State evaluation device for auditory organ

Abstract

To appropriately evaluate the state of an auditory organ.SOLUTION: A state evaluation device 40 for an auditory organ includes: a data acquisition unit 42 which acquires information on a frequency characteristic of a sound pressure level, which is measured as a pressure fluctuation in an external auditory canal 210 of a subject 200 when a stimulus sound is output toward the external auditory canal 210 to indicate an impedance of the auditory organ of the subject 200; a state information calculation unit 44 which includes a learned model 43 created by machine learning configured to output information on a state of the auditory organ when information on the frequency characteristic is input; a post-processing unit 45 which creates information on the state of the auditory organ of the subject 200 on the basis of information related to an output of the state information calculation unit 44 in accordance with the frequency characteristic of the subject 200; and an output unit 46 which outputs the information created by the post-processing unit 45.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to a condition assessment device for the hearing organ, and in particular to a device for non-invasive assessment of the condition of the middle ear. [Background technology]

[0002] Various methods for non-invasively examining the condition of the middle ear are known. Tympanometry is one such method. Tympanometry is known to be particularly effective in examining serous otitis media. For example, Patent Document 1 discloses technology relating to an ear probe that can be used for tympanometry, etc. Middle ear diseases are not limited to serous otitis media, but also include various other conditions such as fixation and disruption of the ossicular chain. There is a demand for a non-invasive and appropriate understanding of the condition of the hearing organs, including the middle ear, which may be in various states. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Publication No. 2023-108617 Summary of the Invention [Problem to be solved by the invention]

[0004] The present invention aims to appropriately evaluate the condition of the hearing organs. [Means for solving the problem]

[0005] According to one aspect of the present invention, a state evaluation device for auditory organs includes a data acquisition unit that acquires information regarding the frequency characteristics of sound pressure levels measured as pressure fluctuations in the ear canal of a subject when a stimulus sound is output into the ear canal of the subject, which indicates the impedance of the subject's auditory organ; a state information calculation unit that includes a trained model created by machine learning and is configured to output information regarding the state of the auditory organ when information regarding the frequency characteristics is input; a post-processing unit that creates information regarding the state of the subject's auditory organ based on information regarding the output of the state information calculation unit in accordance with the frequency characteristics related to the subject; and an output unit that outputs the information created by the post-processing unit. [Effects of the Invention]

[0006] According to the present invention, the state of the hearing organ can be appropriately evaluated. [Brief explanation of the drawings]

[0007] [Figure 1] FIG. 1 is a schematic diagram showing an outline of a configuration example of a hearing organ condition evaluation system according to one embodiment. [Figure 2] FIG. 2 is a functional block diagram showing an outline of a configuration example of a hearing organ condition evaluation system according to one embodiment. [Figure 3A] FIG. 3A is a diagram for explaining the sound pressure in the ear canal. [Figure 3B] FIG. 3B is a diagram for explaining the sound pressure in the ear canal. [Figure 4] FIG. 4 shows an example of the measurement results of the frequency characteristics of sound pressure levels obtained using a condition evaluation system with an adult subject as the subject. [Figure 5] FIG. 5 is a flowchart outlining an example of a method for determining a middle ear condition according to one embodiment. [Figure 6] FIG. 6 is a flowchart showing an outline of an example of the operation of a computer for measuring middle ear characteristics according to one embodiment. [Figure 7] FIG. 7 is a flowchart showing an outline of an example of a method for automatically specifying a feature amount according to an embodiment. [Figure 8A] FIG. 8A is an example of the results of automatic feature identification for an example SPL curve with clear local minima and maxima. [Figure 8B] FIG. 8B is an example of the results of automatic feature identification for an example SPL curve with clear local minima and maxima. [Figure 9A] FIG. 9A is an example of the results of automatic feature identification for an example SPL curve with unclear local minima and maxima. [Figure 9B] FIG. 9B is an example of the results of automatic feature identification for an example SPL curve with unclear local minima and maxima. [Figure 10A] FIG. 10A shows an example of the results of automatic feature identification for an example SPL curve in which the minimum and maximum values ​​representing the characteristics of the middle ear were manually selected. [Figure 10B] FIG. 10B shows an example of the results of automatic feature identification for an example SPL curve in which the minimum and maximum values ​​representing the characteristics of the middle ear were manually selected. [Figure 11] FIG. 11 is a flowchart showing an outline of a process for determining the state of the middle ear according to one embodiment. [Figure 12] FIG. 12 is a conceptual diagram for explaining an example of the Mahalanobis distance. [Figure 13] FIG. 13 is a functional block diagram showing an outline of a configuration example of a state information calculation unit that performs processing related to calculation of the Mahalanobis distance according to one embodiment. [Figure 14] FIG. 14 is a flowchart showing an outline of an example of the operation of a condition evaluation device according to an embodiment that uses Mahalanobis distance. [Figure 15A] FIG. 15A is a diagram showing an example of an index PF related to the probability that the middle ear condition is fixed, obtained by a condition evaluation device according to one embodiment using Mahalanobis distance from known features indicating that the middle ear condition is fixed, normal, or disconnected. [Figure 15B]FIG. 15B is a diagram showing the sensitivity, specificity, and accuracy obtained when the threshold value for determining that adhesion has occurred is changed with respect to the calculation results shown in FIG. 15A. [Figure 15C] FIG. 15C is an ROC curve showing the relationship between sensitivity and specificity shown in FIG. 15B. [Figure 16A] FIG. 16A is a diagram showing an example of an index PS relating to the probability that the middle ear condition is detached, obtained by a condition evaluation device according to one embodiment using Mahalanobis distance from known features indicating that the middle ear condition is fixed, normal, or detached. [Figure 16B] FIG. 16B is a diagram showing the sensitivity, specificity, and accuracy obtained when the threshold value for determining transection is changed with respect to the calculation results shown in FIG. 16A. [Figure 16C] FIG. 16C is an ROC curve showing the relationship between sensitivity and specificity shown in FIG. 16B. [Figure 17A] FIG. 17A shows, as a comparative example, an example of static compliance obtained by tympanometry for subjects whose middle ears are known to be fixed, normal, or disconnected. [Figure 17B] FIG. 17B is a diagram showing the sensitivity, specificity, and accuracy obtained when the threshold value for determining adhesion is changed with respect to the measurement results shown in FIG. 17A. [Figure 17C] FIG. 17C is an ROC curve showing the relationship between sensitivity and specificity shown in FIG. 17B. [Figure 17D] FIG. 17D is a diagram showing the sensitivity, specificity, and accuracy obtained when the threshold value for determining transection is changed with respect to the measurement results shown in FIG. 17A. [Figure 17E] FIG. 17E is an ROC curve showing the relationship between sensitivity and specificity shown in FIG. 17D. [Figure 18] FIG. 18 is a functional block diagram illustrating an outline of a configuration example of a state information calculation unit using principal component analysis and a support vector machine (SVM) according to an embodiment. [Figure 19]FIG. 19 is a flowchart showing an outline of an example of the operation of a condition evaluation device according to an embodiment that uses SVM. [Figure 20] FIG. 20 is a diagram showing an example of plotting the first and second principal components obtained by principal component analysis for each of the data for fixation, detachment, and normality. [Figure 21] FIG. 21 is a diagram showing the factor loadings of the first and second principal components obtained by the principal component analysis shown in FIG. [Figure 22] Figure 22 shows an example of validation of a two-classification task of "disconnection vs. fixation and normal" using SVM. [Figure 23] FIG. 23 shows an example of a comparative example in which the peak pressure (TPP) and static compliance (SC) obtained by tympanometry are plotted for subjects whose middle ears are known to be fixed, normal, or disconnected. [Figure 24] FIG. 24 is a functional block diagram showing an outline of a configuration example of an auditory organ condition evaluation system according to a modified example. DETAILED DESCRIPTION OF THE INVENTION

