JP6299053B2 - Arrayed sound collection sensor device - Google Patents

Description

This invention relates to an array sound receiving sensor unit, about the array voice pickup sensor equipment in particular can be effectively utilized in the shunt stenosis diagnosis support system.

  Many patients with chronic renal failure have undergone artificial blood vessel remodeling procedures called shunts that anastomoses arteries and veins in the forearm. With this shunt, the blood flow circulated from the vein to the outside can be kept stable, and the dialysis efficiency can be increased. The problem with shunting is that blood vessels become constricted or occluded due to shear stress. For this reason, it is necessary to constantly diagnose the state of blood vessels and restore blood flow by percutaneous vasodilatation as necessary.

  An angiography test is generally used as a test method for shunt stenosis, and is a highly reliable test in which the blood flow path is obtained as an image, but it is invasive. In addition, since expensive equipment is required and more than one specialized staff is required, it is not installed unless it is a large hospital. A common non-invasive diagnostic method is to auscultate blood flow sound (shunt sound) generated from the shunt. Shunt sounds may have stenosis, pulsation may be interrupted (intermittent stenosis sound), and high sounds may be mixed (high frequency stenosis sound), and 70 to 80% can be detected by careful auscultation by a skilled person. It is said. However, careful auscultation by skilled workers is necessary, and the number of skilled workers is overwhelmingly smaller than the number of patients who need shunt stenosis diagnosis. Therefore, it is desired to realize an automatic machine that can screen patients who need precise diagnosis by skilled workers in advance.

  So far, when a shunt sound is determined by signal processing, a method of analyzing the sound pressure of the shunt sound (power of the entire sound) (Patent Document 7) and its spectrum information (FFT, LPC, cepstrum) has been proposed. (Patent documents 1, 2, 5, non-patent documents 1, 3). In addition, methods using time-frequency analysis such as wavelet transform have been proposed (Patent Documents 3 and 6, Non-Patent Document 2). However, the method of observing the analysis results requires reading skills. Therefore, in order to quantify the degree of stenosis from these analysis results, a method for calculating the correlation from the stenosis shunt sound waveform (Patent Document 4), and a method for learning sound pressure and spectral peak information as a feature vector using a neural network (Patent Document 2), a method (Patent Document 3) has been proposed in which the wavelet transform result is evaluated from the power of each time-frequency area using an image analysis method, and the correlation between images has been proposed (Patent Document 3), but still not supported There is also sound. The reason is that there are many types of changes associated with stenosis, the timbre varies depending on the degree of stenosis, and there are individual differences even if normal.

JP2011-98090 JP 2010-29434 A JP2009-254678 JP2009-230302 JP-A-2005-328941 JP-A-2005-40518 JP 5-115547

H. A. Mansy et al .: “Computerized analysis of auscultatory sounds associated with vascular patency of haemodialysis access”, Medical & Biological Engineering & Computing, Vol.43, No.1, pp.56-62 (2005) T. Sato et al .: “Evaluation of blood access dysfunction based on a wavelet transform analysis of shunt murmurs”, Journal of Artificial Organs, Vol.9, No.2, pp.97-104 (2006) Nishitani, Yoshi et al .: "Analysis of changes in shunt sound frequency characteristics associated with shunt stenosis", Journal of Dialysis, 43 (3), pp.287-295 (2010)

  On the other hand, there is a problem with sensors that collect shunt sounds. That is, shunt sounds are basically confirmed only on veins where blood flow is caused by shunts. In addition, the stenosis sound is remarkably confirmed in the stenosis part where the shunt sound spreads, and in other places, the result of auscultation and analysis by an expert is often the same as the normal sound. Sensors used to collect shunt sounds have been installed for each channel in principle, and after confirming the shunt sound by human auscultation, the shunt sound can only be obtained by attaching a sensor at that position. Cannot be collected. In particular, in order to detect stenosis, it must be placed on the stenosis. In addition, there are things that are difficult to fix because they do not fit the shape on the skin.

  Furthermore, improvements in signal analysis methods are also necessary. Not only for shunt stenosis diagnosis, but also for medical diagnosis of diseases and abnormal diagnosis of products, the characteristics of the test object are classified into categories based on many acoustic data, acceleration data, X-ray absorption data, etc. A diagnosis for determining the state of the self-feature map is performed, and the self-feature map SOM method is known as an effective means for performing the diagnosis. In order to perform accurate and accurate diagnosis and determination, the data itself tends to increase as the number of data increases. From such a background, there has been a problem that an increase in data amount causes an increase in processing time.

The invention is particularly, assuming the existence of a variety of shunt stenosis sounds, to identify these shunt stenosis sound is effective in Resid Yanto stenosis diagnosis support system to support the screening of shunt stenosis, but a skilled person Both provide an arrayed sound pickup sensor device that can pick up sound easily.

First, the shunt stenosis diagnosis support system will be described.

The first sheet Yanto stenosis diagnosis support system, shunt sound input means for generating at least pulsation component of the test shunt sound waveform data of a shunt sound and sound receiving from the subject, the test sample produced in the shunt sound input means A self-organizing map is created using at least the time-frequency characteristics of the index sample among the shunt sound-form data and the time-frequency characteristics of the shunt sound-form data of a plurality of types of index samples as evaluation indexes, and the created self-organization map is created. A first analysis means for analyzing the position of the test sample relative to the position of the index sample on the quantization map, and a plurality of types of models based on the time frequency characteristics of the shunt sound waveform data of the plurality of types of index samples, Analyzing the probability of belonging to each of the above models of the time-frequency characteristics of shunt sound waveform data Those comprising analyzing means.

  It is preferable to create test shunt sound waveform data for a plurality of beats of the sound signal obtained by the subject's sound collection.

  The index sample by the first analysis means preferably includes a normal sound sample and a stenosis sound sample, and the stenosis sound sample includes a plurality of (various) different stenosis sound samples. Is preferred. The distinction between normal and stenotic sounds and the type of stenotic sound depends on the judgment of an expert. Each sample is preferably for a plurality of persons, and it is preferable to create shunt sound sample waveform data for a plurality of beats from a sound signal obtained from one person.

  The first analysis means may create a self-organizing map by mixing the data of the test sample and the data of the index sample, or on the self-organizing map created by the data of the index sample. It may indicate the position of the occupied test sample.

  Preferably, the apparatus further includes a first determination unit that determines a possibility of shunt stenosis based on an analysis result by the first analysis unit. The apparatus further includes second determination means for determining the possibility of shunt stenosis based on the analysis result by the second analysis means.

  Further, a first determination means for determining the possibility of shunt stenosis based on the analysis result by the first analysis means, and a possibility of the shunt stenosis based on the analysis result by the second analysis means. It is preferable to further include a second determination unit that performs the determination, and an integrated determination unit that performs an integrated determination based on the determination of the first determination unit and the determination of the second determination unit.

  In these cases, the analysis result by the first analysis unit, the analysis result by the second analysis unit, the determination result by the first determination unit, the determination result by the second determination unit, or the integrated determination unit If an output means for outputting the integrated judgment result is provided, the analysis result and judgment result can be known.

  Based on the analysis result by the first analysis means, the person may determine the presence or degree of the shunt stenosis, and similarly, even if the person determines the shunt stenosis based on the analysis result by the second analysis means. Good. A person may make the integrated determination.

