System, program and method for predicting mental and nervous system disorders

Abstract

To provide an estimation system, an estimation program, and an estimation method that estimate the voice spoken by a subject, identify and estimate the disease the subject is suffering from, prevent the disease from becoming serious, and enable the patient to receive appropriate treatment based on accurate identification of the disease.

Description

[Technical field]

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS This application claims priority to Japanese Patent Application No. 2018-133356, filed on July 13, 2018, and Japanese Patent Application No. 2019-057353, filed on March 25, 2019.

[0002] The present invention relates to a system, a program, and a method for predicting psychiatric and neurological disorders. [Background technology]

[0003] Compared to diseases that impair physical health (hereafter referred to as "physical diseases"), diseases that impair mental health (hereafter referred to as "mental diseases") often have unknown causes and mechanisms of onset. In addition, definitive diagnosis, which requires the sampling of a part of the patient's body, and treatment are often difficult.

[0004] Regarding types of mental illness, diagnostic criteria are disclosed, for example, in the Diagnostic and Statistical Manual of Mental Disorders (hereinafter referred to as "DSM-5") published by the American Psychiatric Association.

[0005] Here, "II. Schizophrenia spectrum disorders and other psychotic disorders", "III. Bipolar disorders and related disorders", "IV. Depressive disorders", "V. Anxiety disorders", "XVIII. Personality disorders" and the like are generally recognized as mental disorders. These mental disorders affect a wide range of people, from young people to the elderly. In addition, these mental disorders can cause a decrease in motivation to act due to depression, and a decrease in sociality, such as difficulty communicating with others, which can interfere with daily life.

[0006] In particular, suicides associated with severe depressive states such as major depression are a major problem. In addition, the economic loss caused by people of productive age suffering from such illnesses being unable to work is also significant. The combined cost of illness in Japan for schizophrenia, depression, and anxiety disorders in fiscal 2008 was estimated at approximately 6 trillion yen (Ministry of Health, Labor, and Welfare: Project Report "Estimation of the Social Cost of Mental Illness," March 2011).

[0007] There is still a stigma surrounding these mental illnesses for a variety of reasons. As a result, patients with these mental illnesses tend to hesitate to disclose that they have a mental illness. This leads to delays in the detection and treatment of patients, resulting in problems with measures not being taken until the illness becomes severe. In addition, the extremely limited number of effective biomarkers for mental illnesses also contributes to delays in detection and treatment.

[0008] For example, physical illnesses can often be detected early through typical health checkups, such as blood tests, X-rays, electrocardiograms, and other tests. On the other hand, mental illnesses can only be detected through a questionnaire with a few questions, so it is difficult to say that the possibility of developing a mental illness can be adequately predicted.

[0009] Furthermore, "XVII. Neurocognitive Disorders" in DSM-5 is a disease related to dementia, and the number of patients is on the rise as the population ages. At present, there is no cure for dementia such as Alzheimer's disease and Lewy body dementia. This is problematic as it places a burden on society as symptoms progress, with families in particular bearing a heavy burden. In response to this situation, Japan's Ministry of Health, Labor and Welfare has formulated the New Orange Plan (Comprehensive Strategy for Promoting Dementia Measures) and is taking measures.

[0010] In recent years, as a quantitative measurement method for observing the brain, there have been cases where SPECT (Single Photo Emission CT), which measures the state of cerebral blood flow, is used to diagnose dementia and Parkinson's disease, and optical topography, which measures the blood hemoglobin concentration in the cerebral cortex, is used to diagnose depression, etc. However, these measuring devices are expensive, and therefore only a limited number of medical institutions are able to perform these measurements.

[0011] In recent years, a technique for measuring ethanolamine phosphate (EPA) in blood and diagnosing depression has been developed (see Patent Document 3), and clinical tests are currently being conducted. This test method has been shown to be able to determine whether a subject has depression, based on the significant correlation between "depression" and the concentration of ethanolamine phosphate in plasma. However, it has not been shown to be able to distinguish between multiple diseases that present with depressive symptoms, such as major depression, atypical depression, and bipolar disorder.

[0012] On the other hand, it has been reported that if depression, for example, is detected in a pre-disease state or when it is mild, various mental health services can be used to recover without medical care (see Non-Patent Document 1). Therefore, there is a demand for a simple testing method for predicting mental illness that places a low burden on patients.

[0013] Furthermore, a technique has been proposed for estimating a person's emotions and mental state (exhibiting depressive symptoms) by analyzing the voice of the person (see Patent Documents 2 and 3). However, this technique cannot estimate the type of illness. [Prior art documents] [Patent documents]

[0014] [Patent Document 1] International Publication No. 2013 / 069645 [Patent Document 2] JP 2007-296169 A [Patent Document 3] International Publication No. 2006 / 132159 [Non-patent literature]

[0015] [Non-Patent Document 1] Research Institute of Economy, Trade and Industry, Policy Discussion Paper Series 14-P-001, January 2004, "Proposal for Active Introduction of Private Sector Initiatives in Low-Intensity Mental Health Services for Pre-Symptomatic Depression" Summary of the Invention [Problem to be solved by the invention]

[0016] The DSM-5 diagnostic criteria lists four subtypes of neurocognitive disorders (dementia): Neurocognitive Disorder Due to Alzheimer's Disease, Frontotemporal Neurocognitive Disorder, Neurocognitive Disorder with Lewy Bodies, and Vascular Neurocognitive Disorder. In addition to these types, dementia caused by brain damage or other diseases is also listed.

[0017] In addition, in the diagnostic criteria of DSM-5, Parkinson's disease is included in the category of dementia with Lewy bodies. The distinction between dementia with Lewy bodies and Parkinson's disease depends on the location where the Lewy bodies occur. Therefore, the symptoms may be similar, making it difficult to distinguish between the two.

[0018] Currently, these dementias cannot be cured completely, but there are drugs that can slow down the progression to some extent, and these drugs have been approved. However, because the treatments differ depending on the type of dementia, it is necessary to estimate these types. In addition, estimation requires advanced technology. For example, it is generally said that it is only by observing the brain of the patient after death that a definitive diagnosis of Alzheimer's dementia can be made; it is difficult to estimate the type of dementia.

[0019] Therefore, if it becomes possible to distinguish the type of dementia when the degree of dementia is mild, i.e., when the patient has mild cognitive impairment (MCI), it may be possible to maintain that condition for longer.

[0020] In addition, with regard to mental illnesses accompanied by mood disorders such as schizophrenia, major depressive disorder, depression with atypical features (hereinafter referred to as "atypical depression"), bipolar disorder, generalized anxiety disorder, personality disorders, persistent depressive disorder, and cyclothymic disorder, when the symptoms are at a mild stage, the patient's illness tends to be interpreted as simply a personality problem, simply being lazy, simply feeling unwell, or as some other problem other than the above-mentioned mental illnesses, and the patient himself may not even be aware of the illness.

[0021] Even if patients are aware of their illness, they may not receive medical care due to prejudice against mental illness and reluctance to visit a psychiatrist or psychosomatic medicine clinic. Furthermore, in some countries and regions, there are few specialized counselors, and it is not customary to consult with specialized counselors, etc., and as a result, social services are not fully utilized as a means of recovery from illness due to social reasons, and even if patients are aware of their illness early on, the illness may become chronic and the symptoms may become severe.

[0022] Additionally, mental illnesses can be difficult to identify because they share common symptoms: for example, Alzheimer's disease and frontotemporal dementia, Alzheimer's disease and dementia with Lewy bodies, Lewy body dementia and Parkinson's disease, bipolar disorder and major depression.

[0023] In addition, a patient may suffer from both dementia and geriatric depression. In addition, a patient may often show depressive symptoms as an early symptom of dementia. On the other hand, when a patient has depressive pseudodementia, the patient is actually depressed, but the symptoms appear as if cognitive function is reduced. Therefore, it is an important issue to distinguish whether the patient has a type of dementia, a type of depression, or whether dementia and depression are combined and either symptom is significantly expressed. This is because, for example, when a patient has depressive pseudodementia, if the patient is treated for depression, the cognitive function will also recover, but even if the patient is treated for dementia, the cognitive function will not recover or the symptoms will worsen.

[0024] The present invention has been made in consideration of the above-mentioned circumstances, and aims to provide a system, program, and method for inferring psychiatric / neurological disorders that can identify and infer whether a subject has dementia by analyzing the speech of the subject.

[0025] Another object of the present invention is to provide a system, program, and method for estimating psychiatric / neurological disorders that can identify and estimate psychiatric disorders accompanied by mood disorders, such as bipolar disorder and major depression, by analyzing the voice spoken by a subject.

[0026] Another object of the present invention is to provide a system, program, and method for predicting psychiatric and neurological disorders that can distinguish between psychiatric disorders with very similar symptoms, thereby reducing misdiagnosis, enabling earlier and more accurate diagnosis, and enabling appropriate treatment.

[0027] Furthermore, another object of the present invention is to provide a system, program, and method for predicting psychiatric and neurological disorders that is easy to operate, can be measured without a special environment, places a small burden on patients, and can predict psychiatric and neurological disorders with short measurements.

[0028] Furthermore, another object of the present invention is to provide a system, program, and method for predicting psychiatric and neurological disorders that are capable of measuring and predicting psychiatric and neurological disorders even when the patient and doctor are physically separated. [Means for solving the problem]

[0029] In order to solve the above problems, the present invention provides a system for estimating a psychiatric or neurological disorder, comprising: an acquisition unit for acquiring voice data of a subject; and a first transmission unit for transmitting the voice data to a server, at least one voice input unit; a first receiving unit for receiving data of an estimated result in the server; a result display unit for displaying the data; a second receiving unit for receiving the voice data from the first transmission unit; a calculation unit for calculating a predicted value of a disease based on a combination of at least one acoustic feature extracted from the subject's voice data; an estimation unit for estimating the subject's disease; and a second transmission unit for transmitting the estimated result data to the display unit, wherein the server comprises a processing device for executing an estimation program; and a recording device having the estimation program recorded thereon, and the estimation unit estimates the subject's disease selected from a group consisting of a plurality of diseases using the predicted value of the disease as an input according to the estimation program. Effect of the Invention

[0030] According to the present invention, it is possible to provide an estimation system, an estimation program, and an estimation method that estimate the voice spoken by a subject, identify and estimate the illness from which the subject is suffering, prevent the illness from becoming more severe, and enable the patient to receive appropriate treatment based on an accurate diagnosis of the illness. [Brief description of the drawings]

[0031] [Figure 1] FIG. 1 is a configuration diagram of the present invention. [Diagram 2] FIG. 1 is a configuration diagram of the present invention. [Diagram 3] 1 is a flowchart of the present invention. [Figure 4A] FIG. 13 is a diagram showing an example of scoring. [Figure 4B] FIG. 1 is a conceptual diagram showing the distribution of disease-specific predicted values. [Figure 5A] 1 is an ROC curve showing the accuracy of the estimation of the present invention. [Figure 5B] 1 is an ROC curve showing the accuracy of the estimation of the present invention. [Figure 6] FIG. 11 is an explanatory diagram of a second acoustic parameter. [Figure 7] FIG. 2 is a graph of the regression analysis of the present invention. [Figure 8] FIG. 2 is a graph of the regression analysis of the present invention. [Figure 9] FIG. 2 is a diagram showing an example of the distribution of the zero crossing rate ZCR and the Hurst exponent H of the present invention. [Figure 10] FIG. 1 is a diagram showing the Parkinson's disease index of the present invention. [Figure 11] FIG. 2 is a diagram in which voice data for each disease of the present invention is plotted. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0032] (Embodiment)

[0033] Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

[0034] FIG. 1 shows a configuration diagram of an estimation system 100 of the present invention. The estimation system 100 of FIG. 1 includes an input unit 110 for acquiring the voice of a subject, a display unit 130 for displaying an estimation result to the subject, and a server 120. The server 120 includes a calculation processing device 120A (CPU), a first recording device such as a hard disk on which an estimation program for the calculation processing device 120A to execute processing is recorded, and a second recording device such as a hard disk on which medical examination data of the subject and a collection of messages to be transmitted to the subject are recorded. The server 120 is connected to the input unit 110 and the display unit 130 via a wired or wireless connection. The calculation processing device 120A may be realized by hardware.

[0035] The input unit 110 includes a voice acquisition unit 111 such as a microphone, and a first transmission unit 112 that transmits the acquired voice data to the server 120. The acquisition unit 111 generates voice data of a digital signal from the voice of the subject in the form of an analog signal. The voice data is transmitted from the first transmission unit 112 to the server 120.

[0036] The input unit 110 acquires a voice signal uttered by the subject via a voice acquisition unit 111 such as a microphone, and generates voice data as a digital signal by sampling the voice signal at a predetermined sampling frequency (for example, 11025 Hz).

[0037] The input unit 110 may include a recording unit for recording audio data, separate from the recording device on the server 120 side. In this case, the input unit 110 may be a portable recorder. The recording unit may be a recording medium such as a CD, a DVD, a USB memory, an SD card, or a mini disk.

[0038] The display unit 130 includes a first receiving unit 131 that receives data such as an estimation result, and an output unit 132 that displays the data. The output unit 132 is a display that displays data such as an estimation result. The display may be an organic electro-luminescence (EL) display, a liquid crystal display, or the like.

[0039] The input unit 110 may have a function such as a touch panel in order to input in advance data on the results of a health check, which will be described later, and data on answers regarding a stress check. In this case, the input unit 110 and the display unit 130 may be realized by the same hardware having the functions of both.