[0008] An embodiment will be described with reference to the drawings. This embodiment relates to a system for measuring the characteristics of the auditory organ and determining the state of the auditory organ. This auditory organ condition evaluation system can perform measurements non-invasively. The condition evaluation system outputs a stimulus sound into the ear canal of a subject and acquires pressure fluctuations within the ear canal at the time. The condition evaluation system is configured to analyze the pressure fluctuations to analyze the characteristics of the subject's auditory organ and determine the state of the auditory organ. In particular, the condition evaluation system can acquire information related to the impedance of the auditory organ. This impedance reflects, in particular, the dynamic characteristics of the middle ear system. Based on the acquired information related to the impedance, the condition of the auditory organ, particularly the middle ear, can be evaluated. For example, the presence or absence of fixation or disruption of the ossicular chain, the presence or absence of otitis media including serous otitis media, the presence or absence of thinning or adhesion of the tympanic membrane, and the presence or absence of a middle ear malformation can be determined. The acquired information can be used to evaluate the sound conduction characteristics of the middle ear and diagnose middle ear diseases and conductive hearing loss. In other words, this condition evaluation system can function as a diagnostic device for the auditory organ.

[0009] [Device configuration] Fig. 1 is a schematic diagram showing an outline of a configuration example of a hearing organ condition evaluation system 1 according to this embodiment. Fig. 2 is a functional block diagram showing an outline of a configuration example of a hearing organ condition evaluation system 1 according to this embodiment. As shown in Fig. 1, the condition evaluation system 1 includes a computer 10, an AD / DA converter 60, an amplifier system 70, and a probe 80.

[0010] The probe 80 is configured so that its tip is inserted into the ear canal 210 of the subject 200. The probe 80 includes an earphone 82 for outputting a stimulating sound toward the ear canal 210, and a microphone 84 for acquiring pressure fluctuations within the ear canal 210.

[0011] The computer 10 is a general-purpose computer, and may be, for example, a personal computer. The computer 10 includes, for example, a central processing unit (CPU) 11, various integrated circuits such as a memory 12 and a storage 13, and various interfaces 14. The computer 10 may also include a display device 15. The computer 10 may further include a field programmable gate array (FPGA) or the like prepared according to the processing of the condition evaluation system 1. The computer 10 performs processing such as controlling the operation of each part of the condition evaluation system 1, generating signals related to the stimulus sound, analyzing pressure fluctuations in the ear canal, and evaluating the middle ear condition. The operation of the computer 10 is performed according to a program recorded in the computer 10 as software or hardware or provided from outside the computer 10.

[0012] The AD / DA converter 60 has a DA converter 62 and an AD converter 64. The amplifier system 70 has an earphone amplifier 72 and a microphone amplifier 74. The DA converter 62 of the AD / DA converter 60 converts the digital signal related to the stimulation sound output from the computer 10 into an analog signal and outputs it to the earphone amplifier 72 of the amplifier system 70. The earphone amplifier 72 amplifies the analog signal related to the stimulation sound input from the DA converter 62 and outputs the stimulation sound from the earphone 82 of the probe 80. The microphone amplifier 74 of the amplifier system 70 amplifies the analog signal related to pressure fluctuations in the ear canal 210 acquired by the microphone 84 of the probe 80 and outputs it to the AD converter 64 of the AD / DA converter 60. The AD converter 64 converts the analog signal acquired from the microphone amplifier 74 into a digital signal and inputs it to the computer 10.

[0013] The computer 10 has functions as a control unit 22, an input signal generating unit 32, a signal output unit 52, a signal acquiring unit 54, a recording unit 34, a condition evaluating device 40, etc. The control unit 22 controls each operation of the computer 10.

[0014] The input signal generating unit 32 generates a signal related to the stimulus sound to be output. The generated signal related to the stimulus sound is output to the DA converter 62 via the signal output unit 52. In this embodiment, the stimulus sound is configured to include each frequency component of the measurement target frequency band. For example, the measurement target frequency band may be from 100 Hz to 2000 Hz, in which case the stimulus sound may include frequency components from 50 Hz to 3000 Hz. The stimulus sound may be, for example, a frequency sweep sound, white noise, an M-sequence sound, a swept-sine sound, etc.

[0015] A signal relating to pressure fluctuations in the ear canal obtained by the microphone 84 is acquired by the signal acquisition unit 54 via the AD converter 64, and the data is recorded in the recording unit 34. The condition evaluation device 40 uses the data recorded in the recording unit 34 to perform various analyses, evaluate the condition of the middle ear, etc.

[0016] As described above, the signal output unit 52, DA converter 62, earphone amplifier 72, earphone 82, etc. collectively function as a stimulation sound output unit 91 configured to output a stimulation sound toward the ear canal 210 of the subject 200. Furthermore, the signal acquisition unit 54, AD converter 64, microphone amplifier 74, microphone 84, etc. collectively function as a sound receiving unit 92 configured to acquire a sound pressure signal that indicates pressure fluctuations in the ear canal 210 when the stimulation sound output unit 91 is outputting the stimulation sound.

[0017] The configuration of the condition assessment system 1 shown here is an example, and can be modified as appropriate as long as it performs similar functions. Here, an example has been shown in which the computer 10 performs all of the various controls and data analysis related to the operation of the condition assessment system 1, but this is not limiting. The functions of the computer 10 may be realized by any number of devices. Furthermore, some of the functions of the computer 10 may be performed by other devices located in remote locations and connected to the computer 10 via a network. For example, the operation of each component of the condition assessment system 1 may be controlled from a remote location, and the acquired data may be analyzed by a server located in a remote location. While various physical configurations are possible, the condition assessment system 1 performs the functions of a stimulation sound output unit 91, a sound receiving unit 92, a condition assessment device 40, etc.

[0018] [Measurement principle] The principles of measurement and analysis of hearing organ characteristics using the condition assessment system 1 will now be described. As shown in FIG. 1 , in the human hearing organ, the eardrum 222 at the end of the external auditory canal 210 and the cochlea 232 of the inner ear are connected via ossicles 224, including the malleus 225, incus 226, and stapes 227. The middle ear consists of the eardrum 222, the ossicles 224, and the tympanic cavity 229, a small, air-filled space in which the ossicles 224 are located. Vibrations from the eardrum 222 are transmitted to the cochlea 232, where sensory cells reside, via the chain of ossicles 224. Middle ear diseases, such as those caused by lesions, can cause conductive hearing loss. For example, there are known middle ear diseases in which vibrations from the eardrum 222 cannot be properly transmitted to the cochlea 232 due to a partial rupture of the chain of ossicles 224 or partial fixation of the ossicles 224. Other known middle ear diseases include various types of otitis media, including serous otitis media, thinning and adhesion of the tympanic membrane, and middle ear malformations.