As the shunt sound collection means, the arrayed sound collection sensor device according to the present invention may be used. Details of the arrayed sound pickup sensor device according to the present invention will be described later . The shunt sound input means generates test shunt sound waveform data based on the output signals from the microphones of the arrayed sound collection sensor device . The first analysis means and the second analysis means use test shunt sound waveform data created based on the output signals of the microphones for analysis.

According to the first shunt stenosis diagnosis support system , since two systems of analysis processing of different analysis methods of the first analysis means and the second analysis means are performed, it is possible to detect the possibility of stenosis with high accuracy. It becomes. In the first analysis means, it becomes possible to know a stenosis sound having an intermediate characteristic among a variety of stenosis sound index samples.

  In one embodiment, the first analysis means may use a self-organizing map (SOM) and make a visual diagnosis on the resulting SOM (referred to as a feature map). Is possible. The SOM is a neural network that outputs a feature map in which similar feature data are collected in the vicinity by unsupervised competitive neighborhood learning. A large number of clinical samples are used as diagnostic index samples, and learning is performed by the SOM including the test samples. Alternatively, the test sample is input after learning only the index sample. The resulting feature map can visually determine what index sample the test sample is similar to. Therefore, a support system that can be expected to cope with various shunt sounds derived from individual differences and the shape and degree of stenosis is realized. Of course, the support system may make a decision.

  The feature vector input to the SOM is a signal analyzed from a shunt sound so that the normal and narrow features of the SOM can be easily grasped. In one embodiment, the STMEM method (Short-time Maximum Entropy) which is a time-frequency analysis result is used. Method). Time-frequency analysis has been empirically and experimentally shown for shunt sound analysis, but it has been confirmed that the wavelet transform is undesirable because of the large amount of information other than differences due to normality and stenosis. On the other hand, the STMEM method can obtain a spectral envelope every short time. The order for obtaining the optimum spectral envelope can be made a condition in which normal and stenotic information is not lost and excessive information does not enter by analyzing many shunt sounds.

  As a preferred embodiment of the second analysis means, a mel cepstrum method (MFCC: Mel Frequency) used as a speech recognition technique in which the time change of the frequency spectrum and the sound pressure is an important feature as well as the feature of the shunt sound. Machine learning using Cepstrum Coefficient (Hidden Markov Model) and Hidden Markov Model (HMM) can be used. As a result of the experimental examination, it is shown that it is effective to improve the detection of stenosis by using both the diagnosis by SOM and the diagnosis by HMM. In other words, through another approach regarding the performance of SOM stenosis diagnosis support, the evaluation can be performed and the accuracy can be improved.

Oite above support system preferably includes means for the first analysis unit to create a sequential subdivision self-organizing map (the details will be described later).

The second sheet Yanto stenosis diagnosis support system is characterized by using a sequential subdivision self-organizing map creating process. That is, the second sheet Yanto stenosis diagnosis support system, generated by the shunt sound input means, the shunt sound input means for generating at least pulsation component of the test shunt sound waveform data of a shunt sound and voice pickup from the subject test Self-organization by sequential subdivision self-organizing map creation processing using at least the time-frequency characteristics of the index sample among the shunt sound waveform data of the sample and the time-frequency characteristics of the shunt sound waveform data of multiple types of index samples as evaluation indices An analysis means for creating an integrated map and analyzing the position of the test sample with respect to the position of the index sample on the created self-organizing map, and an output means for outputting an analysis result by the analysis means.

  In one embodiment, the analyzing means includes means for creating category data in which the index sample data is hierarchized from the learning data as learning data, and the creating means creates learning data at the stage of creating the first hierarchy. Create multiple representative data to classify and assign the learning data to the representative data, and classify the learning data belonging to each representative data in the previous hierarchy at the stage of creating the next and subsequent layers Creation of new representative data, processing to assign learning data belonging to each representative data of the previous hierarchy to the new representative data, and creation of attribution information indicating which new representative data was created from each representative data of the previous hierarchy The processing to be performed is sequentially repeated up to the designated upper limit number of times.

  The sequential subdivision self-organizing map creation method is a method that can create a map at high speed while maintaining the global structure of the initial arrangement by performing principal component analysis while gradually subdividing the map. Since this method is a method of creating initial values by principal component analysis, the global rearrangement of learning data is only data belonging to cells including boundary subdivision points. Therefore, a map can be created while maintaining the global structure of the initial arrangement obtained by principal component analysis. Furthermore, when a map is created again using the same learning data, only about 1% of points move to adjacent cells, and local reproducibility is also high.

Also in the second support system , it is preferable to create test shunt sound waveform data for a plurality of beats of a sound signal obtained by the subject's sound collection.

  The index sample in the analyzing means preferably includes a normal sound sample and a stenosis sample, and the stenosis sample preferably includes a plurality of (various) different stenosis samples. . The distinction between normal and stenotic sounds and the type of stenotic sound depends on the judgment of an expert. Each sample is preferably for a plurality of persons, and it is preferable to create shunt sound sample waveform data for a plurality of beats from a sound signal obtained from one person.

  The analysis means may be a method for creating a self-organizing map by mixing test sample data and index sample data, or on the self-organizing map created by the index sample data. You may superimpose the map by data.

  Preferably, the information processing apparatus further includes a determination unit that determines the possibility of shunt stenosis based on the analysis result of the analysis unit.

  The analysis result output means includes a display device, a printer, and a communication device that transmits the analysis result to another device (such as a server). A person may determine the presence or absence and the degree of shunt stenosis based on the analysis result by the analysis means output to the output means.

As a shunt sound collection means, the arrayed sound collection sensor device according to the present invention can be used . The upper SL shunt sound input means for generating a test shunt sound waveform data based on the output signals from each microphone of the array voice pickup sensor device. The analysis means uses test shunt sound waveform data created based on the output signals of the microphones for analysis.

In the arrayed sound collection sensor device according to the present invention , a plurality of cavities are formed in a sheet-like soft support, and a microphone is provided at the bottom of each of these spaces.

  In a preferred embodiment, the cavities are formed in a plurality of columns and a plurality of rows. However, the cavities need not be regularly arranged in the vertical and horizontal directions, and each column does not have to have the same length. As a result, an arrangement that looks random may be used. The number of microphones should be large, but at least two microphones (cavities) are sufficient. The above number and arrangement of microphones are collectively referred to as an array.

  In this array-type sound pickup sensor device, the sheet-like soft support is directed so that the cavity opening faces the skin of the subject's forearm, and the forearm shunt portion of the subject and its front, rear, left and right Place it so that it covers the nearby skin, and fix it by some means. For example, a rubber or cloth presser band (holding band) provided with a hook-and-loop fastener may be wound on the subject's arm from above the support and fixed with the hook-and-loop fastener. Such a rubber band or cloth band may be fixed to the support.

  The arrayed sound collection sensor device is provided with a plurality of microphones, and it can be expected that a sound signal that appropriately represents the constriction state of the subject's shunt is collected from any one of the microphones. Even if it does not necessarily perform auscultation, it can be easily put on the subject and sampled. In addition, since the sheet-like soft support is used, it can be deformed according to the unevenness of the skin due to the running of blood vessels in the forearm of the subject and can be brought into close contact without disturbing the blood flow. That is, it is possible to minimize the pressure on the blood vessels of the subject. The space in the cavity where the microphone is arranged is sealed without gaps by the close contact of the soft support, and a high S / N ratio of the sound signal is obtained. If a large number of microphones are arranged two-dimensionally, it is possible to examine many stenosis locations over a wide range, and to investigate the spread of shunt sounds, the running of veins, and the like. If the presser band as described above is used, the microphone does not come off during sound collection.

