TW202135048A - Method of extracting acoustic features in a disease estimation program, and disease estimation program and an apparatus using the acoustic features - Google Patents

Abstract

In an apparatus for estimating a plurality of psychiatric and neurological disorders by means of speech analysis, an estimation apparatus equipped with means for extracting acoustic features unaffected by the location of speech acquisition, and a method for operating the estimation apparatus are provided.

Description

Method for extracting sound feature quantity in disease estimation program, and disease estimation program and device using the sound feature quantity

The present invention relates to a method for extracting sound feature quantities that are not related to the environment in a disease estimation program, and a disease estimation program and a disease estimation device that use sound feature quantities that are not related to the environment.

Techniques for estimating emotions by analyzing the voice of the subject are gradually becoming popular. Patent Document 1 discloses a technique that converts the subject's voice to a frequency spectrum, slides on the frequency axis to obtain an autocorrelation waveform, and calculates the fundamental frequency therefrom to estimate the emotional state. Prior art literature Patent literature

Patent Document 1: International Patent Publication No. 2006/132159

The problem to be solved by the invention

However, when a user inputs a voice in a room such as a home or a medical institution, depending on the place where the voice is obtained, sound interference may occur due to the reflected sound generated by the walls, floor, ceiling, etc. constituting the room. This kind of sound interference may change the amount of sound features extracted from the input sound and reduce the accuracy of disease estimation, but Patent Document 1 does not mention this problem.

In addition, the device of Patent Document 1 only ends with estimating the emotional state of the user, and does not mention a program for estimating mental system diseases or neurological diseases (hereinafter sometimes referred to as neuropsychiatric diseases). Generally, since there are no effective biomarkers, etc., it is difficult to estimate diseases from various types of neuropsychiatric diseases.

For example, according to the diagnostic criteria in the DSM-5 manual published by the American Psychiatric Association (APA), the difference between Levitic Cognitive Impairment and Parkinson's Disease depends on the location of Levitic, so the symptoms may be similar. In addition, Alzheimer’s disease and frontotemporal dementia, Alzheimer’s disease and Levitic cognitive impairment, Levitic cognitive impairment and Parkinson’s disease, and the relationship between bipolar disorder and severe depression Diseases are difficult to distinguish.

In addition, patients often show symptoms of depression, which is an early symptom of dementia. On the other hand, when the patient suffers from melancholic pseudodignosis, the patient is actually melancholic, but its symptoms are manifested as cognitive decline. Therefore, it is difficult to distinguish whether the patient is suffering from a certain type of dementia, depression, or a combination of dementia and depression, and one of the symptoms is obvious.

If the disease cannot be inferred from multiple types of psychiatric and neurological disease candidates, it is possible for the patient to make a mistake in selecting the medical institution to be treated, thereby inadvertently worsening the symptoms.

Therefore, the object of the present invention is to provide an estimation device and a method for operating the estimation device in an apparatus for estimating various mental and neurological diseases through speech analysis. The estimation device has a means for extracting sound feature quantities that are not affected by the place where the speech is obtained. . Technical means to solve the problem

As a result of in-depth research to solve the above-mentioned problems, the inventor of the present application discovered an estimation device and a method of operating the estimation device in a device for estimating a variety of mental and neurological diseases, wherein the estimation device has the ability to extract Voice is a means to obtain the voice feature quantity affected by the place.

In other words, the present invention includes the following aspects. [1] An estimation device for mental and nervous system diseases, including: Based on the voice feature quantity (A) and the voice feature quantity related to various diseases (B) that do not significantly differ with the recording environment, the common voice feature quantity (A) and the voice feature quantity (B) are extracted ( C) the extraction unit; A calculation unit that calculates the disease prediction value based on the above-mentioned voice feature quantity (C); and The estimation unit of the disease is estimated by inputting the above-mentioned disease prediction value. [2] The estimating device according to [1] above, wherein the estimable candidates for the disease include Neurocognitive Disorder Due to Alzheimer's Disease, Inrevi Neurocognitive Disorder with Lewy Bodies, Parkinson’s Disease, Bipolar Disorder, and Depressive Disorder with atypical features. with Atypical Features) and Major Depressive Disorder. [3] The estimation device described in [1] above, wherein one of the disease candidates that can be estimated is severe depression. [4] An operating method of an estimation device, including: The extraction unit of the estimation device extracts the above-mentioned sound characteristic quantities (A) and the above-mentioned sound characteristic quantities (B) based on the sound characteristic quantities (A) and various diseases-related sound characteristic quantities (B) that do not significantly differ with the recording environment Common voice feature quantity (C) steps; The step of calculating a disease prediction value based on the voice feature quantity (C) by the calculation unit of the estimation device; and The step of estimating the disease by inputting the predictive value of the disease by the estimating unit of the estimating device. Invention effect

The present invention can provide a method for extracting voice feature quantities that are not affected by the location of voice acquisition in a program for estimating various mental and neurological diseases, and a device using the method.

In the following, the device for estimating various mental and neurological diseases of the present invention will be described in detail, but the configuration requirements described below are an embodiment of the present invention and are described in detail in these contents. In the following description, the disease predictive value can be referred to as the "mental value". ＜1. Program＞

The estimation apparatus 200 of this embodiment is realized by, for example, a computer 100 having the configuration shown in FIG. 1. In the following, a description will be given using examples. FIG. 1 is a hardware configuration diagram, which shows an example of a computer that implements the functions of the estimation device 200. The computer 100 has a central processing unit (Central Processing Unit, CPU) 101, a random access memory (Random Access Memory, RAM) 102, a read-only memory (Read-Only Memory, ROM) 103, and a hard disk (Hard Disk Drive, HDD) 104, communication interface (Interface, I/F) 105, input/output interface 106, and media interface 107.

The CPU 101 operates based on programs stored in the ROM 103 or the HDD 104, and controls each part. When the computer 100 is started, the CPU 101 will execute the startup program stored in the ROM 103, which depends on the program of the hardware of the computer 100 and so on.

The HDD 104 stores programs executed by the CPU 101 and data used by the programs. The communication interface 105 receives data from another device via the network N and sends it to the CPU 101, and sends the data generated by the CPU 101 to the other device.

The CPU 101 controls an output device such as a display, a voice input device such as a microphone, and an input device such as a keyboard and a mouse via the input/output interface 106. The CPU 101 obtains voice data from the input device via the input/output interface 106. In addition, the CPU 101 outputs the generated data to the output device via the input/output interface 106.

The media interface 107 reads programs or data stored in the recording medium 108 and provides them to the CPU 101 via the RAM 102. The CPU 101 loads the program from the recording medium 108 to the RAM 102 via the media interface 107, and executes the loaded program. The recording medium 108 is, for example, an optical recording medium such as a DVD (Digital Versatile Disc) or a phase change rewritable Disk (PD), and an optical recording medium such as a magneto-optical disk (MO). Magnetic recording media, magnetic tape media, magnetic recording media or semiconductor storage, etc.

