JP2020127703A - Estimation system of mild cognitive dysfunction and estimation device of mild cognitive dysfunction - Google Patents

Abstract

To estimate presence or absence of mild cognitive dysfunction in a subject at high validity.SOLUTION: An estimation system of mild cognitive dysfunction includes: a discriminator being made to learn to output the presence or absence of the mild cognitive dysfunction on the basis of the learning data comprising the correct answer mark of the answer to question concerning the orientation of age and date/time and one or more voice feature values extracted from audio data at the answer about a healthy person and a mild cognitive dysfunction onset person; a data acquisition unit for acquiring subject information containing the age information on a subject, the information on the answer to the question and the audio data at the answer to the question; a data analysis unit for acquiring the correct answer mark of the answer on the basis of the subject information, and extracting the one or more voice feature values from the audio data; and an estimated result output unit for inputting the age on the subject, the correct answer mark and the one or more voice feature values in the discriminator, and outputting the presence or absence of the mild cognitive dysfunction of the subject outputted from the discriminator as the estimated result.SELECTED DRAWING: Figure 2

Description

The present invention relates to a system for estimating mild cognitive impairment and a device for estimating mild cognitive impairment, which estimates the presence or absence of mild cognitive impairment using voice data of a subject.

As we enter a super-aging society, the number of people with dementia who visit medical institutions is rapidly increasing. According to the 2015 World Alzheimer's Report, the number of people with dementia in 2015 will reach 46.8 million worldwide. There are people with dementia who are expected to increase to 131.5 million by 2050. Although no definitive treatment for dementia is currently established, early detection of dementia may occur due to the possibility of slowing the progression of subsequent dementia by discovering it at the stage of mild cognitive impairment. The development of discovery methods is strongly desired.

As a typical dementia screening method that has been implemented so far, the revised Hasegawa-type simplified intelligence evaluation scale (HDS-R), the mini mental state test (MMSE), and the like can be mentioned. Although these methods are mainly based on voice consultation, it has been pointed out that it takes a long time to carry out screening, and it is difficult for an inspector who has received sufficient training to carry out screening. Therefore, as a technique for solving these problems, an automatic identification technique for estimating a diagnosis result by a specialist by machine learning using a voice feature amount extracted from voice data of a subject has been studied. These automatic identification techniques do not determine whether a subject has dementia, and the diagnosis of dementia is rather a diagnostic manual by the American Psychiatric Association, which is a Diagnostic Manual and Statistical Manual of Mental Disorders-5 ( Although it should be carefully done by an expert based on DSM-5) etc., it is possible to start early treatment by notifying the subject or his related person that there is a suspicion of cognitive impairment early. become.

For example, Non-Patent Document 1 (Information Processing Society of Japan, Research Report Vol. 2017-SLP-117 No. 8, pp1-6 (2017)) has a clinical dementia scale (CDR) for the purpose of early detection of dementia. Targeting subjects with a score of 0 (healthy), 0.5 (suspected of dementia), and 1 (mild dementia), and correct or incorrect answer of the subject to each question of HDS-R, answer voice Diagnosis results by specialists using various machine learning methods using the voice features defined by the utterance style analysis and the language features defined by the voice recognition result for the answer voice (CDR score 0, 0.5, 1) The result of the experiment to predict is reported. In this document, as a voice feature, a voice feature called “embase” of openSMILE (see Non-Patent Document 2 (http://auding.com/technology/opensmile)) that is widely used in emotion recognition and speech style recognition from voice. A quantity set is used, a keyword feature for a recognition result by CSJ-Kaldi is used as a language feature, and eight types of classifiers are used for machine learning. Emobase is a total of 26 types of features such as voice intensity, sound pressure, 12 mel frequency cepstrum coefficient, pitch, pitch envelope, voice probability, 8 line spectrum vs. 0 crossing ratio, and the time difference between these features. 19 kinds of statistics (maximum/minimum/range/average/maximum/minimum absolute value/slope/intercept/error of linear approximation/error of parabolic approximation/standard deviation/skewness/kurtosis of distribution -The k-th quartile (k=1, 2, 3)-The 1st-2nd quartile range-The 2nd-3rd quartile range-The 1st-3rd quartile range) were found, totaling 988. It is a statistical voice feature amount set including short-term voice feature amounts of types (26×2×19). Then, although the classification accuracy varies depending on the type of the classifier used, the accuracy of 0.56 to 0.63 was obtained by the classification using only the correct answer score regarding the question regarding the orientation of the date and time in HDS-R. On the other hand, the identification using only the voice features extracted from the answer voice to the question gives only the accuracy of 0.47 to 0.55, and the identification using only the language feature extracted from the answer voice to the question. However, it was shown that the accuracy of 0.43 to 0.56 was obtained, and the accuracy of 0.34 to 0.64 was maintained even in the discrimination in which the voice and language features were added to the correct answer score of the above question. ing. Therefore, as far as it is judged from the above results, it can be seen that the addition of the voice feature amount did not effectively act to improve the accuracy of the CDR score identification by machine learning.

Non-Patent Document 3 (Proceedings of the 2018 Acoustic Research Conference of the Acoustical Society of Japan 1-Q-44 (2018)) is not a study limited to persons with dementia with mild severity, but statistical parametric speech. A modulation spectrum (see IEEE/ACM Trans, on Audio, Speech and Language Processing, Vol. 24, No. 4, pp 755-767 (2016)) effective for synthetic speech quality of synthesis was introduced. We propose automatic identification technology. The modulation spectrum is defined as a logarithmic power spectrum of the time series of the voice feature quantity, and is a long-term voice feature quantity that represents the time variation of the time series of the voice feature quantity using a Fourier base. The modulation spectrum can be calculated from the time change of an arbitrary short-term speech feature amount, but in this document, the modulation spectrum of the mel frequency cepstrum coefficient is adopted as the modulation spectrum with high identification performance, and the mel frequency cepstrum called modcep is adopted. The low-order component of the cepstrum (modulation cepstrum) of the coefficient modulation spectrum is used as the audio feature set together with the above-mentioned emobase. Then, by using the voice feature amount and four types of discriminators included in the voice feature amount set extracted from the voice data at the time of answering the question of HDS-R, 2 of the Alzheimer-type dementia onset person and the healthy person are used. When the value classification task was performed, the identification rate was 47.2 to 54.2% when using Emobase, whereas the identification rate was 51.9 to 56.5 when using Modcep. %, the discrimination rate was 47.2 to 55.1% when the combination vector of the embase and modcep was used, and it is reported that modcep improved the discrimination rate. However, the identification rate in the automatic identification using only the voice feature amount is less than 57% even in this document.