[0019] The sound pressure inside the ear canal 210 will now be described. During measurement using the condition evaluation system 1, the diaphragm 83 of the earphone 82 vibrates at one end of the ear canal 210, and the eardrum 222 vibrates at the other end of the ear canal 210. Figures 3A and 3B are diagrams that schematically show this state.

[0020] When the frequency of the stimulation sound output from earphone 82 is lower than the resonance frequency of the middle ear, as shown in FIG. 3A, vibrating membrane 83 of earphone 82 and eardrum 222 are displaced in the same phase. Therefore, pressure P in ear canal 210 is expressed as follows: P=K(ΔV-ΔV TM ) / V Here, K is the bulk modulus of air, ΔV is the volume change caused by the diaphragm 83 of the earphone 82, and ΔV TM is the volume change caused by the eardrum 222, and V is the volume of the ear canal 210. At this time, the more the eardrum 222 vibrates, the smaller the volume change becomes, and the smaller the sound pressure becomes.

[0021] When the frequency of the stimulation sound output from the earphone 82 reaches the resonant frequency of the middle ear, the phase of the eardrum 222 is inverted, as shown in FIG. 3B. Therefore, the pressure P in the ear canal 210 is P=K(ΔV+ΔV TM ) / V At this time, the more the eardrum 222 vibrates, the greater the volume change and the greater the sound pressure.

[0022] FIG. 4 shows an example of measurement results using the condition evaluation system 1 with an adult as subject 200. The measurement shown in FIG. 4 was performed by outputting a frequency sweep sound, the frequency of which varied from 100 Hz to 2000 Hz, from earphone 82 over 10 seconds. When this frequency sweep sound was output, the sound pressure in ear canal 210 was measured using microphone 84. The solid line in FIG. 4 shows the sound pressure level (SPL) acquired using microphone 84 versus the frequency of the stimulus sound. Here, SPL is defined as: SPL = 20 log | P / P REF | where P is the sound pressure measured by microphone 84 and P REF is the reference sound pressure, 2.0×10 -5 The sound pressure level is expressed as a frequency characteristic of the eardrum 222. ...

[0023] The example shown in Figure 4 is an example in which measurements were performed using a frequency sweep sound, but the method is not limited to using a frequency sweep sound. Other methods may be used. For example, an SPL curve similar to that shown in Figure 4 can be obtained by using a stimulus sound containing each frequency component and performing various frequency analyses on the obtained sound pressure level.

[0024] The SPL curve is known to represent the characteristics of the middle ear of the subject 200. Large changes in SPL observed in the SPL curve are known to represent resonance in the middle ear. In FIG. 4, the intermediate value between the frequency at which the SPL curve shows a minimum value and the frequency at which the SPL curve shows a maximum value is indicated by a dashed-dotted arrow as the resonance frequency (RF) of the middle ear. In addition, ΔSPL, which is the difference between the minimum value and the maximum value of the SPL curve, is known to represent the mobility of the eardrum 222.

[0025] For example, by evaluating the resonant frequency (RF) and ΔSPL, information such as whether the middle ear is normal or whether there is a middle ear disease due to a lesion or the like can be obtained. For example, when the ossicles 224 are fixed and difficult to move, the ΔSPL may be smaller and the RF may be slightly higher than in a normal ear. Also, when the ossicles 224 are detached and unable to transmit sound to the inner ear, the ΔSPL may be larger and the RF may be slightly lower than in a normal ear. In this way, feature quantities such as RF and ΔSPL can be identified from the SPL curve, and the condition of the hearing organ can be evaluated based on these feature quantities. For example, the sound conduction characteristics of the middle ear of the subject 200 can be evaluated based on the evaluated condition of the hearing organ, and this can be used for diagnosis, etc. If the condition evaluation system 1 can output a diagnosis result, the condition evaluation system 1 can become a diagnostic device.

[0026] In this embodiment, a trained model 43 obtained by machine learning is used to evaluate the state of the hearing organs based on feature quantities obtained from an SPL curve. In this embodiment, for example, the trained model 43 is created in advance by supervised learning using training data in which the relationship between feature quantities obtained from an SPL curve and the state of the hearing organs is known. By inputting feature quantities obtained from an SPL curve of a subject 200 whose state of the hearing organs is unknown into this trained model 43, information on the state of the hearing organs of the subject 200 is output from the trained model 43. In this embodiment, the accuracy of the evaluation of the state of the hearing organs using the machine learning model is improved by using feature quantities, etc., that can be theoretically explained to indicate the state of the hearing organs in the SPL curve, in particular.

[0027] [How to determine the condition of the middle ear] The condition of the middle ear is determined using the condition evaluation system 1 of this embodiment, for example, according to the following procedure. FIG. 5 is a flowchart showing an outline of this method. First, the condition evaluation system 1 is used to measure the characteristics of the middle ear of the subject 200, and data related to the SPL curve is obtained (step S101). Next, feature quantities such as the minimum value, maximum value, RF, and ΔSPL are identified from the obtained SPL curve (step S102). Based on the obtained feature quantities, the condition of the middle ear of the subject 200 is determined (step S103).

[0028] [Measurement of middle ear characteristics] The measurement of the middle ear characteristics of the subject 200 by the condition evaluating system 1 in step S101 will now be described. During measurement, the probe 80 is inserted into the ear canal 210 of the subject 200. Fig. 6 is a flowchart showing an outline of an example of the operation of the computer 10. The description will be made with reference to this flowchart.

[0029] In step S201, the computer 10 generates an input signal. In step S202, the computer 10 outputs the generated input signal to the DA converter 62. Based on this input signal, a stimulus sound is output from the earphone 82 toward the ear canal 210 of the subject 200 via the DA converter 62 and the earphone amplifier 72.

[0030] A signal indicating the sound pressure in the ear canal 210 at this time is generated by the microphone 84. The signal generated by the microphone 84 is input to the computer 10 via the microphone amplifier 74 and the AD converter 64. In step S203, the computer 10 acquires the sound pressure signal from the microphone 84. Although there are various procedures depending on the measurement method, the computer 10 acquires data relating to the sound pressure level in the ear canal 210 for each frequency over a predetermined time.

[0031] In step S204, the computer 10 analyzes, organizes, and performs predetermined processing on the obtained data to obtain data on sound pressure levels for each frequency. For example, if the stimulus sound is a frequency sweep sound, the sound pressure level for each frequency can be obtained from the acquired sound pressure levels for each time. If the stimulus sound contains each frequency component, the sound pressure level for each frequency can be obtained by, for example, performing a Fourier transform. In step S205, the computer 10 records the obtained data on sound pressure levels for each frequency in the recording unit 34. The data recorded in the recording unit 34 is data representing an SPL curve.

[0032] The measurement is not limited to being performed only once, but may be performed repeatedly. For example, the measurement may be performed repeatedly while adjusting the insertion of the probe 80 into the ear canal 210, and the measurement may be terminated when an appropriate measurement is obtained.

[0033] [Identifying the features of the SPL curve] An example of a method for identifying feature quantities such as the minimum value, maximum value, RF, and ΔSPL from the SPL curve in step S102 will be described.

[0034] The feature quantities may be manually determined by an operator based on a graph showing an SPL curve such as that shown in Fig. 4. The feature quantities manually determined by the operator may be recorded in the recording unit 34. The recorded feature quantities may be read out as appropriate when evaluating the condition of the middle ear.