For example, when the computer 100 executes the estimation apparatus 200 according to the embodiment, the CPU 101 of the computer 100 implements the function of the control unit by executing the program loaded on the RAM 102. In addition, the data in the memory unit is stored in the HDD104. The CPU 101 of the computer 100 reads and executes these programs from the recording medium 108, but as another example, these programs can be obtained from another device. ＜2. Configuration of estimation device＞

Next, the configuration of the estimation apparatus 200 according to the embodiment will be described with reference to FIG. 2. As shown in FIG. 2, the estimation device 200 is communicably connected to the client 201 via the network N in a wired or wireless manner. The estimation device 200 can also be connected to multiple client terminals 201.

As shown in FIG. 2, the estimation device 200 includes a communication unit 202, a voice feature extraction unit 203, a first voice feature extraction unit 204, a second voice feature extraction unit 205, a calculation unit 206, An estimation unit 207 and a memory unit 208. The sound feature extraction unit 203, the calculation unit 206, and the estimation unit 207 are executed by a calculation processing device (CPU), and cooperate with each other as a control unit (not shown).

The communication unit 202 is, for example, implemented by a network interface card (NIC) or the like. The communication unit 202 is connected to the network N in a wired or wireless manner, and sends/receives information to/from the client 201.

The control unit can be realized by, for example, a central processing unit (CPU) or a micro processing unit (Micro Processing Unit, MPU), etc., and executes various programs stored in the memory unit 208 by using RAM as a work area. In addition, the control unit is realized by, for example, a body circuit such as an Application Specific Integrated Circuit (ASIC) or a Field Programmable Gate Array (FPGA).

The memory unit 208 is realized by, for example, a semiconductor storage device such as RAM (Random Access Memory) or Flash Memory, or a storage device such as a hard disk or an optical disk. The user terminal 201 includes a voice input unit and an estimation result output unit. The estimation device 200 obtains the user's voice from the input unit, converts the user's voice from an analog signal to a voice data of a digital signal, and stores the voice data in the memory unit 208 via the communication unit 202.

The input unit obtains the voice signal spoken by the subject through a voice acquisition unit such as a microphone, and generates digital signal voice data by sampling the voice signal at a predetermined sampling frequency (for example, 11025 Hz, Hertz). The input unit may include a memory unit for separately recording voice data with the memory unit 208 of the estimation device 200. In this case, the input unit may be a portable recorder. The memory unit of the input unit may be a recording medium such as CD, DVD, USB storage, SD card, or compact disc.

The output unit includes a receiving unit that receives data such as the estimation result and a display unit that displays the data. The display unit is a display that displays data such as estimation results. The display may be an organic electro-luminescence (Organic Electro-Luminescence) or a liquid crystal display. ＜＜Extraction unit 203＞＞

The extraction unit 203 has a first sound feature extraction unit 204 and a second sound feature extraction unit 205. Here, the extraction unit 204 of the first sound feature quantity is used to create a set of the first sound feature quantity. Multiple healthy subjects are preliminarily speaking the same content in multiple places, the acquired speaking voice is labeled and normalized, then the voice analysis is performed to extract multiple feature quantities, and the multiple feature quantities are transformed into The paired t-test is compared, and a group of sound feature quantities that are not significantly different in any place are defined as the set of first sound feature quantities. As an example of a collection of voice feature quantities with no significant difference, in the paired t test, a collection of voice feature quantities with a P value of more than 0.05 is preferable, and a collection of voice feature quantities with a P value of more than 0.1 is more preferable. The theoretical numerical range of the P value is 0 to 1, and the significance level of the P value is usually set to 0.05.

The set of first sound feature quantities is stored in the memory unit 208. The set of first voice feature amounts may be used together with the set of second voice feature amounts described later, or only the set of first voice feature amounts may be used as a feature amount that has nothing to do with the environment.

The extraction unit 205 of the second sound feature quantity creates a set of second sound feature quantities. A number of healthy subjects are pre-said different speech content in multiple places, the obtained speech speech is labeled and normalized, and then the speech analysis is performed to extract multiple feature quantities, and the multiple feature quantities are transmitted through non- Paired t-tests (Unpaired t-test) are compared, and a group of sound feature quantities that are not significantly different in any place are defined as the set of second sound feature quantities. As an example of a collection of voice feature quantities with no significant difference, in the paired t test, a collection of voice feature quantities with a P value of more than 0.05 is preferable, and a collection of voice feature quantities with a P value of more than 0.1 is more preferable.

The set of second sound feature quantities is stored in the memory unit 208. The set of second sound feature amounts may be used together with the set of first sound feature amounts, or only the set of second sound feature amounts may be used as a feature amount that has nothing to do with the environment.

Next, the basis for setting the threshold of the P value will be explained. FIG. 6 shows an example of voice analysis based on the utterance of a healthy subject to extract voice feature amounts, and there is a significant difference in voice feature amounts in paired t-tests or t-tests. On the other hand, FIG. 7 shows an example in which voice analysis is performed based on the utterance of a healthy subject to extract voice feature amounts, and there is no significant difference in voice feature amounts in paired t-tests or t-tests. Let healthy subjects speak the same content or different content in different places to obtain the spoken voice, and when comparing a certain voice feature amount, there is a significant difference as shown in Figure 6, indicating the difference in the attributes of the voice It only lies in the environment, and it is strongly suspected that it is the sound characteristic amount related to the environment. Therefore, when the set of sound feature quantities exceeds 0.05 at the P value, that is, there is no significant difference as shown in FIG. 7, it can be selected as a sound feature quantity that has nothing to do with the environment.

In addition, if the P value of the set of sound feature quantities exceeds 0.1, it means that when a healthy subject is going to each place, the sound feature quantity is not affected by slight physical conditions, and it can be selected as a sound feature quantity that is not related to the environment. . In addition, if the P value of the set of sound feature quantities exceeds 0.1, it means that it is difficult to be affected by at least one sound feature quantity used for disease estimation (the feature quantity F(a) is described below). From the perspective of creating a disease estimation program Look is better.

Next, a method of creating a set of first sound feature amounts will be described in more detail. Here, in order to eliminate the difference caused by the environment of the place, the significant difference in the sound characteristic amount between the places is measured. For example, for the voices collected in seven locations (respectively called location 1 to location 7), like location 1 and location 2, location 1 and location 3 generally create ₇ C ₂ pairs, and the extraction is not significant in any pair Differential sound characteristics (paired t test). This paired t test is to obtain the speech voice of one or more healthy subjects in all places as the target. Here, a healthy subject refers to a person who has not suffered from the disease targeted for analysis.

This paired t test allows only one healthy subject to participate. It is better to allow more than 2 people to participate in order to increase credibility, and it is even better to allow more than 3 people to participate. In addition, when performed by multiple healthy subjects, voices acquired in the same place can be processed collectively or individually for multiple persons. In the separate processing, the detected logarithm becomes ₇ C ₂ × the number of people.

In addition, when a healthy subject utters multiple phrases in each place to obtain speech, those phrases can be processed collectively or individually. In the separate processing, the voice feature quantity with no significant difference is extracted for each phrase.

