calculation step, second feature calculation step) shown in FIG. 2 , the feature calculation unit 33 calculates feature quantities of the bowel sounds extracted in step S3. Specifically, the first feature calculation unit 331 of the feature calculation unit 33 calculates a first feature quantity, which is a feature quantity of the bowel sounds extracted before intestinal stimulation, and the second feature calculation unit 332 of the feature calculation unit 33 calculates a second feature quantity, which is a feature quantity of the bowel sounds extracted after intestinal stimulation. The durations of the bowel sounds corresponding to the first and second feature quantities (the durations of the acoustic data extracted from the bowel sounds before and after intestinal stimulation) are not particularly limited, but are preferably, for example, 1 to 60 minutes, 1 to 30 minutes, 1 to 10 minutes, or 1 to 3 minutes. Even acoustic data obtained from such a short measurement period can accurately evaluate the state of the bowel, as will be shown in the examples described below. This reduces the burden on the subject.

[0029] In this embodiment, the feature amount is preferably a feature amount related to the time domain of bowel sounds, and specifically, the frequency of bowel sounds. Other time domain feature amounts include, for example, the duration of bowel sounds, bowel sound signal level, maximum bowel sound amplitude (BS Max. amplitude), bowel sound power (BS power), maximum amplitude spectral density of bowel sounds, zero crossing rate of bowel sounds, kurtosis of bowel sounds, and the ratio of largest absolute to root mean squared value of bowel sounds.

[0030] The feature amount is not particularly limited as long as it is an acoustic feature amount of bowel sounds, and may be an acoustic feature amount related to the frequency domain, etc. Examples of frequency domain features include the spectral bandwidth of bowel sounds (the width from the peak to an arbitrary attenuation point), the peak frequency of bowel sounds, the spectral centroid of bowel sounds, the spectral flatness of bowel sounds, the median frequency of bowel sounds, the mean frequency of bowel sounds, the first formant of bowel sounds, the second formant of bowel sounds, the spectral entropy of bowel sounds, the spectral spread of bowel sounds, the spectral rolloff of bowel sounds, the spectral slope of bowel sounds, the spectral kurtosis of bowel sounds, and the spectral skewness of bowel sounds. Alternatively, the number of segments per unit time that are likely to contain bowel sounds (BS segments) or the number of non-BS segments per unit time (also known as Bowel-Sound Duration or Silence Duration) may be used as features. Furthermore, the recorded (time-series) data before and after stimulation itself may be used as features, and features related to the time-frequency domain may include cepstrum coefficients based on a filter bank, spectrograms, scalograms, and spectrograms using a filter bank (e.g., log-mel spectrograms).

[0031] Note that the first feature amount calculation unit 331 and the second feature amount calculation unit 332 may calculate a single first feature amount and a single second feature amount, respectively, but preferably calculate a plurality of first feature amounts and a plurality of second feature amounts, respectively, and the types and numbers of the calculated first feature amounts and second feature amounts are the same.

[0032] In step S5 (evaluation step), the evaluation unit 34 evaluates the intestinal condition of the subject based on the first feature amount and the second feature amount. In this embodiment, the evaluation unit 34 evaluates the intestinal condition using the machine-learned evaluation model M2.

[0033] The evaluation model M2 is generated by extracting bowel sounds from acoustic data obtained from multiple subjects through prior verification experiments, etc., calculating a first feature value, which is a feature value of the bowel sounds of each subject before stimulation of the subject's intestines, and a second feature value, which is a feature value of the bowel sounds of each subject after stimulation of the subject's intestines, generating training data by labeling the state of each subject's intestines to the calculated first feature value (and either its logarithm) and the second feature value (a value obtained by arithmetic operations on the first feature value and the second feature value (e.g., a ratio), and either their logarithms), and then performing machine learning based on the training data. The evaluation unit 34 inputs the first feature (or any of its logarithms) calculated by the first feature calculation unit 331 and the second feature (or any of the values ​​(e.g., ratios) obtained by arithmetic operations on the first feature and the second feature, and their logarithms) calculated by the second feature calculation unit 332 into the evaluation model M2, and evaluates the intestinal condition of the subject based on the output of the evaluation model M2 in response to the input.