The program for executing the above-described sequential subdivision self-organizing map creation process is a program for delivering a plurality of learning data to a computer and creating category data hierarchized from the learning data. In the step of obtaining data, at the stage of creating the first hierarchy, creating multiple representative data that classifies the learning data, assigning the learning data to the representative data, and creating the next and subsequent hierarchies In this stage, the creation of new representative data that classifies the learning data belonging to each representative data of the previous hierarchy, the process of assigning the learning data belonging to each representative data of the previous hierarchy to the new representative data, and the previous hierarchy The process of creating attribution information indicating which new representative data has been created from each representative data of is repeated sequentially up to the specified maximum number of times. It is characterized by, category data output is characterized in that it contains representative data of the learning data and the next hierarchy belonging to the representative data.

  In one embodiment, a program for handing over input data to the computer and determining to which category data the input data belongs is determined by the program, the program handing over the learning data to the program. It is characterized by having a procedure for creating category data, a procedure for judging to which category the input data belongs, and a procedure for transmitting the judgment result to the display unit.

  In another embodiment, the program is operated by passing the input data and the reference data group to the computer, the procedure for creating the learning data by integrating the input data and the reference data, and the passing of the learning data to the program It is characterized by having a procedure for creating data and a procedure for transmitting to the display section as a judgment result which category the input data belongs to.

Exiled following subdivision SOM manufacturing method, passing a plurality of learning data to the computer, a method for creating a category data layered from the learning data, the procedure to obtain the category data hierarchy of the can, At the stage of creating the first hierarchy, creating multiple representative data that classifies the learning data, and assigning the learning data to the representative data, and at the stage of creating the next hierarchy, Create new representative data that categorizes the learning data belonging to each representative data, process to assign the learning data belonging to each representative data in the previous hierarchy to the new representative data, and which new data from each representative data in the previous hierarchy It is characterized by the process of creating attribution information that indicates whether representative data has been created. The category data that is output includes the representative data, the learning data belonging to it, and the next level. And it is characterized in that it contains representative data.

Exiled following subdivision SOM generating apparatus is an apparatus for creating a category data hierarchized from a given training data, it means for obtaining a category data hierarchy of the can, creating a first level stage In the step of creating a plurality of representative data that classifies the learning data and assigning the learning data to the representative data, the learning data belonging to each representative data of the previous hierarchy at the stage of creating the next and subsequent layers Create new representative data to classify the data, process to assign the learning data belonging to each representative data in the previous hierarchy to the new representative data, and which new representative data was created from each representative data in the previous hierarchy The process of creating attribution information is iteratively repeated up to the specified maximum number of times, and the category data to be output is representative data, the learning data belonging to it, That it contains a table data is characterized in.

According to the sequential subdivision self-organizing map creation program, method or apparatus , in order to solve the problem of increase in processing time, the number of categories to be grouped is gradually increased and the number of data belonging to the categories is gradually decreased. Processing time can be shortened by realizing a typical process.

It is a block diagram which shows the whole structure of a shunt stenosis diagnosis support system. It is a perspective view which shows an array-like sound collection sensor apparatus. It is an expanded sectional view which follows the III-III line of FIG. The use state of a sensor apparatus is shown. It is a flowchart which shows the flow of the whole process. It is a wave form diagram which shows the mode of signal processing. The classification result of shunt sound based on a MEM spectrum is shown. The classification result of shunt sound based on a MEM spectrum is shown. The classification result of shunt sound based on a MEM spectrum is shown. The classification result of shunt sound based on a MEM spectrum is shown. The classification result of shunt sound based on a MEM spectrum is shown. An example of an SOM created based on an index sample is shown. Another example of the SOM created based on the index sample is shown. It is an example which shows the position of the test sample on SOM created based on the parameter | index sample. The characteristics of the test sample on the SOM of the index sample are shown. It is a flowchart which shows the procedure of sequential subdivision SOM. It is a flowchart which shows the process which determines attribution of input data, after creating the category data based on learning data. It is a flowchart which shows the procedure which unifies reference data (learning data) and input data, and produces a category. 13 is a flowchart showing details of the process of S43 of FIG. 10, S53 of FIG. 11, or S73 of FIG. 14 is a flowchart showing another procedure of the flow of FIG. 15 is a flowchart showing details of the process of S82 of FIG. 13 and S92 or S94 of FIG. 12 is a flowchart showing details of the process in S62 of FIG. The data structure of memory space is shown. It shows how the grouping gradually becomes more detailed.

1 Overall Configuration of Shunt Stenosis Diagnosis Support System FIG. 1 shows the overall configuration of a shunt stenosis diagnosis support system.

  The sensor unit 1 collects a shunt sound from a subject and includes a plurality of microphones as will be described later, and converts the sound signal into an electrical signal. The electric signal output from the microphone is amplified by the amplifier 2. The amplifier 2 is provided for each of the plurality of microphones. Specific configurations of the sensor unit 1 and the amplifier 2 will be described in detail later.

  The electrical signal representing the sampled shunt sound sampled is then converted into a digital signal by the A / D converter 3. One channel is assigned to each microphone. The A / D converter 3 encodes the test shunt sound signal of all channels into a digital signal and inputs it to the signal processing device 4 as a serial signal. The A / D converter 3 starts sampling of the test shunt sound (A / D conversion) in response to the trigger signal from the operation unit 11.

  Specifically, the signal processing device 4 is realized by a computer (computer system). The signal processing device 4 includes a subject data storage unit 7, a reference data group storage unit 8, and an arithmetic processing unit 10, and a program for executing processing shown in flowcharts (FIGS. 5 and 10 to 16) described later. Stored. If the arithmetic processing unit 10 is largely functionally divided, a feature extraction unit 5 and an analysis determination unit 6 are included. The arithmetic processing unit 10 performs two types of feature extraction / analysis processing according to different methods. The first feature extraction / analysis process uses a short-time maximum entropy method (STMEM) result of the shunt sound as a feature vector, and learns by using a self-organizing map (SOM). Is determined or diagnosed by the SOM obtained as follows (referred to as a feature map). In the second feature extraction / analysis process, a mel cepstrum method (MFCC) and a hidden Markov model (HMM), which are used as a speech recognition technique in which the time change of the frequency spectrum and the sound pressure is an important feature, are used. Judgment or diagnosis is performed using machine learning by Hidden Markov Model). The feature extraction unit 5 executes time-frequency feature extraction by the STMEM method or the MFCC method, and the analysis determination unit 6 executes analysis by the SOM method or the HMM method. In the SOM method, the sequential subdivided SOM method is used. Details of the feature extraction and analysis determination will be described later. The determination or diagnosis in the analysis determination unit 6 includes display of an analysis result for diagnosis by a person and automatic determination as described later. Comprehensive determination or diagnosis based on the first and second analysis determination results is also possible. Judgment or diagnosis reveals normality, stenosis, individual characteristics, degree of stenosis, and the like.

  The digital signal of the shunt sound output from the A / D converter 3 is stored in the subject data storage unit 7. The data can also be incorporated into a reference data group in the reference data group storage unit 8. By storing the subject information (name, etc.) together with these data, it can be left as a history of the subject, and it is also possible to make a judgment incorporating the characteristics of the individual subject.

