KR20190021896A - System And Method For Detecting And Predicting Brain Disease - Google Patents

Abstract

In the present invention, disclosed are a system and a method for detecting and predicting a brain disease such as dementia or the like. In an embodiment, disclosed is a method for detecting and predicting a brain disease, in a method for detecting and predicting a brain disease such as cognitive disorders and dementia based on cognitive ability of a user, which comprises the steps of: obtaining response of a user to questions related to memory, attention, computation, linguistic, and perceptive abilities through an input unit; extracting, a control unit, data related to memory, attention, computation, linguistic, and perceptive abilities from the response of the user; extracting, by the control unit, feature data related to a cognitive ability of the user from the extracted data; learning, by the control unit, a brain disease detection and prediction model using the feature data; and inferring, by the control unit, a cognitive state of the user from the learned brain disease detection and prediction model.

Description

Technical Field [0001] The present invention relates to a system and method for detecting and predicting brain diseases,

The present invention relates to a system and method for detecting and predicting a user's brain disease such as dementia.

As the aging society continues, the number of patients suffering from brain diseases such as dementia and cognitive decline is increasing. In particular, in the case of dementia, accurate detection and prediction of a high-risk group of dementia (potential dementia patients) can prevent the progression to dementia or facilitate continuous management, since the treatment response is limited after the progress of dementia have. Conventional cognitive intelligence tests for predicting dementia have been used as screening tools such as PET and fMRI, which have high testing costs to diagnose dementia and other cognitive disorders. Especially, the simple mental state examination tool such as MMSE-K, which is a scale form, is actively used in the current screening test for dementia.

However, dementia is diagnosed when subjective and objective cognitive decline is accompanied by significant impairment of daily living performance. However, current cognitive screening has limitations in capturing minute cognitive change signals of high - risk dementia. In addition, the existing cognitive status check tool exists in the form of a scale, and it is a handwritten tool, which makes it difficult to manage the data of the test procedure and the test result, and the diagnosis result may vary depending on the examiner. In addition, since the cognitive status check of this method is mainly performed in the offline mode, it is difficult to consider the continuous change of the cognitive status of the user.

The present invention provides a system and method for detecting and predicting a user ' s brain disease such as dementia.

A method for detecting and predicting brain diseases according to the present invention is a method for detecting and predicting brain diseases including cognitive dysfunction and dementia based on cognitive ability of a user. The method includes the steps of: inputting questions related to memory, attention, Obtaining a user's response to the user; Extracting data related to memory, attention, computation power, linguistic ability, and perception ability from the user's response; Extracting feature data related to a cognitive power of the user from the extracted data; Learning the brain disease detection and prediction model using the feature data; And inferring the perception state of the user from the learned brain disease detection and prediction model by the controller.

Here, the brain disease detection and prediction model includes an intra-model using past and current play feature data and a perceived state of a user, and an inter-model using play data of a user community having demographic data similar to a user .

And the intra-model and inter-model may be configured to include a probability that the user belongs to the cognitive state normal, the cognitive disorder, and the dementia, respectively, with probability related to the cognitive state class of the user.

In addition, the step of learning the brain disease detection and prediction model may perform the simultaneous learning through the partial teacher learning of the intra model and the inter model.

In addition, the intra-model may be configured using a linear-chain conditional random field in which a set of previous feature vectors of the user and a set of current feature vectors are input.

Also, the inter-model may be composed of an input generating part and an artificial neural network.

In addition, the probability score vector generated through the intra-model and the inter-model can be externally determined to determine the perceived state.

Further, the learning of the brain disease detection and prediction model may include, among the conditions that the user does not know the recognition state, or the condition that the number of iterations for learning becomes equal to a predetermined threshold, One can be repeated until it is first achieved.