It should be noted that there are voice feature set that can be used for the automatic identification technique in addition to the demobase and the modcep, and many voice feature sets such as Tsanas features and YAAFE features are known (for example, non-patent document 1). 5 (see Table 2 of PLos ONE 12(10):e0185613 (2017))).

IPSJ Research Report Vol. 2017-SLP-117 No. 8, pp1-6 (2017) http: //audering. com/technology/opensmile Acoustical Society of Japan 2018 Spring Research Presentation Lecture Proceedings 1-Q-44 (2018) IEEE/ACM Trans, on Audio, Speech and Language Processing, Vol. 24, no. 4, pp755-767 (2016) PLos ONE 12(10):e0185613(2017) https://www. chiba. med. or. jp/person/nursing/download/text2012_10.pdf

Although the automatic identification technology that estimates the diagnosis result by the specialist by machine learning using voice data is extremely excellent in terms of simplicity and speed, as described above, the value of the correct answer rate that can be satisfied in the studies so far is It cannot be said that it has been obtained. In addition, since it may be possible to delay the progression of subsequent dementia by the discovery at the stage of mild cognitive impairment, a technology that enables automatic identification at the stage of mild cognitive impairment is desired. Be done.

Therefore, an object of the present invention is a mild cognitive impairment estimation system that simply and quickly estimates the presence or absence of mild cognitive impairment in a subject by using voice data, with an improved correct answer rate. It is to provide an estimation system having.

In order to find a question suitable for identifying the presence or absence of suspicion of mild cognitive impairment in a short time and with a high correct answer rate, the inventors have used a typical dementia screening method and have a long-term The study was conducted by focusing on HDS-R, which has a track record. Below, each question and correct answer score of HDS-R are shown. The total score of HDS-R is 30 points, and it is said that the suspicion of dementia increases when the score is 20 points or less. Further, the rule for calculating the correct answer score for each question is set for each question. For example, when the answer to question 1 is the actual age ±2, the correct answer score is 1 point, and in other cases. As an incorrect answer, the correct answer score is 0.

The inventors first selected a person having a total HDS-R score of 14 or more as a "mild cognitive impairment onset person" from among persons who were diagnosed as dementia onset persons by an expert. That is, as far as the present invention is concerned, the term "mild cognitive impairment" means "cognitive impairment having a total HDS-R score of 14 or more in dementia". HDS-R is not used to judge the severity of dementia, but on the other hand, there is a significant difference in the overall score of HDS-R between the severity groups, indicating that the severity is mild. It is known that the average score of the group is 19±5 (see Non-Patent Document 6 (https://www.chiba.med.or.jp/personal/nursing/download/text2012_10.pdf)). In the present invention, based on this information, the term "mild cognitive impairment" was defined as described above.

Next, regarding the healthy person and the person with a mild cognitive impairment, the inventors next performed automatic identification of a healthy/mild cognitive impairment using only the correct answer score for each question of HDS-R, and HDS-R. We evaluated the correct answer rate in automatic identification of healthy/mild cognitive dysfunction using only the voice features extracted from the voice data when answering each question. The correct answer rate is a value calculated from the following formula.
Correct answer rate = (Number of true positives + Number of true negatives) / (Number of subjects)

As a result, as will be described in detail below, in the automatic identification using only the correct answer score for each question of HDS-R, the correct answer rate in the case of using question 7, that is, the question regarding delayed reproduction of three words. Was the highest, followed by Question 2, that is, the correct answer rate when using the question regarding the orientation of the date and time. In automatic identification using only the voice feature amount extracted from the voice data at the time of answering each question of HDS-R, the influence of the type of question is small and the correct answer rate is around 0.6. Similar to the result in Non-Patent Document 1, the value of the correct answer rate when only the voice feature amount was used was lower than the value of the correct answer rate when only the correct answer score was used.

The correct answer rate in the automatic identification using the correct answer score for the question 7, that is, the question regarding the delayed reproduction of three words, reached 0.9, which was an extremely high value. Therefore, it can be seen that question 7 is extremely excellent for estimating the presence or absence of mild cognitive impairment. However, since it takes a relatively long time from the start of this question until the answer is obtained, there is a problem in the ease and speed of evaluation. Therefore, in the automatic identification using only the correct answer score, the inventors have the next highest correct answer rate, and the time required from the start of the question to obtaining the answer is short, that is, the question 2, that is, the date and time. Focusing on the question about orientation, we investigated a method of improving the correct answer rate in automatic identification using voice data when answering this question. For questions regarding the orientation of the date and time, the test server for estimating the presence or absence of mild cognitive impairment in the subject and the subject terminal used by the subject to input the answer to the question by voice are available on the Internet. This is also preferable in that the inspection server can automatically determine whether the answer is correct or incorrect, even when connected via a communication line such as a line.

As a result, the inventors conducted machine learning using the age of the subject in addition to the correct answer score of the subject's question 2 and the voice feature amount extracted from the voice data at the time of answering this question. The present inventors have completed the present invention by discovering that automatic identification improves the correct answer rate to the level of the correct answer rate for question 7. As will be described in detail below, the correct answer rate in automatic identification using only age is lower than the correct answer rate in automatic identification using only correct answer points for question 2. However, the introduction of age effectively contributed to the improvement of the correct answer rate. Further, in the case of automatic identification using the correct answer score for question 2 and the voice feature amount extracted from the voice data at the time of answering this question, and automatic identification using the correct answer score and age for the answer to question 2. Did not get a satisfactory improvement in correct answer rate.

Therefore, the present invention first
An estimation system for a mild cognitive impairment, which includes an examination server and a subject terminal connected via a communication line, for estimating the presence or absence of a mild cognitive impairment in a subject,
The inspection server
One kind extracted from the sound data at the time of answering the question regarding the age and the orientation regarding the age and the orientation regarding the age and the orientation regarding each of the healthy person and the person with the mild cognitive impairment. The above speech feature amount, and a discriminator trained to output the presence or absence of mild cognitive impairment based on learning data consisting of:
About the subject, transmitted from the subject terminal, age information, information on the answer to the question regarding the orientation of the date and time, and voice data at the time of answering the question regarding the orientation of the date and time, A data acquisition unit that receives subject information including
A data analysis unit that obtains the correct answer score based on the subject information, and extracts the one or more types of voice feature amounts from the voice data;
The age of the subject, the correct answer score, and the one or more types of voice feature quantities are input to the discriminator, and the presence or absence of mild cognitive impairment in the subject output from the discriminator is estimated. As a result, the present invention relates to an estimation system for a mild cognitive impairment, comprising: an estimation result output unit for transmitting to the subject terminal.