  The reference data group is a data group that serves as an index when determining the test shunt sound in the determination algorithm in the analysis determination unit 6. The storage unit 8 stores digital data of shunt sounds labeled according to various characteristics (normal, stenosis, stenosis type, degree of stenosis) examined in detail. Like the test shunt sound, it is converted into a feature vector and used in the decision algorithm. It is also possible to prepare it as feature vectorized data for speeding up. This reference data group can also be increased by labeling and incorporating various features into the stored test shunt sound. In addition, by labeling the information of the individual subject, it is also possible to make a determination incorporating the characteristics of the individual subject. Also, the reference data group can be shared between other devices via a network group or the like.

  The result determined by the analysis determination unit 6 is displayed on the display unit 9. The display can be a display or paper. Simultaneously with the determination result, the test shunt sound waveform, frequency analysis result, time frequency result, and past diagnosis data can be displayed. These displays can also be observed remotely via a network group. In addition to the display unit 9, the output device includes a printer, a transmission device that transmits to other devices, and the like.

  The operation unit 11 is a part for inputting various settings (personal information input, determination condition setting, display condition setting), and a diagnosis start command. Input is via a button or touch panel display. Various setting commands are passed to the signal processing device 4. The start command is passed to the A / D converter 3 to start collecting the test shunt sound, and is passed to the signal processing device 4 to start signal processing.

2 Arrayed Sound Collection Sensor Device Specific examples of the sensor unit 1 and the amplifier 2 described above are shown in FIGS.

  As shown in FIGS. 2 and 3, the arrayed sound collection sensor device 20 has a sheet-like soft support 21, and a plurality of cavities (holes) 22 are formed on one surface of the support 21. The cavity 22 has an inner bottom surface, and a small microphone (for example, a silicon microphone) 23 is fixed to the inner bottom surface. As will be described later, the support 21 is disposed so that the opening of the cavity 22 is in contact with the forearm 27 of the subject. At this time, the depth of the cavity is determined so that a space is maintained in the cavity 22 between the microphone 23 and the skin of the forearm of the subject. The shape of the cavity 22 may be cylindrical or rectangular. The microphone 23 is preferably smaller than the size of the cavity 22 so as to be in contact with the inner bottom surface of the cavity 22 (not necessarily in contact) and not in contact with the inner (circumferential) surface of the cavity.

  As an example, a hole that becomes a cavity 22 is formed in a soft resin (for example, a soft urethane molding resin with high adhesiveness) through the thickness direction, and a rubber sheet (rubber plate) is bonded to one surface of the soft resin 21a. Thus, one opening of the cavity 22 is closed to form the inner bottom surface.

  An amplifier circuit IC24 is fixed to a position corresponding to each microphone 23 on the surface opposite to the cavity 22 of the support 21 (rubber sheet 21b), and is electrically connected to the corresponding microphone 23. The output side of the IC 24 is connected to the A / D converter 3 (both wires and cords for electrical connection are not shown).

  As shown in FIG. 4, such an array-like sound collection sensor device 20 has a sheet-like soft support 21 facing the skin of the subject's forearm so that the opening of the cavity 22 faces the subject's forearm. Place the shunt on the front, back, front and back of the skin, and fix it by some means. For example, a rubber or cloth presser band (holding band) 25 provided with hook-and-loop fasteners 26a and 26b may be wound on the subject's arm from above the support 21 and fixed with the hook-and-loop fasteners 26a and 26b.

  In the sheet-like soft support 21, cavities 22 and microphones 23 are preferably arranged in a plurality of columns and a plurality of rows as shown in the figure. The size of the support body 21 should cover half or all of the subject's forearm from the wrist to the elbow, and should cover about half to 2/3 in the circumferential direction of the forearm. A larger number of microphones (cavities) is preferable. As shown in the figure, the microphones (cavities) do not need to be regularly arranged in the vertical and horizontal directions, and each row does not need to have the same length. As a result, an arrangement that looks random may be used. Further, it is sufficient that there are at least two microphones (cavities). The number and arrangement of microphones as described above are collectively referred to as an array in this specification.

  As described above, the arrayed sound collection sensor is provided with a plurality of microphones, and it can be expected that a sound signal appropriately representing the constriction state of the subject's shunt is collected from any one of the microphones. Even if a person does not necessarily perform prior auscultation, it can be attached to the subject and easily picked up. In addition, since the sheet-like soft support is used, it can be deformed according to the unevenness of the skin due to the running of blood vessels in the forearm of the subject and can be brought into close contact without disturbing the blood flow. That is, it is possible to minimize the pressure on the blood vessels of the subject. The space in the cavity where the microphone is arranged is sealed without gaps by the close contact of the soft support, and a high S / N ratio of the sound signal is obtained. If a large number of microphones are arranged two-dimensionally, it is possible to examine many stenosis locations over a wide range, and to investigate the spread of shunt sounds, the running of veins, and the like. If the presser band as described above is used, the microphone does not come off during sound collection.

3 Overall Process Flow The overall process flow in the signal processing device 4 will be described with reference to FIGS. 5 and 6.

  As described above, the amplifier 2 and the A / D converter 3 are obtained from the microphones 23 of the arrayed sound collection sensor device 20, and the shunt sound waveform is input (S1). The number of shunt sound waveforms equal to the number of microphones 23 included in the sensor device 20 is obtained.

  Next, a signal waveform for one beat of each shunt sound is extracted (S2). The envelope is obtained by extracting only the low-frequency component of the half-wave of the time series waveform of the shunt sound, and the time between the maximum and minimum points is extracted as the starting point of pulsation. The signal to be cut out may be equal to or less than one beat, for example, about 3/4. A plurality (for example, 10) of such signal waveforms are extracted per channel (one microphone).

Calculating the STMEM vector y _p for each beat (S11) (first feature extraction process). P feature vectors are obtained per beat.

  The test sample thus obtained and the index sample prepared in advance are simultaneously learned by (sequential subdivision) SOM (S12) (first analysis determination process). The test sample may be passed through the learning result of the index sample. Here, the test sample is p STMEM vectors per beat. Since the SOM outputs a map in which data of similar characteristics are collected in the vicinity, it is possible to visually determine to which index sample group the diagnostic object belongs, or to obtain a determination result by calculation ( S13).

  On the other hand, a feature vector is obtained by MFCC for each beat (S21) (second feature extraction processing), and the HMM is diagnosed (S23) (second analysis determination processing).

  Both determination results can be integrated and determined (S30).

4. Index Sample An index sample is required in the SOM process (S12) described above. Also in the HMM process (S22), an index sample for model learning (supervised learning) is required in advance.

  Classify normal sounds and stenosis sounds (intermittent stenosis sounds and high-frequency stenosis sounds) by auditioning shunt sounds with specialist doctors, and further investigate the spectrogram and the shape of the MEM spectrum for one beat, and classify them into five types. The results obtained are shown in FIGS. 7a to 7e (in these figures, PSD = Power Spectrum Density (spectrum intensity), Frequency (frequency), Time (time), Spectrogram). It is preferable that there are many index samples. For each of the following five types, those collected from at least a plurality of people (equivalent number) are used.

  Five characteristic MEM spectrum shapes and spectrogram characteristics of normal sound, high-frequency stenosis sound (3 types A, B, and C), and intermittent stenosis sound are as follows.