In addition, the brain disease detection and prediction system according to the present invention is a system for detecting and predicting brain diseases including cognitive dysfunction and dementia based on the cognitive power of a user. The system detects and predicts brain diseases such as memory, attention, A display presented to the user; An input unit for inputting a user's response to questions related to memory, attention, calculation, linguistic ability, and perception; And a control unit for extracting data related to memory, attention, computation power, linguistic ability, and perception ability from the user's response. The control unit extracts feature data related to the perceived power of the user from the extracted data, A disease detection and prediction model is learned, and the user's perception state can be deduced from the learned brain disease detection and prediction model.

Here, the brain disease detection and prediction model includes an intra-model using past and current play feature data and a perceived state of a user, and an inter-model using play data of a user community having demographic data similar to a user .

And the intra-model and inter-model may be configured to include a probability that the user belongs to the cognitive state normal, the cognitive disorder, and the dementia, respectively, with probability related to the cognitive state class of the user.

In addition, learning the brain disease detection and prediction model can perform simultaneous learning through partial teacher learning of the intra-model and the inter-model.

In addition, the intra-model may be configured using a linear-chain conditional random field in which a set of previous feature vectors of the user and a set of current feature vectors are input.

Also, the inter-model may be composed of an input generating part and an artificial neural network.

In addition, the probability score vector generated through the intra-model and the inter-model can be externally determined to determine the perceived state.

In addition, learning of the brain disease detection and prediction model can be performed by learning a condition in which there is no user who does not know the recognition state, or a condition in which the number of iterations for learning becomes equal to a preset threshold value Can be repeated until it is first achieved.

The brain disease detection and prediction system according to the present invention detects and predicts brain diseases (cognitive dysfunction and dementia) of a user using cognitive-strength measurement contents and log data-based machine learning models measured from the contents, Or the problem of data management can be solved, and the cost can be lowered.

1 is a schematic block diagram of a brain disease detection and prediction system according to an embodiment of the present invention.
FIG. 2 is a flow chart showing an overall flow in a brain disease detection and prediction system according to an embodiment of the present invention.
FIG. 3 illustrates a flow of a machine learning model for inferring a user's perceptual state in a brain disease detection and prediction system according to an embodiment of the present invention.
FIG. 4 illustrates an exemplary structure of an intra-model used in a brain disease detection and prediction system according to an embodiment of the present invention.
5 illustrates an exemplary structure of an intermodel used in a brain disease detection and prediction system according to an embodiment of the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings, so that those skilled in the art can easily carry out the present invention.

1 is a schematic block diagram of a brain disease detection and prediction system according to an embodiment of the present invention.

Referring to FIG. 1, a brain disease detection and prediction system 10 according to an exemplary embodiment of the present invention may include a smartphone or a tablet, and may include an application to detect and predict a brain disease . In addition, the brain disease detection and prediction system 10 may be equipped with a separate server to process data received from a user's PC, a smartphone, or a tablet to perform brain disease detection and prediction.

More specifically, the brain disease detection and prediction system 10 may include a display 100, an input unit 200, a storage unit 300, and a control unit 400. When the brain disease detection and prediction system 10 is implemented as an application, the display 100, the input unit 200, the storage unit 300, and the control unit 400 may be provided in a single smartphone or a tablet. have. In addition, if the brain disease detection and prediction system 10 is configured based on an internet with a separate server, the display 100 and the input unit 200 may be a display and an input unit of a user's PC, a smartphone or a tablet The storage unit 300 and the control unit 400 may be implemented as separate servers independent of the user.

The display 100 is for providing a user with a test for detecting and predicting brain disease, and may be a display of a user's PC, smartphone or tablet. The display 100 can display items to be tested and results for items such as memory, attention, linguistic ability, calculating power, and perception strength to be described later. Accordingly, each user can conduct a test on-line to continuously and quickly confirm brain disease detection and prediction results.

The input unit 200 may be a variety of input devices such as a keypad of a user's PC, a smartphone or a tablet, a virtual keypad, or a mouse. May be used by the user to submit an answer for the test displayed on the display 100. Through the input unit 200, the user can submit an answer to the test item of the display 100 on-line and confirm the result accordingly.