[0034] Furthermore, as a variant, the evaluation model M2 may be generated by extracting bowel sounds from acoustic data obtained from a plurality of subjects with healthy intestinal conditions through prior verification experiments, etc., calculating, from the bowel sounds of each subject, a first feature quantity that is a feature quantity of the bowel sounds before stimulation of the subject's intestines and a second feature quantity that is a feature quantity of the bowel sounds after stimulation of the subject's intestines, labeling either the calculated first feature quantity or its logarithm with either the second feature quantity, a value obtained by arithmetic operations on the first feature quantity and the second feature quantity (for example, a ratio), or their logarithms, to generate training data, and performing machine learning based on the training data. In this case, the evaluation unit 34 inputs either the first feature calculated by the first feature calculation unit 331 or its logarithm into the evaluation model M2, and evaluates the intestinal condition of the subject by comparing the value output by the evaluation model M2 in response to this with either the second feature calculated by the second feature calculation unit 332, a value (e.g., a ratio) obtained by arithmetic operations on the first feature and the second feature, or their logarithms. More specifically, the evaluation unit 34 calculates the difference between the output value of the evaluation model M2 and either the second feature calculated by the second feature calculation unit 332, a value (e.g., a ratio) obtained by arithmetic operations on the first feature and the second feature, or their logarithms, and evaluates that the intestinal condition is healthy if the difference is less than a threshold, and that the intestinal condition is unhealthy if the difference is equal to or greater than the threshold.

[0035] Furthermore, as another variant, the evaluation model M2 may be generated by extracting bowel sounds from acoustic data obtained from a plurality of subjects through prior verification experiments or the like, calculating, of the bowel sounds of each subject, a first feature quantity which is a feature quantity of the bowel sounds before stimulation of the subject's intestines, and a second feature quantity which is a feature quantity of the bowel sounds after stimulation of the subject's intestines, and generating training data by labeling the intestinal condition of each subject to either the calculated first feature quantity or its logarithm, and a value (e.g., a ratio) obtained by arithmetic operations between the second feature quantity, the first feature quantity, and the second feature quantity, and either the difference (or any of the sum, product, or quotient) of their logarithms, and then performing machine learning based on the training data. In this case, the evaluation unit 34 inputs into the evaluation model M2 the first feature calculated by the first feature calculation unit 331 and either its logarithm, the second feature calculated by the second feature calculation unit 332, a value (e.g., a ratio) obtained by arithmetic operations between the first feature and the second feature, and either the difference (or sum, product, or quotient) between their logarithms, and evaluates the intestinal condition of the subject based on the output of the evaluation model M2 in response to the input.

[0036] In this embodiment, the evaluation model M2 is, for example, an artificial neural network (ANN) model, but is not limited to this. Examples of neural network models include general shallow networks, convolutional neural networks (CNNs), recurrent neural networks (RNNs), and long short-term memory (LSTMs). Traditional machine learning models such as random forests, support vector machines (SVMs), and XGBoost can also be used as the evaluation model M2. Furthermore, innovative models (e.g., multimodal AI, AutoML, TinyML, next-generation transformers, etc.) can also be used as the evaluation model M2.

[0037] As will be shown in the examples described below, the use of the evaluation model M2 allows for accurate evaluation of the state of the intestines, and in particular, by combining multiple feature amounts, the performance of the evaluation model M2 can be further improved.

[0038] In this embodiment, the intestinal condition is evaluated as to whether the intestines are healthy or not, but the present invention is not limited to this. For example, the intestinal condition may be evaluated as to the motor function (motility) of the intestines, or the type of stimulation (strong stimulation or weak stimulation) to which the intestines are subjected (daily or temporarily).

[0039] Examples of the present invention will be described below, but the present invention is not limited to the following examples.

[0040] In Example 1, 37 female subjects were administered a questionnaire reflecting the Rome IV diagnostic criteria, and based on their responses, the subjects were classified into functional diarrhea (FD), functional constipation (FC), and healthy individuals. The classification results are shown in Table 1.