[0040] The arithmetic processing device 120A includes a second receiving unit 121 that receives the voice data transmitted from the first transmitting unit 112, a calculation unit 122 that calculates a predicted value of a disease based on acoustic features extracted from the voice data of the subject, an estimation unit 123 that estimates the disease of the subject using an estimation model learned by machine learning using the predicted value of the disease as an input, and a second transmitting unit 124 that transmits data related to the estimation result to the display unit 130. Note that the calculation unit 122 and the estimation unit 123 are described separately to explain the functions, but the functions of the calculation unit and the estimation unit may be performed simultaneously. Note that in this specification, the term "mental value" is used synonymously with the predicted value of the disease.

[0041] Fig. 2 shows another configuration diagram of the estimation system 100 of the present invention. In the estimation system 100 in Fig. 2, the server 120 of the estimation system 100 in Fig. 1 is connected to the input unit 110 or the display unit 130 via a network NW. In this case, the input unit 110 and the display unit 130 are a communication terminal 200. The communication terminal 200 is, for example, a smartphone, a tablet terminal, or a notebook computer or desktop computer equipped with a microphone.

[0042] The network NW connects the communication terminal 200 and the server 120 via a mobile phone communication network or a wireless LAN based on a communication standard such as Wi-Fi (Wireless Fidelity) (registered trademark). The estimation system 100 may connect a plurality of communication terminals 200 and the server 120 via the network NW.

[0043] The estimation system 100 may be realized by a communication terminal 200. In this case, an estimation program stored in the server 120 is downloaded via the network NW and recorded in a recording device of the communication terminal 200. A CPU included in the communication terminal 200 may execute an application recorded in the recording device of the communication terminal 200, causing the communication terminal 200 to function as the calculation unit 122 and the estimation unit 123. The estimation program may be recorded on an optical disk such as a DVD or a portable recording medium such as a USB memory and distributed. (First embodiment)

[0044] Fig. 3 shows an example of estimation processing of the estimation system 100 shown in Fig. 1. The processing shown in Fig. 3 is realized by the arithmetic processing device 120A of the estimation system 100 executing an estimation program recorded in the first recording device 120B of the estimation system 100. The functions of the second receiving unit 121, the calculating unit 122, the estimating unit 123, and the second transmitting unit 124 of the arithmetic processing device 120A will be described with reference to Fig. 3. (Calculation unit 122)

[0045] When the process starts, in step S101, the calculation unit 122 determines whether or not voice data has been acquired by the acquisition unit 111. If voice data has already been acquired, the process proceeds to step S104. If voice data has not been acquired, in step S102, the calculation unit 122 commands the output unit 132 of the display unit 130 to display a predetermined template message.

[0046] This estimation program does not estimate psychiatric and neurological disorders based on the meaning and content of the subject's speech. Therefore, the acquired voice data may be any data with a speech time of 15 to 300 seconds. There is no particular restriction on the language used, but it is desirable that the language used by the population when the estimation program was created matches the language used by the population. Therefore, the fixed phrase displayed on the output unit 132 may be any fixed phrase in the same language as the population and with a speech time of 15 to 300 seconds. Preferably, the fixed phrase has a speech time of 20 to 180 seconds.

[0047] For example, it may be "Iroha Ni Ho He To" or "Ai Ue Oka Ki Ku Ke Ko" which do not include a specific emotion, or it may be a question such as "What is your name" or "When is your birthday?".

[0048] There are no particular limitations on the environment in which the voice is acquired as long as it is an environment in which only the voice spoken by the subject can be acquired, but it is preferable that the voice is acquired in an environment of 40 bB or less, and more preferably, 30 dB or less.

[0049] When the subject reads out the template, in step S103, voice data is acquired from the voice uttered by the subject, and the process proceeds to step S104.

[0050] Next, in step S104, the calculation unit 122 issues a command to transmit the voice data from the input unit 110 to the second receiving unit 121 of the server 120 via the first transmitting unit 112.

[0051] Next, in step S105, the calculation unit 122 determines whether or not the predicted value of the subject's disease has been calculated. The predicted value of the disease is a feature F(a) consisting of a combination of acoustic features generated by extracting one or more acoustic parameters, and is a predicted value of a specific disease used as an input to a machine learning estimation algorithm. The acoustic parameters are parameterized features of sound transmission. If the predicted value of the disease has already been calculated, the process proceeds to step S107. If the predicted value of the disease has not been calculated, in step S106, the calculation unit 122 calculates the predicted value of the disease based on the subject's voice data and the trained estimation program.

[0052] The following steps S107 to S112 will be described in detail in the following paragraphs, but will be briefly described here. In step S107, the calculation unit 122 acquires the subject's medical examination data acquired in advance from the second recording device 120C. Step S107 may be omitted, and the disease may be estimated from the predicted value of the disease without acquiring the medical examination data.

[0053] Next, in step S108, the estimation unit 123 estimates a disease based on the predicted value of the disease calculated by the calculation unit 122 and / or the medical examination data.

[0054] By providing individual thresholds for distinguishing between specific and other diseases with respect to the predicted value of a disease, it is possible to distinguish between patients whose predicted value of a disease has been calculated and those who should be identified and those who should not. In the embodiment described below, the cases are classified into those in which the threshold is exceeded and those in which it is not exceeded.

[0055] Next, in step S109, the estimation unit 123 determines whether or not advice data corresponding to a disease has been selected. The advice data corresponding to a disease is advice for preventing a disease or avoiding its aggravation when the subject receives the advice data. If the advice data has been selected, the process proceeds to step S111.

[0056] If the advice data has not been selected, in step S110, advice data corresponding to the subject's symptoms is selected from the second recording device.

[0057] Next, in step S111, the estimation unit 123 issues an instruction to transmit the disease estimation result and the selected advice data to the first receiving unit 131 of the display unit 130 via the second transmitting unit 124.

[0058] Next, in step S112, the estimation unit 123 commands the output unit 132 of the display unit 130 to output the estimation result and advice data. Finally, the estimation system 100 ends the estimation process. Second embodiment

[0059] The estimation program of the present invention can analyze voices even from remote locations, so it can be used in online medical treatment and online counseling. When diagnosing psychiatric and neurological disorders, doctors observe the patient's facial expressions, movements, and conversation status through interviews and consultations. However, patients may feel prejudice against psychiatric and neurological disorders, and may hesitate to visit a psychiatric hospital or clinic.

[0060] Online medical treatment and counseling allows patients to meet with a doctor or counselor without visiting a facility. Therefore, compared to other illnesses other than neuropsychiatric disorders, mental and neurological disorders are highly compatible with online medical treatment.

[0061] Doctors, counselors, and clinical psychologists can use this estimation program to perform analysis when interviewing patients (or clients) online. This makes it very easy to determine whether a patient is suffering from a psychiatric or neurological disorder and what type of disorder that is. In addition, during the interview, various psychological tests and cognitive function tests such as MMSE, BDI, and PHQ-9 can also be conducted.

[0062] In this case, the patient needs computer hardware capable of transmitting audio, as well as a monitor screen for the interview and a microphone for recording audio.

[0063] If the patient does not have these devices at home, they can be provided at, for example, a primary care physician's office, and the patient can attend the primary care physician's office and be interviewed there via the device.

[0064] For example, if a patient visits a hospital for treatment of a physical illness, and the doctor diagnoses the patient and determines that he or she may have a mental or neurological disorder, it is possible to capture the voice on the spot and analyze it using the program of the present invention.

[0065] In other locations, if psychiatrists and neurologists are able to provide online consultations, it is possible for a patient's family doctor, psychiatrist, or neurologist to work together online to make a diagnosis. Third embodiment

[0066] The prediction program of the present invention can be used as a screening device by increasing the sensitivity for predicting a particular disease (in this case, the specificity is generally decreased).

[0067] By using this test as part of health checkups conducted by companies and local governments, and medical examinations conducted by medical institutions, it can contribute to the early detection of psychiatric and neurological disorders that have been difficult to detect until now and for which there have been no easy tests.

[0068] For example, audio capture can be performed as part of a series of tests, similar to eye exams, vision tests, and hearing tests, and the program can provide estimated results on the spot or together with the results of other tests. (Fourth embodiment)

[0069] The estimation program of the present invention does not require any special equipment, so anyone can use it easily. On the other hand, since its use is limited to mental and neurological disorders, it is not necessarily used frequently. Therefore, a set of the estimation device of the present invention can be installed in specialized hospitals equipped with expensive testing equipment, so that a family doctor or the like can request testing from the specialized hospital when a target patient visits the hospital.

[0070] Devices used for psychiatric and neurological disorders include optical topography, myocardial scintigraphy, cerebral blood flow scintigraphy, CT, MRI, electroencephalography, etc. These are used for disease estimation and exclusion diagnosis, but the estimation device of the present invention is extremely minimally invasive and can be used in conjunction with these tests or prior to these tests. Fifth embodiment

[0071] The estimation program of the present invention can be easily used at home, so it can be used as a monitoring device after diagnosis. For example, in the case of mood disorders, drugs and psychotherapy are administered according to the patient's illness, and the effectiveness of these therapies can be measured. In addition, by using it continuously, it is possible to observe daily whether symptoms are stable and whether there are any signs of recurrence. Sixth embodiment

[0072] Since the estimation program of the present invention analyzes spoken voice, it can be applied as a monitoring device for elderly people.

[0073] Whether an elderly person living alone is doing well is a concern for close relatives. By implementing the estimation program of the present invention in an elderly care system using communication means such as telephone or video telephone, it is possible to not only observe vital signs, but also to measure whether there is a tendency for dementia or depression, making it possible to take appropriate measures even if the elderly person lives alone.

[0074] In these various embodiments, the method of acquiring the voice is not particularly limited, but examples include: (1) the subject sends recorded voice via telephone or the Internet; (2) the examiner contacts the subject via telephone or the Internet and acquires the voice by talking to the subject; (3) a voice acquisition device is installed in the subject's residence and the subject records on the device; and (4) the voice acquisition device automatically starts up at regular intervals and acquires the subject's voice by talking to the subject.

[0075] When acquiring the speech, it is preferable to display the sentence to be spoken on a display provided in the estimation device, or to play the sound of the sentence to be spoken from a speaker, so that the subject can speak smoothly. Recording begins with a mechanical sound indicating the start of recording, and when the speaker finishes speaking, the recording can be stopped by pressing a switch, allowing the spoken voice to be captured for each sentence. (Calculation of disease prediction values)

[0076] An outline of the calculation of a disease prediction value in the present invention will be described. As a pre-processing before the estimation process in the estimation unit 123 in FIG. 1, the calculation unit 122 goes through a stage of extracting a plurality of acoustic parameters from the voice data of the subject. The acoustic parameters are extracted from the voice data of the patient. The acoustic parameters are parameterized characteristics when sound is transmitted. For example, the acoustic parameters include the zero-crossing rate and the Hurst exponent. The zero-crossing rate is calculated as the number of times per unit time that the waveform of the sound pressure of the voice crosses the reference pressure as a degree of the intensity of the change in the waveform in the voice. The Hurst exponent indicates the correlation of the change in the waveform of the voice.

[0077] The acoustic parameters consist of a first acoustic parameter and a second acoustic parameter. The first acoustic parameter is an acoustic parameter extracted from the voice of a patient from whom a specific disease is to be estimated. The second acoustic parameter is an acoustic parameter recorded in advance in database 120B. The second acoustic parameter is extracted from the voice data of a patient with Alzheimer's disease, Lewy body dementia, Parkinson's disease, major depression, atypical depression, or bipolar disorder, and each acoustic parameter is linked to each disease in advance.

[0078] The acoustic parameters used in the present invention include the following items: 1) Volume envelope (attack time, decay time, sustain level, release time) 2) Waveform fluctuation information (Shimmer, Jitter) 3) Zero-crossing rate 4) Hurst exponent 5) VOT (Voice Onset Time) 6) Statistics of the intra-utterance distribution of certain coefficients of the Mel-frequency cepstrum (1st quartile, median, 3rd quartile, 95th percentile, arithmetic mean, geometric mean, difference between the 3rd quartile and the median, etc.) 7) Statistics of the intrautterance distribution of the rate of change of the frequency spectrum (1st quartile, median, 3rd quartile, 95th percentile, arithmetic mean, geometric mean, difference between the 3rd quartile and the median, etc.) 8) Within-utterance distribution statistics for the time variation of a certain coefficient of the Mel-frequency cepstrum (1st quartile, median, 3rd quartile, 95th percentile, arithmetic mean, geometric mean, difference between the 3rd quartile and the median, etc.) 9) Statistics on the intra-utterance distribution of the time variation of a certain coefficient of the Mel-frequency cepstrum (1st quartile, median, 3rd quartile, 95th percentile, arithmetic mean, geometric mean, difference between the 3rd quartile and the median, etc.) 10) Squared error for quadratic regression approximation of intra-utterance time variation of frequency spectrum with 90% roll-off 11) Arithmetic errors for quadratic regression approximation of intra-utterance time variation of frequency spectrum centroid Other parameters include pitch rate, probability of being voiced, power in a given range of frequencies, scale, speech rate (number of morae in a given period of time), pauses and intervals, and volume.

[0079] The estimation program has a learning function based on artificial intelligence and performs estimation processing using the learning function. Neural network-type deep learning may be used, or reinforcement learning that partially strengthens the learning field may be used, or other techniques such as genetic algorithms, cluster analysis, self-organizing maps, and ensemble learning may be used. Of course, other artificial intelligence techniques may be used. In ensemble learning, a classification algorithm may be created using a method that combines boosting and decision trees.

[0080] In the stage of creating the estimation program, the algorithm creator uses a stepwise method to select one or more of the above-mentioned second acoustic parameters to be used as the variable f(n) in order to obtain the best combination. Next, a coefficient is assigned to the selected acoustic parameter to create one or more acoustic features. Furthermore, these acoustic features are combined to create the feature F(a).