The storage unit 300 may be provided in a user's PC, a smartphone, or a tablet, or may be configured as a separate server. The storage unit 300 may store test items on cognitive power measurement items such as memory, attention, linguistic power, computation power, and perceived power to be described later, and answers of each user. In addition, the storage unit 300 may store the answers that each user has submitted in the past for a predetermined period of time, and may continuously detect brain diseases and predict results for each user.

The controller 400 uses the correct ratio for each item of the answer submitted by the user stored in the storage unit 300 to calculate an intra model using the previous data t-1 of the user and the current data t RC) and the output of the inter-model (EC) using the response data of the user community with demographic data similar to the user. In addition, the controller 400 multiplies the inferred results from the models RC and EC to derive final brain disease detection and prediction results. The specific operation of the control unit 400 will be described later.

Hereinafter, the operation of the brain disease detection and prediction system according to an embodiment of the present invention will be described in more detail.

FIG. 2 is a flow chart showing an overall flow in a brain disease detection and prediction system according to an embodiment of the present invention. FIG. 3 illustrates a flow of a machine learning model for inferring a user's perceptual state in a brain disease detection and prediction system according to an embodiment of the present invention. FIG. 4 illustrates an exemplary structure of an intra-model used in a brain disease detection and prediction system according to an embodiment of the present invention. 5 illustrates an exemplary structure of an intermodel used in a brain disease detection and prediction system according to an embodiment of the present invention.

2 and 3, a brain disease detection and prediction system according to an embodiment of the present invention displays and presents to a user about cognitive performance measurement items such as memory, attention, linguistic ability, computation power, and perception ability, Extracts the data characteristic of each content from the log-play data submitted by the answer. In addition, a data feature vector is generated based on data characteristics, and an output is generated for each model (intra model RC, inter model EC) through a simultaneous learning algorithm using each feature vector as an input value . In addition, it can obtain a final product by such an inference result (C _EC) of the intra-model, the speculative result of _(RC) (C RC) and the inter-model (EC), brain disease detection and prediction results.

Hereinafter, each step of FIG. 2 will be described in detail.

Ⅰ. To measure cognitive ability By contents  Configuration

More specifically, contents of the contents (memory, attention, linguistic ability, calculating power, perception) that measure the cognitive ability of each user through the display 100 and the input unit 200 are as follows.

1. Memory

Counts for 3 seconds before the first problem is exposed, and then the problem will proceed immediately without counting.

1) In one problem, a card with numbers starting from 1, such as 1, 2, 3, 4, is exposed without any overlap on any position on the screen.

2) If the user decides that he / she has memorized the order of the card (the order from 1 to the card with the greatest number of cards), he can press the "Start" button and the number written on the card disappears Only the position of.

3-1) After the number of the card disappears, if the user touches the card in the order of the numbers written on the card, the problem is judged as "Correct" and the "Ding Dong" sound is played and "" appears on the screen.

3-2) On the contrary, when touching the card in the wrong order, vibration alarm and "X" display are displayed on the screen and it is judged as "wrong answer". In this case, incorrect judgment is immediately judged when the wrong number of consecutive numbers is selected. In other words, if you touch the position where "is displayed on the third touch, you have to correct the number from 1 to 4, and it will be wrong.

4) If the problem is judged as "Correct" or "Incorrect", the following problem will be resolved.

1 minute, 30 seconds, 1) to 4) process (one problem) is repeated, 1 minute and 30 seconds is called one cycle. The number displayed in 1) starts from 3 (1 to 3), and when two or more consecutive alignments are made, the number of displayed numbers increases by one, and seven is the maximum display number. The raw log data measured in a problem of the content are as follows: number of digits displayed, position of displayed digits, time from when the digits are displayed, until the user presses the "start" button, The order in which the card was pressed, the time it took the user to find the number on the card (touch time), and the number of problems the user performed during one cycle.

2. Attention

The contents are given a task of judging whether the contents of the color box and the text match.

1) A color ruler and text (words related to colors: red, blue, etc.) are displayed

2) Press the "O" button when the color of the color matches the text and the "X" button if it does not match.