[0041] Next, each subject underwent a beverage ingestion test. Specifically, after approximately 12 hours of fasting, each subject ingested 200 mL of strong carbonated water at approximately 10°C or below. Acoustic data were collected from each subject's body sounds during a resting state and for 10 minutes after ingestion. The instruments used to collect the acoustic data were an electronic stethoscope (E-scope2, Cardionics Inc., Houston, TX, USA), a multi-track recorder (R16 Zoom Co., Ltd., Tokyo, Japan), and an audio interface (UA-1010 Roland Corp., Shizuoka, Japan). During the acoustic data collection, the subject was placed in a supine position, with the electronic stethoscope secured to the right lower abdomen. Audio data were recorded using the multi-track recorder or audio interface. The frequency of the collected acoustic data was downsampled to 4000 Hz, taking into account the frequency characteristics of the electronic stethoscope and bowel sounds.

[0042] In the evaluation device 3, a trained neural network model was used as the extraction model M1 to extract multiple bowel sounds from each acoustic data. Specifically, the acoustic data was segmented with a segment length of 64 ms and a shift size of 16 ms, and the signals within the segments were normalized to a mean value of 0 and a standard deviation of 1. A total of 52 features, consisting of 10-dimensional linear prediction cepstral (LPC) coefficients, their Δ and ΔΔ, and a 22-dimensional MFCC, were extracted from the normalized segments and used as input to the neural network. Each feature was labeled with a binary value indicating whether it was a bowel sound (BS) segment or not to generate training data. The generated training data was used to train the neural network and generate the extraction model M1 for two-class classification of whether it was a bowel sound or not. The number of units in the intermediate layer of the neural network was 40. The performance of the extraction model M1 was evaluated using five-fold cross-validation, and the results were a sensitivity of 80.47±2.83%, a specificity of 97.61±0.74%, a PPV of 86.17±1.88%, an NPV of 96.52±0.74%, an accuracy of 95.02±0.87%, and an F1 score of 83.17±0.80%.

[0043] In addition, taking into account changes in acoustic features due to environmental noise, noise subtraction was performed using the average of the spectra obtained from segments determined not to be bowel sounds.

[0044] Next, a first feature quantity, which is a feature quantity before ingesting a beverage, and a second feature quantity, which is a feature quantity after ingesting a beverage, were calculated from the extracted bowel sounds. More specifically, the first feature quantity was calculated by multiplying the average (x ｂ ) and the second feature is defined as the average (x ａ ) and the first feature x ｂ and the second feature x ａwas calculated for each subject. Three feature quantities were used, which are robust against noise, such as: frequency of bowel sounds (number of BS detected / min), peak frequency of bowel sounds (Peak frequency), and spectral bandwidth of bowel sounds (Bandwidth). The spectral bandwidth was the frequency range from the maximum value of the amplitude spectrum to 90%.

[0045] Next, an evaluation model M2 was generated using a neural network model. The neural network model used had an input layer with dimensions twice the number of features, a 10-dimensional intermediate layer, and a 2-dimensional fully connected network output layer. The output activation function for the intermediate layer was the relu function, and the output activation function for the output layer was the softmax function.

[0046] In Example 1, the first feature x of each subject ｂ and the second feature x ａ The intestinal condition of each subject is labeled (healthy: 0, FD or FC: 1) to generate training data, and the first feature x ｂ and the second feature x ａ The evaluation model M2 was machine-learned using the explanatory variables and the intestinal condition as the objective variable. In the evaluation of the intestinal condition of each subject, the evaluation model M2 was ｂ and the second feature x ａ If the value output by the evaluation model M2 in response to this was below a threshold, the intestinal condition was evaluated as healthy, and if it was above the threshold, the intestinal condition was evaluated as unhealthy. The performance of the evaluation model M2 was evaluated by repeating hold-out cross-validation 100 times, in which the dataset of each subject was randomly divided in a ratio of 7:3 for training and validation, and the mean and variance of the obtained classification performance metrics were used.

[0047] In the machine learning and performance verification of the evaluation model M2, the two explanatory variables were set in the following four patterns: (Pattern 1) Explanatory variable 1: First feature x ｂ , explanatory variable 2: second feature x ａ (Pattern 2) Explanatory variable 1: First feature x ｂ , explanatory variable 2: first feature x ｂ and the second feature xａ (Pattern 3) Explanatory variable 1: First feature x ｂ , explanatory variable 2: logarithm of the second feature (Pattern 4) explanatory variable 1: first feature x ｂ Logarithm of, explanatory variable 2: first feature x ｂ and the second feature x ａ Logarithm of the ratio of

[0048] In either pattern, each explanatory variable was normalized so that the mean was 0 and the standard deviation was 1, and machine learning and performance verification of the evaluation model M2 were performed based on the normalized data group.