The present invention also provides
A mild cognitive impairment estimating device for estimating the presence or absence of mild cognitive impairment in a subject,
One kind extracted from the sound data at the time of answering the question regarding the age and the orientation regarding the age and the orientation regarding each of the healthy person and the person with the mild cognitive impairment The above speech feature amount, and a discriminator trained to output the presence or absence of mild cognitive impairment based on learning data consisting of:
Data acquisition to obtain subject information including age information about the subject, information about answers to questions regarding orientation at the date and time, and voice data when answering questions regarding orientation at the date and time Department,
A data analysis unit that obtains the correct answer score based on the subject information, and extracts the one or more types of voice feature amounts from the voice data;
The age of the subject, the correct answer score, and the one or more types of voice feature quantities are input to the discriminator, and the presence or absence of mild cognitive impairment in the subject output from the discriminator is estimated. An estimation device for mild cognitive impairment, comprising: an estimation result output unit that outputs the result.

Note that in the present invention, the question regarding the orientation of the date and time is not limited to the question 2 of HDS-R, and may include, for example, the question regarding the season, and some of the questions regarding the year, month, day, and day of the week are omitted. It may be. Further, the data acquisition unit, the age information and the information of the answer to the question necessary to specify the correct answer score for the question regarding the orientation and orientation of the date and time for each subject, touch panel display or. It may be acquired as a character string via a keyboard or the like, or may be acquired as voice data. When acquired as voice data, the data analysis unit specifies the age of the subject from the voice data at the time of answering the question asking the age, and answers from the voice data at the time of answering the question regarding the orientation of the date and time. The correct answer points are specified and one or more kinds of voice feature quantities are extracted, and the obtained data is used in the estimation result output section to estimate the presence or absence of mild cognitive impairment in the subject.

The mild cognitive impairment estimation system and the mild cognitive impairment estimation device of the present invention enable the presence/absence of mild cognitive impairment in a subject to be estimated easily and quickly with a high correct answer rate.

It is a figure which shows the outline of a structure of the estimation system of the mild cognitive impairment of 1st Embodiment. It is a functional block diagram of the estimation system of the mild cognitive impairment shown in FIG. 3 is a flowchart showing a machine learning process by the classifier shown in FIG. 2. 3 is a flowchart showing a process of estimating a mild cognitive impairment by the estimation system for a mild cognitive impairment shown in FIG. 1. It is a functional block diagram of the estimation device of the mild cognitive impairment according to the second embodiment. Correct answer rate in automatic identification using only correct answer points for each question of HDS-R, correct answer rate in automatic identification using only voice feature amount extracted from voice data at the time of answering each question, and using only age It is the figure which compared with the correct answer rate in the automatic identification. When modcep is used as the voice feature set and a gradient boosting tree is used as the discriminator, the number of correct answers to the question 2 of HDS-R and the voice feature amount and age extracted from the voice data at the time of answer The correct answer rate in the automatic identification using, the correct answer rate in the automatic identification using only the correct answer score, the correct answer rate in the automatic identification using the correct answer score and the voice feature amount, and the automatic identification using the correct answer score and the age It is a figure compared with the correct answer rate in. When using embase as a voice feature set and using a gradient boosting tree as a discriminator, the number of correct answers to the question 2 of HDS-R and the voice feature amount and age extracted from the voice data at the time of answer The correct answer rate in the automatic identification using, the correct answer rate in the automatic identification using only the correct answer score, the correct answer rate in the automatic identification using the correct answer score and the voice feature amount, and the automatic identification using the correct answer score and the age It is a figure compared with the correct answer rate in. It is extracted from the correct answer score and the voice data at the time of answer to the question 2 of HDS-R when the combination vector of emobase and modcep is used as the voice feature set and the gradient boosting tree is used as the classifier. Correct answer rate in automatic identification using voice feature amount and age, correct answer rate in automatic identification using only correct answer score, correct answer rate in automatic identification using correct answer score and voice feature amount, and correct answer score and age It is a figure compared with the correct answer rate in automatic identification using and.

Hereinafter, embodiments of the estimation system for mild cognitive impairment and the estimation device for mild cognitive impairment according to the present invention will be described. However, the estimation system for mild cognitive impairment according to the present invention and the estimation for mild cognitive impairment The device is not limited to the following embodiments, and modifications can be made without departing from the spirit of the present invention.

1st Embodiment FIG. 1: is the figure which showed the outline of a structure of the estimation system 1 of the mild cognitive impairment according to this Embodiment. The estimation system 1 for mild cognitive impairment includes a subject terminal 20 for inputting/outputting information about a subject who uses the system 1, and the subject terminal 20 based on the information provided from the subject terminal 20. The examination server 10 that estimates the presence or absence of mild cognitive impairment in the examiner and provides the estimation result to the subject terminal 20, and the examination server 10 and the subject terminal 20 are connected to each other via an Internet line or the like. It is connected by a communication line 30. The inspection server 10 is configured by a general computer including an arithmetic processing unit, a communication unit, a storage unit, and the like, and cooperates with software (processing program) stored in the storage unit to serve as the inspection server 10. Is configured to work. The examinee's terminal 20 is composed of a general computer including an arithmetic processing unit, a communication unit, a touch panel display 20a for inputting/outputting data, a voice input microphone 20b, a voice output speaker 20c, and the like. The application software provided from the examination server 10 to the subject terminal 20 is configured to operate as the subject terminal 20.

FIG. 2 is a functional block diagram of the estimation system 1 for the mild cognitive impairment, which includes the examination server 10, the subject terminal 20, and the communication line 30 illustrated in FIG. For easy understanding, only one subject terminal 20 is shown.

The subject terminal 20 has a data input unit 21 and an estimation result display unit 22 as essential components. The data input unit 21 presents a question asking the age of the subject and a question as to the orientation of the date and time through the display on the touch panel display 20a and/or the voice output from the speaker 20c, and the age of the subject is displayed. Information about the subject and the answer to the question regarding the orientation regarding the date and time is acquired through the touch panel display 20a, and the voice data when the subject responds to the question regarding the orientation regarding the date and time is microphoned. 20b, and the acquired data is transmitted to the inspection server 10 via the communication line 30 as subject information. The estimation result display unit 22 displays the estimation result regarding the presence or absence of mild cognitive impairment in the subject, which is transmitted from the examination server 10 via the communication line 30, on the touch panel display 20a and/or the speaker 20c. It is presented to the subject through a voice output from.