[0081] There are three types of stepwise methods: forward, backward, and incremental. Any of them can be used. The regression analysis used in the stepwise method includes linear classification such as linear discriminant and logistic regression analysis. The variables f(n) and their coefficients, i.e., the coefficients xn of the equation F(a) shown below, are called regression coefficients and are weights assigned to the function f(n).

[0082] After the creator of the learning algorithm selects the regression coefficients, the quality may be improved by machine learning to increase the estimation accuracy based on disease information accumulated in a database.

[0083] A predictor of disease for a subject is calculated from one or more acoustic features, for example, based on the following formula F(a):

[0084]

number

[0085] Here, f(n) is any one or more second acoustic parameters arbitrarily selected from the above acoustic parameter items (1) to (11). xn is a regression coefficient specific to the disease. f(n) and xn may be recorded in advance in the recording device 120 of the estimation program. The regression coefficient of the feature F(a) may be improved during the machine learning process of the estimation program.

[0086] 1 calculates a feature quantity for distinguishing between healthy subjects and subjects with a disease, or between diseases, based on the combination of the second acoustic parameters. From the feature quantity, a reference range and a scoring is performed to calculate how far the subject's value is from the reference range, thereby calculating a predicted value of the subject's disease.

[0087] FIG. 4A is an image showing that the intensity of a certain acoustic feature varies for each disease. It shows that the subject had the highest score for disease A. Therefore, the predicted value for the subject's disease A is calculated to be higher than that of the other disease groups. In addition, by setting the intensity of 50 as the threshold, for example, the subject can be classified into a group of diseases A, B, and C, and a group of diseases D and E.

[0088] In Fig. 4A, the predicted value of the disease is calculated based on the intensity of one acoustic feature, but in reality, it is difficult to classify diseases using only one acoustic feature. Therefore, the disease may be classified by calculating the feature F(a) obtained by combining several acoustic features.

[0089] Based on this feature F(a), we calculate the predicted disease value for the labeled subject's voice and obtain the distribution of the predicted values ​​for each disease. This makes it possible to classify each disease.

[0090] FIG. 4B is a distribution image of disease prediction values ​​(denoted as "mental values" in FIG. 4B) obtained by combining three acoustic parameters.

[0091] From Figure 4B, it can be seen that the distribution of predicted values ​​for the Lewy body dementia patient group can be separated from the distribution of predicted values ​​for the patient groups with other diseases and the healthy control group. The present invention sets a combination of acoustic parameters for each disease so that it can be distinguished from other diseases, calculates the feature F(a), and uses the feature F(a) as an input to determine which disease the voice of each subject corresponds to.

[0092] Another method is to extract features F(a) for each disease from the voice of each patient, determine which disease features the patient has more of, and compare the predicted values ​​of the diseases to infer which disease the patient is suffering from.

[0093] In this case, the predicted value of a disease can be regarded as the degree of the disease. By comparing and recalculating the predicted values ​​of each disease, it is possible to express the probability of having each disease.

[0094] In this way, feature (a) related to each disease is extracted from the voices of patients with six diseases, including Alzheimer's disease, dementia with Lewy bodies, Parkinson's disease, major depression, atypical depression, and bipolar disorder, as well as the voices of healthy individuals, and a prediction value for each disease is calculated.

[0095] Furthermore, a prediction program may be created from the voices of patients with 10 diseases, including four more diseases, namely vascular dementia, frontotemporal dementia, cyclothymia, and dysthymia, for the target diseases.

[0096] Finally, by analyzing the voice spoken by the subject, it is estimated whether the subject has one of the above-mentioned 6 to 10 diseases or is healthy. (Estimated flow of the estimated program)

[0097] Regarding the estimation flow of the estimation program, as described above, the feature (a) for each individual disease may be extracted and a disease prediction value may be calculated for each disease. Alternatively, a combination of acoustic features for a group of diseases may be created first, and the feature F(a) for the group of diseases may be used as an input to the estimation unit. Input and estimation may be performed in multiple stages to finally estimate each disease or healthy state. (Estimated flow 1)

[0098] For example, in the first step, patients are estimated into three groups: (1-A) dementia group including Alzheimer's disease, dementia with Lewy bodies, and Parkinson's disease; (1-B) mood disorder group including major depression, atypical depression, and bipolar disorder; and (1-C) healthy subjects.

[0099] Next, in the second step, the disease of the patient in the dementia group is estimated using a program that estimates whether the voice of the patient classified in the (1-A) dementia group has one of the three diseases: (1-A-1) Alzheimer's disease, (1-A-2) Lewy body dementia, or (1-A-3) Parkinson's disease. On the other hand, the disease of the patient in the mood disorder group is estimated using a program that estimates whether the voice of the patient classified in the (1-B) mood disorder group has one of the three diseases: (1-B-1) major depression, (1-B-2) atypical depression, or (1-B-3) bipolar disorder. (Estimated flow 2)

[0100] In another embodiment of the judgment flow, in the first step, patients are estimated into three groups: (2-A) a dementia group including Alzheimer's disease, dementia with Lewy bodies, and Parkinson's disease; (2-B) a mood disorder group including major depression, atypical depression, and bipolar disorder; and (2-C) healthy subjects.

[0101] Next, in the second step, the voices of patients classified into the (2-A) dementia group are judged as having Lewy body dementia or not by a program (2-A-1) that infers whether the patient has Lewy body dementia or not, while the voices of patients classified into the (2-B) mood disorder group are judged as having major depression or not by a program (2-B-1) that infers whether the patient has major depression or not.

[0102] Similarly, by using (2-A-2) a program for estimating whether the disease is Alzheimer's disease or other dementia, (2-A-3) a program for estimating whether the disease is Parkinson's disease or other dementia, (2-B-2) a program for estimating whether the disease is atypical depression or other mood disorder, and (2-B-3) a program for estimating whether the disease is bipolar disorder or other mood disorder, it is possible to ultimately determine whether the disease is Alzheimer's disease, dementia with Lewy bodies, Parkinson's disease, major depression, atypical depression, or bipolar disorder. (Estimated flow 3)

[0103] In another embodiment, in the second step, the disease of a patient classified into the dementia group is estimated by using the voice of the patient classified into the (3-A) dementia group in combination with three estimation programs: (3-A-1) a program for estimating whether the patient has Alzheimer's disease or dementia with Lewy bodies, (3-A-2) a program for estimating whether the patient has Lewy body dementia or Parkinson's disease, and (3-A-3) a program for estimating whether the patient has Parkinson's disease or dementia of the Alzheimer's type.

[0104] On the other hand, the voice of a patient classified as the (3-B) mood disorder group is used in combination with three estimation programs, namely, (3-B-1) a program for estimating whether the patient has major depression or atypical depression, (3-B-2) a program for estimating whether the patient has atypical depression or bipolar disorder, and (3-B-3) a program for estimating whether the patient has bipolar disorder or major depression, to estimate the illness of the patient classified as the mood disorder group. (Estimated flow 4)

[0105] In addition, the first step of dividing the subjects into the three classes of dementia group, mood disorder group, and healthy subjects may be a two-step process in which the subjects are first divided into two classes, (4-A) healthy group and (4-B) other disease group, and then the disease group is divided into (4-B-1) dementia group and (4-B-2) mood disorder group. (Subject of the voice obtained when creating the estimation program)

[0106] Acquisition of voice data for the second acoustic parameter will now be described. It is preferable that subjects from whom voice is to be acquired are selected according to the following criteria.

[0107] (a) The subjects will be fully informed that the captured voice data will be used for disease analysis, and will provide written consent.

[0108] (a) 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 native language. However, since there may be differences between races and languages, it is preferable to make a judgment using the same race and language as a comparison control. For example, from the viewpoint of convenience, when the present invention is implemented in Japan, it is preferable to obtain voice from a person whose native language is Japanese, create this estimation program, and estimate a disease from Japanese speech. Also, when the present invention is implemented in an English-speaking country, it is preferable to obtain voice from a person whose native language is English, create this estimation program, and estimate a disease from English speech.

[0109] (c) There are no particular restrictions on age, so long as the individual is able to speak. However, taking into consideration changes in voice and emotional stability, 15 years or older is preferable, 18 years or older is even more preferable, and 20 years or older is particularly preferable. Also, taking into consideration the physical difficulty of speaking due to age, it is preferable for the individual to be under 100 years old, and even more preferable for the individual to be under 90 years old.

[0110] (d) When implementing the present invention in Japan, it is preferable to have a person who can read (speak) Japanese sentences when acquiring voice. However, when estimating a person whose native language is other than Japanese, it is possible to identify them by reading aloud sentences in the relevant native language.

[0111] (E) When creating the estimation algorithm, it is preferable to use voices of individuals with each of the following six diseases: Alzheimer's disease, dementia with Lewy bodies, Parkinson's disease, major depression, atypical depression, and bipolar disorder. However, this does not include individuals who suffer from these diseases concomitantly. In addition to these six diseases, voices of individuals with vascular dementia, frontotemporal dementia, dysthymia, and cyclothymic disorder may also be used. Furthermore, voices of individuals with schizophrenia, generalized anxiety disorder, and other neuropsychiatric disorders may also be used.

[0112] (f) Healthy individuals are preferably those who have been confirmed to be free of dementia and mood disorders. (Method of acquiring audio)

[0113] The means for acquiring the voice will now be described.

[0114] (1) The microphone is not particularly limited as long as it can capture voice. For example, it may be selected from a handheld microphone, a headset, a microphone built into a mobile terminal, and a microphone built into a personal computer or tablet. From the viewpoint of capturing only the voice of the subject, a lapel microphone, a directional microphone, or a microphone built into a mobile terminal is preferable.

[0115] (2) The recorder may be a portable recorder, a personal computer, a tablet, or a mobile terminal with a built-in or external recording medium.

[0116] (3) There is no restriction on the content of the utterance when inferring a disease. For example, utterances made freely by the subject, utterances in which the subject reads out a prepared sentence, and utterances during a telephone or face-to-face conversation can be used. However, when creating an inference algorithm, it is preferable to use utterance content common to the subjects. Therefore, when creating an inference algorithm, utterances in which the subject reads out a prepared sentence are preferable.

[0117] Regarding the speech time, if it is too short, the accuracy of the estimation result will be low. It is preferably 15 seconds or more, more preferably 20 seconds or more, and particularly preferably 30 seconds or more. Also, if it is longer than necessary, it will take a long time to obtain the result. It is preferably 5 minutes or less, more preferably 3 minutes or less, and particularly preferably 2 minutes or less.

[0118] (4) There are no particular restrictions on the observation and testing items. However, it is preferable to obtain information to verify whether there are differences in voice due to differences in the subject's profile. Information includes general information such as gender, age, height, weight, etc.; medical information includes confirmed diagnosis, severity, complications of physical diseases, medical history, onset time, etc.; testing information such as MRI and CT; interviews and questions such as Patient Health Quastionnaire (PHQ)-9, The MINI-International Neuropsychiatric Interview (MINI screen), Hamilton Depression Rating Scale (HAM-D or HDRS), Young Mania Rating Scale (YMRS), Mini-Mental State Examination (MMSE), Bipolar Spectrum Diagnostic Scale (BSDS), The Unified Parkinson's Disease Rating Scale (MDS-UPDRS) revised by The Movement Disorder Society, etc.

[0119] (5) There are no particular limitations on the environment in which the voice is acquired, so long as it is an environment in which only the patient's speech can be acquired. A quiet environment, specifically 40 dB or less, is preferable, and 30 dB or less is even more preferable. Specifically, there are examination rooms, counseling rooms, conference rooms, hearing test rooms, and examination rooms for CT, MRI, X-ray, etc. Voice may also be acquired in a quiet room in the subject's home.

[0120] Figure 5 is a graph of the ROC curve showing the separation performance between healthy subjects or specific diseases and others. The horizontal axis shows 1-specificity, and the vertical axis shows sensitivity. In other words, the horizontal axis shows the false positive rate, and the vertical axis shows the true positive rate. In both ROC curves in Figure 5, the true positive rate was high at points where the false positive rate was low.

[0121] In addition, the AUC (area under the ROC curve) was high in all cases, and a significant difference was confirmed compared to random classification. The diseases for which the separation performance was verified were dementia with Lewy bodies, Parkinson's disease, and major depression. The AUC in each ROC curve was 0.794 for dementia with Lewy bodies, 0.771 for Parkinson's disease, and 0.869 for major depression. It should be noted that diseases that can be predicted using the present invention are not limited to those mentioned above. The AUC results confirm that the accuracy of disease prediction is high.

[0122] The estimation program created as described above can be used without any restrictions by those suspected of having a psychiatric disorder or those presumed to be healthy. It can also be used as a tool for medical examinations by doctors, or as a test item for health checkups and medical examinations, simply by acquiring the subject's voice.

[0123] Regarding the number of times that the voice is captured, it is possible to determine the presence of the disorder if the voice is captured once. However, even healthy people may experience a depressive state due to a life event, and it is quite possible that the person happened to be in a depressed mood when the voice was captured. In addition, in cases of major depression and atypical depression, the mood may change between morning and night. Therefore, when the estimated result indicates that the person is suffering from some mental illness, it is preferable to capture the voice at least once more and make another estimation. (Disease prediction by combining factors other than voice)

[0124] We will now explain disease estimation by incorporating elements other than voice. The psychiatric / neurological disorder estimation system of the present invention is capable of estimating the type of psychiatric / neurological disorder using only the subject's voice, but in order to further increase the accuracy of the estimation, elements possessed by the subject other than voice may also be used.

[0125] The elements held by these subjects include medical examination data, the results of stress questionnaires, and the results of cognitive function tests.

[0126] In step S107 of FIG. 3, the calculation unit 122 acquires the subject's health check data acquired in advance from the second recording device 120C. The health check data is data recorded by the subject when inputting voice or before inputting voice, such as the results of a health check. The health check data may be acquired in advance in the form of a medical interview via the communication terminal 200 or the like. The health check data includes items such as whether or not each of the body mass index, fasting blood glucose, hemoglobin A1c, blood pressure, total cholesterol, LDL cholesterol, and urea nitrogen is higher than the standard value, and whether or not HDL cholesterol is lower than the standard value. The subject's physical health value may be calculated as a physical value based on the above items.