3-1) If you select the correct button (whether the color of the color box matches the color of the color box) and the "O" button and the "X" button, .

3-2) If the wrong button (whether the color of the color box matches the text of the color box) among the "O" button and the "X" button is selected, it is judged as "wrong answer" and vibration alarm and "X" display are displayed on the screen.

4) If the problem is judged as "Correct" or "Incorrect", the following problem will be resolved.

1 minute, 30 seconds, 1) to 4) process (one problem) is repeated, 1 minute and 30 seconds is called one cycle. 1), the color and text of the color are randomly displayed in red, green, and blue. If you select 5 correct answers in succession, yellow and black colors are added. If the answer is 5 consecutive correct answers, the color of the text is added. The raw log data measured in one question of the content is as follows: After the color of the displayed color, the information of the displayed text, the color of the displayed text, the color, and the text are displayed, The amount of time it takes to touch, the button you have selected, the correct answer, and the number of problems you have performed during a cycle.

3. Linguistic power

The content is processed in such a way as to select the corresponding word for a given picture.

1-1) In any case, a random picture appears on the top of the display.

1-2) At the same time as 1-1), four or six syllables appear at any position on the bottom of the display.

2-1) If the syllable of the word corresponding to the picture in which the syllable set touched by the user among the bottom exposed syllable is the same, the order and the kind are the same, it is judged as "correct answer". "Ding-dong" sound is played back and "O" appears on the screen.

2-2) If the syllable set touched by the user among the six lower syllables in the picture shown is different from the syllable, order, and type of the corresponding word, it is judged as "wrong answer". Vibration alarm and "X" are displayed.

2-3) When the user selects the correct answer syllable set size (the number of syllables) in both 2-1) and 2-2), it is judged as "correct answer" or "wrong answer".

3) If the problem is judged as "correct" or "wrong answer", the following problem will be solved.

1 minute, 30 seconds, 1) to 3) process (one problem) is repeated, 1 minute and 30 seconds is called one cycle. Counts for 3 seconds before the first problem is exposed, and then the problem will proceed immediately without counting. The syllables appearing in 1-2) appear randomly in the syllable defined in the database. The resource in the picture shown in 1-1) is a picture of the self-produced object. 1-2), the initial number of syllables is 4, and if the correct answer is 5, the number of syllables is increased to 6. The raw log data measured in one question of the content is as follows: The type of displayed picture, the displayed syllable, the position of the displayed syllable, the display of the syllable by the size of the set of correct answers The time taken to select, the syllable set selected by the user, the correct answer, and the number of problems the user performed during a cycle.

4. Computational Power

Calculation power is the content for measuring how much the user can adjust a given computation problem.

1) Computation problem is given on the screen, and the space where the correct answer is entered is displayed in blank.

2-1) If you press the "OK" button after inputting the correct answer for the given calculation problem, the problem is judged as "correct" and "Ding Dong" sound is played and "O" is displayed.

2-2) If you click the "OK" button after inputting the correct answer and the correct answer for the given calculation problem, it is judged as "wrong answer" and vibration alarm and "X" are displayed.

3) If the problem is judged as "correct" or "incorrect", the next problem will be repeated from 1).

1 minute, 30 seconds, 1) to 3) process (one problem) is repeated, 1 minute and 30 seconds is called one cycle. A given computation problem consists only of addition and subtraction, and the correct answer consists of only 0 and positive integers. Counts for 3 seconds before the first problem is exposed, and then the problem will proceed immediately without counting. 1), the problem arises from two numbers and one operator. In case of five consecutive correct answers, the number of numbers and the number of operators increase by one, the maximum number is four, and the number of speakers is three. The number of the first problem is in the range of 1 ~ 9. The raw log data measured in a problem of the content is as follows: the number of displayed digits, the displayed digits, the displayed operator, the correct answer, the time from the display of the number until the user selects the "OK" Whether the answer is correct, the answer you entered, and the number of problems the user performed during a cycle.

5. Perception

The content is processed by locating the remaining objects among the plurality of objects appearing at an arbitrary position.