The inspection server 10 has a learning processing unit 11 having a learning data storage unit 12 and a discriminator 13, a data acquisition unit 14, a voice feature amount extraction unit 16, a correct answer score confirmation unit 17, and an age information confirmation unit 18. The analysis unit 15 and the estimation result output unit 19 are included as essential components.

The learning processing unit 11 stores in the discriminator 13 the correct answer score for the question regarding the age and date orientation regarding the subject, and one or more types of voice feature amount extracted from the voice data at the time of answering the question. When input, the machine learning is performed on the discriminator 13 so that the presence or absence of mild cognitive impairment in the subject can be output. Identification of the age of the subject, identification of the correct answer score for the question, and extraction of the one or more types of voice feature amount from the voice data at the time of answering the question are performed by the data analysis unit 15 described later. Be seen. The learning data storage unit 12 stores, as learning data for machine learning, the age and the same date and time as that presented to the subject for each of a healthy person and a person with mild cognitive impairment. The data consisting of the correct answer score for the question regarding orientation and the one or more types of voice feature amounts extracted from the voice data at the time of answering the question are associated with the data regarding the presence or absence of mild cognitive impairment. Remembered

The voice feature amount extracted from the voice data at the time of answering the question regarding the orientation of the date and time may be one or more types, and a known voice feature amount set including a plurality of types of voice feature amounts is also used without particular limitation. be able to. As such an audio feature amount set, in addition to the demobase adopted in Non-Patent Documents 1 and 3 and the modcep adopted in Non-Patent Document 3, avec2011, avec2013, emo_lage, emobase2010, IS09_emotion, IS10_paraling, IS10_paring. , IS11_speaker_state, IS12_speaker_trait, IS12_speaker_trait_compat, IS13_ComParE, Essentia descriptors, MPEG7 descriptors, KTu featuresEatures, jAudi, and so on. It is also possible to use a set of a combination vector of voice feature quantities included in these voice feature quantity sets, for example, a combination vector of emobase and modcep, as the voice feature quantity set.

As the classifier, known classifiers that can be used for the binary classification task can be used without particular limitation, and logistic regression, support vector machines, decision trees, random forests, gradient boosting trees, random forests, etc. Discriminator used in supervised learning methods such as XG boosting, perceptron, convolutional neural network, recursive neural network, residual network, naive Bayes, k-approximation method, which are ensemble learning combined with gradient boosting tree Is mentioned. It is preferable to use random forest, gradient boosting, or XG boosting as the classifier because a high correct answer rate can be obtained even with a relatively small number of learning data.

The learning processing unit 11 operates in a preparatory stage before the estimation system 1 for mild cognitive impairment is activated, and plays a role of optimizing the discriminator 13 by machine learning using the learning data. FIG. 3 shows a flowchart showing the processing performed by the learning processing unit 11. In step S101, the learning processing unit 11 recognizes the unprocessed learning data not used for the learning process, that is, the age and the date and time of each of the healthy person and the mild cognitive impairment onset person. Whether there is data that correlates the correct answer score for the question, the one or more types of voice features extracted from the voice data at the time of answering the question, and the presence/absence of mild cognitive impairment. If there is unprocessed learning data, the process proceeds to step S102. If there is no unprocessed learning data, the process proceeds to step S104, and the series of processes ends. When there is unprocessed learning data, that is, the learning data stored in the learning data storage unit 12 before the first operation of the learning processing unit 11 or the learning data added to the learning data storage unit 12 after the first operation. If there is data, the learning processing unit 11 acquires unprocessed learning data from the learning data storage unit 12 in step S102, and in step S103, the normal/mild cognitive impairment of the healthy/mild cognitive impairment is output. Learning processing is performed so that the difference between the output result and the correct value is minimized. Prior to the learning process, the whitening process with an average value of 0 and a standard deviation of 1 is performed on the data of the age, the number of correct answers, and each voice feature amount forming the learning data. After step S103, the learning processing unit 11 returns to step S101 again and continues the processing.

The data acquisition unit 14 includes information regarding the age of the subject, which is transmitted from the subject terminal 20 via the communication line 30, information regarding the subject's answer to the question regarding the orientation regarding the date and time, and the date and time described above. The subject information consisting of voice data at the time of the subject's answer to the question regarding the orientation is received and transmitted to the data analysis unit 15.

The data analysis unit 15 includes a voice feature amount extraction unit 16, a correct answer score confirmation unit 17, and an age information confirmation unit 18. The voice feature amount extraction unit 16 extracts one or more types of voice feature amounts from voice data at the time of a subject's answer to a question regarding orientation regarding date and time. The correct answer score confirming unit 17 identifies the correct answer score based on the information of the answer of the subject to the question regarding the orientation of the date and time. The age information confirmation unit 18 identifies the age of the subject based on the age information of the subject. The data analysis unit 15 causes the speech feature amount extraction unit 16, the correct answer score confirmation unit 17, and the age information confirmation unit 18 to execute respective processes, receives the execution results, and then estimates that the processes have ended. Send to 19.

The estimation result output unit 19 receives the signal indicating the end of processing from the data analysis unit 15 and then calculates the age of the subject, the correct answer score, and the one or more types of voice feature amounts acquired by the data analysis unit 15. Each data is read, the read data is input to the discriminator 13, the discriminator 13 is made to execute the binary classification task of the healthy/mild cognitive impairment for the subject, and the output result by the discriminator 13 is obtained. The estimation result is transmitted to the subject terminal 20.

Next, a specific process in the estimation system 1 for mild cognitive impairment will be described. FIG. 4 is a flowchart showing a process of estimating the presence or absence of mild cognitive impairment in a subject by the estimation system 1 for mild cognitive impairment. In the following, description will be made assuming that the subject uses the estimation system 1 for mild cognitive impairment with the assistance of an assistant, but that the subject can independently use the system 1. Of course.