Next, a method of creating a set of second sound feature amounts will be described in more detail. Here, in order to eliminate the difference caused by the patient group (and the healthy subject group), the significant difference in the voice feature amount of the patient group was measured. For example, if the voices of multiple patients with severe depression (severe depression group A) are acquired within a certain period of time, the voices of multiple healthy subjects (group A of healthy subjects) are acquired at the same time, and the voices of multiple healthy subjects are acquired in another period. If the voices of multiple patients with severe depression (severe depression group B) are obtained, and if multiple healthy subjects’ voices are obtained in the same period (healthy subjects group B), then t-test (unpaired t-test) Measure the significant differences in voice characteristics of each group (severe depression group A and severe depression group B, healthy subject A and healthy subject B) in the same disease (or healthy subject). In addition, when patients in each group utter multiple phrases to obtain speech, those phrases can be processed collectively or individually. In the separate processing, the voice feature quantity with no significant difference is extracted for each phrase.

The voice feature value extraction unit 203 compares the first voice feature value set and the second voice feature value set that exceed the desired P value, and defines the common voice feature value set as the first voice feature set that is not affected by the location where the voice is obtained A collection of three sound feature quantities. In addition, it is also possible to define only the set of first voice feature amounts that exceed the desired P value as a set of third voice feature amounts that are not affected by the location where the voice is acquired.

The third set of sound feature quantities is used when extracting at least one set of sound feature quantities (feature quantity F(a)) for calculating the predictive value of multiple diseases. For example, extracting at least one set of common feature values used to calculate at least one set of voice feature amounts for multiple diseases and the third set of voice feature amounts as a predictive value for real multiple diseases The sound feature quantity (feature quantity F(a)). ＜＜Processing flow of extraction unit 203＞＞

Here, the processing flow of the extraction unit 203 will be described with reference to FIG. 3. When the operation starts, in step S1001, the extraction unit 203 performs a speech marking operation on the speech data stored in the storage unit 208 after the speech is obtained in advance. Next, in step S1002, the extraction unit 203 performs normalization processing on the speech data for which the speaking mark has been completed. After the normalization process is performed, the pre-processing operation is completed. Next, in step S1003, the extraction unit 203 extracts the voice feature amount from the previously processed speech data.

Next, in step S1004A, the extraction unit 204 of the first voice feature amount of the extraction unit 203 compares the extracted voice feature amount with the speech obtained by multiple healthy subjects uttering the same content in multiple places in advance. The voice features created by speech are compared with Paired t-test. Next, in step S1005A, the extraction unit 204 of the first sound feature quantity defines a set of sound feature quantities that are not significantly different in any place as the set of first sound feature quantities according to the desired threshold of the P value.

On the other hand, in step S1004B, the second voice feature value extraction unit 205 of the extraction unit 203 obtains the extracted voice feature value from the different speech content of a plurality of healthy subjects uttered in a plurality of places in advance. The voice feature volume created by the speaking voice of, is compared with the unpaired t-test. Next, in step S1005B, the second voice feature amount extraction unit 205 defines a set of voice feature amounts with no significant difference between any places as a second voice feature set based on the desired threshold of the P value.

Next, in step S1006, the voice feature value extraction unit 203 compares the first voice feature value set exceeding the desired P value with the second voice feature value set, and defines the common voice feature value set as not receiving The collection of the third voice feature quantity affected by the location where the voice is acquired, and the operation ends. In addition, when only the set of first sound feature amounts based on the value exceeding the desired P value is defined as the set of third sound feature amounts, step S1006 may be omitted.

By performing the above processing, the set of third voice feature quantities that are not affected by the location where the voice is acquired is combined with at least one set of voice feature quantities (feature quantity F(a)) used to calculate the predictive value of multiple diseases. Make more accurate disease estimates. ＜＜Calculation unit 206 and estimation unit 207＞＞

The calculation unit 203 calculates predictive values of multiple diseases based on the following inference models of diseases and based on a combination of at least one voice feature amount. The estimation unit 207 takes the predicted value of the disease as input to estimate a variety of neuropsychiatric diseases. Next, the calculation unit 206 and the estimation unit 207 are described in detail. ＜＜Calculation of predictive value of disease＞＞

The calculation of the disease predictive value will be briefly described. The calculation unit 206 undergoes a step of extracting a plurality of voice feature quantities from the voice data of the subject. The voice feature quantity is extracted from the patient's voice data. The sound characteristic amount refers to the characteristic amount of the characteristic when the sound is transmitted. For example, the sound feature quantity includes a zero-point crossing rate and a Hurst index. The zero crossing rate is calculated by calculating the number of times the sound pressure waveform of the voice intersects the reference pressure per unit time, as the severity of the waveform change in the voice. The Hurst index represents the correlation of changes in the voice waveform.

From here, explain the formula for estimating the disease. Since it is necessary to distinguish it easily from the set of the aforementioned first voice feature quantity to the third voice feature quantity in order to describe the voice feature quantity, the "sound feature quantity" is referred to as the "sound parameter" for explanation. However, in the description of the present invention, the "sound feature amount" and the "sound parameter" are essentially the same, and both are used as the input of the inference device, and have the meaning and degree of the column representing the entity feature.

The sound parameters used by the disease estimation device include a first sound parameter and a second sound parameter. The first sound parameter is a parameter of a sound extracted from the voice of a subject whose specific disease should be estimated. The second sound parameter is a parameter of the sound recorded in the memory unit 208 in advance. The second voice parameter is extracted from the voice data of patients suffering from Alzheimer’s disease, Levitic cognitive impairment, Parkinson’s disease, severe depression, atypical depression or bipolar disorder, and preliminarily Each sound parameter is linked to each disease.

The sound parameters used in the present invention include the following items: 1) Volume envelope (rise time, decay time, maintenance level, release time) 2) Waveform change information (shimmer, jitter) 3) Zero crossing rate 4) Hurst index 5) Voice Onset Time (Voice Onset Time, VOT) 6) Statistical value of sound distribution of Mel frequency cepstral coefficient (first quartile, median, third quartile, 95% point, arithmetic mean, geometric mean, third quartile Difference from the median, etc.) 7) Intra-phonetic distribution statistics of the rate of spectral change (first quartile, median, third quartile, 95% point, arithmetic mean, geometric mean, third quartile Difference from the median, etc.) 8) The statistic value of the voice distribution of the Mel frequency cepstral coefficient over time (the first quartile, the median, the third quartile, the 95% point, the arithmetic mean, the geometric mean, the third The difference between the quartile and the median, etc.) 9) The statistical value of the voice distribution relative to the time change of the time change of the Mel frequency cepstral coefficient (first quartile, median, third quartile, 95% point, arithmetic mean, geometric Mean, the difference between the third quartile and the median, etc.) 10) Approximate square error of quadratic regression between time change in utterance and 90% spectrum roll-off 11) The arithmetic error of the quadratic regression approximation of the time change of the spectral gravity center sound In addition, other factors include pitch rate, probability of utterance, frequency power within any range, scale, speaking speed (number of voices in a certain period of time), pause/interval, volume, etc.