1) Multiple objects appear at arbitrary positions on the screen, and only one of the displayed objects has a different shape or color from the rest of the objects.

2) After a certain period of time, the screen is blinded with eight equal-sized cards.

3-1) When selecting the partition card at the position where the remaining objects and other objects shown in 1) are displayed on the blind processed screen, the "correct answer" is determined, and a "ding dong" sound is reproduced and "O" is displayed .

3-2) When one of the objects shown in 1) is selected and one of the remaining cards other than the partition card at the position where the other object is displayed is selected, it is judged as "wrong answer" and vibration alarm and "X" Is displayed.

4) If it is judged as "correct" or "wrong answer" for the problem, show the correct screen.

5) The next problem is going to be repeated from 1).

1 minute, 30 seconds, 1) to 5) process (one problem) is repeated, 1 minute and 30 seconds is called one cycle. Counts for 3 seconds before the first problem is exposed, and then the problem will proceed immediately without counting. The initial 1) object starts with a 6-second exposure and the exposure time is reduced by 1 second for 3 consecutive correct answers. In the case of 3 consecutive correct answers in 1 second exposure, it is reduced to 0.5 second, and the decrease in exposure time no longer occurs. The other one of the exposed objects in the initial 1) is different in color from the rest of the object, and the color and shape may be changed in five consecutive correct answers. The raw log data measured in a problem of the content are as follows: the position where the remaining and other objects are exposed, the number of partition cards, the remaining and other objects have different attributes, the exposure time of the object, The time it takes for the user to select one of the split cards, the correct answer, the split card chosen by the user, and the number of problems the user has performed during a cycle.

Ⅱ. Feature extraction method by contents

On the other hand, the control unit 400 extracts the following features from the above-mentioned non-processed log data stored in the storage unit 300.

1. Memory

)을 수행했을 때 9차원의 벡터(

)를 추출한다.Data related to the memory feature allows the user to select one cycle

), The 9-dimensional vector (

).

)를 다음과 같이 계산한다.From the given memory task, measure the probability of an answer by the number of appearances. That is, a 7-dimensional vector (

) Is calculated as follows.

[Equation 1]

는 나타난 숫자의 개수가

개 일 때 사용자가 올바르게 숫자의 순서를 찾은 횟수를,

는 나타난 숫자의 개수가

인 횟수를 의미한다. here

Is the number of

The number of times the user correctly found the sequence of numbers,

Is the number of

.

norm)를

로 사용한다.Also, to reflect that the degree of memory that can be stored depends on the distance of successive numbers, the average distance of successive numbers in a cycle

norm)

.

로 사용한다.Finally,

.

2. Concentration

)을 수행했을 때 8차원의 벡터(

)를 추출한다.Concentration-related data feature allows users to select one cycle

), The 8-dimensional vector (

).

로 사용된다.The probability of correct answer depends on whether the color of the text and the color of the palette are the same.

.

&Quot; (2) "

&Quot; (3) "

와

는 각각 해당 속성의 문제가 나온 횟수와 사용자가 해당 속성의 문제의 정답을 맞힌 횟수를 의미한다.

과

는 각각 나타난 텍스트와 파레트의 색이 동일한 경우와 다른 경우의 문제 속성을 의미한다. here

Wow

Means the number of times the problem of the attribute has occurred and the number of times the user has met the correct answer to the problem of the property.

and

Refers to the problem attribute when the color of the displayed text and the palette are the same or different, respectively.

은 각각 빨간색, 파란색, 노란색, 초록색, 검은색이 파레트에 나타났을 때 사용자가 답을 맞춘 확률이다.

Is the probability that the user answered when red, blue, yellow, green, or black appeared on the pallet.

&Quot; (4) "

와

는 각각 색상별 (

) 문제가 나온 횟수와 사용자가 해당 속성의 문제의 정답을 맞춘 횟수를 의미한다.here

Wow

Respectively.

) The number of times the problem occurred and the number of times the user matched the correct answer to the problem in the property.

로 사용한다.Finally,

.