[0035] Furthermore, the feature quantity may be automatically identified from the SPL curve based on a predetermined algorithm. That is, the condition evaluation device 40 may have a function as a feature quantity identification unit 41. An example of a method for automatically identifying a feature quantity will be described. Fig. 7 is a flowchart showing an outline of an example of the method for automatically identifying a feature quantity.

[0036] In step S301, the computer 10 acquires data relating to an SPL curve, which is the relationship between the frequency to be analyzed and the sound pressure level. The acquired data may be data in a predetermined frequency range to be analyzed, for example, data from 200 Hz to 2500 Hz only. In step S302, the computer 10 performs a smoothing process. For example, the computer 10 sets a value equivalent to 10% of the data length acquired in step S302 as the frame length and performs Savitzky-Golay filter processing.

[0037] In step S303, the computer 10 detects minimum and maximum values ​​from the smoothed SPL curve data. Here, the minimum and maximum values ​​are local minimum and maximum values, and any number of minimum and maximum values ​​may be detected. In step S304, the computer 10 selects 0 to 2 pairs of minimum and maximum values ​​from the minimum and maximum values ​​detected in step S303. For example, the computer 10 extracts 0 to 2 pairs of minimum and maximum values ​​using the array of minimum and maximum values ​​detected in step S303. For example, thresholds are set for the RF and ΔSPL obtained from the minimum and maximum values, and appropriate minimum and maximum values ​​are extracted by eliminating cases in which the obtained RF and ΔSPL are inappropriate as values ​​representing middle ear characteristics. For example, if the obtained RF or ΔSPL is very large or very small, the minimum and maximum values ​​are excluded.

[0038] In step S305, the computer 10 calculates the RF value as the intermediate value between the frequency of the minimum value and the frequency of the maximum value. In step S306, the computer 10 calculates the ΔSPL value as the difference between the sound pressure level of the minimum value and the sound pressure level of the maximum value. In step S307, the computer 10 records the calculated feature quantities, such as the minimum value, maximum value, RF, and ΔSPL, in the recording unit 34.

[0039] <Analysis example> Figures 8A and 8B show examples of SPL curves with clear local minimum and maximum values. Figures 9A and 9B show examples of SPL curves with unclear local minimum and maximum values. Figures 10A and 10B show examples of SPL curves that were previously manually selected from multiple observed minimum and maximum values ​​by examining which of the minimum and maximum values ​​represent the characteristics of the middle ear. In each figure, circles (◯) and squares (□) represent the minimum and maximum values ​​identified by the above-mentioned algorithm, respectively, and downward triangles (▽) and upward triangles (△) represent the correct minimum and maximum values ​​that should have been identified. In the examples shown in Figures 9B and 10B, the minimum and maximum values ​​identified by the above-mentioned algorithm were incorrect. In the other examples, the above-mentioned algorithm identified the appropriate minimum and maximum values. Table 1 shows the number of correct and incorrect answers for the analysis examples.

[0040] [Table 1]

[0041] As described above, it has become clear that, for example, the minimum and maximum values ​​of the SPL curve can be automatically identified using the above-mentioned algorithm in many cases, and RF and ΔSPL can be identified.

[0042] [Determining middle ear condition] The determination of the middle ear condition in step S103 will now be described. The condition evaluation device 40 has functions as a data acquisition unit 42, a condition information calculation unit 44 including a trained model 43, a post-processing unit 45, and an output unit 46.

[0043] The data acquiring unit 42 is configured to acquire information related to the feature quantities of the SPL curve of the subject 200 recorded in the recording unit 34. The data acquiring unit 42 acquires, for example, values ​​to be used for analysis from the feature quantity values ​​such as the minimum value, maximum value, RF, and ΔSPL obtained from the SPL curve acquired by measuring the subject 200 and recorded in the recording unit 34.

[0044] The trained model 43 is configured to output information about the state of the auditory organs when information related to measurement data is input. As described above, the trained model 43 is created in advance by supervised learning using training data in which the relationship between the feature amounts of the SPL curve and information about the state of the auditory organs is known, for example. Various learning models can be used for machine learning. The state information calculation unit 44, which includes the trained model 43, calculates information about the state of the auditory organs of the subject 200 based on information obtained when the feature amounts acquired by measuring the subject 200 are input to the trained model 43.

[0045] The post-processing unit 45 performs predetermined processing on the information relating to the state of the hearing organs output from the state information calculation unit 44, and calculates information representing the state of the hearing organs. The output unit 46 outputs the information representing the state of the hearing organs calculated by the post-processing unit 45. This information is, for example, recorded in the recording unit 34 or displayed on the display device 15.

[0046] 11 is a flowchart showing an outline of an example of processing for determining the condition of the middle ear. In step S401, the computer 10 acquires data related to the feature quantities of the SPL curve recorded in the recording unit 34. In step S402, the computer 10 inputs the acquired data related to the feature quantities into the trained model 43. In step S403, the computer 10 acquires information related to the condition of the middle ear using the trained model 43. In step S404, the computer 10 performs predetermined post-processing on the acquired information related to the condition of the middle ear to calculate information representing the condition of the middle ear. In step S405, the computer 10 outputs the acquired information related to the condition of the middle ear.

[0047] [First aspect of trained model - Use of Mahalanobis distance] Various models can be used as the machine learning model used in the trained model 43. As an example, a model using Mahalanobis distance calculation can be used as this machine learning model.

[0048] When points in a population are represented by a multivariate vector, the Mahalanobis distance is

number

[0049] In this embodiment, the variables are feature quantities such as minimum values, maximum values, RF, and ΔSPL. For example, the Mahalanobis distance between each group of middle ears with different middle ear conditions is calculated, and the current middle ear condition is estimated based on the calculated Mahalanobis distance. The current middle ear condition can be determined from various conditions, such as normal, "fixed" in which some of the ossicles 224 are fixed, "disconnected" in which some of the chain of ossicles 224 is disconnected, "serous otitis media" (a type of otitis media), and "thinning of the eardrum."

[0050] FIG. 12 is a conceptual diagram illustrating an example of the Mahalanobis distance. In this example, RF and ΔSPL are used as feature quantities. White circles represent data obtained by measurement of a subject whose middle ear condition is known to be normal (N). Dark gray circles represent data obtained by measurement of a subject whose middle ear condition is known to be fixed (F). Light gray circles represent data obtained by measurement of a subject whose middle ear condition is known to be transection (S). The cross marks indicate the center of gravity of each group, and the gray elliptical regions represent 50% equal probability ellipses. As shown in this figure, it is difficult to determine whether the middle ear condition is normal, fixed, or transection based solely on the values ​​of RF and ΔSPL. Therefore, in this embodiment, the Mahalanobis distance is used for this determination.

[0051] For example, RF and ΔSPL values ​​labeled with whether the middle ear state is normal, fixed, or disconnected are used as training data, and a probability distribution such as the mean value and covariance matrix of variables related to each state is obtained by machine learning. One embodiment of the trained model 43 has information about such a probability distribution.

[0052] The condition information calculation unit 44 of the condition evaluation device 40 uses the trained model 43 to calculate the Mahalanobis distance related to the feature amount of the SPL curve related to the middle ear of the subject 200, who is the discrimination target.