(a) High-frequency constriction sound A (high-pitch A): A spectral component of 500 Hz or higher is not attenuated as much as normal sound, and there is a frequency component with a gentle slope.
(b) High-frequency stenosis sound B (high-pitch B): A peak is locally confirmed in the spectrum. In the spectrogram, stripes appearing in the time axis direction appear. A sound like a whistle is heard.
(c) High-frequency stenosis sound C (high-pitch C): A spectrum shape that can be judged to be clearly different from the normal sound, which is different from both the high-frequency stenosis sounds A and B, and is designated C for convenience.
(d) Intermittent: After pulsation, the sound pressure is significantly attenuated and cannot be heard as a continuous blood flow sound.
(e) Normal sound (normal): Frequency components concentrate in a low band of 500 Hz or less, and gradually attenuate as the frequency increases. In the spectrogram, the low frequency components are not interrupted.

5 Processing details and features
(1) Extraction of pulsation As described above, an envelope is obtained by applying half-end rectification of the shunt sound to a low-pass filter, and the time between the maximum and minimum points of the envelope is set as the start point of the pulsation.

(2) STMEM method After extracting one beat by the STMEM method, a feature vector that can be easily determined by the determination processing unit (SOM) can be obtained by obtaining a spectral envelope every short time. That is, a feature vector is obtained in which information that is redundant in obtaining normal / stenosis is omitted as much as possible.

(3) SOM
SOM is unsupervised learning, which is a technique for comprehensively searching for features from many test samples, and data having similar acoustic features is visualized in a form gathered in the vicinity. That is, it can be judged how similar to the shunt sound as an index. This facilitates individual calibration. According to this method, it is possible to deal with various types such as individual differences of shunt sounds and the degree of stenosis, and it helps doctor's judgment.

More specifically, SOM is a two-layer neural network composed of an input layer and an output layer (competitive layer) using an unsupervised competitive neighborhood learning algorithm. The output layer is usually set in two dimensions or three dimensions for visualization, and the nodes (cells) are arranged in a grid. When the input is an n-dimensional vector, each node on the output layer has an n-dimensional reference vector. Given an input vector y _{p, the} node with the reference vector that is most similar to y _p , that is, with the smallest Euclidean distance, becomes the winning node and becomes the output of y _p . SOM learning is to change the reference vector, the reference vector of the winning node approached to y _p, closer to also y _p reference vectors neighboring nodes. Thus, by learning a set of high-dimensional input vectors, an output map is obtained in which similar sample groups are gathered at the same position (node) or neighborhood on the map.

  As an example, the condition of the feature vector STMEM input to the SOM is created by dividing one beat into five frames from the beat starting point with 0.15 s as one frame and obtaining a MEM spectrum for each frame. The number of output points of the intensity of the MEM spectrum was 200 points at equal frequency intervals in the frequency band of 0 to 2 kHz. Therefore, the feature vector has 1000 elements, and the number of input layer nodes to SOM is 1000.

  Mix index sample and test sample to learn SOM. At this time, the test sample is P vectors corresponding to P beats, and is treated as independent data in the calculation of SOM. Only the index sample may be learned, and then the test sample may be input.

  FIG. 8a shows an example of SOM results obtained using the five types of index samples described above.

  Further, FIG. 9 shows the result of SOM in which samples of shunt sounds (diagnostic shunt sounds) obtained from a subject are added to five types of index samples.

  In FIGS. 8a and 9 (the same applies to FIGS. 8b and 8c), these are a large number of hexagonal lattice maps, each hexagon corresponding to an output node and 1 for input vector 1 data (1 beat). One node is selected and output. Since the input vector is mapped to the feature space, there is no concept of the vertical and horizontal axes, data with a similar input vector is output to nearby output nodes, and the more distant output nodes the input vector Dissimilar data is output. All outputs are displayed at the same time. In FIG. 8a, normal sound, high-frequency stenosis sound ABC, and intermittent stenosis sound are shown by different gray levels, and in FIG. 9, hatching direction and gray and black are shown. . In addition, in the node from which data is output, the number is represented by the number of symbols such as circle, triangle, and square in FIG. 8a, and hatched hexagon, gray, and black hexagons in FIG. It is expressed in size.

  An area where normal sound data is output (area surrounded by a black line in FIG. 9) is formed, and it can be seen that there is no stenosis sound data in that area. It can also be seen that even in the case of stenosis sound data, each of the high-frequency stenosis sounds A, B, C, and the intermittent stenosis sound data roughly forms a region.

  In FIG. 9, the diagnostic shunt sound (test shunt sound) is represented within the normal sound range (near the normal sound) (black hexagon), and it can be seen that the shunt sound of this subject is normal. .

  By displaying the learning result as shown in FIG. 9 on the display unit 9, it is possible to support the diagnosis of the shunt sound of the subject. In other words, it can be determined which of the five types of index samples the subject's shunt sound is similar to according to which of the five types of index samples is represented on the SOM.

(4) Sequential subdivision SOM
Sequential subdivision SOM is a self-organization technique that combines principal component analysis and SOM to subdivide learning data (shunt sounds). Since the conventional SOM performs the initial arrangement of data at random, even if the same learning data is used, the arrangement of the map obtained each time it is created (learned) changes. According to the sequential subdivision SOM, the principal component analysis is performed first, the data is arranged, and the map obtained by subdividing by using the principal component analysis again is created while maintaining the global structure of the initial arrangement. And reproducibility is high. (That is, if the same learning data is used, it is very likely that the map changes will be similar. However, the same map is not necessarily obtained, and the characteristics of the SOM algorithm are not lost. (It is expressed as “global”.) With the same learning data, it has been experimentally obtained that only about 1% of data is arranged in adjacent cells.

  In addition, it is fast and the experimental results of learning 511 7774-dimensional vector data show a 42% reduction compared to the conventional SOM.

  The processing of the sequential subdivision SOM will be described in detail later.

(5) Quantification of stenosis level by SOM By calculating the Euclidean distance from the “cell where normal shunt sound data is placed” closest to the “cell (node) where collected diagnostic shunt sound data is placed” The stenosis level can be determined quantitatively.

In the SOM as shown in FIGS. 8a, 8b, and 8c, a cell with a white background is a normal sound. A gray (gray level) cell includes learning data extracted from a narrowed blood vessel. The gray density of the cell is divided into a plurality of levels (for example, 5 levels in FIG. 8a and 32 levels in FIGS. 8b and 8c) so that the gray level increases as the distance from the white cell increases. The density level _{d c (= 0,1,2, ‥‥,} 32) (Figure 8b, the case of FIG. 8c) is defined as follows. The model vector assigned to each cell is M _c (c = 1, 2,..., C), and the function dis (M _c ) is between the white cell containing the normal sound closest to the c-th cell. Of Euclidean distance. That is,
And Here, W _g is a set created by an index of a cell with a white background (that is, a cell containing only normal sound) in the g-th generation map. If the cell c contains only normal sound, dis (M _c ) = 0. The value of the function dis (M _c ) can be used as a parameter indicating the difference between the shunt sound and the normal sound mapped to the c th cell. However, since its value changes depending on the dimension of the input vector, etc., it is converted into a normalized quantity d _c as follows:
Here, [x] represents the maximum integer not exceeding the real number x, and H _g is a set of all indices of the cells included in the g-th generation map. The constant u is an empirical parameter and is determined so that the gray background color clearly surrounds the cell containing the normal sound. As an example, u = 3. The density level d _c represents the difference from the normal sound by a non-linear function of M _c . Since the value of d _c is increased by including a signal shunt sound is due to stenosis, call this parameter the stenosis level.

  By calculating the stenosis level expressed by equation (2) and discriminating it at a predetermined threshold level (may be in multiple stages), it is possible to automatically determine whether or not there is a suspicion of stenosis. . This is an example of the determination process in S13 of FIG.