3. Linguistic power

)을 수행했을 때 4차원의 벡터(

)를 추출한다.The linguistic-related data feature allows the user to select one cycle

), A 4-dimensional vector (

).

로 사용된다.When the number of exposed words is 4 or 6, respectively,

.

&Quot; (5) "

&Quot; (6) "

와

는 각각 노출된 단어의 개수가 각각 4개 일 때를 의미하며

와

는 각각 해당 속성의 문제가 나온 횟수와 사용자가 해당 속성의 문제의 정답을 맞힌 횟수를 의미한다.here

Wow

Means that the number of exposed words is 4, respectively

Wow

Means the number of times the problem of the attribute has occurred and the number of times the user has met the correct answer to the problem of the property.

은 전체 게임 횟수 중 사용자가 선택한 음절 집합이 정답 음절 집합과 순서는 다르나 구성이 같을 확률을 사용한다.

Uses the probability that the syllable set selected by the user is equal to the set of correct syllables but the order is the same.

&Quot; (7) "

와

는 각각 문제 총 횟수와 사용자가 선택한 음절 집합이 정답 음절 집합과 순서는 다르나 구성이 같은 횟수를 의미한다.

Wow

Represents the number of times the total number of problems and the number of syllables selected by the user are the same, although the order is different from the set of the correct syllables.

로 사용한다.Finally,

.

4. Computational Power

)을 수행했을 때 6차원의 벡터(

)를 추출한다.Data features related to computational power allow the user to select one cycle

), The vector of six dimensions (

).

로 사용된다.The probability of correct answer depends on the number of

.

&Quot; (8) "

과

은 각각 나타난 숫자 개수가

개일 때의 총 문제 수와 정답을 맞힌 회수를 의미한다.here

and

Is the number of

It means the total number of problems and correct answers.

은 나타난 숫자의 개수가 3개일 때, 두 연산자가 서로 다를 경우의 정답확률을 사용한다.

Uses the correct answer probability when the two operators are different when the number of the displayed numbers is three.

&Quot; (9) "

과

는 각각 숫자의 개수가 3개일 때 두 연산자가 서로 다를 때의 총 문제 수와 정답을 맞힌 회수를 의미한다.here

and

Means the total number of problems when the two operators differ from each other when the number of the numbers is three and the number of correct answers.

으로 사용된다.The probability of correct answers based on the number of two digits in a given query

.

&Quot; (10) "

과

은 질의 내에 두 자릿수 숫자가

개인 문제의 총 개수와 사용자가 정답을 맞힌here

and

Is a two-digit number within the query.

The total number of personal issues and the

로 사용된다.Finally, the average correct answer probability

.

5. Perception

)을 수행했을 때 9차원의 벡터(

)를 추출한다.Data pertinent to perceptive power allows the user to select one cycle

), The 9-dimensional vector (

).

은 각각 8분할된 영역에서 찾아야 하는 객체가 노출된 위치에 따른 정답확률을 사용한다.

Uses the probability of correct answer according to the position where the object to be searched in each of the eight divided areas is exposed.

&Quot; (11) "

과

은 8분할 중 n번째 위치에 찾아야 하는 객체가 노출된 총 문제 수와 사용자가 맞힌 회수를 의미한다. 본 발명의 1)에서 사용된 지각력 콘텐츠에서는 왼쪽 첫 번째 행부터 1영역, 오른쪽 첫 번째 행이 2영역, 왼쪽 두 번째 행이 3영역이며 오른쪽 네 번째 행이 8영역을 의미한다.here

and

Is the number of the total number of problems exposed by the object to be searched at the n-th position and the number of times the user has met. In the perceptual content used in 1) of the present invention, 1 region from the first left row, 2 regions on the right first row, 3 regions on the left second row, and 8 regions on the right fourth row.

Finally, the average correct answer probability .