[0053] For example, when a trained model obtained by machine learning using RF and ΔSPL values ​​labeled with whether the above-mentioned middle ear state is normal, fixed, or disconnected as training data is used, the following occurs: When the RF and ΔSPL of the SPL curve for the middle ear of subject 200 are input into the trained model, Mahalanobis distances for each of the normal, fixed, and disconnected states are output. The state information calculation unit 44 may output the Mahalanobis distances for each of the normal, fixed, and disconnected states. Information about the Mahalanobis distances may be presented, for example, on the display device 15 via the post-processing unit 45 and the output unit 46 as information about the state of the middle ear of subject 200, i.e., the middle ear to be identified.

[0054] In one embodiment, the condition evaluation device 40 may calculate an index indicating the probability or other probability that the middle ear to be discriminated is in each state based on the calculated Mahalanobis distance, and output the result. For example, the post-processing unit 45 may calculate an index indicating the probability or other probability that each state is in based on the Mahalanobis distance for each state of normal, fixed, and disconnected output by the condition information calculation unit 44. For example, the Mahalanobis distance D N and the Mahalanobis distance D to the group whose middle ear state is fixation (F). F and the Mahalanobis distance D to the group whose middle ear state is transection (S). S Using the above, the post-processing unit 45 calculates an index P F and the index P S And,

number

[0055] In one embodiment, the condition evaluation device 40 may determine whether or not the middle ear condition to be determined is a predetermined state based on the probability that the middle ear condition is a predetermined state or an index related to the probability, and output the result. In this case, a threshold for determining whether or not the state is a predetermined state may be set by machine learning in a machine learning model separate from the machine learning model for calculating the Mahalanobis distance. For example, the threshold may be set by using the training data described above, and may be an index P related to the probability that the middle ear condition to be determined is a fixed state. F A threshold value that maximizes the accuracy of determining whether or not a sticking has occurred may be set for the input exponent P F The result of the determination as to whether or not the middle ear condition is fixation can be output depending on whether or not the threshold value is greater than the threshold value. Similarly, the threshold value can be set by using the training data described above and by using an index P S A threshold value that maximizes the accuracy of determining cleavage may be set for the input exponent P S Depending on whether or not the difference is greater than the threshold value, a determination result indicating whether or not there is a disconnection can be output.

[0056] An example configuration of the state information calculation unit 44 in this case is shown in FIG. 13. The state information calculation unit 44 includes a first trained model 431, an index calculation unit 441, and a second trained model 432. The first trained model 431 is a machine learning model that outputs a Mahalanobis distance. The first trained model 431 is prepared by machine learning using feature quantities labeled with the middle ear condition as training data. The index calculation unit 441 is configured to calculate an index related to the probability that the middle ear condition is a predetermined state based on the Mahalanobis distance, and has, for example, a predetermined calculation formula. The second trained model 432 is a machine learning model that outputs a determination result as to whether the middle ear condition is a predetermined state based on the index. The second trained model 432 is prepared by machine learning using the index labeled with the middle ear condition as training data.

[0057] Feature quantities such as RF and ΔSPL are input from the data acquisition unit 42 to the state information calculation unit 44. The feature quantities such as RF and ΔSPL are input to the first trained model 431. At this time, the first trained model 431 calculates, for example, the Mahalanobis distance D N , the Mahalanobis distance D to the group whose middle ear state is fixation (F) F , and the Mahalanobis distance D to the group whose middle ear state is transection (S). S The Mahalanobis distance to each group is output.

[0058] The output Mahalanobis distance is input to the index calculation unit 441. The index calculation unit 441 uses the input Mahalanobis distance to calculate, as described above, for example, an index P F , and an index P relating to the probability that the middle ear condition being discriminated is transection. S An index related to the middle ear being in a predetermined state is calculated.

[0059] The calculated index is input to the second trained model 432. For example, the index P F and the index P SThe second trained model 432 to which the above has been input outputs to the post-processing unit 45 a discrimination result indicating whether the middle ear condition to be discriminated is a predetermined state, such as whether the middle ear condition to be discriminated is fixated or not, and whether the middle ear condition to be discriminated is detached or not.

[0060] 14 is a flowchart showing an example of the operation of the condition evaluation device 40 that uses the Mahalanobis distance. In this example, the trained model 43 has information on the probability distribution, such as the mean value and covariance matrix of variables related to each state, obtained by machine learning using training data having RF and ΔSPL values ​​labeled to indicate whether the state of the middle ear is normal, fixed, or disconnected.

[0061] In step S501, the data acquiring unit 42 of the condition evaluating device 40 acquires data on the feature quantities of the SPL curve related to the middle ear to be discriminated.

[0062] In step S502, the state information calculation unit 44 of the state evaluation device 40 calculates the Mahalanobis distance for the acquired feature quantity using the trained model 43. For example, as shown in FIG. 12, when data 201 to be discriminated is given, the state information calculation unit 44 calculates the Mahalanobis distance D N and the Mahalanobis distance D to the group whose middle ear state is fixation (F). F and the Mahalanobis distance D to the group whose middle ear state is transection (S). S and calculate.

[0063] In step S503, the condition information calculation unit 44 of the condition evaluation device 40 calculates the probability of each condition or an index related to the probability based on the obtained Mahalanobis distance. For example, the condition evaluation device 40 uses the obtained Mahalanobis distance to calculate an index P related to the probability that the middle ear condition to be determined is fixation, as described above. F and the index P S and calculate.

[0064] In step S504, the state information calculation unit 44 of the state evaluation device 40 determines whether the middle ear state is one of the states based on the thresholds set in the trained model 43. The state information calculation unit 44 calculates, for example, an index P F The state information calculation unit 44 also uses, for example, an index P S is compared with a set threshold value to determine whether or not there is a disconnection.

[0065] In step S505, the post-processing unit 45 of the condition evaluation device 40 performs post-processing on the Mahalanobis distance calculated by the condition information calculation unit 44, the probability of each state or an index related to the probability, and the determination result of whether each state exists. In step S506, the output unit 46 of the condition evaluation device 40 outputs the post-processed data. As a result, the computer 10 can present, for example, the probability or an index related to the probability that the middle ear condition to be determined is normal, fixed, or detached. The computer 10 can also present the determination result, such as whether the middle ear condition is fixed or detached.

[0066] The computer 10 may be configured to calculate only either the probability or index of each state or the determination result of whether or not each state exists, and to display only either one of them.

[0067] As in this example, in this embodiment, the feature quantities of the SPL curve obtained by measurement of a subject whose middle ear condition is known are used as training data, and the Mahalanobis distance related to the feature quantities of the SPL curve to be discriminated is calculated, thereby creating a trained model 43 that can be used to output information about the condition of the middle ear, such as the state of the middle ear to be discriminated or the probability that it is in that state. Condition evaluation device 40 can output this information about the condition of the middle ear using the created trained model 43.

[0068] <Analysis example> The method for determining the middle ear condition of this embodiment was evaluated using leave-one-out cross validation (LOOCV). Measurement data from a total of 56 ears, including 14 ears with ossicular fixation, 10 ears with ossicular disruption, and 32 normal ears, from which SPL curves could be obtained, were used for the evaluation.

[0069] One of the 56 data was used as test data, and the other 55 data were used as training data. The training data was labeled as normal, fixated, or detached, and used as training data. The mean values ​​and covariance matrices of the variables for normal, fixated, and detached were obtained. Using these, the Mahalanobis distances of the test data to the respective training data groups for normal, fixated, and detached were calculated, and the index P related to the probability that the middle ear condition of the test data was fixated was calculated as described above. F and the exponent P for the probability of disconnection S and was calculated.