[0127] Although there is currently no definitive biomarker for psychiatric and neurological disorders, the search for biomarkers for each disorder is ongoing. Therefore, by measuring these biomarkers and combining the results with the voice analysis results of the present invention, the disorder can be more reliably diagnosed.

[0128] For example, biomarkers for Alzheimer's disease include (1) amyloid beta 42 (Aβ42) in cerebrospinal fluid, (2) amyloid PET imaging, (3) total tau protein or phosphorylated tau protein in cerebrospinal fluid, and (4) atrophy of the temporal and parietal lobes in fMRI examinations.

[0129] Biomarkers for Lewy body dementia include (1) a decrease in dopamine transporters as determined by SPECT, (2) decreased MIBG accumulation as determined by MIBG myocardial scintigraphy, and (3) REM sleep without decreased muscle activity as determined by polysomnography.

[0130] Biomarkers for Parkinson's disease include (1) a decrease in dopamine transporters as determined by SPECT, (2) a decrease in MIBG accumulation as determined by MIBG myocardial scintigraphy, and (3) an increase in alpha-synuclein oligomers in cerebrospinal fluid.

[0131] Biomarkers of vascular dementia include lacunar infarcts, white matter lesions, cerebral hemorrhage, multiple cortical infarcts, and other deficits detected by MRI.

[0132] On the other hand, although various research is being conducted on PEA (ethanolamine phosphate) and ribosomal genes, no biomarkers for mood disorders such as major depression have yet been established.

[0133] The second recording device 120C records, in addition to the health data, response data on stress that describes the subject's response to a questionnaire on stress. The calculation unit 122 executes a scoring process on the response data and calculates a stress value that indicates the degree to which the subject is affected by stress based on the response data.

[0134] The stress questionnaire, for example, executes a scoring process defined according to a questionnaire such as a simple occupational stress questionnaire on the subject's response data to obtain information about the stress the subject is experiencing. The stress information includes, for example, at least one of a scale of the stress intensity experienced by the subject, a scale of the subject's subjective symptoms, a scale of the subject's tolerance to stress, a scale of the subject's ability to recover from a stressful state, and a scale of the subject's support from those around him / her.

[0135] The measure of the stress intensity experienced by the subject indicates the intensity of the stress experienced by the subject at the time the subject answers the questions. For example, in the case of the Simple Occupational Stress Questionnaire, the measure of stress intensity is obtained from the answers to 17 questions included in the area "A" out of a total of 57 questions. The measure of stress intensity may be obtained not only from the causes of psychological burden in the workplace as in the Simple Occupational Stress Questionnaire, but also from stress factors in general private life. The measure of stress intensity may also be obtained using, for example, an interpersonal stress event scale (Tsuyoshi Hashimoto, "An Attempt to Classify Interpersonal Stress Events in University Students," Social Psychology Research, Vol. 13, No. 1, pp. 64-75, 1997).

[0136] The scale of subjective symptoms indicates the psychological state of the subject at the time of answering the questionnaire. For example, in the case of the Simple Occupational Stress Questionnaire, it is obtained from the answers to the 29 questions in area "B". The scale of subjective symptoms checks the subjective symptoms of the mind and body caused by psychological burden, and indicates the state resulting from stress. Note that the factors of stress that cause subjective symptoms are not limited to work.

[0137] In addition, the subjective symptom scale may be determined using a scale for measuring a depressive state, such as the Beck Depression Inventory (BDI: Beck et al., "An Inventory for Measuring Depression", Arch. Gen. Psychiatry, Vol. 4, pp. 561-571, 1961), the Self-Rating Depression Scale (SDS: Zunk et al., "Self-Rating Depression Scale in an Outpatient Further Validation of the SDS", Arch. Gen. Psychiatry, Vol. 13, pp. 508-515, 1965), the Geriatric Depression Rating Scale (GDS), the Self-Rating Questionnaire for Depression (SQR-D), the Hamilton Depression Rating Scale (HAM-D), or the Patient Health Questionnaire-9 (PHQ-9). These tests are scored, and depending on the score of each test, it is possible to evaluate the patient as normal, borderline, suspected of depression, etc. In addition, these tests are merely evaluations of depressive symptoms and cannot identify diseases by themselves, but they can assist in the diagnosis of diseases by the estimation system of the present invention.

[0138] For example, the BDI has a maximum score of 63, with 0-10 being in the normal range, 11-16 being neurotic or mildly depressed, 17-20 being on the borderline of depression and requiring specialist treatment, 21-30 being moderately depressed and requiring specialist treatment, 31-40 being severely depressed and requiring specialist treatment, and 41 or above being extremely depressed and requiring specialist treatment.

[0139] The PHQ-9 has a maximum score of 27 points, with 0-4 representing none, 5-9 representing mild, 10-14 representing moderate, 15-19 representing moderate to severe, and 20-27 representing severe depressive symptoms.

[0140] The scale of support from the surroundings for the subject indicates the state of cooperation from the people who are in contact with the subject, such as superiors and colleagues at work, family, friends, neighbors, etc. For example, in the case of the Simple Occupational Stress Questionnaire, the scale of support from the surroundings for the subject is obtained from the answers to the nine questions included in the area "C". The scale of support from the surroundings indicates that the stress experienced by the subject is more easily alleviated and the subject's stress tolerance increases as the support from the surroundings increases. The scale of support from the surroundings may be obtained using the Social Support Scale for University Students (Kataju Yasushi, Onuki Naoko, "Development of a Social Support Scale for University Students and Examination of Reliability and Validity", Rissho University Psychology Research Annual Report, No. 5, pp. 37-46, 2014) or the like.

[0141] The measure of the subject's stress tolerance is a measure indicating whether the subject is likely to develop a mental disorder when exposed to the same level of stress. For example, in the case of the simple occupational stress questionnaire, it is calculated using some of the answers to the questions in the areas "A" to "C". It is preferable to appropriately determine which questions to use when calculating the measure of stress tolerance. That is, as explained in the stress vulnerability model, there are people who develop the disease and people who do not develop the disease even when exposed to the same stress, and if the subject is more vulnerable to stress (if the tolerance is low), the subject is more likely to develop the disease even with small stress. Therefore, the second calculation unit 260 can more accurately calculate the risk of developing a disease by calculating the subject's stress tolerance at the same time as the stress load.

[0142] Stress tolerance includes two elements, vulnerability and resilience, and it is believed that the two elements overlap. In determining a measure of stress tolerance, one or both of the two elements may be used. Since multiple risk factors are involved in the onset process, determining a measure of vulnerability among the measures of stress tolerance has been studied from the perspective of various risk factors. For example, a measure of vulnerability is the Narcissistic Vulnerability Scale (Yuichiro Uechi, Kazuhiro Miyashita, "Relationship between Narcissistic Vulnerability, Self-Inconsistency, and Self-Esteem as Factors of Social Phobia Tendency", Personality Research, Vol. 17, pp. 280-291, 2009). Other measures of vulnerability include the Japanese version of the Rumination-Reflection Questionnaire (Takano Keisuke, Tanno Yoshihiko, "Attempt to Create a Japanese Version of the Rumination-Reflection Questionnaire", Personality Research, Vol. 16, pp. 259-261, 2008) and the Japanese version of the Brief Core Schema Scale (Yamauchi Takashi, Sudo Anju, Tanno Yoshihiko, "Reliability and Validity of the Japanese Version of the Brief Core Schema Scale", Psychological Research, Vol. 79, pp. 498-505, 2009).

[0143] In psychology, resilience usually means a defense factor against stress, and indicates resistance to stress and recovery from illness. Resilience is thought to be divided into innate resilience and acquired resilience acquired after birth. Scales of resilience include, for example, the Two-Dimensional Resilience Factor Scale (Hirano Mari, "An Attempt to Classify Resilience into Innate and Acquired Factors," Personality Research, Vol. 19, pp. 94-106, 2010). In addition, innate resilience is considered to have four subscales: optimism, control, sociability, and initiative, while acquired resilience is considered to have three subscales: problem-solving orientation, self-understanding, and understanding of the psychology of others. Since resilience differs from person to person, the risk of illness can be determined more accurately by measuring the resilience of the person at the same time as the stress load.

[0144] As cognitive function tests, the Hasegawa Dementia Scale-Revised (HDS-R) and the Mini-Mental State Examination (MMSE) are commonly used. In addition to these, the Wechsler Memory Scale-Revised (WMS-R), the N-type Mental Function Test (N-type), the National Institute of Advanced Industrial Science and Technology Screening Test (National Institute of Advanced Industrial Science and Technology), the Montreal Cognitive Assessment (MoCA), the Dementia Assessment Sheet for Community-based Integrated Care System-21 items (DASC-21), etc. can also be used. These tests are scored, and the scores of each test can be used to evaluate normal, borderline, mild dementia, severe dementia, etc. In addition, this is merely an evaluation of cognitive function and cannot distinguish diseases by itself, but it can assist in disease judgment by the estimation system of the present invention.

[0145] For example, the HDS-R has a maximum score of 30, with a score of 20 or less indicating suspicion of dementia, with a sensitivity of 93% and a specificity of 86%. The MMSE has a maximum score of 30 points, and scores of 23 or less are considered to be indicative of dementia, with a sensitivity of 81% and a specificity of 89%. Score of 27 or less is considered to be indicative of mild cognitive impairment (MCI). (About the estimation part)

[0146] The estimation unit 123 in Fig. 1 will be described with reference to the steps in Fig. 3. In step S108, the estimation unit 123 estimates a disease based on the predicted value of the disease and / or the medical checkup data acquired by the calculation unit 122. In step S107, if the calculation unit 122 also calculates a stress value, the disease may be estimated based on the predicted value of the disease, the medical checkup data, the physical value, and the stress value.

[0147] Next, in step S109, the estimation unit 123 determines whether or not advice data corresponding to a disease has been selected. Advice data corresponding to a disease is advice for preventing a disease or avoiding aggravation when the subject receives the advice data. Advice data is created for each disease and recorded in the second recording device 120C. If advice data corresponding to a disease has been selected, the process proceeds to step S111. If advice data has not been selected, in step S110, advice data corresponding to the subject's symptoms is selected from the second recording device.

[0148] Next, in step S111, the estimation unit 123 issues an instruction to transmit the disease estimation result and the selected advice data to the first receiving unit 131 of the display unit 130 via the second transmitting unit 124.

[0149] Next, in step S112, the estimation unit 123 instructs the output unit 132 of the display unit 130 to output the estimation result and advice data.

[0150] Finally, the estimation system 100 ends the estimation process. Every time the estimation system 100 receives the subject's voice data from the communication terminal 200, it repeatedly executes the processes of steps S101 to S112.

[0151] In the above embodiment, the calculation unit 122 calculates a predicted value of the subject's disease based on the feature F(a) using the voice data of the subject acquired from the input unit 110. The estimation unit 123 estimates the subject's health condition or disease based on the calculated predicted value of the subject's disease and / or the medical examination data recorded in the second recording device 120C, and can show the estimation result together with advice data to the subject.

[0152] As described above, by performing steps S101 to S112 shown in FIG. 3, the estimation system 100 can provide an estimation system 100, an estimation program, and an estimation method that can estimate a psychiatric or neurological disorder suffered by a subject from a voice uttered by the subject, prevent the disease from becoming severe, and enable the patient to receive appropriate treatment based on an accurate diagnosis of the disease.

[0153] The estimation system 100 of the present invention is used in such a way that a doctor, clinical psychologist, nurse, laboratory technician, counselor, or other person who operates the device of the present invention uses it while facing the subject whose voice is to be acquired. In a room where a quiet environment is maintained, such as a treatment room or counseling room, one or more people who operate the device of the present invention use it while directly explaining how to acquire voice to the subject whose voice is to be acquired in an open state. In addition, the subject whose voice is to be acquired enters a hearing test room or other various types of test room, and the aforementioned operator uses it while watching the subject through glass or by a monitor image. When the subject is in a remote location such as at home, there is a method in which the method of acquiring voice is explained to the subject in advance, and the subject himself / herself records by a specified date and time. When performing it in a remote location, it is also possible to acquire voice while confirming the person by a camera image or the like using a separate communication line.

[0154] It is also possible to obtain estimates by collecting voice data when a subject visits a medical institution or counseling room, or by collecting voice data as part of a medical checkup or medical examination conducted by a company or local government. (Associating multiple diseases with voice data - Voice acquisition)

[0155] The procedure for creating an estimation program is described below.

[0156] In order to match multiple diseases with voice data, we obtained voice samples from the following patients and healthy individuals between December 25, 2017 and May 30, 2018. 20 examples of Alzheimer's disease patient voices 20 examples of voices from patients with Lewy body dementia 20 examples of Parkinson's disease patient voices 20 examples of voices from patients with major depression - 16 examples of voices from patients with bipolar disorder 19 examples of voices from patients with atypical depression 20 examples of normal voices

[0157] These patients were diagnosed with the respective disorders according to the criteria of DSM-5 or ICD-10 by doctors in specialized fields such as psychiatry or neurology. Doctors also confirmed that they did not have other psychiatric or neurological disorders by administering PHQ-9, MMSE, etc.

[0158] Healthy subjects were confirmed to have no depressive symptoms or decline in cognitive function by administering PHQ-9, MMSE, etc.

[0159] An Olympus lapel microphone and a Roland portable recorder were used to capture the audio. The audio data was recorded on an SD card.

[0160] The subjects read out the 17 sentences shown in Table 1 below, with sentences 1 to 13 read twice each and sentences 14 to 17 read once each.