When an assistant who supports a subject who desires to use the system 1 for estimating mild cognitive impairment operates the subject terminal 20 to access the system 1, the data input unit 21 of the subject terminal 20 , After explaining the system 1 through the display on the touch panel display 20a and/or the voice output from the speaker 20c, "Please tell me your age first." or "Tell me your birth date first." A question asking the age of the subject such as "Please." is presented to the subject and the assistant through the display on the display 20a and/or the voice output from the speaker 20c (S1). On the other hand, the assistant inputs information corresponding to the question about the age of the subject to the data input unit 21 according to the display of the input image displayed on the touch panel display 20a (S2). It is preferable that the assistant confirms accurate information about the age and date of birth of the subject by referring to the insurance card or the like in advance. Next, the data input section 21 explains, “Voice the answer to the next question,” via the display on the touch panel display 20a and/or the voice output from the speaker 20c, and then “today Questions about orientation such as "what day of the year and what day of the year?" and "what day of the week?" are presented to the subject and the assistant (S3). On the other hand, the subject answers to the above-mentioned questions such as "It is XX month XX day of Heisei xx" and "It is a day of the week" in the data input section 21 through the microphone 20b. Input (S4). At this time, it is preferable that the assistant prompts the examinee to input an answer by pointing the microphone 20b at the examinee.

Then, the data input unit 21 further says, “For confirmation, enter the answer to the question above according to the display on the screen” via the display on the touch panel display 20a and/or the voice output from the speaker 20c. Such as, prompting the input of information for confirming the subject's answer to the question regarding the orientation of the date and time, the assistant is the information of the subject's answer according to the display of the input image displayed on the touch panel display 20a. Is input to the data input unit 21 (S5). The confirmation information may be input by repeating the voice response such as "Heisei xx year xx month xx day" and "○ day of the week" as is, for example, year, month, day, day of the week. If only one answer is wrong, the correctness of the answer such as "○, ○, ○, ×" may be input. The data input unit 21 collects the information obtained in steps S2, S4, and S5 as subject information and transmits it to the inspection server 10 via the communication line 30 (S6).

The data acquisition unit 14 in the inspection server 10 receives the subject information about the subject from the subject terminal 20 and transmits it to the data analysis unit 15 (S7). The data analysis unit 15 first transmits to the voice feature amount extraction unit 16 voice data at the time of answering a question regarding the orientation of the date and time in the subject information, and from this voice data, One or more types of voice feature quantities are extracted and the extraction result is received (S8). The voice feature amount extraction unit 16 extracts a voice feature amount of the same type as one or more types of voice feature amounts stored in the learning data storage unit 12 according to an extraction rule corresponding to this voice feature amount. Next, the data analysis unit 15 transmits the information of the answer to the question regarding the orientation of the date and time in the subject information to the correct score confirmation unit 17, and identifies the correct answer score for the subject. Then, the identification result is received (S9). The correct answer score confirmation unit 17 specifies the correct answer score according to a confirmation rule according to the format of the information of the transmitted answer. When HDS-R question 2 is adopted as a question regarding the orientation of the date and time, 1 point is assigned to each of the year, month, day, and day of the week, and the total score of the answers that have obtained the correct answer is the correct answer score. However, if the response information is sent in a format such as “Heisei xx year xx month xx day” or “xx day of the week”, the year, month, day, day of the week in the response and year of inspection day , Month, day, day of the week are compared to determine the correct answer, the correct answer score is specified, and if the answer information is sent in a format such as “○, ○, ○, ×”, the correct answer (○ ) Is specified as the correct answer score. Next, the data analysis unit 15 transmits the age information of the subject information to the age information confirmation unit 18, causes the age of the subject to be identified, and receives the identification result (S10). The age information confirmation unit 18 specifies the age according to a confirmation rule according to the format of the transmitted information on the age. When the age itself is sent, the sent age is directly specified as the age of the subject, and when the age information is sent in the form of the date of birth, the date of examination and the date of birth are combined. Age is specified from the difference.

Next, when the data analysis unit 15 transmits a signal indicating that the processing has been completed to the estimation result output unit 19, the estimation result output unit 19 causes the data analysis unit 15 to acquire the age of the subject and the above. The correct answer score and the one or more kinds of speech feature quantities are input to the learned discriminator 13 (S11), and the discriminator 13 is caused to execute a binary classification task for healthy/mild cognitive impairment, and the discriminator The output result by 13 is received (S12), and the output result is transmitted to the subject terminal 20 via the communication line 30 (S13).

The estimation result display unit 22 of the subject terminal 20 receives the output result regarding the presence or absence of mild cognitive impairment in the subject (S14), and outputs the output result in the estimation system 1 for mild cognitive impairment. As a result, it is presented to the subject and the assistant through the display on the touch panel display 20a of the subject terminal 20 and/or the voice output from the speaker 20c (S15), and the process is ended. If an output result that the cognitive function of the subject is healthy is obtained, an estimation result such as "There is no problem with your cognitive function" is presented to the subject and the assistant, If you get the output that the subject has mild cognitive impairment, "Your cognitive decline may be present. We recommend that you consult a specialist." Such an estimation result is presented to the subject and the assistant.

Second Embodiment A second embodiment of the present invention is a device for estimating mild cognitive impairment, which works as a single device without going through a communication line. The device 40 for estimating a mild cognitive impairment comprises a general computer system having an arithmetic processing unit, a storage unit, a touch panel display for inputting/outputting data, a voice input microphone, a voice output speaker, and the like. In cooperation with software (processing program) stored in the storage unit, the apparatus is configured to operate as the estimation device 40 for mild cognitive impairment.

FIG. 5 is a functional block diagram of the estimation device 40 for mild cognitive impairment. The apparatus 40 for estimating mild cognitive impairment includes a learning processing unit 41 having a learning data storage unit 42 and a discriminator 43, a data acquisition unit 44, a voice feature amount extraction unit 46, a correct answer score confirmation unit 47, and age information. It has a data analysis unit 45 having a confirmation unit 48 and an estimation result output unit 49 as essential components. Of these, the learning processing unit 41 having the learning data storage unit 42 and the discriminator 43 is the learning data storage unit 12 and the discriminator in the examination server 10 of the estimation system 1 for the mild cognitive impairment according to the first embodiment. The data analysis unit 45, which has the same function as the learning processing unit 11 including 13 and includes the voice feature amount extraction unit 46, the correct answer score confirmation unit 47, and the age information confirmation unit 48, is the same as that of the first embodiment. It has the same function as the data analysis unit 15 including the voice feature amount extraction unit 16, the correct score confirmation unit 17, and the age information confirmation unit 18 in the inspection server 10 of the estimation system 1 for the mild cognitive impairment.