(6) Diagnosis and determination by HMM The MFCC coefficient represents the spectral envelope in the analysis frame. By calculating ΔMFCC, ΔΔMFCC, logarithmic power, Δlogarithmic power, and ΔΔ logarithmic power, which will be described later, these are time-frequency analysis (Processing in FIG. 5 S21). These are input vectors to the HMM.

  As an example, the MFCC derivation condition was set to a width of 20 ms using a frame with a Hanning window and a frame shift interval of 10 ms. The dimension of the MFCC coefficient is calculated as 24, and the lower 12 dimensions are used for diagnosis. Therefore, MFCC (12) is obtained for each frame, and a total of 39-dimensional vectors of ΔMFCC (12), ΔΔMFCC (12), logarithmic power (1), Δlogarithmic power (1), and ΔΔlogarithmic power (1) were obtained. . Since 67 frames per beat (0.67 s), the HMM learns 2613 feature vectors.

  The HMM uses a “left-to-right” type model and uses the Baum-Welch algorithm as a learning algorithm. The number of HMM states was set to 8. Output of 10 data (10 beats) per shunt sound can be obtained, but if more than 3 data are output as stenosis sound, the shunt sound is diagnosed as stenosis, otherwise it is diagnosed as normal . It can also be determined using a plurality of different threshold values. In any case, this determination is the determination in FIG.

(7) Mel Cepstrum (MFCC)
The cepstrum is defined as the inverse Fourier transform of the logarithm of the short-time Fourier spectrum of the waveform, and has the feature that the spectral envelope and the fine structure can be separated approximately. MFCC is a cepstrum analysis that incorporates a model of the human auditory mechanism.

  The human auditory sense has a non-logarithmic non-linear characteristic called mel scale with respect to the pitch of the sound, with fine frequency resolution at low frequencies and coarse frequency resolution at high frequencies. It is obtained by combining a filter bank that takes into account human auditory characteristics in the cepstrum derivation process. Mel scale is defined by equation (1) for frequency f [Hz].

Since the shunt sound signal is an unsteady signal whose signal characteristics change with time, in the mel cepstrum analysis, the signal is cut out in short time units (frames), and a Hanning window is used for this framing process. Subsequently, fast Fourier transform is performed on each frame, and this amplitude spectrum is applied to a 24-dimensional filter bank equally spaced on the mel scale to extract spectral components m _j of each band. The output from the filter bank is logarithmically transformed and further subjected to discrete cosine transformation to be converted into a cepstrum of dimension N = 24. The MFCC coefficient C _i obtained from this is defined by equation (4). However, since the 12th dimension of the low-order component of the MFCC coefficient is used here, i is calculated up to 12 in equations (4) to (7).

The MFCC coefficient represents the spectral envelope in an analysis frame. In general, in addition to the parameters used for feature vectors, dynamic features corresponding to temporal changes in the spectral envelope are used. The dynamic feature is defined by equation (5). The i-th value of the MFCC coefficient in an analysis frame n is C _i (n), and the slope when a straight line is applied to the value of C _i (n) in the interval [n−δ, n + δ] centered on n is ΔC _i (n). Here, δ = 2 is set, and calculation is performed for five frames including two frames before and after a certain analysis frame n. The calculation is performed assuming that the −1st and −2th frames are the same as the 0th frame for the start (0th) frame.

The end is also handled in the same way. ΔC _i (n) represents a temporal change amount of C _i (n). For the MFCC coefficient C _i , the parameter obtained by equation (5) is called ΔMFCC. Furthermore, for the ΔMFCC coefficient, the parameter obtained by equation (6) is called the ΔΔMFCC coefficient.

  Further, the logarithmic power of each frame is obtained as E for each frame from Equation (7), the parameter ΔE obtained from Equation (8) is called Δlogarithmic power, and the parameter ΔΔE obtained from Equation (9) is called ΔΔlogarithmic power. . Here, M = 12 in equation (7). For each frame, MFCC, ΔMFCC, ΔΔMFCC, logarithmic power, Δlogarithmic power, and ΔΔlogarithmic power are collectively used as a feature vector.

(8) Overall Determination Although the overall determination in FIG. 5S30 can be performed by a human, the signal processing device 4 can also perform it automatically.

  As described above, the determination in FIG. 5S13 can be performed by a human by looking at the display as shown in FIG. 9, or by comparing the stenosis level obtained by equation (2) with one or more threshold values. The computer can determine the presence / absence and degree. In the determination of FIG. 5S23, the computer can determine whether or not there is a stenosis.

  In S30 of FIG. 5, the determination is made using the determination results of S13 and S23. For example, the integrated determination result may be obtained by OR logic of the determination of S13 and the determination of S23 (whether there is stenosis). When the determinations at S13 and S23 indicate the degree of stenosis, the higher degree of stenosis can be used as the overall determination result.

  In this way, by performing overall judgment based on the analysis and judgment of the two systems, a highly reliable support system can be realized.

6 Sequential subdivision SOM
The procedure of the sequential grouping process for grouping the learning data into categories and the determination process for determining the category to which the input data belongs (sequential subdivision SOM: FIG. 5 S12, S13) will be described below. This processing is performed by the analysis determination unit 6 in FIG. 1 in the above embodiment, but is described here in a generalized manner. The analysis determination unit 6 is provided with a storage unit.

  For example, a hard disk or a semiconductor memory device can be used as the storage unit. In this storage unit, data obtained by extracting a feature vector using a sensor or the like (a test sample as an example) or data prepared in advance (an index sample as an example) is held as learning data A (n, f) To do. Here, n is an index for designating individual learning data. For example, if the total number of learning data is N, it is any one of numbers 1, 2, 3 to N, or a variable that can designate these numbers. Further, f is an index for designating the components of the learning data, and if the total number is F, it is a variable that can designate numbers 1, 2 to F or these numbers. For example, when a time-series signal of acoustic data is processed and the signal intensity for each frequency is used as the data, the frequency or a number designating this is f.

  Although it is stored in the storage unit, the component of the learning data is not limited to the signal strength specified by the frequency, and the component may be the signal strength for each time of the time series signal. Furthermore, components may be specified by both time and frequency as in the case of performing time-frequency analysis. In this case, the index f is a number that specifies both time and frequency. In addition, although it is not possible to list all data formats, it is a requirement of learning data that it has a signal component that can capture the characteristics of the acquired data.

Using the data A (n, f) and A (m, f) specified by two indices n and m as an amount indicating the distance of the learning data, for example,
Or, for example, a quantity defined using w (f), which is a positive real number, as a coefficient
Can be used. Besides this,
And so on. In this way, the amount that can be used as the distance is unlimited and not all are listed, but the distance between the same data is 0, or the distance is defined as a function value having a rule that can be associated with this numerical value.

  FIG. 10 shows a program procedure according to the present invention. In addition, FIG. 13 and FIG. 14 show the hierarchical category creation procedure (S43 in FIG. 10). As shown in these figures, the feature of this process is that the process of grouping the learning data into categories is performed and created sequentially multiple times. Can be created. The upper limit of the number of groupings is set to G (confirmation or input in S42 in FIG. 10), and the number of times up to that time is designated by any one of 1, 2 to G, or a number g corresponding thereto. The procedure of the grouping process will be described below using this symbol g. Note that acquisition of learning data (S41 in FIG. 10) includes only the acquisition of the index sample in the above embodiment (FIG. 11) and includes both the index sample and the test sample (FIG. 12). Information G specifying the upper limit number and learning data are stored in the storage unit (see FIG. 17).