[0161] [Table 1]

[0162] When acquiring the voice, the subjects were informed that the data would be used in a study to analyze the relationship between the voices of patients with neuropsychiatric disorders and their illnesses, as well as the content of the speech and the method of acquiring the voice, and were asked to sign a written consent form. In addition, the acquired data, including the voice, was encoded and managed in a format that did not identify individuals. (Creating Estimation Program 1)

[0163] Of the 17 types of speech content mentioned above, for each subject, a total of 30 utterances were obtained, including utterances 1 to 13 (26 utterances per subject, each twice) and utterances 14 to 17 (4 utterances per subject, each once).By dividing the longer utterances into two and excluding unclear utterances, a total of 4,186 utterances were obtained, including 613 utterances from patients with Alzheimer's disease, 573 utterances from patients with dementia with Lewy bodies, 608 utterances from patients with Parkinson's disease, 619 utterances from patients with major depression, 484 utterances from patients with bipolar disorder, 573 utterances from patients with atypical depression, and 615 utterances from healthy individuals.

[0164] For each of these utterances, multiple acoustic parameters were extracted, and the characteristics of the acoustic parameters, such as frequency and intensity, were associated with the subject's disease, resulting in the six diseases and seven types of healthy subjects and multiple acoustic parameters associated with them. For each disease, multiple acoustic features were extracted by selecting combinations of multiple acoustic parameters. Furthermore, we created an estimation program 1 based on the feature F(a) to combine these to estimate whether the subject has one of the six diseases or is a healthy subject.

[0165] The disease inference program used in the disease inference system of the present invention does not analyze the contents of the words spoken in relation to the acoustic features. This inference program calculates a disease prediction value from the features created by extracting acoustic features from the utterance. Therefore, it has the advantage of being language independent. However, when a subject or test subject actually speaks, the acoustic characteristics may be affected if the subject cannot speak fluently unless the sentence is in his or her native language. Therefore, for example, when inferring a disease of a subject whose native language is English, it is preferable to first collect and analyze the voices of patients and healthy people whose native language is English to create an inference program for English, and then use this to infer a disease from the English utterance. In the same manner, an inference program for languages ​​other than Japanese and English can be created.

[0166] When creating a prediction program for English and using it to predict a disease, examples of the texts that subjects or subjects read include the English texts shown below.

[0167] For example, the following is 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 particular limitations on the sentences to be spoken in any language, but it is preferable that the sentences be well-known sentences so that anyone can read them easily. Also, long vowels such as "ahhhh", "ehhhh", and "uhhhhh" are preferable because they can be pronounced by anyone, regardless of their native language.

[0168] A commercially available feature extraction program, such as openSMILE, can be used to extract acoustic features from speech.

[0169] Although such an acoustic feature extraction program can extract many acoustic features from many utterances (one phrase), in the present invention, 7440 acoustic features are used as a base.

[0170] When creating each estimation program, several acoustic features suitable for diagnosing the disease are selected from the 7,440 acoustic features, and then a formula is set up to combine them and calculate the predicted value for each disease so that each patient can be distinguished.

[0171] Therefore, the number and combination of acoustic features to be used vary depending on the type of discrimination (that is, depending on each estimation program). (Estimation results based on Estimation Program 1)

[0172] In estimation program 1, acoustic features were selected for each of the following: healthy individuals, dementia with Lewy bodies, Alzheimer's disease, Parkinson's disease, major depression, bipolar disorder, and atypical depression, and prediction values ​​for healthy individuals and each disease were calculated.

[0173] Of the 615 utterances made by healthy subjects, 383 were estimated to be from healthy subjects, 246 were estimated to be from dementia with Lewy bodies, 72 were estimated to be from Alzheimer's disease, 51 were estimated to be from Parkinson's disease, 18 were estimated to be from major depression, 30 were estimated to be from bipolar disorder, and 15 were estimated to be from atypical depression. Therefore, the proportion of utterances made by healthy subjects that were actually estimated to be from healthy subjects (sensitivity, recall) was 62.3%.

[0174] In addition, of the 600 utterances estimated to be those of healthy individuals, 383 were actually from healthy individuals, 38 were from dementia with Lewy bodies, 67 were from dementia of the Alzheimer's type, 45 were from Parkinson's disease, 32 were from major depression, 28 were from bipolar disorder, and 7 were from atypical depression. Therefore, the proportion of utterances estimated to be from healthy individuals that were actually from healthy individuals (precision, accuracy rate: Precision) was 63.8%.

[0175] Of the 940 utterances made by patients with major depression, 46 were estimated to be from healthy individuals, 43 were estimated to be from dementia with Lewy bodies, 24 were estimated to be from Alzheimer's disease, 15 were estimated to be from Parkinson's disease, 748 were estimated to be from major depression, 38 were estimated to be from bipolar disorder, and 26 were estimated to be from atypical depression. Therefore, the proportion of utterances actually made by patients with major depression that were estimated to be from patients with major depression (sensitivity, recall) was 79.6%.

[0176] In addition, of the 923 utterances estimated to be those of patients with major depression, 748 were actually those of patients with major depression, 29 were actually those of dementia with Lewy bodies, 8 were actually those of Alzheimer's disease, 33 were actually those of Parkinson's disease, 37 were actually those of healthy individuals, 33 were actually those of bipolar disorder, and 35 were actually those of atypical depression. Therefore, the proportion of utterances estimated to be those of patients with major depression that were actually those of patients with major depression (precision, matching rate) was 81.0%.

[0177] Of the 589 utterances made by patients with dementia with Lewy bodies, 41 were estimated to be from healthy individuals, 347 were estimated to be from dementia with Lewy bodies, 67 were estimated to be from Alzheimer's disease, 61 were estimated to be from Parkinson's disease, 29 were estimated to be from major depression, 24 were estimated to be from bipolar disorder, and 20 were estimated to be from atypical depression. Therefore, the proportion of utterances from patients with dementia with Lewy bodies that were actually estimated to be from patients with dementia with Lewy bodies (sensitivity, recall) was 58.8%.

[0178] In addition, 602 utterances estimated to be those of patients with Lewy body dementia were actually In total, 347 were from patients with Lewy body dementia, 46 were from patients with major depression, and There were 73 cases of dementia of the ulzheimer's type, 68 cases of Parkinson's disease, and 1 case of healthy people. There were 32 utterances from dementia with Lewy bodies, 25 from bipolar disorder, and 14 from atypical depression. Therefore, the percentage of utterances that were actually from patients with dementia with Lewy bodies and major depression (precision, matching rate) among those estimated to be from patients with dementia with Lewy bodies was 57.6%.

[0179] From these results, an AUC curve was obtained for each disease, which shows the degree of prediction for healthy subjects and other diseases. As shown in Figure 5, the AUC for the degree of prediction for each mental illness was about 0.8, indicating high accuracy. (Creating Estimation Program 1-2)

[0180] Prediction program 1-2 extracts and predicts major depression among psychiatric and neurological disorders that at least include Alzheimer's disease, Lewy body dementia, Parkinson's disease, major depression, atypical depression, and bipolar disorder, which have similar symptoms.

[0181] The types of utterances used in the analysis were 17 types of utterances (phrases), similar to those in the estimation program 1.

[0182] First, 7,440 acoustic features were extracted for each spoken phrase using openSMILE, a software that extracts acoustic features. Data with significant differences depending on the facility where the voice was acquired were excluded, and a classification algorithm was created using the features of 1,547 pieces of data by a method that combined boosting and decision trees, and a disease prediction value was generated to distinguish between major depression and other diseases.

[0183] In addition, 2,236 speech phrases from the five patients other than major depression were used, while 940 speech phrases from patients with major depression were used. In order to avoid an imbalance in the amount of data, pseudo-data of the speech of patients with major depression was created using the Synthetic Minority Over-sampling Technique (SMOTE), resulting in 3,139 speech phrases from patients with major depression.

[0184] The data was divided into 10 parts, and 10-fold cross-validation was performed on the classification algorithm of estimation program 1-2. (Estimation results based on Estimation Program 1-2)

[0185] The estimation results based on the estimation program 1-2 are shown in Table 2. When using the estimation program 1-2 to distinguish between major depression and other psychiatric and neurological disorders, the recall rate for major depression was 91.1%, the recall rate for other psychiatric and neurological disorders was 93.8%, and the accuracy rate was 92.7%. The AUC was 0.977.

[0186] [Table 2]

[0187] (Application example of Estimation Program 1-2 / Estimation Program 1-2-2)

[0188] Since the inference program 1-2-2 can distinguish between major depression and other psychiatric and neurological disorders, in an actual clinical setting it can distinguish between major depression and other disorders with similar symptoms. Specifically, it can be used to distinguish between major depression and bipolar disorder, or between major depression and atypical depression. (Creating Estimation Programs 1-3)

[0189] Prediction programs 1-3 extract and predict dementia with Lewy bodies from among psychiatric and neurological disorders that at least include Alzheimer's disease, dementia with Lewy bodies, Parkinson's disease, major depression, atypical depression, and bipolar disorder, which have similar symptoms. The types of utterances used for analysis were 17 types of utterances (phrases), the same as in Prediction Program 1.

[0190] First, 7,440 acoustic features were extracted for each spoken phrase using openSMILE, a software that extracts acoustic features. Features that showed significant differences depending on the facility where the speech was acquired were excluded, and a classification algorithm was created using 1,547 features by combining boosting and decision trees, generating disease prediction values ​​for distinguishing between Lewy body dementia and other diseases.

[0191] In addition, 2,987 speech phrases from the five patients other than Lewy body dementia were used, while 583 speech phrases from patients with Lewy body dementia were used.In order to avoid an imbalance in the amount of data, pseudo-data of the speech of patients with Lewy body dementia was created using the Synthetic Minority Over-sampling Technique (SMOTE), resulting in 2,696 speech phrases from patients with Lewy body dementia.

[0192] The data was divided into 10 parts, and 10-fold cross-validation was performed for the classification algorithms of estimation programs 1-3. (Estimation results based on Estimation Program 1-3)

[0193] The results of estimation based on prediction programs 1-3 are shown in Table 3. When using prediction programs 1-3 to distinguish between dementia with Lewy bodies and other psychiatric and neurological disorders, the recall rate for dementia with Lewy bodies was 88.6%, the recall rate for other psychiatric and neurological disorders was 89.8%, and the accuracy rate was 89.2%. The AUC was 0.959.

[0194] [Table 3]

[0195] (Application example of Estimation Program 1-3 / Estimation Program 1-3-2)

[0196] Since the prediction program 1-3-2 can distinguish between dementia with Lewy bodies and other psychiatric and neurological diseases, in actual clinical situations it can distinguish between dementia with Lewy bodies and other diseases with similar symptoms. Specifically, for example, it can be used to distinguish between dementia with Lewy bodies and Alzheimer's disease, or between dementia with Lewy bodies and Parkinson's disease. (Creating Estimation Program 2)

[0197] In estimation program 1, we first selected acoustic parameters associated with each disease, then combined those acoustic parameters to generate features and created an estimation program. In estimation program 2, we first selected two target diseases, created a combination of acoustic parameters that distinguish between the two diseases, and created a program that estimates the disease by calculating the combination of acoustic parameters that distinguish between the two diseases and their accuracy.

[0198] Prediction Program 2 selects two from six diseases, including Alzheimer's disease, Lewy body dementia, Parkinson's disease, bipolar disorder, atypical depression or major depression, and healthy state, estimates which of these the subject is most likely to have, and then performs each combination of the two to estimate the most likely disease (or healthy state).

[0199] For example, a case will be described in which either Alzheimer's dementia or Lewy body dementia, either Lewy body dementia or Parkinson's disease, or either major depression or bipolar disorder, or either of these diseases or a healthy person is estimated. Here, to estimate whether a person has Lewy body dementia or a healthy person, for example, parameters A, B, and C may be used. On the other hand, to estimate whether a person has Alzheimer's dementia or Lewy body dementia, parameters D, E, and F may be used, and the most effective parameters for estimating either of the two combinations may be different.

[0200] Even in this case, by calculating the relative probabilities for the two combinations, it is possible to rank the probabilities of the six diseases and health. If the difference between the first and second ranked diseases is greater than a certain level, it is possible to determine and estimate that the person has the first ranked disease. Also, if the difference between the first and second ranked diseases is not large, it is possible to express the probability of having the two diseases.

[0201] In this way, it is possible to estimate whether a person is likely to have either Alzheimer's disease or dementia with Lewy bodies, either dementia with Lewy bodies or Parkinson's disease, or whether they are likely to have major depression or bipolar disorder, or whether they are healthy. (Creating Estimation Program 2-1)

[0202] As shown in Table 4, speech samples were obtained from healthy subjects and patients with Alzheimer's disease from several hospitals.

[0203] For each subject, of the 17 types of speech content listed in Table 1 above, utterances 1 to 13 (26 utterances per case, 2 times each) were taken for a total of 30 utterances (4 utterances per case, 1 time each). By excluding unclear utterances, we obtained 737 utterances from patients with Alzheimer's disease and 1,287 utterances from healthy subjects.

[0204] [Table 4]

[0205] For each patient's speech, 7,440 acoustic features were extracted using openSMILE, a software for extracting acoustic features.

[0206] Examples of acoustic features include: (1) physical features calculated on a frame-by-frame basis and similar features such as Fourier transform, Mel-frequency cepstral coefficients, voiced sound probability, zero-crossing rate, and fundamental frequency (F0); (2) time-domain processing of the above (1) such as moving average and time change; and (3) statistical quantities covering frames within the same file as in (2) such as mean, maximum, minimum, center of gravity, quartile, variance, kurtosis, and skewness.

[0207] Of the 7,440 acoustic features, we first excluded features that were likely to be dependent on the recording environment, and selected 197 acoustic features.