The estimation program has a learning function of artificial intelligence, and performs estimation processing through its learning function. The reasoning model can use classification algorithms, such as linear model regression, linear regression, rigid regression, Lasso, or logistic regression. You can also use deep learning like neural networks, or you can use reinforcement learning in some areas of reinforcement learning. Others can also use genetic algorithms, cluster analysis, self-organizing maps, integrated learning, etc. Of course, other technologies related to artificial intelligence can also be used. In ensemble learning, classification algorithms can be created by using both enhancement and decision tree methods.

In the stage of creating the estimation program, the algorithm creator uses the stepwise regression method to obtain a better combination of the arbitrary sound parameters used as the variable f(n) from the above-mentioned second sound parameter items, and selects one or more Sound parameters. Then, coefficients are assigned to the selected arbitrary sound parameters, and one or more sound parameters are created. Then combine these sound parameters to create F(a).

There are three types of stepwise regression methods: variable increase method, variable decrease method and variable increase and decrease method, any of which can be used. The regression analysis used in the stepwise regression method includes linear classification processing such as linear discriminant and logistic regression analysis. The variable f(n) and its coefficients, that is, the coefficient xn of F(a) in the following formula, are called regression coefficients, which give weight to the function f(n).

After the creator of the learning algorithm selects the regression coefficients, the disease information accumulated in the database can be used to improve the quality through machine learning to improve the accuracy of the estimation.

The disease predictive value of the subject can be calculated based on the following formula F(a) to calculate more than one sound parameter, for example.

Formula 1 F (a) = x1 x f (1) + x2 x f (2) + x3 x f (3) + ··· + xn x f (n)

Here, f(n) is one or more second sound parameters arbitrarily selected from the above-mentioned sound parameter items (1) to (11). xn is the regression coefficient inherent to the disease. f(n) and xn can be recorded in the recording device 120 of the estimation program in advance. The regression coefficient of the parameter F(a) can be improved in the process of machine learning of the estimation program.

The calculation unit 206 in FIG. 2 distinguishes between healthy subjects and disease subjects based on the combination of the second sound parameters, or calculates parameters for distinguishing the disease subjects. From this parameter, by scoring the calculation reference range and the distance between the subject’s value and the reference range, the subject’s disease prediction value is calculated.

Fig. 8 shows a graph of different intensities of a certain sound parameter for each disease, which shows that disease A has the highest score. Therefore, the predicted value of the subject relative to disease A is calculated, which is higher than that of other disease groups. In addition, for example, if the intensity 50 is set as a threshold, the disease groups of disease A, disease D, and disease E can be classified from disease groups of disease B and disease C.

Although Figure 8 calculates the predictive value of a disease based on the intensity of a sound parameter, in fact it is difficult to classify a disease with only one sound parameter. Therefore, it is also possible to classify diseases by combining several sound parameters to calculate the required parameter F(a).

Based on the parameter F(a), the predicted value of the disease is calculated for the voice of the marked subject, and the distribution of the predicted value of each disease is obtained. In this way, each disease can be classified.

Figure 9 is a distribution diagram of disease prediction values (described as "mental value" in Figure 9) obtained through the combination of three sound parameters

It can be seen from Fig. 9 that the distribution of the predictive value in the Levitic dementia group can be distinguished from the distribution of the predictive value in the group of patients with other diseases and the group of healthy subjects. In the present invention, a combination of sound parameters is set for each disease to distinguish it from other diseases, the parameter F(a) is calculated, and the parameter F(a) as the combination of the sound parameters is input, so that each subject can be determined What kind of disease does your voice apply to.

Another method is to extract the parameter F(a) of each disease from the voice of each patient, find out which disease has more parameters, and then compare the disease prediction values with each other to estimate the disease of the patient. .

In this case, the predictive value of the disease can be regarded as the degree of influence by the disease. By comparing the predictive value of each disease, the probability of being affected by the disease can be expressed.

In this way, from the voices of patients with six diseases including Alzheimer’s disease, Levitic dementia, Parkinson’s disease, severe depression, atypical depression, and bipolar disorder, as well as from healthy subjects The parameter F(a) related to each disease is extracted from the voice of, and the predictive value of each disease is calculated.

In addition, regarding the target disease, it is estimated that the program can be created from the speech of patients with 10 diseases, including four additional diseases: Vascular Neurocognitive Disorder, Frontotemporal Neurocognitive Disorder, Cyclic mood disorder (Cyclothymic Disorder) and persistent depression (Persistent Depressive Disorder).

Finally, by analyzing and judging the voice spoken by the subject, the estimation unit 207 estimates whether the subject has any of the above 6 to 10 diseases, or is healthy.

Regarding the estimation process of the estimation program, as described above, the disease prediction value can be calculated for each individual disease by extracting the feature quantity F(a) of each disease, but first, the sound feature quantity related to a group of diseases can be created In combination, the feature quantity F(a) related to the group of diseases can be used as the input of the estimation unit, and then the input and estimation of multiple stages can be divided into two or more stages, and the final estimation is made for each disease or health state . ＜＜Processing of estimation device＞＞

FIG. 4 shows an example of the estimation process of the estimation device 200 shown in FIG. 2. FIG. 4 is implemented by the calculation processing device (CPU) of the estimation device 200 executing the estimation program stored in the memory unit 208 of the estimation device 200.

When the process starts, in step S2001, the control unit obtains voice data. The voice data can be obtained from the input unit of the user terminal 201 or stored in the memory unit 208 before being read by the control unit. Next, in step S2002, the voice feature amount extraction unit 203 extracts the first voice parameter from the voice data. Next, in step S2003, the sound feature quantity related to the environment is excluded from the first sound parameter, and the processed first sound parameter is extracted. For example, by comparing the first sound parameter with the third sound feature quantity obtained by the extraction unit 203, a part that is not in common can be determined as a sound feature quantity related to the environment.

Next, in step S2004, the calculation unit 206 compares the parameter F(a) obtained from the second sound parameter with the processed first sound parameter obtained in step S2003 to calculate the predicted value of each disease.

Next, in step S2005, the estimating unit 207 distinguishes the multiple patients for which the disease prediction value is calculated from the patients to be designated and other patients through each threshold for distinguishing the specific disease from other diseases, and ends the process. In the following embodiment, the judgment is made by classifying the cases where the threshold value is exceeded and the cases where the threshold value is not exceeded. ＜3. Application area of the program＞

Since the estimation program of the present invention can even analyze voices from remote locations, it can be used in online medical and online consultation scenarios. When diagnosing neuropsychiatric diseases, doctors will observe the patient's facial expressions, movements and conversations through consultations or interviews. However, patients may hesitate to go to psychiatric hospitals and clinics due to their prejudice against neuropsychiatric diseases.

You can conduct online diagnosis, treatment and consultation with doctors and counselors without going to a medical institution. Therefore, compared with other diseases other than neuropsychiatric diseases, online diagnosis and treatment of neuropsychiatric diseases has a high affinity.

When talking about patients (or clients) online, doctors, consultants and clinical psychologists can use this estimate for analysis. In this way, it is very easy to estimate whether there is a neuropsychiatric disease and the type of the disease. In addition, during the interview, various psychological tests and cognitive function tests, such as MMSE, BDI, PHQ-9, etc., can be taken at the same time.