[0070] The index P for the probability that the middle ear condition obtained when each of the 56 data is used as test data is fixed. F 15A shows the index P for each known middle ear condition, i.e., fixed, normal, or transection. F showed.

[0071] FIG. 15B shows the sensitivity, specificity, and accuracy obtained when the threshold for determining fixation is changed for the calculation results shown in FIG. 15A. Sensitivity is the percentage of test data that is fixed that can be determined to be fixed. Specificity is the percentage of test data that is not fixed that can be determined to be not fixed. Accuracy is the percentage of test data that is fixed that can be determined to be fixed and test data that is not fixed that can be determined to be not fixed. FIG. 15C is a receiver operating characteristic (ROC) curve showing the relationship between sensitivity and specificity.

[0072] The exponent P with the highest accuracy FWhen the threshold value was set at 0.52, the sensitivity was 0.64 (9 / 14), the specificity was 0.88 (37 / 42), and the accuracy was 0.82 (46 / 56).The AUC (area under the curve) of the obtained ROC curve was 0.86.

[0073] In addition, the index P for the probability that the middle ear condition is detached obtained when each of the 56 data is used as test data is S The index P S showed.

[0074] Figure 16B shows the sensitivity, specificity, and accuracy obtained when the threshold for determining transection is changed for the calculation results shown in Figure 16A. Figure 16C shows the ROC curve for this case.

[0075] The exponent P with the highest accuracy S When the threshold value was set at 0.49, the sensitivity was 1.00 (10 / 10), the specificity was 1.00 (46 / 46), and the accuracy was 1.00 (56 / 56). The AUC of the obtained ROC curve was 1.00.

[0076] As a comparative example, we compared the results with those obtained by tympanometry, a known diagnostic method for middle ear problems. Tympanometry measures the ease with which the middle ear vibrates as an equivalent volume (compliance) while changing the pressure in the ear canal from -200 daPa to 200 daPa. One method of evaluating tympanometry is static compliance, which is the difference between the maximum compliance and the compliance when the pressure in the ear canal is set to 200 daPa. We evaluated the distinction between fixation and detachment using measurement data from a total of 60 ears for which static compliance (SC) could be measured by tympanometry: 16 ears with ossicular fixation, 8 ears with ossicular detachment, and 36 normal ears.

[0077] Figure 17A shows the measured static compliance for each known fixation, normal, or rupture. Figure 17B shows the sensitivity, specificity, and accuracy obtained when the threshold for determining fixation is changed for the measurement results shown in Figure 17A. Figure 17C shows the ROC curve for this case. When the threshold for static compliance = 0.22 mL, which maximizes accuracy, was used as the threshold, the sensitivity was 0.19 (3 / 16), the specificity was 1.00 (44 / 44), and the accuracy was 0.78 (47 / 60). The AUC shown by the obtained ROC curve was 0.50. Figure 17D shows the sensitivity, specificity, and accuracy obtained when the threshold for determining rupture is changed for the measurement results shown in Figure 17A. Figure 17E shows the ROC curve for this case. When the threshold static compliance of 1.57 mL was used, which maximized accuracy, the sensitivity was 0.50 (4 / 8), the specificity was 1.00 (52 / 52), and the accuracy was 0.93 (56 / 60).The AUC of the ROC curve obtained was 0.88.

[0078] The above results are summarized in Table 2. For the tympanometry data of the comparative example, the values ​​according to the standards reported by Ichimura et al. in 1976 are also shown.

[0079] [Table 2]

[0080] It was found that high classification performance could be achieved by obtaining an SPL curve through measurement, calculating a probability index using the Mahalanobis distance using the RF and ΔSPL features obtained from the SPL curve, and then using this to make a distinction.Compared to tympanometry, it was found that high classification performance could be achieved using the SPL curve.

[0081] The analysis example shown here uses RF and ΔSPL as feature quantities, but is not limited to this. Using one or more feature quantities, such as the minimum value, maximum value, RF, and ΔSPL, the Mahalanobis distance can be calculated, and an index relating to the probability of each state can be calculated using this. Furthermore, by setting an appropriate threshold for the index, it is possible to determine whether each state exists. While the example shown here illustrates the determination of the presence or absence of fixation and detachment, the calculation of the probability and determination of the presence or absence of otitis media, including serous otitis media, thinning or adhesion of the tympanic membrane, middle ear malformation, and other middle ear diseases can also be performed in a similar manner.

[0082] [Second aspect of trained models: using SVM] A support vector machine (SVM) can be used as an example of a machine learning model used in the trained model 43. This SVM can use feature quantities such as the minimum and maximum values, RF, and ΔSPL obtained from the SPL curve. For example, a two-class support vector machine can be used to classify, for example, the presence or absence of fixation or detachment of the ossicular chain, the presence or absence of otitis media including serous otitis media, the presence or absence of thinning or adhesion of the tympanic membrane, and middle ear malformation.

[0083] For example, a trained model that determines whether an input feature value corresponds to the fixation state or not is created in advance by machine learning using, as training data, feature values ​​based on SPL curves obtained by measuring a subject in a fixation state and feature values ​​based on SPL curves obtained by measuring a subject not in that state.Similarly, a trained model that determines whether an input feature value corresponds to the fixation state or not is created in advance by machine learning using, as training data, feature values ​​based on SPL curves obtained by measuring a subject in each state and feature values ​​based on SPL curves obtained by measuring a subject not in that state.

[0084] For example, when a feature based on an SPL curve obtained by measuring subject 200, whose middle ear condition is unknown, is input to the trained model, information on the determination result as to whether the middle ear condition according to the trained model is a predetermined condition or not is output. This information can be presented, for example, on display device 15 via post-processing unit 45 and output unit 46, as information on the condition of subject 200's middle ear, i.e., the middle ear to be determined.

[0085] As feature quantities obtained from the SPL curve, values ​​such as the minimum value, maximum value, RF, and ΔSPL may not be used as they are, but rather principal components obtained by performing principal component analysis on these values ​​may be used. By performing principal component analysis, appropriate analysis may be performed using values ​​that should be emphasized among multiple feature quantities. An example configuration of the state information calculation unit 44 in this case is shown in FIG. 18. The state information calculation unit 44 includes a third trained model 433 and a fourth trained model 434.

[0086] The third trained model 433 is a machine learning model related to principal component analysis. The third trained model 433 has eigenvectors for calculating principal components obtained by machine learning based on data of multiple feature quantities. When multiple feature quantities are input, the third trained model 433 outputs principal component scores using the eigenvectors. For example, feature quantities such as minimum values, maximum values, RF, and ΔSPL based on an SPL curve obtained by measuring a subject 200 whose middle ear condition is unknown are input from the data acquisition unit 42 to the state information calculation unit 44. These feature quantities are input to the third trained model 433. At this time, the third trained model 433 outputs principal component scores, such as the first principal component and the second principal component.

[0087] The fourth trained model 434 is a machine learning model using an SVM. The fourth trained model 434 is prepared by machine learning using, as training data, principal component scores such as the first principal component and the second principal component labeled with the middle ear condition, and has an SVM separating hyperplane obtained by machine learning. When the principal component scores output from the third trained model 433 are input, the fourth trained model 434 outputs, based on this separating hyperplane, to the post-processing unit 45, a discrimination result indicating whether the middle ear condition to be discriminated is a predetermined state, such as whether the middle ear condition to be discriminated is fixated or disconnected.