The data acquisition unit 44 presents a question regarding the age of the subject and a question regarding the orientation of the date and time through the display on the touch panel display and/or the voice output from the speaker, and the information regarding the age of the subject. And the information about the response of the subject to the question regarding the orientation of the date and time is obtained through the touch panel display, and the voice data at the time of the response of the subject to the question regarding the orientation of the date and time is input via the microphone. The acquired data is transmitted to the data analysis unit 45 as subject information. The estimation result output unit 49 receives the signal indicating the end of processing from the data analysis unit 45, and then obtains the number of correct answers to the question regarding the age and the orientation regarding the subject acquired by the data analysis unit 45. At least one type of voice feature amount extracted from the voice data at the time of answering the question regarding the orientation regarding the date and time is read, the read data is input to the discriminator 43, and the discriminator 43 receives the normal sound of the subject. / The binary classification task of mild cognitive impairment is executed, and the output result by the discriminator 13 is presented as the estimation result via the display on the touch panel display and/or the voice output from the speaker.

The estimation device 40 for the mild cognitive impairment according to the second embodiment is different from the estimation system 1 for the mild cognitive impairment according to the first embodiment in that the data acquisition unit 44 receives data via a communication line. The point that the data acquisition unit 44 acquires the subject information by itself without receiving the subject information provided from the examiner terminal, and the estimation result output unit 49 indicates whether the subject has a mild cognitive impairment. The estimation result is not transmitted to the subject terminal via the communication line, and the estimation result output unit 49 itself estimates the estimation result via the touch panel display and/or the speaker provided in the estimation device 40 for the mild cognitive impairment. Since they are the same except that they are presented, a further description will be omitted.

Modification In the modification of the estimation system 1 for the mild cognitive impairment according to the first embodiment and the estimation device 40 for the mild cognitive impairment according to the second embodiment, the data acquisition units 14 and 44 are used to perform the examination. Age information about the person and the answer of the subject to the question regarding the orientation of the date and time are acquired by the voice data. In this case, in step S2 of FIG. 4, a process of obtaining age information about the subject by voice instead of the process of obtaining by input from the touch panel display is executed, and the process of step S5 of FIG. That is, there is no need for the process of obtaining the answer to the question regarding the orientation of the date and time by voice data in step S4 and then confirming the answer by input from the touch panel display. In this modification, the correct answer confirmation units 17 and 47 apply a voice recognition device for the elderly based on S-JNAS or the like to voice data at the time of the answer of the subject to the question regarding the orientation of the date and time to convert the voice data into voice data. After converting to a character string, the correct answer score of the subject is specified, and the age information confirmation units 18 and 48 apply the same voice recognizer to the voice data at the time of answering the question asking the age of the subject. After converting the voice data into a character string, the age of the subject is specified.

The present invention will be described using the following examples, but the present invention is not limited to the following examples.

(1) Experimental conditions The distribution of the total HDS-R score among the participants and participants of this experiment is shown below. The persons with mild cognitive impairment in the table are composed of persons whose HDS-R score is 14 to 25 among persons diagnosed as having dementia by an expert based on DSM-5. ing.

Recording of voice data at the time of answering each question of HDS-R is performed by two persons, a recording participant (healthy person or a person with mild cognitive impairment) and a diagnostic person, or a recording participant in a diagnostic room of a medical institution. It was carried out by 3 people including the attendants. The sampling frequency of the recorded voice was 48 kHz, the file format was RIFF WAV, and the voice was normalized by downsampling to 16 kHz during voice analysis. The audio data of the recording participants was manually extracted from the recorded audio data.

Ages of healthy subjects and persons with mild cognitive impairment shown in Table 2, correct answer scores for each question of HDS-R, and voice feature amount extracted from voice data at the time of answering each question Based on the data consisting of, the correct answer rate in the automatic discrimination of healthy/mild cognitive impairment was evaluated. As the audio feature amount set including the audio feature amount extracted from the audio data, there are an embbase adopted in Non-Patent Documents 1 and 3, a modcep adopted in Non-Patent Document 3 (however, Fourier transform length is 2048, frame The shift was set to 5 msec. The number of types of voice feature amount was 20.), and three types of a set (hereinafter, referred to as “emobase+modcep”) composed of a combined vector of the above-mentioned emobase and the above-mentioned modcep were used. Further, three types of gradient boosting tree (GBC), random forest (RF), and support vector machine (SVM) were used as discriminators for automatic discrimination using data including voice feature amounts. These discriminators were prepared using the scikit-learn library (https://github.com/scikit-learn/scikit-learn, version 0.19.1). However, for the purpose of tuning the hyperparameters of each classifier, the cost parameter of the linear support vector machine, the cost parameter and the gamma parameter of the RBF kernel support vector machine, the number and class weight of the estimator of the random forest, the gradient boosting tree A grid search was used to optimize each of the number of estimators and the maximum depth.

In the evaluation of the automatic discrimination using the above-mentioned discriminator, the learning data for dividing the healthy subjects and the mild cognitive impairment onset shown in Table 2 into arbitrary 5 groups and allowing the discriminator to learn 4 groups of data The data of the remaining one group was used for calculating the correct answer rate. When inputting the learning data to the discriminator, the data of age, correct answer score, and voice feature amount were whitened and then input. In addition, the calculation of the correct answer rate was performed by cross validation in which evaluations with different groups for calculating the correct answer rate were repeated 5 times.

(2) Preliminary examination: discrimination using only correct answer points/discrimination using only age/discrimination using only voice feature amount Prior to the evaluation of the automatic discrimination of healthy/mild cognitive impairment according to the present invention, only age is evaluated. Correct answer rate in automatic identification when used, correct answer rate in automatic identification when using only the correct answer score for each question of HDS-R, and extracted from voice data when answering each question of HDS-R The correct answer rate in the automatic identification when only the speech feature amount was used was evaluated. Table 3 shows the obtained values of the correct answer rate.

但し、機械学習は本来一次元のデータを用いた機械学習には適さないため、正解点数のみを用いた自動識別における各質問についての正答率の値及び年齢のみを用いた自動識別における正答率の値は、最も正答率が高くなるように識別のカットオフ値を定めた上で得た。仮に一次元のデータを用いて機械学習により正答率を得たとしても、表中の正答率を超えることは無いと考えられる。

However, since machine learning is originally not suitable for machine learning using one-dimensional data, the value of the correct answer rate for each question in automatic identification using only correct answer points and the correct answer rate in automatic identification using only age The value was obtained after defining the cutoff value for discrimination so that the correct answer rate was the highest. Even if the correct answer rate is obtained by machine learning using one-dimensional data, it is considered that the correct answer rate in the table is not exceeded.