In this case, the patient needs computer hardware capable of transmitting voice, as well as a monitor screen for interviews and a microphone for voice recording.

If these devices are not in the patient's home, they can be installed in, for example, a home clinic. Patients can go to the family clinic and have an interview through the device there.

In addition, for example, when a patient goes to a family clinic to treat a physical disease, if the family doctor is diagnosed with a neuropsychiatric disease, he can obtain the voice on the spot, and then use the program of the present invention for analysis.

Also in other places, as long as psychiatrists and neurologists are in a state where they can perform online diagnosis and treatment, family doctors, psychiatrists, and neurologists can collaborate online for diagnosis.

By increasing the sensitivity for estimating specific diseases (in this case, the specificity is usually reduced), the estimation program of the present invention can be used as a screening device.

By using it as a test item for health checkups conducted by companies and local governments and human body checkups conducted by medical institutions, it is helpful for early detection of neuropsychiatric diseases that have been difficult to detect so far and there is no simple detection method.

For example, like fundus examinations, vision tests, hearing tests, etc., the acquired voice can be used as one of a series of tests, and the estimated results of the program can be notified on the spot, or together with other test results.

Since the estimation program of the present invention does not require a special device, anyone can easily use it. On the other hand, the frequency of use is not always very high because it is limited to occasions of neuropsychiatric diseases. Therefore, as long as the set of estimating devices of the present invention is installed in a specialized hospital equipped with an expensive examination device, when the target patient comes to the hospital for treatment, the family doctor or the like can request the specialized hospital to perform the examination.

Devices used in neuropsychiatric diseases include optical topography, myocardial scintigraphy, cerebral blood flow scintigraphy, CT, MRI, brain wave, etc. These devices are used for disease estimation and exclusion diagnosis, but since the estimation device of the present invention is extremely low in invasiveness, they can be used in combination with these examinations or used before these examinations.

Since the estimation program of the present invention can be easily used at home, it can be used as a monitoring device after diagnosis. For example, in the case of a disease in the mood disorder group, medication and psychotherapy are performed against the patient's disease, and the effectiveness of these therapies can be evaluated. In addition, through continuous use, you can observe whether the symptoms are stable and whether there are signs of recurrence every day.

Since the estimation program of the present invention analyzes the speech produced by speaking, it can be used as a caregiving device for the elderly.

Whether the living conditions of elderly people living alone are good is a concern of their close relatives. By implementing the presumptive program of the present invention in the elderly care system using telephone or videophone, etc., it is possible not only to watch life reactions, but also to estimate whether there is a tendency for dementia and depression. Therefore, even if a person lives alone, he can take appropriate measures. Responses.

In these various embodiments, there are no special restrictions on the method of obtaining the voice, such as (1) the method of sending the recorded voice from the subject through the telephone or the Internet, (2) the inspector contacting the subject through the telephone or the Internet , And obtain the voice through conversation, (3) install the voice acquisition device in the subject’s residence, and use the device to record the subject; (4) the voice acquisition device is automatically and regularly activated, And the method of obtaining the subject’s voice through dialogue with the subject and so on.

When the voice is acquired, preferably, the sentence to be spoken is displayed on the display provided in the estimation device, or the sound of the sentence to be spoken is played from the speaker, so that the subject can read aloud smoothly. Start recording through the mechanical sound at the beginning of the recording, and when the speech ends, end the recording through the switch, and the spoken voice can be obtained for each sentence. ＜4. Estimation program flow＞ ＜＜Estimation process-1＞＞

For example, in the first step, three groups are estimated: (1-A) cognitive impairment group, including Alzheimer’s disease, Levitic cognitive impairment and Parkinson’s disease; (1-B) mood disorder group, including severe Depression, atypical depression and bipolar disorder; and (1-C) health group.

Next, in the second step, the disease of patients in the cognitive impairment group is estimated by a program, which can be estimated from the voices of patients classified as (1-A) cognitive impairment group (1-A-1) Ah Which of the three diseases: Zheimer's disease, (1-A-2) Levitic body cognitive impairment, and (1-A-3) Parkinson's disease. On the other hand, a program is used to estimate the disease of patients in the mood disorder group. This program can estimate (1-B-1) severe depression, ( 1-B-2) Which of the three diseases, atypical depression and (1-B-3) bipolar disorder? ＜＜Estimation process-2＞＞

As another form of the judgment process, in the first step, three groups are estimated: (2-A) Cognitive impairment group, including Alzheimer's disease, Levitic cognitive impairment and Parkinson's disease; (2-B) Mood disorder group, including severe depression, atypical depression and bipolar disorder; and (2-C) healthy group.

Next, in the second step, the voice of patients classified as (2-A) cognitive impairment group is estimated by a program to estimate whether it is (2-A-1) Levitic cognitive impairment, and the program estimates whether the patient is Is suffering from Levitic cognitive impairment or other cognitive impairment. On the other hand, the voice of patients classified into the (2-B) mood disorder group is estimated to be (2-B-1) severe depression through a program that estimates whether the patient has severe depression or other Emotional disorders.

And similarly, estimate whether you have (2-A-2) Alzheimer’s disease or other dementia programs, and estimate whether you have (2-A-3) Parkinson’s disease or other dementia programs , Estimate whether you have a program of atypical depression or other mood disorders, and estimate whether you have a program of bipolar disorder or other mood disorders. By using this program, you can finally determine whether you have Alzheimer's disease, Levitic cognitive impairment, Parkinson's disease, severe depression, atypical depression, or bipolar disorder. ＜＜Estimation process-3＞＞

In another embodiment, in the second step, the voices of patients classified as (3-A) cognitive impairment group are used to estimate whether they have (3-A-1) Alzheimer’s disease or Levitic cognition Disordered program, estimated whether you have (3-A-2) Levitic cognitive impairment or Parkinson’s disease, estimated whether you have (3-A-3) Parkinson’s disease or Alzheimer’s disease Program. The three estimation programs are used together to estimate the disease of patients classified as cognitive impairment group.

On the other hand, by using the voice of patients classified as (3-B) mood disorder group, using the program that estimates whether they have (3-B-1) severe depression or atypical depression, and whether they have (3-B) -2) A program of atypical depression or bipolar disorder, and a program that estimates whether you have (3-B-3) bipolar disorder or severe depression. The three estimation programs are used together to estimate the disease of the patients classified as the mood disorder group. ＜＜Estimation process-4＞＞

In addition, as described above, it is divided into three categories: cognitive impairment group, mood disorder group and healthy group. The first step is to divide it into (4-A) healthy group and (4-B) other disease group. The next step is to divide the disease group into (4-B-1) cognitive impairment group and (4-B-2) ) Two stages for the mood disorder group. (The subject of the sound obtained when creating the estimation program)

The acquisition of the voice data of the second voice parameter will be described. Preferably, the subject whose voice is to be obtained is selected according to the following criteria.

(A) After obtaining a sufficient explanation, the subject should agree in writing to use the obtained speech for disease analysis.