[0049] First, Table 2 shows the evaluation results of the performance of the evaluation model M2 when one feature amount is used.

[0050] Table 2 suggests that for any feature, performance tends to improve by using the first feature and the logarithm of the ratio between the first feature and the second feature (Pattern 4).

[0051] Next, the performance of the evaluation model M2 was evaluated when a plurality of feature quantities were combined. First, the evaluation results when the explanatory variables were set in the above-mentioned pattern 1 are shown in Table 3.

[0052] Table 4 shows the evaluation results when the explanatory variables were set according to the above pattern 2.

[0053] Table 5 shows the evaluation results when the explanatory variables were set according to the above pattern 3.

[0054] Table 6 shows the evaluation results when the explanatory variables were set according to the above pattern 4.

[0055] When there are three features, the input layer of the evaluation model M2 is six-dimensional. Tables 3 to 6 confirm that combining three features results in the highest accuracy and F1 score compared to using only one feature. Therefore, we were able to confirm the effectiveness of combining multiple features on performance.

[0056] Finally, Table 7 shows the evaluation results of the performance of the evaluation model M2 when three feature quantities are used and the setting pattern of the explanatory variables for each feature quantity is made variable.

[0057] From Table 7, it can be seen that by varying the setting pattern of the explanatory variables for each feature amount, higher performance was achieved than the performance when the setting pattern of the explanatory variables for each feature amount was the same (bottom rows of Tables 3 to 6). This suggests that it is possible to further improve the performance of the evaluation model M2 by varying the setting pattern of the explanatory variables for the feature amounts.

[0058] In this embodiment, the logarithm is a natural logarithm, but the base of the logarithm is not particularly limited. ｂ and the ratio (x ａ / x ｂ ) are logarithmized, but it is also possible to take the logarithm of only one of them. ｂ Logarithms and ratios of (x ａ / x ｂ ) or based on the first feature x ｂ and the ratio (x ａ / x ｂ ) may be used to evaluate the state of the intestine. ｂ and the ratio (x ａ / x ｂ ), instead of taking the logarithm, a linear transformation such as a Hilbert transform, or a nonlinear transformation such as a Yeo-Johnson transform, a Box-Cox transform, an arcsin transform, a square root transform, dynamic range compression, exponential scaling, or a custom transformation thereof may be performed.

[0059] Example 2 In Example 2, machine learning and performance verification of the evaluation model M2 were performed using a method different from that of Example 1. The neural network model used as the evaluation model M2 used a fully connected network in which the input layer had the same dimensions as the number of features, the intermediate layer had 10 dimensions, and the output layer had one dimension. The output activation function of the intermediate layer was a hyperbolic tangent function (tanh function), and the output activation function of the output layer was a linear function. Other aspects were the same as in Example 1, so detailed description will be omitted.

[0060] In Example 2, the first feature value x of each of the 17 subjects whose intestinal conditions are healthy (hereinafter referred to as healthy subjects) shown in Table 1 was calculated. ｂ The second feature x of each healthy subject is ａ to generate training data, and the first feature x ｂ is the explanatory variable, and the second feature x ａ The evaluation model M2 was machine-learned using the objective variable x. In the evaluation of the intestinal condition of each subject, the evaluation model M2 was ｂ is input, and the value (predicted value) output by the evaluation model M2 in response to this is used as the second feature x ａ The performance of the evaluation model M2 was evaluated using the mean and variance of the classification performance metrics obtained by randomly dividing the datasets of each subject shown in Table 1 into 100 hold-out cross-validations with a training:validation ratio of 7:3.

[0061] In the machine learning and performance verification of the evaluation model M2, the explanatory variables and the objective variables were set in the following four patterns: (Pattern 5) Explanatory variable: First feature value x ｂ , objective variable: second feature x ａ (Pattern 6) Explanatory variable: first feature x ｂ , objective variable: first feature x ｂ and the second feature x ａ (Pattern 7) Explanatory variable: First feature x ｂ logarithm of, objective variable: second feature x ａ (Pattern 8) Explanatory variable: First feature x ｂlogarithm of, objective variable: first feature x ｂ and the second feature x ａ Logarithm of the ratio of

[0062] In both patterns, the explanatory variables and objective variables were normalized so that the mean was 0 and the standard deviation was 1, and machine learning and performance verification of the evaluation model M2 were performed based on the normalized data group. In addition, the threshold value used for comparison with the output value of the evaluation model M2 was set so as to maximize performance.