  As shown in FIG. 13, in the first grouping, that is, the grouping procedure with g = 1, C (g) representative data B (g) having the same components as the learning data held in the storage unit. , C, f) are created (S81). The index c has a function of designating representative data and changes from c = 1, 2 to C (g). The index f is an index that performs the same function as the index f in the learning data A (n, f), and assumes any value of f = 1, 2 to F. For each piece of data A (n, f) designated by the index n held in the storage unit, the distance is set to the representative data B (g, c, f), and c is set to 1, 2 to 2. Obtained by changing to C (g). This value is expressed as dis (n, g, c).

  According to equation (10), this value is expressed by the following equation (of course, it may be obtained as in equation (11) or equation (12)).

  The index c that gives the minimum value among these distances is expressed by cmin (n, g). Using this symbol, the g-th category to which the learning data held in the storage unit designated by the index n belongs can be designated by the index cmin (n, g). In this manner, the N learning data in the storage unit can be associated with or belong to any one of the representative data prepared in C (g). The total number of learning data indices assigned to the c-th representative data is indicated by M (g, c), and those indices designated by a new index i are indicated by the symbol m (g, c, i). To do. Here, the index i is one of 1, 2 to M (g, c), or a variable that can be associated with these. These data are stored in the storage unit (see FIG. 17).

  The procedure for creating the representative data B (g, c, f) is an array created from M (g-1, c ′) learning data belonging to the c-th category of the g-1th grouping. Using the variable a (i, f) = A (m (g−1, c ′, i), f), for example, the principal component analysis which is a known method can be used as follows. When g = 1, which means the first grouping, g-1 = 0, but in this case, the category is composed of all the learning data held in the storage unit, and the category index is only c ′ = 1. Therefore, it is assumed that C (g-1) = 1, M (g-1, c ') = N, and m (g-1, c', i) = i.

  It is assumed that aa (f, f ') is new array data created by calculating the sum for the index i of the product a (i, f) · a (i, f'). The array data aa (f, f ') is mathematically a matrix, and its eigenvalue s (k) and eigenvector v (k, f) can be obtained by a known method. Here, k is an index for designating an eigenvector, and is a number that takes one of values 1, 2 to F. Further, it is assumed that the sum of the squares of v (k, f) for all f is 1. If, for example, two of these eigenvalues s (k) are selected in the order of their magnitudes, and these indices are expressed as k1 and k2, vectors v (k1, f) and v (k2, f) are represented in the vector space of the learning data. The principal component analysis teaches that fluctuation is the largest direction. The coefficient set (p1 (k), p2 (k)) is changed to 1, 2, 3, and 4 by changing k to (0, 0), (0, 1), (1, 0), (1, 1 ), B (k, f) = p1 (k) · v (k1, f) / √ {s (k1)} + p2 (k) · v (k2, f) / √ {s (k2)} Are added to a part of the representative data of the g-th grouping process, and categories are added accordingly. Further, the index of the added category is displayed as d (g, c, k) (k = 1, 2, 3, 4). Here, the number of changes of the index k is limited to four. However, since a different category may be created for each category of g−1th c ′, this number is generally represented by D (g−1, c ′). The expansion method can be realized not only by adding (0, 2), (2, 0), (1, 2), (2, 1) to the above (p1, p2), but also other methods. But it can be easily guessed that this is possible. Since these procedures can be performed for each of the (g−1) th categories, the representative data is displayed as B (g, c, f) by collectively specifying them using the index c. From the above procedure, D (g−1, c ′) new gth grouping categories are created from the learning data included in the c′th category of the g−1th grouping (S82). ). Details of S82 are shown in FIG.

  In the procedure related to the grouping, the category data of the g-th hierarchy, that is, the representative data B (g, c, f) related to the c-th category of the g-th grouping is converted into the index m ( g.c, i) (i = 1, 2 to M (g, c)) is replaced with an average for the index i of the storage unit data A (m (g, c, i), f) specified. it can. With this replacement, the category data of the previous hierarchy, that is, the M (g−1, c ′) data of the category created from the c′th category of the g−1th grouping is changed to the above. It is also possible to rearrange using the representative data obtained by replacement. Further, this process can be repeated a desired number of times. The present invention does not exclude such a processing procedure. This procedure is shown in FIG. Details of the processing of S92 of FIG. 14 are also shown in FIG.

  When M (g-1, c) = 0 in the procedure related to the above grouping, use v (k1, f), v (k2, f) used when creating the g-2th category. Create representative data with. Further, when g−1 = 0, the total number of categories is considered as 1, c ′ = 1, and M (g−1, c ′) = N.

  In the procedure related to the above grouping, there is a possibility that the same one exists among those designated by different indices c of the newly created representative data B (g, c, f). It does not exclude sex. Furthermore, if the same representative data is intentionally prepared, the learning data separated and attributed to the different category by the g-1th grouping process is put into the common category described above by the gth grouping. It may be attributed, and has the effect of avoiding excessive separation of learning data.

  By re-reading g-1 and g as g and g + 1 in the procedure relating to the grouping, the learning data belonging to the category of the gth grouping can be grouped again to create the g + 1th category. . Furthermore, any number of groupings can be performed sequentially by repeating this procedure. It is also possible to provide a procedure for designating or confirming this value, where G is the upper limit number of groupings. Furthermore, as shown below, it is possible to expand the range of numerical values that G can take and generally define G as information for specifying the upper limit number of times.

  If the sequential grouping procedure is continued, the number of data belonging to all categories may eventually become 1 or 0, so this number can be set as the upper limit number of grouping times g. For example, by specifying G = -1, the grouping can be repeated up to the upper limit. Furthermore, when G = −2, for example, the sequential grouping can be repeated until the maximum value of the learning data belonging to each category of the g-th grouping becomes 10, for example. The present invention does not exclude such a procedure.

  In the procedure related to the grouping, the g-th grouping, that is, the c-th category of the g-th hierarchy is the index m (indicating the learning data belonging to the representative data B (g, c, f). g, c, i) (i = 1, 2 to M (g, c)) and an index d (g, c, k) (k = 1) that specifies the category of the (g + 1) th grouping created sequentially. , 2 to D (g, c)). The structure of this data is summarized in FIG. These pieces of information are stored in the storage unit every time they are created.

  The procedure for grouping data is a procedure for creating representative data and indicators related to the category. In general, the grouping becomes more detailed as the value of g increases. FIG. 18 shows a figure similar to the SOM feature map known as an example of displaying the hierarchical category data in the figure. This is category data when there are four types A, B, C, and D as learning data. One map is determined by the grouping number g or the layer number g. In each map, each category is displayed as a net-like figure, and different types of learning data are assigned to the category with different markers. In addition, the category of the previous hierarchy is overwritten by the meshes drawn with white lines. The darkness of the background color of the mesh representing the category is increased in six levels of 0, 1, to 5 as the distance between the representative data of the category to which the learning data of type A belongs and the representative data of each category increases. Is shown in Note that the known SOM feature map cannot be hierarchized.

  Input data having the same components as the learning data is represented by I (f) and a symbol. A value obtained by obtaining the distance between the input data I (f) and the representative data B (g, c, f) in accordance with the definition described above is represented by dis (I; g, c). At this time, the index c giving the minimum value among the values of dis (I; g, c) obtained by changing c from 1, 2 to C (g) is expressed as cmin (I; g). The data I (f) or the acquired input data is defined as a determination result that determines which of the categories classified in the g-th group belongs. This determination procedure is shown in FIG. S51 to S53 are the same as S41 to S43 in FIG.