(B) This estimation system does not perform disease estimation analysis based on the meaning or content (text) of words. Therefore, there are basically no restrictions on nationality or mother tongue. However, because there may be differences between race and language, it is best to identify the same race and language as a comparison. For example, from a convenient point of view, when the present invention is implemented in Japan, it is better to create the estimation program by obtaining the voice of a person whose mother tongue is Japanese, and to estimate a disease through the Japanese voice. In the case of implementing the present invention in an English-speaking country, it is better to create an estimation program by obtaining the voice of a native English speaker, and to estimate the disease through the English voice.

(C) As long as the person has language ability, there is no special restriction on age. However, in consideration of changes in voice and emotional stability, it is preferably 15 years or older, more preferably 18 years or older, and particularly preferably 20 years or older. In addition, taking into account the difficulty of speaking due to age, it is better to be younger than 100 years old, and it is better to be younger than 90 years old.

(D) When the present invention is implemented in Japan, the person who can read the Japanese sentence (pronunciation) is the best when acquiring the voice. However, when it is estimated that there is a person with a native language other than Japanese, it can be distinguished by reading the sentence of the corresponding native language.

(E) When creating the estimation algorithm, it is best to use the voices of people belonging to the following six diseases: Alzheimer’s disease, Levitic cognitive impairment, Parkinson’s disease, severe depression, atypical depression, and bipolar disorder Affective disorder. However, people who also suffer from these diseases are excluded. In addition to the above six diseases, voices belonging to vascular dementia, anterior temporal lobe dementia, mood regulation disorder, and cyclic mood disorder can also be used. In addition, the voices of people corresponding to schizophrenia, generalized anxiety disorder, and other neuropsychiatric diseases can also be used.

(F) Healthy subjects should preferably be identified as those who belong to neither the cognitive impairment group nor the emotional impairment group. ＜＜How to get voice＞＞

The method for acquiring voice will be described.

(1) There are no special restrictions on the microphone, as long as the sound can be obtained. For example, you can choose a handheld microphone, a headset, a microphone built into a portable terminal, or a microphone built into a personal computer or tablet, etc. From the viewpoint that only the voice of the target person can be acquired, a pin microphone, a directional microphone, and a microphone built into a mobile terminal are preferable. From the point of view that only the voice of the subject can be obtained, pin microphones, directional microphones, and microphones built in portable terminals are better.

(2) The recorder can be a portable recorder, a personal computer, a tablet computer, or a recording medium built-in or externally placed in a portable terminal.

(3) When estimating diseases, there is no restriction on the content of the voice. For example, it is possible to use a voice freely uttered by the subject, a phrase read from a text prepared by the subject in advance, or words from a telephone or face-to-face conversation. However, when creating an estimation algorithm, it is best to use speech content that is common to the subject. Therefore, when creating the estimation algorithm, it is best to use the corpus of the text prepared by the examinee in advance.

If the speaking time is too short, the accuracy of the estimation result will be lower. It is preferably 15 seconds or more, more preferably 20 seconds or more, and particularly preferably 30 seconds or more. In addition, if it takes longer, it will take some time to get results. It is preferably 5 minutes or less, more preferably 3 minutes or less, and particularly preferably 2 minutes or less.

(4) Observation and inspection items are not restricted. However, it is best to obtain information to verify whether there is a difference in voice due to the condition of the subject. This information includes general information such as gender, age, height, weight, clear diagnosis name, severity, complications of physical diseases, medical history, time of onset and other medical information, examination information such as MRI, CT, and patient health questionnaires (Patient Health Quastionnaire). , PHQ)-9 as medical interviews and questions, MINI-International Neuropsychiatric Interview (mini screen), Hamilton Depression Rating Scale (HAM-D or HDRS), Young’s Mania Rating Scale (Young Mania Rating Scale, YMRS), Mini-Mental State Examination (MMSE), Bipolar Spectrum Diagnostic Scale (BSDS), from the Movement Disorder Society (Movement Disorder Society) ) Revised Unified Parkinson’s Disease Rating Scale (MDS-UPDRS), etc.

(5) Regarding the environment for obtaining the voice, there is no particular limitation as long as it is an environment in which only the voice of the patient can be obtained. A quiet environment, specifically, 40 decibels or lower is better, and 30 decibels or lower is more ideal. Specific examples include examination rooms, consultation rooms, conference rooms, hearing examination rooms, CT, MRI, X-ray and other examination rooms. In addition, the voice can also be obtained in a quiet room of the subject's home.

The estimation program created as described above can be used without any special restrictions, regardless of whether it is suspected of having a neuropsychiatric disease or is estimated to be healthy. If the voice of the examinee is obtained, the program can also be used as a medical examination tool for doctors, or as an examination item in a physical examination or physical examination.

As for the number of voice acquisitions, it can be distinguished once. However, for example, even a healthy subject may feel melancholy due to a certain event in life, and it is very likely that they are just in the mood of melancholy when they get the voice. In addition, the mood of severe depression, atypical depression, etc. may change in the morning and evening. Therefore, if the estimation result shows that the person suffers from a certain mental illness, it is best to obtain the voice at least one more time and perform the estimation again.

Doctors, clinical psychologists, nurses, laboratory technicians, consultants, or anyone else who manipulates the device of the present invention when facing a subject who obtains speech can use the estimation system 100 of the present invention. In a room where a quiet environment is maintained, such as a treatment room or a consultation room, one or more people operating the device of the present invention use the device while directly explaining the method of voice acquisition to the subject in an open state. In addition, the subject who obtained the sound enters the hearing test room or other various test rooms, and the above-mentioned processor uses the device when observing the subject through glass or through a monitor image. If the subject is in a remote area, such as at home, the method of voice collection can be explained to the subject in advance, and the subject can record the voice before the designated date and time. If you are in a remote location, you can also use a separate communication line to obtain the voice, and at the same time use camera images to confirm the identity of the subject.

In addition, when the subject goes to a medical institution or consultation room, the voice can be obtained and estimated, and it can also be used as one of the inspection items of the company or local government's health check or human health check to obtain and estimate the voice. ＜5. Example of creating estimation program 1＞ ＜＜Excluding the example of the estimation process of the sound feature quantity depending on the environment＞＞

As shown in FIGS. 3 and 4, the third set of sound feature quantities and the subsequent estimation are used to test the process of excluding sound feature quantities related to the environment. ＜＜Creation and result of estimation program 1＞＞

An example of verification when the estimation process is performed after extracting the first group or the third set of sound feature amounts will be introduced. First of all, as the same speech content to be used in the first sound feature extraction, the inventor takes the same pronunciation content as an example, using 13 fixed phrases (each phrase is repeated twice), 3 long vowels and short vowels. The phrase "pa ta ka" (Japanese Roman pinyin) is used as an example of the same pronunciation. A total of 30 pronunciations were prepared, and multiple healthy subjects were asked to speak the same content in seven different places. Regarding the different utterance content used in extracting the second voice feature amount, the healthy subject is allowed to speak freely and obtain the voice.