[0088] Although an example of reducing dimensions by principal component analysis has been shown here, analysis that increases dimensions by, for example, a kernel method may also be performed.

[0089] FIG. 19 is a flowchart showing an outline of an example of the operation of the condition evaluation device 40 that uses SVM.

[0090] In step S601, the data acquisition unit 42 of the condition evaluation device 40 acquires data on the features of the SPL curve related to the middle ear to be discriminated. In step S602, the condition information calculation unit 44 of the condition evaluation device 40 uses the trained model 43 to discriminate whether the middle ear condition is one of the respective conditions based on the acquired features.

[0091] In step S603, the post-processing unit 45 of the condition evaluation device 40 performs predetermined post-processing on the determination results of whether or not each condition is present as determined by the condition information calculation unit 44. In step S604, the output unit 46 of the condition evaluation device 40 outputs the post-processed data. As a result, the computer 10 can present the determination results, such as whether or not the condition of the middle ear to be determined is fixated or detached.

[0092] As in this example, in this embodiment, a trained model 43 can be created that outputs information about the condition of the middle ear, such as the condition of the middle ear to be discriminated, using the feature quantities of the SPL curve obtained by measuring a subject whose middle ear condition is known as training data. The condition evaluation device 40 can output this information about the condition of the middle ear using the created trained model 43.

[0093] <Analysis example> Using measurement data from 11 ears with ossicular fixation, 12 ears with ossicular disruption, and 54 normal ears for which SPL curves could be obtained, we evaluated the ability of SVM to distinguish between ossicular fixation and disruption.

[0094] The four feature quantities used in the analysis were the minimum value, maximum value, RF, and ΔSPL obtained from the SPL curve. In order to visualize the data, in this example, principal component analysis was performed to obtain the first principal component (PC1) and second principal component (PC2), which have increasing variance, from the four feature quantities of the minimum value, maximum value, RF, and ΔSPL. In other words, the four-dimensional feature quantities of the minimum value, maximum value, RF, and ΔSPL were compressed into two dimensions.

[0095] Figure 20 is a plot of the data for fixation, detachment, and normality based on the first principal component (PC1) and the second principal component (PC2). From this figure, it can be seen that it is possible to classify whether or not there is detachment based on the first principal component. It can also be seen that it is possible to classify whether or not there is fixation based on the second principal component.

[0096] Figure 21 is a graph showing the factor loadings (principal component loadings) of the first and second principal components obtained by the principal component analysis described above. It was revealed that the contributions of RF and ΔSPL were large for the first principal component, while the contributions of the minimum and maximum values ​​were small. It was revealed that the contributions of the minimum and maximum values ​​were large for the second principal component, while the contributions of RF and ΔSPL were small. This suggests that transection tends to cause large changes in RF and ΔSPL. It also suggests that fixation, which reduces the elasticity of the eardrum, tends to cause large changes in the minimum and maximum values.

[0097] Classification analysis was performed using a two-class linear SVM on the data containing the first and second principal components shown in Figure 20. The classification tasks were "dissection vs. fixation and normal" and "fixation vs. dissection and normal." To evaluate the generalization performance of the SVM classification results, 1,000 validation runs were performed using the holdout validation method.

[0098] In this verification, data from seven ears were randomly selected for each of fixation, transection, and normal, and these were used as training data. In addition, in the verification of "transection vs. fixation and normal," data from four ears of transection and data from two ears each of fixation and normal were selected, and these were used as test data. Similarly, in the verification of "fixation vs. transection and normal," data from four ears of fixation and data from two ears each of transection and normal were selected, and these were used as test data.

[0099] Figure 22 shows an example of a single validation run of a two-classification task: "dissection vs. fixation and normal." In Figure 22, the straight line shown in the figure indicates the separating hyperplane of the SVM obtained by machine learning. In this example, four pieces of data on dissection and four pieces of data on fixation and normal are correctly classified.

[0100] Sensitivity, specificity, and accuracy were calculated for each validation run, and the mean and standard deviation of 1000 validation runs were calculated.

[0101] As a comparative example, a classification analysis using a two-class linear SVM was also performed on the data of peak pressure (TPP) and static compliance (SC) obtained from tympanometry measurement results, as shown in Figure 23, and the generalization performance of the classification results was evaluated. Peak pressure (TPP) is the ear canal pressure at which the maximum compliance value was obtained. For performance evaluation, 1,000 validation runs were similarly performed using the holdout validation method. Sensitivity, specificity, and accuracy were calculated for each validation run, and the average and standard deviation of the 1,000 validation runs were calculated.

[0102] These results are shown in Table 3. As shown in Table 3, it was clear that this embodiment can achieve high classification performance for both transection and fixation. In this embodiment, an SPL curve is obtained by measurement, and classification is performed using a two-class linear SVM that uses the feature quantities obtained from the SPL curve, namely, the minimum value, the maximum value, RF, and ΔSPL, and their first and second principal components. These classification performances were higher than those of classification using a two-class linear SVM that uses tympanometry measurement results.

[0103] [Table 3]

[0104] Various methods using SVMs may be employed, without being limited to the analysis examples shown here. For example, the above example is an example of a two-dimensional SVM using the first and second principal components after performing principal component analysis of the feature quantities, i.e., minimum and maximum values, RF, and ΔSPL. However, two or more feature quantities, such as minimum and maximum values, RF, and ΔSPL, may be used, and a two-dimensional or higher-dimensional SVM may be used. Furthermore, while the example described here is an example of determining whether or not a fixation or detachment is present, similar determinations may also be made regarding the presence or absence of otitis media, including serous otitis media, thinning or adhesion of the tympanic membrane, middle ear malformations, and other middle ear diseases.

[0105] [Variations] In the above embodiment, the sound pressure level for each frequency was measured when the pressure in the ear canal was atmospheric pressure. The same measurement may be performed by varying the pressure in the ear canal. For example, the pressure in the ear canal may be varied from +200 dPa to -200 dPa, and the relationship between the input frequency and the sound pressure level at each pressure may be measured.

[0106] 24 is a functional block diagram showing an outline of a configuration example of a condition evaluation system 2 according to a modified example having a mechanism for changing the static pressure in the ear canal. In addition to the configuration of the condition evaluation system 1 of the above-described embodiment, the condition evaluation system 2 includes a pressure control unit 110. The pressure control unit 110 operates under the control of the computer 10. The pressure control unit 110 includes a syringe pump 112 and a pressure sensor 114.

[0107] One end of a tube 122 is connected to the syringe of the syringe pump 112, and the other end of the tube 122 is connected to a hole 86 provided at the tip of the probe 80. In addition, a pressure sensor 114 is connected to the tube 122. The pressure sensor 114 is configured to measure the pressure inside the tube 122.

[0108] During measurement, the probe 80 is inserted into the ear canal 210 of the subject 200 to seal the ear canal 210. At this time, the condition evaluation system 2 can change the static pressure in the ear canal 210 by operating the syringe pump 112. The operation of the syringe pump 112 is controlled in accordance with the pressure in the tube 122 measured by the pressure sensor 114, i.e., the static pressure in the ear canal 210 of the subject 200. The recording unit 34 records the sound pressure measured by the microphone 84, as well as the static pressure in the ear canal 210 measured by the pressure sensor 114 or the target static pressure.