[0208] Using these 197 acoustic features, the J48 decision tree algorithm (Weka implementation of C4.5) was used to distinguish between healthy subjects and Alzheimer's disease patients, and speech obtained from three hospitals was used for training, and speech obtained from three other hospitals was used for validation. (Estimation results based on Estimation Program 2-1) Result 1: Classification performance for each utterance

[0209] For each utterance, we used a prediction program to predict whether the utterance was from a patient with Alzheimer's disease or a healthy person. The results are shown in Table 5.

[0210] [Table 5]

[0211] Result 2: Classification performance per subject For each subject's utterances (approximately 30 each), we counted whether the utterance was estimated to be from Alzheimer's disease or from a healthy individual, and the result with the majority vote was taken as the estimation result by the estimation program. The results are shown in Table 6.

[0212] [Table 6]

[0213] (Creation of Estimation Program 2-2)

[0214] Twenty 3-second utterances of "ah" (from 20 individuals) were obtained from each patient with Parkinson's disease and dementia with Lewy bodies. Characteristic acoustic parameters were extracted from these utterances, and various acoustic parameter groups associated with each utterance of patients with Parkinson's disease and dementia with Lewy bodies were obtained by correlating the characteristics of the acoustic parameters, such as frequency and intensity, with the subject's disease. Characteristic acoustic parameters for distinguishing between Parkinson's disease and dementia with Lewy bodies were selected, and a feature-based estimation program 2-2 was created to estimate whether the subject has Parkinson's disease or dementia with Lewy bodies by combining two of the selected acoustic parameters.

[0215] In creating the estimation program, 10-fold cross-validation was performed on the utterances. (Estimation results based on Estimation Program 2-2)

[0216] For 20 utterances of "ah" from patients with Parkinson's disease (from 20 individuals) and 20 utterances of "ah" from patients with dementia with Lewy bodies (from 20 individuals), prediction program 2-2 was used to predict whether the patient had Parkinson's disease or dementia with Lewy bodies.

[0217] As a result, of the 20 utterances from patients with Parkinson's disease, 17 were presumed to be from Parkinson's disease and 3 were presumed to be from dementia with Lewy bodies.Furthermore, of the 20 utterances from patients with dementia with Lewy bodies, 18 were presumed to be from dementia with Lewy bodies and 2 were presumed to be from Parkinson's disease.

[0218] Therefore, the accuracy of prediction program 2-2 was 87.5%. For dementia with Lewy bodies, the recall was 90% and the precision was 85.7%. For Parkinson's disease, the recall was 85% and the precision was 89.5%. (Estimated Program 2-3)

[0219] Using a similar method to that used in Prediction Program 2-1, we created Prediction Program 2-3, which predicts whether a subject has Lewy body dementia or is healthy.

[0220] Nineteen "ah" utterances (from 19 individuals) from healthy subjects and 20 "ah" utterances (from 20 individuals) from patients with Lewy body dementia were used to predict whether they were healthy or had Lewy body dementia using prediction program 2-3.

[0221] As a result, of the 19 utterances from healthy subjects, 19 were estimated to be normal and none were estimated to be dementia with Lewy bodies. In addition, of the 20 utterances from patients with dementia with Lewy bodies, 17 were estimated to be dementia with Lewy bodies and three were estimated to be normal.

[0222] Therefore, the accuracy of the prediction program 2-3 was 92.3%. In addition, the sensitivity for Lewy body dementia was 85%, the specificity was 100%, the positive precision was 100%, and the negative precision was 86.4%. (Estimated Program 2-4)

[0223] Using a similar method to that used for Prediction Program 2-1, we created Prediction Program 2-4, which predicts whether a subject is healthy or suffering from major depression.

[0224] For 19 "ah" utterances (from 19 individuals) from healthy subjects and 18 "ah" utterances (from 18 individuals) from patients with major depression, prediction was made as to whether the subjects were healthy or had major depression using prediction program 2-4.

[0225] As a result, of the 19 utterances from healthy subjects, 17 were estimated to be healthy and 2 were estimated to be major depression. In addition, of the 18 utterances from patients with major depression, 17 were estimated to be major depression and 1 was estimated to be healthy.

[0226] Therefore, the accuracy of the estimated program 2-4 was 91.9%. In addition, the sensitivity for major depression was 94.4%, the specificity was 89.5%, the positive precision was 89.5%, and the negative precision was 94.4%. (Estimated Program 2-5)

[0227] Prediction program 2-5 is a program that distinguishes between healthy individuals and patients with major depression. First, 7,440 acoustic features were extracted from each phrase (approximately 30 utterances from each person) by 20 healthy subjects and 20 patients with major depression using openSMILE, a software that extracts acoustic features. Features that showed significant differences depending on the facility where the voice was acquired were excluded, and 1,547 features were selected. Then, using a method that combined boosting and decision trees, a disease prediction value was generated from the features to distinguish between healthy subjects and patients with major depression, and a classification algorithm was created. The estimation results for each phrase are shown in Table 7. The prediction program 2-5 had a sensitivity of 81.4%, a specificity of 85.5%, and an accuracy rate of 83.5% for diagnosing major depression.

[0228] [Table 7]

[0229] Table 8 shows the results of judging each phrase as either major depression or healthy for each person based on the estimation program 2-5. Furthermore, the healthy judgment rate for each person was calculated, and those with a healthy judgment rate of 60% or more were presumed to be healthy, and those with a healthy judgment rate of less than 60% were presumed to be major depression patients. As a result, as shown in Table 9, all 20 healthy people were presumed to be healthy, and all 20 major depression patients were presumed to be major depression patients. In other words, the specificity, sensitivity, and accuracy rate were all 100%.

[0230] [Table 8] TIFF0007662168000010.tif141162

[0231] [Table 9]

[0232] Therefore, by using Prediction Program 1-2 and Prediction Program 2-5 in combination, it is possible to predict major depression with a high probability among patients with Alzheimer's disease, Lewy body dementia, Parkinson's disease, major depression, atypical depression, and bipolar disorder, as well as healthy individuals. (Estimated Program 2-6)

[0233] The estimation program 2-6 is a program that distinguishes between healthy subjects and patients with dementia with Lewy bodies. A classification algorithm was created using the same procedure as the estimation program 2-5, except that the speech phrases of 20 healthy subjects and 20 patients with dementia with Lewy bodies were used as the voices. The estimation results for each phrase are shown in Table 10.

[0234] The prediction program 2-6 had a sensitivity of 81.5%, a specificity of 83.1%, and an accuracy rate of 82.2% in determining whether or not someone had Lewy body dementia.

[0235] [Table 10]

[0236] Table 11 shows the results of each person's estimation of whether they were Lewy body dementia or healthy based on the estimation program 2-6. Furthermore, the healthy judgment rate for each person was calculated, and those with a healthy judgment rate of 60% or more were estimated to be healthy, and those with a healthy judgment rate of less than 60% were estimated to be patients with Lewy body dementia. As a result, as shown in Table 12, all 20 healthy people were estimated to be healthy, and 19 of the 20 Lewy body dementia patients were estimated to have Lewy body dementia and 1 was estimated to be a healthy person. In other words, the specificity was 100%, the sensitivity was 95%, and the accuracy rate was 97.5%.

[0237] [Table 11] TIFF0007662168000014.tif132153

[0238] [Table 12]

[0239] Therefore, by using Prediction Program 1-3 and Prediction Program 2-6 in combination, it is possible to predict Lewy body dementia with a high probability among patients with Alzheimer's disease, Lewy body dementia, Parkinson's disease, major depression, atypical depression, and bipolar disorder, as well as healthy individuals.

[0240] Furthermore, by combining a prediction program that classifies a specific disease from the other five diseases with a prediction program that classifies the specific disease from healthy individuals, similar to prediction programs 2-5 and 2-6, it is possible to predict with a high probability whether a patient has Alzheimer's disease, dementia with Lewy bodies, Parkinson's disease, major depression, atypical depression, or bipolar disorder, or whether the patient is a healthy individual.

[0241] (Estimated Program 2-7) Prediction program 2-7 is a program that predicts whether a person is a healthy individual, a patient with major depression, or a patient with dementia with Lewy bodies.

[0242] A classification algorithm was created using the same procedure as in estimation program 2-5, except that the speech phrases used were those of 20 healthy subjects, 20 patients with major depression, and 20 patients with dementia with Lewy bodies. The estimation results for each phrase are shown in Table 13.

[0243] The specificity was 67.8%, the sensitivity for dementia with Lewy bodies was 70.3%, the sensitivity for major depression was 69.5%, and the accuracy rate was 69.2%.

[0244] [Table 13] (Creating Estimation Program 3)

[0245] An estimation is made for each subject's utterance, and the most frequently estimated disease (or healthy state) is estimated to be the disease (or healthy state) that the subject suffers from. For example, if the estimation results for 32 utterances obtained from subject A are 4 healthy, 16 dementia with Lewy bodies, 3 Alzheimer's disease, 3 Parkinson's disease, 2 major depression, 2 bipolar disorder, and 2 atypical depression, subject A is estimated to suffer from dementia with Lewy bodies, which was the most frequently estimated. (Creating Estimation Program 4)

[0246] For each subject's utterance, an estimate is made for each utterance, and the frequency of the estimates is used to estimate the probability that the subject has the disease (or is healthy). In this case, estimates with low frequency can be excluded.

[0247] For example, if the estimation results of 32 utterances obtained from subject A are 4 healthy, 16 dementia with Lewy bodies, 3 Alzheimer's disease, 3 Parkinson's disease, 2 major depression, 2 bipolar disorder, and 2 atypical depression, then we will exclude estimates of 3 or less, leaving 4 healthy and 16 dementia with Lewy bodies, and estimate that the probability of dementia with Lewy bodies is 80% and the probability of healthy is 20%. (Creating Estimation Program 5)

[0248] The prediction program 5 can predict two specific diseases even when they are mild. For example, mild cognitive impairment (MCI) is defined in the DSM-5 as mild dementia due to Alzheimer's disease, mild frontotemporal dementia, mild dementia with Lewy bodies, mild vascular cognitive impairment, mild dementia due to Parkinson's disease, and several other mild dementias. When dementia is mild, the characteristic symptoms of the disease are unclear and diagnosis is often difficult.

[0249] Therefore, for example, precise discrimination algorithms will be created between two specific diseases, such as an algorithm to predict Alzheimer's disease and dementia with Lewy bodies, an algorithm to predict Alzheimer's disease and frontotemporal dementia, and an algorithm to predict dementia with Lewy bodies and Parkinson's disease, which will enable prediction of one of the diseases when the condition is mild. (Creation of Estimation Program 6)

[0250] It is presumed that the patient is suffering from one of the following disorders: II. Schizophrenia spectrum disorders and other psychotic disorders, III. Bipolar disorders and related disorders, IV. Depressive disorders, V. Anxiety disorders, XVIII. Personality disorders, XVII. Neurocognitive disorders.

[0251] Such an algorithm for predicting disease groups is suitable for use, for example, when screening for mental illnesses during a health checkup, or when a person is examined for the first time by a general internist or general practitioner.

[0252] For example, we select major depression and atypical depression as the depressive disorder group, bipolar disorder as the III. bipolar disorder and related disorders group, and Alzheimer's disease and dementia with Lewy bodies as the neurocognitive disorder group, and apply characteristic voice parameters to each of these groups to create an algorithm that estimates the disease group. (Creation of Estimation Program 7)

[0253] Depending on how the features are generated and which acoustic parameters are selected, it is possible to estimate which of the three groups, a dementia group including Alzheimer's disease, dementia with Lewy bodies, and Parkinson's disease, a mood disorder group including major depression, bipolar disorder, and atypical depression, and a healthy group, to which disease the subject belongs or whether the subject is healthy. The dementia group may further include two diseases, vascular dementia and frontotemporal dementia, and the mood disorder group may further include dysthymia and cyclothymia.

[0254] In addition, depending on how the features are generated and which acoustic parameters are selected, it is possible to estimate which of the four disease groups a person belongs to: a group consisting of Alzheimer's disease and dementia with Lewy bodies, a group consisting of Parkinson's disease, a group of mood disorders including major depression, bipolar disorder, and atypical depression, and a group of healthy people.

[0255] For each subject, of the 17 types of utterances mentioned above, a total of 30 utterances were obtained, including utterances 1 to 13 (26 utterances per case, each twice) and utterances 14 to 17 (4 utterances per case, each once).By dividing the longer utterances into two and excluding the unclear utterances, a total of 4183 utterances were obtained, including 2356 utterances from patients with Alzheimer's disease, dementia with Lewy bodies, and Parkinson's disease (collectively referred to as the dementia group), 1342 utterances from patients with major depression, bipolar disorder, and atypical depression (collectively referred to as the mood disorder group), and 585 utterances from healthy individuals (healthy subject group).

[0256] Characteristic acoustic parameters were extracted from each of these utterances, and the characteristics of the acoustic parameters, such as frequency and intensity, were associated with the subject's disease, resulting in three groups: dementia group, mood-related gargling group, and healthy subjects, and various associated acoustic parameter groups. By selecting characteristic acoustic parameters for each disease and combining the selected acoustic parameters, we created a feature-based estimation program7 for determining whether a subject is in the dementia group, mood disorder group, or healthy subjects. When creating the estimation program, the utterances of each group were divided into five parts and cross-validation was performed five times. (Estimation results based on Estimation Program 7)

[0257] Of the 585 utterances made by healthy subjects, 364 were estimated to be from healthy subjects, 180 were estimated to be from the dementia group, and 41 were estimated to be from the mood disorder group. Therefore, the proportion of utterances made by healthy subjects that were actually estimated to be from healthy subjects (sensitivity, recall) was 62.2%.