After marking and normalizing the acquired voice, 7,440 voice features were extracted with the analysis software OpenSMILE. Next, a paired t-test is used for comparison, the P value is set to be greater than 0.5, and 170 voice feature quantities are obtained as the first set of voice feature quantities. Next, through t-test (non-paired t-test) for comparison, the P value is set to be greater than 0.5, and 549 sound feature quantities are obtained as the set of second sound feature quantities. Next, compare the set of the first voice feature quantity with the set of the second voice feature quantity, and obtain the common voice feature quantity as the set of the third voice feature quantity. This set of voice feature quantity is not affected by the voice collection place , A total of 169 voice feature quantities are obtained. 99.4% of the 169 voice feature quantities are included in the first set of voice feature quantities. From the above results, it is also possible to define it as the third set of voice feature amounts only based on the set of first voice feature amounts exceeding the desired P value.

Next, the above-mentioned 169 sound feature quantities that do not contain environmental factors are compared with the parameter F(a) obtained from the second sound parameter, and an inference program that excludes sound feature quantities related to the environment is obtained and verified, and authenticating. As learning materials, 30 patients with severe depression, a total of 963 phrases, and 30 healthy subjects, a total of 965 phrases were used. 14 patients with severe depression and 30 healthy subjects were used as verification data. In terms of verification data, about 30 phrases of each person were judged as severe depression or health, and most of these 30 phrases were judged as the final estimation result.

Figure 5 is a mixed matrix of verification data, where HE means healthy and MDD means severe depression. As shown in Figure 5, it is estimated that the correct diagnosis rate of the program is 79.5%. ＜6. Example of creating estimation program 2＞ ＜＜Work to associate multiple diseases with voice data-voice collection＞＞

The following describes the program used when creating the estimation program. In order to attach voice data to a variety of diseases, the voices of the following patients and healthy subjects were obtained from December 25, 2017 to May 30, 2018.

・20 voice cases of patients with Alzheimer's disease ・20 voice cases of patients with Levitic dementia ・20 voice cases of patients with Parkinson's disease ・20 voice cases of patients with severe depression (severe depression group A) ・16 voice cases of patients with bipolar disorder ・19 voices of patients with atypical depression ・20 voice cases of healthy subjects (healthy subjects A group)

In addition, the voices of the following patients and healthy subjects were obtained from June 28, 2019 to October 31, 2019.

・37 voices of patients with Alzheimer's disease ・57 cases of voices of patients with Levi body dementia ・28 voices of patients with other dementias (including vascular dementia and anterior temporal lobe dementia) ・35 voice cases of Parkinson's disease patients ・57 voices of patients with severe depression (group B of severe depression) ・34 voices of patients with bipolar disorder ・30 voice cases of patients with atypical depression ・38 voices of patients with other depression (including mood regulation disorder and cyclic mood disorder) ・Sounds of healthy subjects 60 cases + 28 cases (Acquisition of 4 people's voices in seven different places: healthy subjects B group)

Moreover, these patients are confirmed by doctors in specialized fields such as psychiatry and neurology according to the standards of DSM-5 or ICD-10 to have their respective diseases. In addition, by performing PHQ-9, MMSE, etc., the doctor confirmed that there are no complications of other neuropsychiatric diseases.

By performing PHQ-9, MMSE, etc., it can be confirmed that healthy subjects have no depressive symptoms or cognitive decline.

Use Olympus (Olympus) needle microphone and Roland (Roland) portable recorder for voice collection. The voice data is recorded on the SD card.

The content of the speech is for the subject to read items 1 to 13 twice and items 14 to 17 once in the 17 sentences shown in Figure 10.

When obtaining the speech, explain to the examinee that it will be used in the study to analyze the relationship between the neuropsychiatric patient's voice and the disease, the content of the speech, and the method of obtaining the speech, and sign a written consent. In addition, symbolize and manage the acquired data, including voice, in a format that cannot identify individuals.

Among the above 17 types of speech content, each subject reads the 1st to 13th types of speech content (2 times each, totaling 26 times) and reads the 14th to 17th types of speech content (1 time each, A total of 4 speeches). In a total of 30 speeches, the longer speech content is divided into two parts, and the unclear speech content is excluded, so as to obtain the speech of each disease patient and healthy subject. ＜＜Estimation Program 2＞＞ ＜＜Extraction of voice features irrelevant to the environment＞＞ For the four healthy subjects in the healthy subject group B, voices were obtained in seven different places (hospital examination rooms and treatment rooms).

After normalizing these voices, OpenSMILE is used for voice analysis and 7440 voice features are extracted. Regarding this feature amount, a paired t-test (Paired t-test) is used to compare each phrase. The results show that in reading the phrase "in the past, once in a certain place", 505 places with no significant difference (p>0.5) of voice features were obtained. In the same way, reading "I slept well yesterday" gets 553, reading "I feel angry" gets 727, and reading "Come on" gets 525 voice features that have no significant difference between places.

In addition, regarding the same feature amount, the voices of the healthy subject A group and the healthy subject B group, and the severe depression group A and severe depression group B were compared by an unpaired t-test. In addition, the unpaired t-test (Unpaired t-test) compared the voices of healthy subjects A group and severe depression group A, healthy subjects B group and severe depression group B. As a result, 246 voice feature quantities were selected in the reading "Once upon a time in a certain place". These voice feature quantities were not significantly different between the severe depression group and the healthy subject group (P>0.1), but in the severe There is a significant difference between the depression group and the healthy subject group. In addition, in the same way, 336 voice features were selected in the reading "I slept well yesterday", 231 voice features were selected in the reading "I feel angry", and 363 voice feature values were selected in the reading "Come on".

Then, as the voice feature quantities selected by the paired t test and the unpaired t test, read "Are you in a certain place before" to get 21, read "I slept well yesterday" to get 14, and read "I feel angry." Obtain 28, read "Come on" aloud to obtain 46 voice features.

In the same way, in each phrase, the selected feature quantities that are not significantly different by the paired t test and the selected feature quantities that are not significantly different by the unpaired t test are extracted. Then, the selected feature quantity that has no significant difference between the paired t test and the unpaired t test is selected as the common feature quantity. Fig. 16 summarizes the results of extracting sound feature quantities irrelevant to the environment in this way. ＜＜6. Create estimation program 2-1 (machine learning)＞＞

Then, 30 severe depression patients and 30 healthy subjects were used to speak "in a certain place before," and 21 speech features that were not related to the environment were used as learning data, and based on the estimation An estimation formula 2-1 is created for the characteristic quantity F(a) of one of the severely depressed and healthy persons. ＜＜Estimation of severe depression through estimation program 2-1＞＞

The voices of 25 patients with severe depression and 52 healthy subjects were used as verification data, and the results were shown in Figure 11 (the confusion matrix in the Uden Index, the same below). ＜＜Estimation program 2-2＞＞

Same as the estimation program 2-1, using a voice that reads "I slept well yesterday." At the same time, in addition to using 14 voice feature quantities that are not significantly different between the paired t test and the unpaired t test for "I slept well yesterday", the estimation program was created in the same way as the estimation program 2-1. 2-2 and verify, the results are shown in Figure 12. ＜＜Estimation program 2-3＞＞

The same as the estimation formula 2-1, use the voice that reads "I feel angry" aloud. At the same time, except for using 28 voice feature quantities that are not significantly different between the paired t test and the unpaired t test for "feeling angry", the estimation program 2-3 was created in the same way as the estimation program 2-1. And verify, the results are shown in Figure 13. ＜＜Estimation program 2-4＞＞

It is the same as the estimation program 2-1, using the voice that reads "Come on" aloud. At the same time, except that 46 voice feature quantities that are not significantly different between the paired t test and the unpaired t test are used for "Come on", the estimation program 2-4 is created in the same way as the estimation program 2-1. And to verify, the results are shown in Figure 14. ＜7. Example of creating estimation program 3＞

Estimation programs 2-1 to 2-4, use the same "today is sunny" voice to create an estimation program 5, use the "exhausted" voice to create an estimation program 6, and use the "mood very calm" Voice creation estimation program7. Using the above seven estimation programs, according to the corresponding speech, you can judge whether you are suffering from severe depression or healthy. Then, according to the majority decision of seven judgments, the final estimation result of each person is obtained, and the result is shown in Figure 15.