In FIG. 6, voice features are set and modcep is used, and gradient boosting tree (GBC) is used as a discriminator to extract the voice data at the time of answering each question of HDS-R. The value of the correct answer rate in the automatic identification using only the voice feature amount, together with the value of the correct answer rate for each question in the automatic identification using only the correct answer score and the value of the correct answer rate in the automatic identification using only the age. Indicated.

As can be seen from Table 3 or FIG. 6, in the automatic identification using only the correct answer score for each question of HDS-R, the correct answer rate greatly changed depending on the type of question. Question 7 had the highest percentage of correct answers when using the question regarding delayed playback of three words, followed by question 2, that is, the highest percentage of correct answers when using the question regarding orientation of the date and time. It was higher than the correct answer rate when using only. The correct answer rate when using only the correct answer score for question 7 reaches 0.9, which shows that this question is a very excellent question regarding automatic identification of healthy/mild cognitive impairment. However, since it takes a relatively long time from the start of this question until the answer is obtained, there is a problem in terms of simplicity and speed of evaluation.

Further, as can be understood from Table 3 or FIG. 6, regarding automatic identification using only the voice feature amount extracted from the voice data at the time of answering each question of HDS-R, the type of voice feature amount set, the discriminator The effect of differences in the type of question and the type of question was relatively small, and the correct answer rate was around 0.6. As can be seen from Table 3, most of the correct answer rates in the automatic identification using only the voice feature amount exceed 0.5. Therefore, the voice feature amount is used to identify the normal/mild cognitive impairment. Can be said to be effective, but has not reached a satisfactory correct answer rate. As for the questions 2 to 9, like the results in Non-Patent Document 1, the value of the correct answer rate when only the voice feature amount is used tends to be lower than the value of the correct answer rate when only the correct answer score is used. there were.

(3) Discrimination Using Age, Correct Answer Score, and Voice Feature Amount As described above, the correct answer rate is extremely high when only the correct answer score for the question 7 is used, but it is easy and quick to evaluate. Since there is a problem, the correct answer rate in the automatic identification using only the correct answer score is next high, and the time required from the start of the question to obtaining the answer is short. Considered identification. Hereinafter, the result of setting the audio feature amount and using modcep or emobase will be described.

In Table 4, the correct answer rate in the automatic identification (the present invention) using the age of the subject, the correct answer score for the question 2 and the voice feature amount extracted from the voice data at the time of the answer to the question 2 is asked. Using the correct answer rate in automatic identification using only the correct answer score for 2, the correct answer rate in automatic identification using the answer score and the voice feature amount for question 2, and the correct answer score and age for the answer to question 2. It shows in comparison with the correct answer rate in the automatic identification.

Further, Table 5 shows the results of the comparative experiment using Question 7 instead of Question 2.

As understood from Table 5, when the answer to the question 7 is used, the automatic identification using the correct answer score and the voice feature amount, the automatic identification using the correct answer score and the age, and the age and the correct answer score. The rate of correct answers in any of the automatic identifications using the voice feature amount was lower than that in the automatic identifications using only the correct answer points. In particular, when a support vector machine (SVM) is used as a classifier, the correct answer rate of automatic identification using data including a voice feature amount is remarkably decreased, and when embase is used as a voice feature amount set, when modcep is used. The rate of correct answers was significantly lower than that of. It is generally known that a large amount of learning data is required to perform machine learning stably, but in the number of healthy people and mild cognitive impairment onset shown in Table 2, This is insufficient for stable machine learning using a vector machine, and the effect of this shortage is that emobase including 988 kinds of voice feature quantities is more than the case where modcep including 20 kinds of voice feature quantities is used. It is considered that the more remarkable appearance was observed when using. If the number of healthy subjects and those with mild cognitive impairment increases, it is expected that the correct answer rate will be obtained using the support vector machine as well as when using the random forest or the gradient boosting tree.

However, as can be understood from Table 4, when the answer to question 2 is used, when the random forest (RF) or the gradient boosting tree (GBC) is used as the discriminator, the age, the correct answer score, and the voice feature amount are used. The correct answer rate in the automatic identification using and improved, and the correct answer rate improved to the level of the correct answer rate (0.903) in the automatic identification using only the correct answer score for the question 7. However, when the support vector machine (SVM) is used as the discriminator, as in the case where the answer to the question 7 is used, the lack of the number of healthy persons and those with mild cognitive impairment causes the lack of voice characteristics. The correct answer rate of the automatic identification using the data including the amount is decreased, and the correct answer rate is significantly decreased when the emote is used as the voice feature amount as compared with the case where the modcep is used. If the number of healthy subjects and those with mild cognitive impairment increases, it is expected that the correct answer rate will be obtained using the support vector machine as well as when using the random forest or the gradient boosting tree.

Emobase is a set of statistical voice features that is composed of 988 types of short-term voice features, and modcep is a set that includes 20 types of long-term voice features that represent time variations of time series of voice features. Although the two are different in terms of temporal characteristics in particular, as can be seen from Table 4, regardless of the difference in the characteristics of the above-mentioned voice features, for automatic identification of healthy/mild cognitive impairment. It can be seen that the automatic identification using the age of the subject, the correct answer score for the question 2, and the voice feature amount extracted from the voice data at the time of answering the question 2 is effective. In particular, when a gradient boosting tree (GBC) is used as a classifier, the correct answer rate is significantly improved in automatic classification using age, correct answer score, and voice feature amount, as shown below.

(4) Effect of voice feature set In FIG. 7, when the voice feature set and modcep are used and the gradient boosting tree is used as the discriminator, the correct answer score and the answer to the question 2 of HDS-R are shown. The correct answer rate in the automatic identification (the present invention) using the voice feature amount and the age extracted from the voice data at the time was used as the correct answer rate, the correct answer score and the voice feature amount in the automatic identification using only the correct answer score. The result of comparison with the correct answer rate in automatic identification and the correct answer rate in automatic identification using correct answer score and age is shown.

FIG. 8 shows the number of correct answers to the question 2 of HDS-R and the speech extracted from the speech data at the time of the answer when the emote is used as the speech feature set and the gradient boosting tree is used as the classifier. The correct answer rate in the automatic identification (the present invention) using the feature amount and the age, the correct answer rate in the automatic identification using only the correct answer score, the correct answer rate in the automatic identification using the correct answer score and the voice feature amount, and the correct answer score. The results of comparison with the correct answer rate in automatic identification using age and age are shown.