[0109] In this modification, instead of the SPL curve expressed in two dimensions of frequency and sound pressure level in the above-described embodiment, an SPL surface expressed in three dimensions of static pressure in the ear canal, frequency, and sound pressure is obtained. Various feature quantities can be acquired from such an SPL surface. For example, as in the above, the minimum value, maximum value, RF, ΔSPL, etc. of the SPL curve at each ear canal pressure can be acquired as feature quantities. Furthermore, values ​​based on changes in the minimum value, maximum value, RF, ΔSPL, etc. of the SPL curve depending on the ear canal pressure can be acquired as feature quantities. The acquired feature quantities can be used to determine the condition of the middle ear. For this determination, various trained machine learning models can be used.

[0110] For example, in the above-described embodiment, the accuracy of determining whether the ear is normal or stuck is lower than the accuracy of determining whether the ear is normal or disconnected. In this regard, in the case of normal ears, the change in the SPL curve in response to a change in static pressure in the ear canal 210 is relatively large, whereas in the case of fixation, the change in the SPL curve in response to a change in static pressure in the ear canal 210 is relatively small. Therefore, according to this modification, the determination of whether the ear is normal or stuck can be made more accurately based on information on, for example, changes in RF or ΔSPL due to changes in static pressure.

[0111] In addition, measurements that change the static pressure in the ear canal can generally provide more accurate judgments because they provide more information than measurements that do not change the static pressure in the ear canal. On the other hand, when measuring newborns or infants, for example, changing the static pressure in the ear canal may cause the sleeping subject to wake up, move, or cry, making accurate measurements impossible. Therefore, whether it is better to change the static pressure in the ear canal may vary depending on the subject.

[0112] [About machine learning models] Although the Mahalanobis distance and SVM have been described above as machine learning models, other machine learning models, such as neural networks including deep neural networks and random forests, may also be used. In the above-described embodiment, an example was given in which feature quantities such as the minimum and maximum values, RF, and ΔSPL of the SPL curve were used, but this is not limiting. For example, various middle ear conditions may be determined using a trained model obtained by deep learning using the SPL curve itself as input. Furthermore, various machine learning methods, not limited to supervised learning, may also be used, such as clustering using unsupervised learning.

[0113] The present invention has been described above by showing preferred embodiments, but it goes without saying that the present invention is not limited to the above-described embodiments, and various modifications can be made within the scope of the present invention. [Explanation of symbols]

[0114] 1, 2: Condition evaluation system 10: Computer, 11: CPU, 12: Memory, 13: Storage, 14: Interface, 15: Display device 22: control unit, 32: input signal generating unit, 34: recording unit, 52: signal output unit, 54: signal acquiring unit 40: State evaluation device, 41: Feature identification unit, 42: Data acquisition unit, 43: Trained model, 44: State information calculation unit, 45: Post-processing unit, 46: Output unit 431: First trained model, 432: Second trained model, 433: Third trained model, 434: Fourth trained model, 441: Index calculation unit 60: AD / DA converter, 62: DA converter, 64: AD converter 70: Amplifier system, 72: Earphone amplifier, 74: Microphone amplifier 80: Probe, 82: Earphone, 83: Vibration membrane, 84: Microphone, 86: Hole 91: Stimulus sound output section, 92: Sound receiving section 110: Pressure control unit, 112: Syringe pump, 114: Pressure sensor, 122: Tube 200: subject, 210: ear canal, 222: eardrum, 224: ossicles, 225: malleus, 226: incus, 227: stapes, 229: tympanic cavity, 232: cochlea

Claims

1. a data acquisition unit that acquires information about frequency characteristics of a sound pressure level measured as a pressure fluctuation in the ear canal when a stimulus sound is output into the ear canal of the subject, the information indicating the impedance of the subject's auditory organ; a state information calculation unit including a trained model created by machine learning configured to output information about the state of the auditory organ when information about the frequency characteristics is input; a post-processing unit that generates information about the state of the auditory organs of the subject based on information about the output of the state information calculation unit in accordance with the frequency characteristics of the subject; an output unit that outputs the information created by the post-processing unit; A condition assessment device for the hearing organ comprising:

2. the data acquisition unit is configured to acquire information relating to a feature amount of the frequency characteristics, The state information calculation unit is configured to receive information about the feature amount. The condition assessment device according to claim 1 .

3. The condition evaluation device according to claim 2 , wherein the condition information calculation unit is configured to perform processing related to calculation of a Mahalanobis distance between each group having a different middle ear condition and the group, the Mahalanobis distance being a variable of the feature amount.

4. 4. The condition evaluation device according to claim 3, wherein the condition information calculation unit is configured to calculate an index relating to a probability that the middle ear of the subject is in a predetermined condition, using the Mahalanobis distance.

5. The condition evaluation device according to claim 4 , wherein the condition information calculation unit is configured to determine whether or not the predetermined condition exists based on an index relating to a probability that the predetermined condition exists.

6. The condition assessment device according to claim 2 , wherein the trained model is configured using a support vector machine.

7. The state assessment device according to claim 6 , wherein the state information calculation unit is configured to perform principal component analysis on the information related to the feature amount and input a result of the principal component analysis to the trained model.

8. The condition evaluation device according to claim 6 , wherein the condition information calculation unit is configured to determine whether or not the middle ear of the subject is in a predetermined condition by using a two-class support vector machine.

9. 9. The condition evaluation device according to claim 2, wherein the feature quantity includes a value related to at least one of a minimum value and a maximum value of an SPL curve indicating the frequency characteristics of the sound pressure level measured in the subject, RF, which is a frequency related to resonance of the middle ear calculated from the frequency indicating the minimum value and the frequency indicating the maximum value, ΔSPL calculated from the difference between the minimum value and the maximum value, and a value based on a change in the minimum value, the maximum value, the RF, or the ΔSPL depending on the static pressure in the ear canal.

10. 9. A condition evaluation device according to claim 1, wherein the information regarding the state of the auditory organ includes at least one of information regarding the subject's middle ear, including normal, ossicular fixation, ossicular detachment, otitis media, thinning of the tympanic membrane, and adhesion of the tympanic membrane.

11. further comprising a feature quantity specifying unit configured to calculate the feature quantity from data indicating the frequency characteristics of the sound pressure level, 9. The condition evaluation device according to claim 2.

12. A hearing organ condition assessment device according to any one of claims 1 to 8; a stimulus sound output unit configured to output a stimulus sound including at least each frequency component of a measurement target frequency band toward the ear canal of the subject; a sound receiving unit configured to acquire a sound pressure signal indicating a pressure fluctuation in the ear canal when the stimulation sound is output; a recording unit configured to record information about the sound pressure signal acquired by the sound receiving unit as data of the measured sound pressure level; A condition assessment system for hearing organs, comprising:

13. the condition evaluation device includes a feature quantity specifying unit configured to calculate a feature quantity from data indicating frequency characteristics of the sound pressure level, The condition assessment system of claim 12.

14. Computer, a data acquisition unit that acquires information about frequency characteristics of a sound pressure level measured as a pressure fluctuation in the ear canal when a stimulus sound is output into the ear canal of the subject, the information indicating the impedance of the subject's auditory organ; a state information calculation unit including a trained model created by machine learning configured to output information about the state of the auditory organ when information about the frequency characteristics is input; a post-processing unit that generates information about the state of the auditory organs of the subject based on information about the output of the state information calculation unit in accordance with the frequency characteristics of the subject; an output unit that outputs the information created by the post-processing unit; A condition assessment program for the hearing organs to function as.

Images