[0063] First, Table 8 shows the evaluation results of the performance of the evaluation model M2 when one feature amount is used.

[0064] Table 8 suggests that, for features excluding the peak frequency, performance tends to improve by using the first feature and the logarithm of the ratio between the first feature and the second feature (Pattern 8).

[0065] Next, the performance of the evaluation model M2 was evaluated when a plurality of feature quantities were combined. First, the evaluation results when the explanatory variables and the objective variables were set in the above-mentioned pattern 5 are shown in Table 9.

[0066] Table 10 shows the evaluation results when the explanatory variables and objective variables were set according to the above pattern 6.

[0067] Table 11 shows the evaluation results when the explanatory variables and objective variables were set according to the above pattern 7.

[0068] Table 12 shows the evaluation results when the explanatory variables and objective variables were set according to the above pattern 8.

[0069] Tables 9 to 12 confirm that combining three features results in the highest performance compared to using only one feature. Therefore, the effectiveness of combining multiple features on performance was confirmed.

[0070] Finally, Table 13 shows the evaluation results of the performance of the evaluation model M2 when three feature quantities are used and the setting patterns of the explanatory variables and objective variables of each feature quantity are varied.

[0071] From Table 13, it can be seen that by varying the setting patterns of the explanatory variables and the objective variables for each feature quantity, higher performance could be achieved than the performance when the setting patterns of the explanatory variables for each feature quantity were the same (bottom rows of Tables 9 to 12). This suggests that it is possible to further improve the performance of the evaluation model M2 by varying the setting patterns of the explanatory variables and the objective variables for the feature quantities.

[0072] In Example 3, machine learning and performance verification of the evaluation model M2 were performed using a method different from that used in Examples 1 and 2. The neural network model used as the evaluation model M2 was the same as that used in Example 1, except that the input layer had the same dimensions as the number of features. Since the rest of the model is the same as in Example 1, detailed description thereof will be omitted.

[0073] In the machine learning and performance verification of the evaluation model M2, the explanatory variables were set in the following four patterns: (Pattern 9) First feature x ｂ and the second feature x ａ The difference (x ｂ -x ａ ) (Pattern 10) First feature x ｂ and the first feature x ｂ and the second feature x ａ The difference between the ratio (x ｂ -x ｂ / x ａ ) (Pattern 11) First feature x ｂ and the second feature x ａ The difference between the logarithm of (log(x ｂ ) -log(x ａ )) (Pattern 12) First feature x ｂ and the first feature x ｂ and the second feature x ａ The difference between the logarithm of the ratio (log(x ｂ ) -log(x ｂ / x ａ ))

[0074] In both patterns, the explanatory variables were normalized so that the mean was 0 and the standard deviation was 1, and machine learning and performance verification of the evaluation model M2 were performed based on the normalized data group.

[0075] First, Table 14 shows the evaluation results of the performance of the evaluation model M2 when one feature amount is used.

[0076] Table 14 suggests that performance tends to improve, particularly by using the first feature amount and the logarithm of the ratio between the first feature amount and the second feature amount (Patterns 11 and 12).

[0077] Next, the performance of the evaluation model M2 was evaluated when a plurality of feature quantities were combined. First, the evaluation results when the explanatory variables were set in the above-mentioned pattern 9 are shown in Table 15.

[0078] Table 16 shows the evaluation results when the explanatory variables were set according to the above pattern 10.

[0079] Table 17 shows the evaluation results when the explanatory variables were set according to the above pattern 12.

[0080] Next, for patterns other than those shown in Tables 15 to 17, the first feature x ｂ and the second feature x ａ The evaluation results when the explanatory variables obtained by the four arithmetic operations above were set are shown in Table 18.

[0081] Tables 15 to 18 confirm that combining three features results in the highest performance compared to using only one feature. Therefore, the effectiveness of combining multiple features on performance was confirmed.