The disease estimation program used in the disease estimation system of the present invention does not analyze the content of the voice feature amount of the spoken speech. The estimation program calculates the disease prediction value from the feature amount formed by extracting the voice feature amount in the speech. Therefore, it has the advantage of not relying on language. However, when one or more subjects actually speak, the voice characteristics may be affected because they cannot speak fluently unless the sentence is in their native language. Therefore, for example, when estimating the disease of a subject whose native language is English, it is best to collect and analyze the voices of patients and healthy subjects whose native language is English, create an English estimation program, and then use this program to pass the English voice To estimate the disease. In the same way, estimation programs in languages other than Japanese and English can be created.

When an estimation program for English is created and diseases are estimated accordingly, examples of the subject or sentences read by the subject include the following English sentences.

For example, the following can be used as an example of an English sentence. (1) A, B, C, D, E, F, G (2) Prevention is better than cure. (3) Time and tide wait for no man. (4) Seeing is believing. (5) A rolling stone gathers no moss. (6) One, Two, Three, Four, Five, Six, Seven, Eight There are no special restrictions on the sentences spoken in any language, but from the point of view of being easy for anyone to read, these sentences are best known. In addition, long vowels such as "a~~~", "e ~~~", "u~~~", etc., should preferably be read by anyone, no matter which language is their mother tongue.

It is also possible to use a commercially available feature quantity extraction program to extract the voice feature quantity from the spoken voice. A specific example is openSMILE, etc.

The estimation device 200 can be applied to, for example, robots, artificial intelligence, automobiles, or call centers, the Internet, mobile terminal device applications and services, such as smart phones and tablet devices, and retrieval systems. In addition, the estimation device 200 can also be applied to a diagnosis device, an automatic medical consultation device, disaster classification, and the like.

Moreover, the above detailed description has clearly clarified the features and advantages of the embodiments. The purpose is to extend the scope of the patent application to the features and advantages of the above-mentioned embodiments without departing from the spirit and scope of the scope of the patent application. In addition, a person with ordinary knowledge in the field should be able to easily imagine all improvements and modifications. Therefore, it is not intended to limit the scope of the embodiment with the present invention to the above-mentioned scope, and may rely on appropriate improvements and equivalents included in the scope disclosed in the embodiment. Industrial availability

Provide an estimation system, estimation program, and estimation method that can estimate the speech of the examinee, recognize and estimate the disease that the examinee suffers to prevent the disease from getting worse, and based on the accurate confirmation of the disease, so that the patient can receive appropriate treatment .

This patent application claims the priority of Japanese Patent Application No. 2020-001943 filed on January 9, 2020, and cites all the contents described in the Japanese Patent Application.

100: Computer 101: CPU 102: RAM 103: ROM 104: HDD 105: Communication interface 106: input/output interface 107: Media Interface 108: recording media 200: estimation device 201: Client 202: Communication Unit 203: Sound feature extraction unit 204: Extraction unit of the first sound feature quantity 205: Extraction unit of the second sound feature amount 206: Computing Unit 207: estimation unit 208: memory unit N: Network S1001, S1002, S1003, S1004A: steps S1004B, S1005A, S1005B, S1006: steps S2001, S2002, S2003, S2004, S2005: steps

Fig. 1 is an example diagram of the hardware configuration of the present invention. Fig. 2 is a diagram showing an example of the configuration of the present invention. Fig. 3 is a flowchart of the present invention, which explains in detail the extraction of voice feature quantities that are not affected by the location where the voice is obtained. Figure 4 is a flow chart of the present invention. FIG. 5 shows an exemplary diagram of the accuracy of disease estimation after extracting sound feature quantities that are not affected by the voice acquisition location according to the present invention. Fig. 6 shows an example diagram of voice feature quantities that have significant differences in paired t-tests or t-tests. Fig. 7 shows an example diagram of voice feature quantities that are not significantly different in paired t-tests or t-tests. Figure 8 shows an example graph of disease predictive value. Fig. 9 shows an example graph of disease prediction values distributed by disease. Fig. 10 shows an example diagram of the content of a phrase read aloud by the subject. FIG. 11 shows another example diagram of the accuracy of disease estimation after extracting sound feature quantities that are not affected by the voice acquisition location according to the present invention. FIG. 12 shows another example diagram of the accuracy of disease estimation after extracting the sound feature amount that is not affected by the voice acquisition location according to the present invention. FIG. 13 shows another exemplary diagram of the accuracy of disease estimation after extracting the sound feature amount that is not affected by the voice acquisition location according to the present invention. FIG. 14 shows another example diagram of the accuracy of disease estimation after extracting the sound feature amount that is not affected by the voice acquisition location according to the present invention. FIG. 15 shows another example diagram of the accuracy of disease estimation after extracting the sound feature amount that is not affected by the voice acquisition location according to the present invention. Fig. 16 shows a table of the results of extracting sound feature quantities independent of the environment.

100: Computer

101: CPU

102: RAM

103: ROM

104: HDD

105: Communication interface

106: input/output interface

107: Media Interface

108: recording media

N: Network

Claims (4)

A mental and nervous system disease estimation device, comprising: an extraction unit, based on the sound feature quantity (A) that does not significantly differ with the recording environment and the sound feature quantity (B) related to various diseases, and extract the sound feature quantity ( A) and the common voice feature amount (C) of the voice feature amount (B); a calculation unit that calculates a disease prediction value based on the voice feature amount (C); and an estimation unit that inputs the disease prediction Value to estimate disease. The estimation device of claim 1, wherein the candidates for the disease that can be estimated include Alzheimer's disease, Levitic dementia, Parkinson's disease, bipolar disorder, atypical depression, and severe Depression. The estimation device of claim 1, wherein one of the candidates for the disease that can be estimated is severe depression. An operating method of an estimation device includes: extracting the voice feature amount (A) based on the voice feature amount (A) and various disease-related voice feature amounts (B) by the extraction unit of the estimation device (A) ) And the common voice feature amount (C) of the voice feature amount (B); the step of calculating the disease prediction value based on the voice feature amount (C) by the calculation unit of the estimation device; and the step of The estimating unit of the estimating device estimates the step of the disease by inputting the predictive value of the disease.

Images