[0258] In addition, of the 558 utterances estimated to be those of healthy individuals, 364 were actually from healthy individuals, 148 were actually from the dementia group, and 46 were actually from the mood disorder group. Therefore, the proportion of utterances estimated to be from healthy individuals that were actually from healthy individuals (precision, matching rate: Precision) was 65.2%.

[0259] Of the 2,356 utterances made by the dementia patients, 148 were estimated to be from healthy individuals, 2,126 were estimated to be from the dementia group, and 82 were estimated to be from the mood disorder group. Therefore, the proportion of utterances made by the dementia patients that were actually estimated to be from the dementia patients (sensitivity, recall) was 90.2%.

[0260] In addition, of the 2,389 utterances estimated to be from dementia patients, 2,126 were actually from dementia patients, 180 were actually from healthy individuals, and 83 were actually from mood disorder patients. Therefore, the proportion of utterances estimated to be from dementia patients that were actually from dementia patients (precision, matching rate: Precision) was 89.0%.

[0261] In addition, of the 1,342 utterances made by patients in the mood disorder group, 46 were estimated to be from healthy individuals, 83 were estimated to be from the dementia group, and 1,213 were estimated to be from mood disorders. Therefore, the proportion of utterances made by patients in the mood disorder group that were actually estimated to be from patients in the mood disorder group (sensitivity, recall) was 90.4%.

[0262] In addition, of the 1,336 utterances estimated to be from patients with mood disorders, 1,213 were actually from patients in the mood disorder group, 41 were actually from healthy individuals, and 82 were actually from patients in the dementia group. Therefore, the proportion of utterances estimated to be from patients with mood disorders that were actually from patients in the mood disorder group (precision, matching rate: Precision) was 90.8%.

[0263] The overall accuracy rate was 86.5%. This prediction algorithm makes it possible to predict which group a person is suffering from a disease.

[0264] An algorithm for predicting disease groups, such as prediction program 7, is also suitable for use in screening for mental disorders during health checkups, or when a patient is examined for the first time by a general internist or general practitioner, similar to prediction program 6. After predicting disease groups, a specific disease among psychiatric and neurological disorders can be predicted by using the above-mentioned prediction program 2 or the later-described prediction program 8 for each disease group. (Creation of Estimation Program 8)

[0265] Depending on how the features are generated and which acoustic parameters are selected, it is possible to predict whether a person has a psychiatric or neurological disease in a binary format, such as whether it is Alzheimer's disease or another disease, or whether it is Lewy body dementia or another disease.

[0266] For example, by analyzing voice data of dementia with Lewy bodies and voice data of other diseases such as Alzheimer's disease and Parkinson's disease and selecting acoustic parameters that can distinguish between them, a program can be created that estimates dementia with Lewy bodies from among dementia groups. Then, by calculating the voice data of a person to be estimated using this estimation program, it can be estimated whether the person has dementia with Lewy bodies or another type of dementia.

[0267] In addition, for example, by analyzing voice data of major depression and voices of other diseases such as dementia with Lewy bodies, Alzheimer's disease, Parkinson's disease, atypical depression, and bipolar disorder and selecting acoustic parameters that can distinguish between them, a program can be created that estimates major depression among psychiatric and neurological diseases. Then, by calculating the voice data of a person to be estimated using this estimation program, it can be estimated whether the person has major depression or other conditions. (Creation of Estimation Program 9)

[0268] As with estimation program 1, depending on how features are generated and which acoustic parameters are selected, it is possible to distinguish between Alzheimer's disease, dementia with Lewy bodies, Parkinson's disease, major depression, bipolar disorder, atypical depression, and health. However, after first distinguishing between the six diseases and health, the three diseases of Alzheimer's disease, dementia with Lewy bodies, and Parkinson's disease are combined into a dementia group, and major depression, bipolar disorder, and atypical depression are combined into a mood disorder group, and it is then possible to estimate whether the patient has either the dementia group or the mood disorder group, or whether they are healthy. (Creating Estimation Program 10)

[0269] Like Prediction Program 9, this program predicts whether a person has either the dementia group or the mood disorder group, or whether they are healthy. However, for each person, a disease is assigned to each of multiple sentences, and the disease assigned to the greatest number of sentences is predicted to be the person's disease, and it predicts whether the person has either the dementia group or the mood disorder group, or whether they are healthy.

[0270] The speech samples used for training were 15 patients with Alzheimer's disease, 12 patients with bipolar disorder, 14 patients with atypical depression, 15 patients with Lewy body dementia, 15 patients with major depression, 15 patients with Parkinson's disease, and 15 healthy individuals.

[0271] The test data used included speech samples from five patients with Alzheimer's disease, four patients with bipolar disorder, five patients with atypical depression, five patients with Lewy body dementia, five patients with major depression, five patients with Parkinson's disease, and five healthy individuals. The data was randomly divided into training data and test data.

[0272] The speech used for the training data was obtained by each patient and healthy subject speaking the 17 sentences mentioned above (numbers 1 to 13 spoken twice each, and numbers 14 to 17 spoken once each), yielding 30 utterances for each person.

[0273] Characteristic acoustic parameters were extracted from each of these utterances, and the characteristics of the acoustic parameters, such as frequency and intensity, were correlated with the subject's disease, resulting in the six diseases and seven types of healthy subjects, as well as various associated acoustic parameter groups.

[0274] By selecting characteristic acoustic parameters for each disease and further combining the selected acoustic parameters, we created an estimation program 10 based on features to infer for each person's spoken sentence whether they have one of the six diseases or are healthy, and then to infer whether they have a dementia group, a mood disorder group, or are healthy. (Estimation results based on Estimation Program 10)

[0275] As test data, eight phrases (16 in total, two for each phrase) were used from the 17 sentences in Table 1: 2 "aiueokakikukeko", 3 "It's a fine day today", 6 "I'm very healthy", 8 "I have an appetite", 9 "I'm calm", 10 "I'm angry", 12 "Let's look up and walk", and 13 "Let's do our best". For each utterance, we estimated whether it was a disease or healthy. Among the estimated diseases, those that were estimated to be Alzheimer's disease, Lewy body dementia, and Parkinson's disease were determined to be diseases in the dementia group. In addition, those that were estimated to be major depression, bipolar disorder, and atypical depression were determined to be diseases in the mood disorder group.

[0276] Here, the 16 utterances from each person were tallied to see whether they were judged to be dementia disease, mood disorder disease, or healthy, and the one that received the most judgments was regarded as the final judgment result. (Result 1) Patients with Alzheimer's disease (patient code AD03):

[0277] Thirteen were judged to be dementia group, one was judged to be mood disorder group, and two were judged to be healthy. Therefore, they were ultimately estimated to be dementia group. This was consistent with the dementia group to which the actual disease Alzheimer's disease belongs. The results of similar assessments of 34 subjects are shown in Table 14.

[0278] [Table 14]

[0279] The accuracy rate of the estimation program 10 was 85.3%. (Creation of Estimation Program 11)

[0280] The probability of having each disease is expressed, and the disease with the highest probability is assumed to be the one affected.

[0281] The speech samples of 20 patients with dementia with Lewy bodies, 20 patients with major depression, and 20 healthy subjects were used for both training and testing. The long vowel "aaaaa," item 14 in Table 1, was used as the speech sample.

[0282] The 7,440 acoustic features were narrowed down to the top 13 acoustic features with the highest correlations using logistic regression analysis, and these were used to calculate the disease prediction values ​​for dementia with Lewy bodies, major depression, and healthy individuals, all of which summed to 1. (Estimation results based on Estimation Program 11)

[0283] Table 15 shows whether each person was diagnosed with Lewy body dementia, major depression, or healthy based on their voice. The predicted disease value in the table is listed as the "mental value."

[0284] [Table 15] TIFF0007662168000019.tif165162

[0285] The accuracy rate of the estimation program 11 was 76.7%.

[0286] Below are the disorder groups in the Diagnostic and Statistical Manual of Mental Disorders (DSM) published by the American Psychiatric Association. As of 2018, the latest edition is DSM-5. I.Neurodevelopmental Disorders II.Schizophrenia Spectrum and Other Psychotic Disorders III. Bipolar and Related Disorders IV. Depressive Disorders V. Anxiety Disorders VI.Obsessive Compulsive and Related Disorders Obsessive-Compulsive Disorder and Related Disorders VII. Trauma and Stressor Related Disorders VIII. Dissociative Disorders IX. Somatic Symptom and Related Disorders X. Feeding and Eating Disorders XI.Elimination Disorders XII. Sleep-Wake Disorders XIII. Sexual Dysfunctions XIV. Gender Dysphoria XV. Disruptive, Impulse-Control, and Conduct Disorders XVI. Substance-Related and Addictive Disorders XVII.Neurocognitive Disorders XVIII. Personality Disorders XIX.Paraphilic Disorders

[0287] Below are the diagnostic criteria for major depression and dementia with Lewy bodies in the DSM-5. As such, it is still difficult to observe the pathology of psychiatric and neurological disorders directly or through some objective indicator, and diagnosis must be made based on symptoms. 1. Major Depressive Disorder A. Five (or more) of the following symptoms are present during the same 2-week period and represent a change from premorbid functioning, with at least one of these symptoms being (1) depressed mood or (2) loss of interest or pleasure: Note: This does not include symptoms clearly attributable to other medical conditions. (1) A depressed mood most of the day or almost every day, as indicated by the person's own words (e.g., feeling sad, empty, or hopeless) or by observation by others (e.g., appearing to be in tears). Note: Irritable mood may also occur in children and adolescents. (2) A marked decrease in interest or pleasure in all or almost all activities for most of the day, or for most of the day, or for most of the day (as indicated by the person's description or by the observations of others). (3) Significant weight loss or gain in the absence of dietary therapy (e.g., a change of more than 5% of body weight in 1 month) or loss or increase in appetite almost every day. Note: In children, consider failure to gain weight as expected. (4) Insomnia or hypersomnia almost every day. (5) Psychomotor agitation or retardation on most days (observable by others and not merely a subjective feeling of restlessness or slowness). (6) Feeling tired or lack of energy almost every day. (7) Almost daily feelings of worthlessness or excessive or inappropriate guilt (which may be delusional; not simply self-blame or guilt about being ill). (8) Diminished ability to think or concentrate or difficulty making decisions on most days (as reported by the person or observed by others). (9) Recurrent thoughts about death (not just fear of death). Recurrent suicidal thoughts or suicide attempts without a specific plan, or a definite plan to commit suicide. B. The symptoms cause clinically significant distress or impairment in social, occupational, or other important areas of functioning. C. The episode is not due to the physiological effects of a substance or to another medical disorder. Note: Criteria A–C constitute a depressive episode. Note: Reactions to significant losses (e.g., death of a loved one, financial ruin, disaster loss, serious medical illness or disability) may include intense sadness, rumination about the loss, insomnia, loss of appetite, and weight loss as described in Criterion A and may resemble a depressive episode. Although these symptoms may be understandable and appropriate in the face of loss, the presence of a depressive episode should be carefully considered in addition to normal reactions to significant loss. This determination requires the exercise of clinical judgment based on each individual's life history and cultural norms regarding how to express distress about loss. D. The depressive episode is not better explained by schizoaffective disorder, schizophrenia, schizophreniform disorder, delusional disorder, or other specified and unspecified schizophrenia spectrum disorders and other psychotic disorders. E. There has never been a manic or hypomanic episode.

[0288] 2. Major Neurocognitive Disorder with Lewy Bodies A. Meets the criteria for dementia B. The disorder develops insidiously and progresses slowly C. The disturbance meets criteria for probable or possible Lewy body dementia based on a combination of core and suggestive features. Probable dementia with Lewy bodies has two core features or one or more core features and one suggestive feature. Possible Lewy body dementia may have only one core feature or one or more suggestive features (1) Core Features (a) Cognitive perturbations with marked fluctuations in attention and alertness (b) Well-structured, detailed, and recurring visual hallucinations (c) Spontaneous parkinsonism secondary to cognitive decline (2) Suggestive features (a) Meets criteria for REM sleep behavior disorder (b) Severe neuroleptic hypersensitivity D. The disturbance is not better explained by a cerebrovascular disorder, another neurodegenerative disorder, the effects of a substance, or another psychiatric, neurological, or systemic disorder. Seventh embodiment

[0289] Next, an embodiment in which the zero-crossing rate and the Hurst exponent are selected as the second acoustic parameter will be described in detail. Note that the description of the same parts as in the first embodiment will be omitted.

[0290] The calculation unit 122 calculates the zero-crossing rate as the degree of intensity of the change in the waveform of the voice. The calculation unit 122 also calculates the Hurst exponent indicating the correlation of the change in the waveform of the voice. The calculation unit 122 outputs the calculated zero-crossing rate and Hurst exponent of the subject to the estimation unit 123.

[0291] The calculation unit 122 sets a standard range of health indicating a healthy state free from illnesses such as depression, so that the estimation unit 123 can estimate a predicted value of the subject's illness (sometimes referred to as a mental value) from the subject's zero crossing rate and Hurst exponent.

[0292] For example, the calculation unit 122 reads out voice data of a plurality of people whose health conditions, such as whether they suffer from an illness such as depression, are known from a recording device of the estimation system 100, and calculates second acoustic parameters including a zero-crossing rate and a Hurst exponent for each of the plurality of people from the read voice data.

[0293] Furthermore, the calculation unit 122 performs linear classification processing such as linear discriminant analysis or logistic regression analysis on the distribution of the zero point crossing rates and Hurst exponents of multiple people calculated by the calculation unit 122 in the two-dimensional space of the zero point crossing rates and the Hurst exponent, and creates features based on these linear models.

[0294] Next, the calculation unit 122 sets a boundary line that separates the area of ​​people suffering from depression, etc. from the reference range of healthy people who do not suffer from depression, etc. based on the feature amount. The calculation unit 122 calculates how far the zero crossing rate and Hurst exponent score of the subject are from the reference range including the determined boundary line as a predicted value of the disease, and outputs it to the estimation unit 123.