[0082] Finally, Table 19 shows the evaluation results of the performance of the evaluation model M2 when three feature quantities are used and the setting pattern of the explanatory variables for each feature quantity is made variable.

[0083] From Table 19, it can be seen that by varying the setting pattern of the explanatory variables for each feature amount, higher performance could be achieved than the performance when the setting pattern of the explanatory variables for each feature amount was the same (bottom rows of Tables 15 to 17, Table 18). This suggests that it is possible to further improve the performance of the evaluation model M2 by varying the setting pattern of the explanatory variables for the feature amounts.

[0084] [Examples 4 to 6] In Examples 4 to 6, the acoustic data obtained from each subject in Example 1 was used to evaluate the intestinal condition at different analysis times. Specifically, in Example 4, the acoustic data from 5 minutes before ingestion and 5 minutes after ingestion were used to evaluate the intestinal condition. In Example 5, the acoustic data from 3 minutes before ingestion and 3 minutes after ingestion were used to evaluate the intestinal condition. In Example 6, the acoustic data from 1 minute before ingestion and 1 minute after ingestion were used to evaluate the intestinal condition.

[0085] In Examples 4 to 6, the frequency of bowel sounds (number of BS detected / min) was used as a feature, and bowel sound extraction from acoustic data, machine learning of evaluation model M2, performance verification, and other processes were performed in the same manner as in Example 1. Table 19 shows the evaluation results of the performance (accuracy) of evaluation model M2 in Examples 4 to 6. For comparison, Table 20 also includes the accuracy of evaluation model M2 in Example 1 (Table 2).

[0086] The length (before stimulation / after stimulation) of the acoustic data from which bowel sounds were extracted for analysis was 5 minutes / 5 minutes in Example 4, 3 minutes / 3 minutes in Example 5, 1 minute / 1 minute in Example 6, and 5 minutes / 10 minutes in Example 1. In Table 19, although the tendency varied somewhat depending on the input (explanatory variables), the accuracy in Examples 4 to 6 was generally equivalent to that of Example 1. That is, it was found that the length of each acoustic data from which bowel sounds before and after bowel stimulation were extracted is preferably 1 to 10 minutes, and that equivalent evaluation accuracy could be obtained even with lengths of 1 to 3 minutes. Therefore, it was found that the state of the bowel can be evaluated with high accuracy even with acoustic data obtained from short-term measurements.

[0087] Example 7 In Example 7, the performance of the evaluation model M2 was evaluated when the features extracted from the data for 5 minutes before ingestion and the acoustic data for 5 minutes after ingestion used in Example 4 were combined with the features extracted from the data for 3 minutes before ingestion and the acoustic data for 3 minutes after ingestion used in Example 5. The frequency of bowel sounds was used as the feature. The evaluation results are shown in Table 21.

[0088] The values ​​shown in Table 21 are higher than the values ​​shown in Table 20. Therefore, it is suggested that the performance of the evaluation model M2 can be further improved by combining acoustic data from different times.

[0089] Example 8 In Example 8, the frequency of bowel sounds was used as a feature, and, except for the explanatory variables, bowel sound extraction from acoustic data, machine learning of evaluation model M2, performance verification, and other processes were performed in the same manner as in Example 1. Table 22 shows the evaluation results of the performance of evaluation model M2 in Example 8.

[0090] In Table 22, taking the logarithm of at least one of the explanatory variables tends to increase the accuracy and F1 score. In other words, it was found that taking the logarithm of at least one of the explanatory variables can further improve the performance of the evaluation model M2.

[0091] [Summary] From the above, it has been found that the present invention can accurately evaluate whether a subject's intestines are healthy, taking into account that the diagnosis based on the Rome IV diagnostic criteria may include functional constipation and functional diarrhea without abnormalities in intestinal motility.In addition, intestinal motility (motility) is related to functional constipation and functional diarrhea, suggesting that the present invention is also effective in accurately evaluating intestinal motility.Furthermore, intestinal motility abnormalities are related not only to functional constipation and functional diarrhea, but also to all intestinal diseases, including inflammatory bowel disease (IBD) and irritable bowel syndrome (IBS), suggesting that the present invention is effective in evaluating the presence or absence of all intestinal diseases.