Further, in FIG. 9, when the emboss+modcep is used as the voice feature amount set and the gradient boosting tree is used as the discriminator, the correct answer points to the question 2 of HDS-R and the voice data at the time of the answer are extracted. Correct answer rate in automatic identification using the voice feature amount and age (the present invention), correct answer rate in automatic identification using only correct answer score, correct answer rate in automatic identification using correct answer score and voice feature amount, and The result compared with the correct answer rate in automatic identification using a correct answer score and age is shown.

As can be understood from FIGS. 7 to 9, the accuracy of the correct answer in the automatic identification using the correct answer score and the voice feature amount and the automatic identification using the correct answer score and the age is improved by the automatic identification using only the correct answer score. It cannot be said that it is remarkable compared with the correct answer rate in the above, and the correct answer rate was rather lowered in the automatic identification using the correct answer score and the voice feature amount in FIG. The correct answer rate in the used automatic identification (present invention) is remarkably improved, and the correct answer rate in the automatic identification using the age, the correct answer score, and the voice feature amount is different from the voice feature in terms of temporal characteristics. The values were almost the same regardless of whether the quantity set (emobase, modcep) or the set of combined vectors (emobase+modcep) was used. Therefore, the automatic identification (the present invention) using the correct answer score for the question 2 of HDS-R, the voice feature amount extracted from the voice data at the time of the response, and the age (the present invention) is the automatic identification of the healthy/mild cognitive impairment. Therefore, it was found that it is extremely effective, and that a high correct answer rate can be obtained without depending on the difference in the characteristics of the voice features.

(5) Field test Using modcep as a voice feature set, we constructed a system for estimating mild cognitive impairment using a trained gradient boosting tree as a discriminator, and this system was used to recognize new subjects. A field test was conducted to evaluate whether functional changes could be correctly judged. Subjects in this test are those who are 75 years of age or older who are informed by DSM-5 and are diagnosed as having Alzheimer's type mild dementia or mild cognitive impairment (MCI). And healthy people who voluntarily participated through the local elderly group. Clinical dementia scale (CDR) was previously performed on these subjects, and based on the CDR score, CDR0 group (healthy, 62 people), CDR0.5 group (suspected dementia or mild cognitive impairment, 13 people), And CDR1 group (mild dementia, 15 persons). The clinical validity of the system is evaluated by comparing the classification by CDR and the determination result by the system. The CDR1 group is included in “subjects of mild cognitive impairment” defined in the present specification.

As a result of the field test, the above system judged that about 90% of the CDR0 group was “good cognitive function” and about 90% of the CDR1 group was “changed cognitive function”, and the high correct answer in any group. Showed the rate. In addition, although the above system is an estimation system for mild cognitive impairment, it is interesting that about 70% of the CDR0.5 group was determined to have "change in cognitive function", and the mild cognitive impairment (MCI) of the above system was determined. It was confirmed to be effective for estimating the presence or absence of.

From the above results, it can be seen that the above system gave a clinically valid judgment in an extremely short time of tens of seconds.

By the estimation system of the mild cognitive impairment of the present invention and the estimation device of the mild cognitive impairment, it is possible to easily and quickly estimate the presence or absence of mild cognitive impairment in the subject with a high correct answer rate, It can be linked to the start of treatment at the stage when mild cognitive impairment is observed.

1 Estimating system for mild cognitive impairment 10 Examination server 20 Subject terminal 30 Communication line 13 Discriminator 14 Data acquisition unit 15 Data analysis unit 19 Estimated result output unit 40 Estimator for mild cognitive impairment 43 Discriminator 44 data Acquisition unit 45 Data analysis unit 49 Estimated result output unit

Claims (4)

An estimation system for a mild cognitive impairment, which includes an examination server and a subject terminal connected via a communication line, for estimating the presence or absence of a mild cognitive impairment in a subject,
The inspection server is
One kind extracted from the sound data at the time of answering the question regarding the age and the orientation regarding the age and the orientation regarding the age and the orientation regarding each of the healthy person and the person with the mild cognitive impairment. The above speech feature amount, and a discriminator trained to output the presence or absence of mild cognitive impairment based on learning data consisting of:
Age information about the subject transmitted from the subject terminal, information on answers to questions regarding the orientation regarding the date and time, and voice data at the time of answering questions regarding the orientation regarding the date and time. A data acquisition unit that receives subject information including
A data analysis unit that acquires the correct answer score based on the subject information, and extracts the one or more types of voice feature amounts from the voice data;
The age of the subject, the correct answer score, and the one or more kinds of voice feature amounts are input to the discriminator, and the presence or absence of mild cognitive impairment in the subject output from the discriminator is estimated. As a result, an estimation result output unit for transmitting to the subject terminal is provided, and the estimation system for mild cognitive impairment is provided. The data acquisition unit receives, as the subject voice information, voice data at the time of answering each of the question about the age of the subject and the question regarding the orientation regarding the date and time, instead of the subject information. ,
The data analysis unit, based on the voice information of the subject, specifies the age of the subject from the voice data at the time of answering the question asking the age, and from the voice data at the time of answering the question regarding orientation of the date and time, The estimation system for mild cognitive impairment according to claim 1, wherein the correct answer score is specified and the one or more kinds of voice feature quantities are extracted. The estimation system for mild cognitive impairment according to claim 1 or 2, wherein the classifier is random forest, gradient boosting tree, or XG boosting. A mild cognitive impairment estimating device for estimating the presence or absence of mild cognitive impairment in a subject,
One kind extracted from the sound data at the time of answering the question regarding the age and the orientation regarding the date and time, and the correct answer score for the question regarding the orientation regarding the age and the date regarding each of the healthy person and the person with the mild cognitive impairment. The above speech feature amount, and a discriminator trained to output the presence or absence of mild cognitive impairment based on learning data consisting of:
Data acquisition for obtaining subject information including age information about the subject, information about answers to questions regarding orientation at the date and time, and voice data at the time of answering questions regarding orientation at the date and time Department,
A data analysis unit that acquires the correct answer score based on the subject information, and extracts the one or more types of voice feature amounts from the voice data;
The age of the subject, the correct answer score, and the one or more kinds of voice feature amounts are input to the discriminator, and the presence or absence of mild cognitive impairment in the subject output from the discriminator is estimated. An estimation device for mild cognitive impairment, comprising: an estimation result output unit that outputs the result.

Images