[0295] The estimation unit 123 estimates a disease from the predicted value of the disease in the subject. Then, the estimation unit 123 outputs information indicating the estimated health state to the communication terminal 200.

[0296] Fig. 6 is an example of voice data acquired via communication terminal 200 shown in Fig. 2. Fig. 6 shows the change over time in sound pressure of a voice uttered by a subject acquired via communication terminal 200. The horizontal axis of Fig. 6 indicates time t, and the vertical axis indicates sound pressure.

[0297] 6 shows data of an utterance unit in which "thank you" was spoken, among the voice data of the utterance by the subject. Times t0, t1, t2, t3, and t4 indicate the start times of utterance of each of the words "a", "ri", "ga", "to", and "u" included in the utterance unit. Note that, although the calculation process of the calculation unit 122 for the voice data in which the word "ri" was spoken among the utterance unit of "thank you" will be described, the calculation unit 122 also performs the same or similar calculation process for other words and other utterance units of "thank you".

[0298] The calculation unit 122 uses the voice data acquired from the communication terminal 200 to calculate the zero-crossing rate and the Hurst exponent for each window WD having a number of samples, such as 512. As shown in Fig. 6, since the sound pressure varies greatly in the utterance of each word, for example, the calculation unit 122 calculates the average value of the sound pressure for each window WD1 having a number of samples, such as 30, smaller than the window WD in order to calculate the zero-crossing rate, and sets the average value calculated for each window WD1 as the reference pressure for each window WD1. The calculation unit 122 measures the number of times the sound pressure of the subject crosses the calculated reference pressure (average value) in each window WD1, and calculates the zero-crossing rate.

[0299] The calculation unit 122 calculates the average value of the zero crossing rates calculated in each window WD1 as the zero crossing rate ZCR of the window WD.

[0300] On the other hand, the standard deviation σ(τ) of the difference between the sound pressure x(t) at time t and the sound pressure x(t+τ) a time τ away from time t is related as shown in formula (1). It is also known that there is a power law relationship between the time interval τ and the standard deviation σ(τ) as shown in formula (2). In formula (2), H is the Hurst exponent.

[0301]

number

[0302] For example, in the case of audio data such as white noise, there is no temporal correlation between each data item in the audio data, so the Hurst exponent H is 0. Moreover, as the audio data changes from white noise to pink noise or brown noise, that is, as the audio waveform acquires temporal correlation, the Hurst exponent H becomes greater than 0.

[0303] For example, when the audio data is Brownian noise, the Hurst exponent H is 0.5. Furthermore, as the audio data has stronger correlation than Brownian noise, that is, as the degree to which the audio data depends on the past state increases, the Hurst exponent H indicates a value between 0.5 and 1.

[0304] For example, the calculation unit 122 calculates the standard deviation σ(τ) of the audio data for each time interval τ between 1 and 15 in the window WD, and calculates the Hurst exponent H by performing a regression analysis on the standard deviation σ(τ) of each time interval τ.

[0305] The calculation unit 122 moves the window WD by a predetermined interval, such as a quarter of the width of the window WD, and calculates the zero point crossing rate ZCR and the Hurst exponent H in each window WD. Then, the calculation unit 122 averages the zero point crossing rates ZCR and the Hurst exponent H of all the calculated windows WD, and outputs the averaged zero point crossing rate ZCR and Hurst exponent H to the estimation unit 123 as the zero point crossing rate and the Hurst exponent H of the subject.

[0306] Fig. 7 shows an example of the distribution of the zero crossing rate ZCR and the Hurst exponent H of a plurality of people calculated by the calculation unit 122 shown in Fig. 1. In Fig. 7, the vertical axis indicates the zero crossing rate ZCR, and the horizontal axis indicates the Hurst exponent H.

[0307] In addition, in Fig. 7, the zero crossing rates ZCR and Hurst exponent H of people suffering from illnesses such as depression are indicated by crosses, and the zero crossing rates ZCR and Hurst exponent H of healthy people are indicated by circles. The distribution of the zero crossing rates ZCR and Hurst exponent H shown in Fig. 7 was generated using voice data of a total of 1218 people. Of the 1218 people, 697 people suffer from illnesses such as depression, and 521 people are healthy.

[0308] The calculation unit 122 executes linear classification processing such as linear discriminant or logistic regression analysis on the distribution of the zero crossing rate ZCR and the Hurst exponent H of a plurality of people shown in Fig. 7. The calculation unit 122 determines a boundary line, indicated by a dashed line, that separates people suffering from illnesses such as depression from healthy people.

[0309] 7, the boundary line indicated by the dashed line is expressed as ZCR=-0.299H+0.299. The calculation unit 122 sets the region below the boundary line indicated by the dashed line as the reference range, outputs information on the reference range including the determined boundary line to the estimation unit 123, and sets the reference range in the estimation unit 123.

[0310] In FIG. 7, the vertical axis of the zero-crossing rate ZCR and the horizontal axis of the Hurst exponent H are linear axes, but when the boundary line shown by the dashed line is expressed by an exponential function, a power function, or the like, it is preferable to use logarithmic axes in order to show the boundary line as a straight line.

[0311] Fig. 8 shows an example of the distribution of the zero-point crossing rate ZCR and the Hurst exponent H according to the acquisition environment of the audio data. In Fig. 8, similar to Fig. 7, the vertical axis indicates the zero-point crossing rate ZCR, and the horizontal axis indicates the Hurst exponent H. Fig. 8 also shows, with dashed lines, the boundary determined by the calculation unit 122 from the distribution of the zero-point crossing rate ZCR and the Hurst exponent H shown in Fig. 7.

[0312] FIG. 8 shows, by black triangles, the distribution of the zero crossing rate ZCR and the Hurst exponent H calculated using voice data obtained by sampling the subject's voice at a sampling frequency of 11 kHz by the communication terminal 200, for example.

[0313] FIG. 9 shows the zero crossings calculated in each of the windows WD using the speech data of one person. 9 shows an example of the distribution of the difference rate ZCR and the Hurst exponent H. In FIG. 9, the vertical axis is the zero crossing rate Z The horizontal axis indicates the Hurst exponent H.

[0314] A least squares method is performed using the linear function ZCR=γH+ε shown by the dashed line for the distribution of the zero crossing rate ZCR and the Hurst exponent H shown in Fig. 9 to calculate the slope γ and intercept ε. For example, using the slope γ and intercept ε calculated for each of multiple people, a threshold value indicating the boundary for classifying the voices of people suffering from depression from the voices of healthy people is determined as a healthy region.

[0315] In addition, the zero crossing rate ZCR and the Hurst exponent H calculated in each window WD are calculated using the voice data of one person, and the average values ​​and are obtained from the data of HE and ZCR for the entire section. For example, the Parkinson's disease index PDVI = +α -β for classifying the voices of people suffering from Parkinson's disease from the voices of healthy people is determined using and calculated for each of multiple people. In the case of Figure 10, α = 0.23 and β = 0.215.

[0316] Here, by using the intercept ε and PDVI, the speech data of a healthy subject, a subject with mild depression, a subject with severe depression, and a subject with Parkinson's disease are plotted as shown in FIG.

[0317] On the other hand, the communication terminal 200 downsamples the subject's voice data sampled at 11 kHz to a sampling frequency of 8 kHz in order to transmit the voice data via the network NW to the estimation system 100. Fig. 8 shows, with white rectangles, the distribution of the zero-crossing rate ZCR and the Hurst exponent H calculated using the voice data downsampled to 8 kHz.

[0318] As shown in Fig. 8, the zero-crossing rate ZCR and Hurst exponent H of the subjects are affected by the deterioration of sound quality (increased noise) due to downsampling. That is, the zero-crossing rate ZCR of the downsampled voice data is larger than the zero-crossing rate ZCR of the voice data sampled at 11 kHz because noise increases and the number of times the sound pressure of the voice crosses the reference pressure increases.

[0319] On the other hand, the Hurst exponent H of the downsampled voice is smaller than that of the voice data sampled at 11 kHz because the increased noise makes the voice data closer to white noise.

[0320] However, the zero-crossing rate ZCR and the Hurst exponent H are affected by downsampling, but they do not change independently of each other, but change in a related manner. That is, as shown in Fig. 8, the zero-crossing rate ZCR and the Hurst exponent H change along the boundary line shown by the dashed line while having a mutual correlation with each other in response to the deterioration of sound quality caused by downsampling or the like.

[0321] Therefore, deterioration of sound quality due to downsampling or the like does not affect the operation of the estimation unit 123 that determines whether the zero-crossing rate ZCR and the Hurst exponent H of the subject are within the reference range. In other words, the zero-crossing rate ZCR and the Hurst exponent H are robust against deterioration of sound quality due to downsampling, etc. Furthermore, the estimation system 100 can estimate the health condition of the subject with higher accuracy than conventional systems, regardless of the environment in which the voice data is acquired.

[0322] The estimation system 100 may be applied to, for example, a robot, an artificial intelligence, an automobile, a call center, the Internet, a mobile terminal device application or service such as a smartphone or a tablet terminal, or a search system. The estimation system 100 may also be applied to a diagnostic device, an automatic interview device, a disaster triage, etc.

[0323] The features and advantages of the embodiments will be apparent from the above detailed description. It is intended that the claims cover the features and advantages of the above-described embodiments without departing from the spirit and scope of the claims. In addition, any improvements and modifications can be easily made by a person having ordinary skill in the art. Therefore, it is not intended to limit the scope of the inventive embodiments to the above, and appropriate improvements and equivalents within the scope of the embodiments disclosed can be used. [Industrial Applicability]

[0324] It is possible to provide an estimation system, estimation program, and estimation method that estimate the voice spoken by a subject, identify and estimate the illness from which the subject is suffering, prevent the illness from becoming more severe, and enable the patient to receive appropriate treatment based on an accurate diagnosis of the illness. [Explanation of symbols]

[0325] 122...Calculation section 123…Estimation part 100…Estimation system 200...Communication terminal

Claims

1. A system for predicting mental and nervous system disorders, comprising: At least one voice input unit including an acquisition unit that acquires voice data of a subject and a first transmission unit that transmits the voice data to a server; a result display unit including a first receiving unit for receiving data of the estimated result in the server and an output unit for displaying the data; a server including a second receiving unit that receives the voice data from the first transmitting unit, a calculating unit that calculates a predicted value of a disease based on a combination of at least one acoustic feature extracted from the voice data of the subject, an estimating unit that estimates a disease of the subject, and a second transmitting unit that transmits data of the estimated result to the display unit; The server includes a processor for executing an estimation program and a recording device for recording the estimation program, and the estimation program The estimation unit estimates a disease of the subject selected from a group consisting of a plurality of diseases, or estimates whether the subject is healthy, using the predicted value of the disease as an input, In a first step, it is estimated which of a disease group including the plurality of diseases the disease belongs to by extracting and combining one or more acoustic features capable of distinguishing the diseases, and in a second step, it is estimated which of the individual diseases included in the estimated disease group the disease belongs to by extracting and combining one or more acoustic features capable of distinguishing the diseases. A prediction system for mental and neurological disorders.

2. As the first step, a first group including Alzheimer's disease, Lewy body dementia, and Parkinson's disease; The second group includes major depression, bipolar disorder, and atypical depression. The speech recognition system is calculated by extracting and combining one or more acoustic features that can distinguish the speech recognition system from the speech recognition system. The estimation system of claim 1 .

3. As the second step, Alzheimer's disease, Lewy body dementia, and Parkinson's disease The speech recognition system is calculated by extracting and combining one or more acoustic features that can distinguish the speech recognition system from the speech recognition system. The estimation system according to claim 2 .

4. As the second step, Major depression, bipolar disorder, and atypical depression The speech recognition system is calculated by extracting and combining one or more acoustic features that can distinguish the speech recognition system from the speech recognition system. The estimation system according to claim 2 .

5. A system for predicting a psychiatric or neurological disorder as described in any one of claims 1 to 4, comprising a subject terminal having a display unit and acquiring voice data for each displayed sentence.

6. A system for predicting a psychiatric or neurological disorder as described in any one of claims 1 to 5, comprising a subject terminal, the subject terminal having a display unit, displaying psychological tests and / or cognitive function tests, and the subject responding thereto.

7. The subject terminal includes the input unit and the display unit, and is connected to the server via a network. The system for predicting a psychiatric or neurological disorder according to claim 1.

8. a second recording device, the second recording device further recording a set of advice data to be presented to the subject based on the estimated result, The estimation unit selects an advice from the advice data collection of the second recording device based on the result. Selecting data and transmitting the selected data to the display unit together with the result data by the second transmission unit. Characterized by: The system for predicting a psychiatric or neurological disorder according to claim 1.

9. a second recording device, the second recording device recording answer data describing answers of the subject to a questionnaire regarding stress, and further calculating a stress value from the answer data of the subject; The estimation unit estimates the disease of the subject based on the predicted value of the disease of the subject and the stress value. The system for predicting a psychiatric or neurological disorder according to claim 1.

10. A second recording device is provided, the second recording device records the medical examination data of the subject, and further calculates a physical value from the medical examination data of the subject; The estimation unit estimates the disease of the subject based on the predicted value of the disease of the subject and the physical value. The system for predicting a psychiatric or neurological disorder according to claim 1.

11. A second recording device is provided, and the second recording device records answer data describing answers of the subject to tests regarding cognitive function, and further calculates a cognitive score from the answer data of the subject; The estimation unit estimates the disease of the subject based on the predicted value of the disease of the subject and the cognitive score. The system for predicting a psychiatric or neurological disorder according to claim 1.

12. A recording medium on which a program to be stored in the estimation system according to claim 1 is recorded.

Images