[0092] [Additional Notes] In the above examples, the average value at a predetermined time before or after stimulation was used as the feature quantity. However, aggregated statistics such as the median, mode, variance / standard deviation, interquartile range, percentile, skewness, and kurtosis can also be used. In each graph, the horizontal axis represents the measured value, but it may represent a normalized or standardized value, and the vertical axis may also be normalized or standardized. Furthermore, values ​​may be incremented by +1 or the like before logarithmic transformation. In each graph, the horizontal axis represents the measured value before stimulation, but it may represent the measured value after stimulation. In the above examples, the feature quantities are logarithmically transformed. However, nonlinear transformations such as square root transformation, exponential transformation, Box-Cox transformation, z-standardization, and Min-Max normalization may also be used.

[0093] REFERENCE SIGNS LIST 1 Evaluation system 2 Sound collection device 3 Evaluation device 30 Auxiliary storage device 31 Acoustic data acquisition unit 32 Extraction unit 33 Feature amount calculation unit 331 First feature amount calculation unit 332 Second feature amount calculation unit 34 Evaluation unit M1 Extraction model M2 Evaluation model P Evaluation program

Claims

1. An evaluation device for evaluating the intestinal condition of a subject, comprising: an extraction unit that extracts bowel sounds from acoustic data obtained from the subject; a first feature calculation unit that calculates, from the extracted bowel sounds, a first feature that is a feature of the bowel sounds before stimulation of the intestines; a second feature calculation unit that calculates, from the extracted bowel sounds, a second feature that is a feature of the bowel sounds after stimulation of the intestines; and an evaluation unit that evaluates the intestinal condition of the subject based on the first feature and the second feature, wherein the evaluation unit evaluates the intestinal condition using a machine-learned evaluation model.

2. The evaluation device described in claim 1, wherein the first feature calculation unit calculates a plurality of the first feature amounts, the second feature calculation unit calculates a plurality of the second feature amounts, and the evaluation unit evaluates the state of the intestines based on the plurality of the first feature amounts and the plurality of the second feature amounts.

3. The evaluation device according to claim 1 or 2, wherein the evaluation model extracts bowel sounds from acoustic data obtained from a plurality of subjects, calculates, of the bowel sounds of each subject, a first feature quantity which is a feature quantity of the bowel sounds of the subject before the intestines of the subject are stimulated, and a second feature quantity which is a feature quantity of the bowel sounds of the subject after the intestines of the subject are stimulated, generates training data by labeling the intestinal condition of each subject to either the calculated first feature quantity or its logarithm, and either the second feature quantity, a value obtained by arithmetic operations on the first feature quantity and the second feature quantity, or their logarithms, and performs machine learning on the training data; and the evaluation unit inputs to the evaluation model either the first feature quantity calculated by the first feature quantity calculation unit or its logarithm, and either the second feature quantity calculated by the second feature quantity calculation unit, a value obtained by arithmetic operations on the first feature quantity and the second feature quantity, or their logarithms, and evaluates the condition based on the output of the evaluation model in response to the input.

4. The evaluation device according to claim 3, wherein the training data is generated by labeling the intestinal condition of each subject with the logarithm of the calculated first feature amount and the logarithm of the ratio between the first feature amount and the second feature amount, and the evaluation unit inputs the logarithm of the first feature amount calculated by the first feature amount calculation unit and the logarithm of the ratio between the first feature amount and the second feature amount calculated by the second feature amount calculation unit into the evaluation model.

5. The evaluation device according to claim 1 or 2, wherein the evaluation model extracts bowel sounds from acoustic data obtained from a plurality of subjects with healthy bowel conditions, calculates, of the bowel sounds of each subject, a first feature quantity which is a feature quantity of the bowel sounds of the subject before the intestines of the subject are stimulated, and a second feature quantity which is a feature quantity of the bowel sounds of the subject after the intestines of the subject are stimulated, generates training data by labeling either the calculated first feature quantity or its logarithm with either the second feature quantity, a value obtained by arithmetic operations on the first feature quantity and the second feature quantity, or their logarithms, and performs machine learning on the training data, and the evaluation unit inputs either the first feature quantity or its logarithm calculated by the first feature quantity calculation unit into the evaluation model, and evaluates the condition by comparing the value output by the evaluation model in response to the input with either the second feature quantity calculated by the second feature quantity calculation unit, a value obtained by arithmetic operations on the first feature quantity and the second feature quantity, or their logarithms.