  In the procedure shown in FIG. 11, the learning data corresponds to the index sample of the above embodiment, and the input data corresponds to the test sample. This corresponds to a process in which an SOM map is created based on the index sample (for example, as shown in FIG. 8b), and the test sample is applied to this. Thereby, it is possible to know the determination result, that is, the position of the test sample (for example, as shown in FIG. 8c) on the SOM map created by the index sample. In FIG. 8c, the position of the test sample is shown as a black square (diagnostic shunt sound). The details of S62 in FIG. 11 are shown in FIG.

  In order to obtain the determination result cmin (I; G), the following sequential processing can be performed by changing the order of g = 1, 2 to G-1 (S61 to S64 in FIG. 11). First, an index c1 = cmin (I; 1) to which the input data belongs is obtained from C (1) categories of the first grouping with g = 1. Next, cmin (I; 2) is obtained from D (g, c1,1) categories of the g + 1th grouping belonging to the c1th category of the first grouping. Similarly, when cmin (I; g) is sequentially repeated as g = 3, 4 to G-1, an index cmin (I; G) is obtained. In this procedure, the number of times for obtaining the distance between the input data and the representative data is the sum of C (1) and D (g, cmin (I; g)) (g = 1, 2 to G-1). is there.

  It is also possible to perform the input data determination process by incorporating the input data into a part of the learning data. The procedure in this case is shown in FIG. Assuming that the index indicating the input data is n0 in the learning data, the category to which the input data belongs in the g-th grouping is represented by cmin (n0, g) (g = 1, 2˜ G). These are also the determination results obtained by the present invention. In FIG. 12, reference data (learning data) (corresponding to the index sample in the above embodiment) and input data (corresponding to test data) are integrated (mixed together) and these are combined into FIG. 10S43 (FIG. 12). The process of S73) is performed. The determination of attribution of input data (S74) corresponds to, for example, the process of determining cmin (n0, G) by the above procedure and the threshold processing of the stenosis level represented by the above equation (2). Is displayed together with the SOM map as shown in FIG. 8c (S75).

  Sequential processing can significantly reduce the processing time for grouping processing to create classification categories for learning data and the processing time for judgment processing to determine which category the input data belongs to . This is the effect of the present invention. The basis for this is shown below.

  In describing the effect of the present invention, it is assumed that the number of categories created in the g-th grouping in the grouping process is the same C1. That is, when the symbols defined in the above description are used, C1 = C (1) = D (2, c) =... = D (G-1, c) is established for an arbitrary index c. Assume that. Here, G is the upper limit number of sequential groupings. At this time, the total number Ct of categories obtained by the G-th grouping is C1 to the G-th power. It should be noted that the essence of the following discussion does not change even in the general case where this assumption is removed.

  In general, in order to directly divide N learning data into Ct categories, it is necessary to obtain Ct representative data and distance for each learning data, and the number of times is at least N · Ct. .

  On the other hand, the number of times of obtaining the distance required when sequential processing is performed can be estimated by G · N · C1.

  The reason is that in the case of the first grouping, that is, when g = 1, it is necessary to compare N learning data with C1 (= C (1)) representative data, so the number of times for obtaining the distance is N · C1. is there. Next, N pieces of learning data are assigned to C1 categories, and the number can be estimated by N / C1. In general, this number varies from category to category, but it is easy to guess that the essence of the discussion will not change even if the variation is taken into account.

  Next, the number of times of obtaining the distance between the C1 representative data of each category obtained by the g = second grouping and the (N / C1) data is (N / C1) · C1. Yes, since this operation is performed in the category of g-1st C1, the number N · C1 multiplied by C1 is the minimum number of times for obtaining the distance in the second grouping. Similarly, the number of times for obtaining the distance until the G-th grouping is N · C1. Therefore, the number of times G · N · C1 obtained by adding N · C1 G times is the minimum number of times for obtaining the distance.

  From the facts pointed out in the above argument, the number of times for obtaining the distance when the N learning data in the storage unit is grouped is N · Ct times when the sequential processing is not performed, but the sequential processing is performed. It is N · G · C 1 times. Since Ct is the power of C1 to the G power, the number of times when the sequential grouping is performed is significantly smaller than the number of times when the sequential grouping is not performed, and the processing time is accordingly increased. Can be shortened. Moreover, the reduction in processing time becomes more prominent as the number N of learning data increases. This is the effect of the present invention.

  In the above-described determination process of the input data I (f), that is, in the process of determining which of the G-th grouping categories that is the upper limit number of times the input data belongs, a sequential process Otherwise, since it is necessary to obtain the distance from all the representative data of the Gth category, the number of times of obtaining the distance is Ct. On the other hand, when the determination is performed by sequentially changing g = 1, 2 to G and dealing with the categories related to the respective groupings, the distance may be obtained only once for G · C. Also in this case, the number of times of obtaining the distance is smaller when the sequential processing is performed, and the processing time can be remarkably shortened. This is also an effect of the present invention.

  In order to make it easier to understand, explanations of main symbols are given below.

N Total number of learning data held in the storage unit G Variable for specifying the upper limit number of times of performing sequential grouping A (n, f) The nth learning data held in the storage unit has F components (N = 1, 2,... N, f = 1, 2 to F)
B (g, c, f) The data representing the c-th category created at the time of the g-th grouping and having the same components as the storage unit C (g) The category created at the time of the g-th grouping Total number M (g, c) Total number of indices of learning data attributed to the cth category of the gth grouping D (g, c) g + 1th grouping belonging to the cth category of the gth grouping The total number of categories m (g, c, i) The i-th learning data index (i = 1, 2 to M (g, c)) belonging to the c-th category of the g-th grouping
d (g, c, k) The index of the category of the g + 1th grouping belonging to the cth category of the gth grouping (k = 1, 2 to D (g, c))
dis (n, g, c) Distance between learning data A (n, f) and representative data B (g, c, f)
dis (I; g, c) Distance between input data I (f) and representative data B (g, c, f) cmin (n, g) The gth data to which the nth learning data held in the storage unit belongs Index of grouping category cmin (I; g) Index of gth grouping category to which input data I (f) belongs

DESCRIPTION OF SYMBOLS 1 Sensor part 3 A / D converter 4 Signal processing apparatus 5 Feature extraction part 6 Analysis determination part 9 Display part
10 Arithmetic processing section
20 Arrayed sound sensor
21 Sheet-like soft support
22 Cavity (hole)
23 Microphone
25 Presser band

Claims (3)

A plurality of cavities are opened in the sheet-like soft support, and a microphone is provided at the bottom of each of these spaces. The depth of the cavities is maintained between the subject and the microphones when wearing the sound. depth der is, an array voice pickup sensor device that are provided on the opposite side of the cavity bottom for each of the amplifiers is the microphone for amplifying the electric signal output from the microphone. A plurality of cavities spaced on one side of a sheet-like flexible support, a microphone is provided in the bottom of each of these spaces, Ri inside near the cavity than the surface of the microphone is above a surface of said sheet-like flexible support , an array voice pickup sensor device that are provided on the opposite side of the cavity bottom for each of the amplifiers is the microphone for amplifying the electric signal output from the microphone.   The arrayed sound collection sensor device according to claim 1 or 2, wherein the cavities are formed in a plurality of columns and a plurality of rows.