6. The evaluation device according to claim 5, wherein the training data is generated by labeling the intestinal condition of each subject with the logarithm of the calculated first feature amount and the logarithm of the ratio between the first feature amount and the second feature amount, and the evaluation unit inputs the logarithm of the first feature amount calculated by the first feature amount calculation unit to the evaluation model, and compares the value output by the evaluation model in response to this with the logarithm of the ratio between the first feature amount and the second feature amount calculated by the second feature amount calculation unit.

7. The evaluation device according to claim 1 or 2, wherein the evaluation model extracts bowel sounds from acoustic data obtained from a plurality of subjects, calculates, of the bowel sounds of each subject, a first feature quantity which is a feature quantity of the bowel sounds of the subject before the intestines of the subject are stimulated, and a second feature quantity which is a feature quantity of the bowel sounds of the subject after the intestines of the subject are stimulated, generates training data by labeling the state of the intestines of each subject to either the calculated first feature quantity or its logarithm, and any difference between the second feature quantity, a value obtained by arithmetic operations on the first feature quantity and the second feature quantity, and their logarithms, and performs machine learning on the training data, and the evaluation unit inputs to the evaluation model either the first feature quantity calculated by the first feature quantity calculation unit or its logarithm, and any difference between the second feature quantity calculated by the second feature quantity calculation unit, a value obtained by arithmetic operations on the first feature quantity and the second feature quantity, and their logarithms, and evaluates the state based on the output of the evaluation model in response to the input.

8. The evaluation device according to claim 1, wherein the feature amount includes a feature amount related to the time domain of the bowel sounds.

9. The evaluation device according to claim 8, wherein the feature amount includes the frequency of occurrence of the bowel sounds.

10. The evaluation device according to claim 1, wherein the feature amount includes a feature amount related to a frequency domain of the bowel sounds.

11. The evaluation device according to claim 10, wherein the feature amount includes at least one of a peak frequency of the bowel sounds and a spectral bandwidth of the bowel sounds.

12. The evaluation device according to claim 1, wherein the feature amounts include the frequency of occurrence of the bowel sounds, the peak frequency of the bowel sounds, and the spectral bandwidth of the bowel sounds.

13. The evaluation device according to claim 1, wherein the evaluation model is an artificial neural network model.

14. The evaluation device according to claim 1, wherein the length of each acoustic data from which the bowel sounds before and after the bowel stimulation are extracted is 1 to 30 minutes.

15. The evaluation device according to claim 14, wherein the length of each acoustic data piece from which the bowel sounds before and after the bowel stimulation are extracted is 1 to 10 minutes.

16. The evaluation device according to claim 15, wherein the length of each acoustic data piece from which the bowel sounds before and after the bowel stimulation are extracted is 1 to 3 minutes.

17. The evaluation device according to claim 1, wherein the stimulus is a stimulus resulting from the ingestion of a beverage or solid food.

18. An evaluation method for evaluating the intestinal condition of a subject, comprising: an extraction step of extracting bowel sounds from acoustic data obtained from the subject; a first feature calculation step of calculating, from the extracted bowel sounds, a first feature which is a feature of the bowel sounds before stimulation of the intestines; a second feature calculation step of calculating, from the extracted bowel sounds, a second feature which is a feature of the bowel sounds after stimulation of the intestines; and an evaluation step of evaluating the intestinal condition of the subject based on the first feature and the second feature, wherein in the evaluation step, a machine-learned evaluation model is used to evaluate the intestinal condition.

19. An evaluation program for evaluating the intestinal condition of a subject, comprising: an extraction step for extracting bowel sounds from acoustic data obtained from the subject; a first feature calculation step for calculating, from the extracted bowel sounds, a first feature which is a feature of the bowel sounds before stimulation of the intestine; a second feature calculation step for calculating, from the extracted bowel sounds, a second feature which is a feature of the bowel sounds after stimulation of the intestine; and an evaluation step for evaluating the intestinal condition of the subject based on the first feature and the second feature, wherein in the evaluation step, the intestinal condition is evaluated using a machine-learned evaluation model.

Images