Ⅲ. Machine Learning Model for Detection and Prediction of Brain Disease

The ratio between positive and negative sets of learning data for learning brain dementia (cognitive impairment, dementia) detection and prediction models is disproportionate. In other words, while there are a large number of positive sets (user data with normal cognitive ability) in collecting learning data, there is a limit to collecting data belonging to negative set (cognitive disorder, dementia risk group). Therefore, in the present invention, a model based on a co-training framework is used to overcome this problem. 3, the brain disease detection and prediction model includes an intra-model (RC) using the previous data (t-1) and current data (t) of the user and a user community And an inter-model (EC) using the play data (t) of the play data (t). The output of the intra-model (RC) and inter-model (EC) is the probability for the cognitive state class of user (x)

는 각각 사용자(x)가 인지상태 정상, 인지장애, 치매에 속할 확률을 의미한다. 최종적으로 뇌 질환을 구성하는 인지장애·치매 추론 결과는 각 모델로부터 추론된 결과 (

)를 곱함으로써 얻을 수 있다.here

Refers to the probability that the user (x) belongs to the cognitive state normal, the cognitive disorder, and the dementia. Finally, the results of cognitive dysfunction and dementia inference constituting the brain disease were inferred from each model

). ≪ / RTI >

On the other hand, co-training is semi-supervised learning using two points of view of data. That is, each sample must have two different sets of data features for the instance. Ideally, the two feature sets of an instance must be conditional independent operations on a given class. The simultaneous learning framework for learning the model considered in the present invention is as follows.

1. Input

), 인지 상태를 아는 사용자 집합 (

), 인지 상태를 모르는 사용자 집합 (

), 알고리즘의 반복을 제어하는 임계값(θ)Feature vector set (

), A set of users who know the recognition status (

), A set of users who do not know the state of perception

), A threshold value (?) For controlling repetition of the algorithm,

2. Output

The intra-model (RC) and inter-model (EC)

3. Learning behavior

According to this, each of the intra-model (RC) and the inter-model (EC) is configured such that a condition in which there is no user who does not know the perceived state or the number of iterations (i) The learning operation can be repeated until at least one of the conditions becomes equal to the threshold value [theta].

IV. Intra  Model( RC )and Inter  Description of model (EC)

1. Intramodel (RC)

)과 현재 특징 벡터 집합 (

)를 입력으로 하는 선형-연쇄 조건부 무작위장(linear-chain conditional random fields, 이하 선형-연쇄 CRFs)을 사용한다. 여기서, 선형-연쇄 CRFs는 시퀀셜 데이터를 다루기 위한 식별 비방향성 확률적 그래픽 모델(discriminative undirected probabilistic graphical model)이다. 은닉 마르코프 모델(hidden Markov)과 융합된 CRFs는 특징간의 독립 추정의 완화(relaxation)을 가능하게 한다. 또한, 최대 엔트로피 마르코프 모델에서 발생할 수 있는 레이블 바이어스(label bias) 문제를 피할 수 있다.The intra-model (RC) includes the user's previous feature vector set

) And the current feature vector set (

Linear-chain conditional random fields (hereafter referred to as linear-chain CRFs) whose input is the input signal. Here, linear-chained CRFs are discriminative undirected probabilistic graphical models for dealing with sequential data. CRFs fused with the hidden Markov model enable the relaxation of the independent estimation of features. Also, the label bias problem that can occur in the maximum entropy Markov model can be avoided.

은 주어진 관측 시퀀스 노드 (

)로부터 추론돼야할 은닉 상태 값 (hidden state variable)을 의미한다. Y_i∈Y는 인지 상태 레이블 c를 가지며 Y_{i -1}과 Y_i사이에 엣지를 갖는 체인 형태다. x에 대해서 확률 변수 Y_i는 마르코프 성질(Markov property)을 다음과 같이 따른다:Fig. 4 shows the graphical structure G = (X, Y) of RC with two kinds of nodes. Node

A given observation sequence node (

(Hidden state variable) that should be deduced from the current state. Y _i ∈Y is a chain type with a cognitive state label c and an edge between Y _{i -1} and Y _i . For x the random variable Y _i follows the Markov property as follows:

&Quot; (12) "

Here i to j mean that i and j are adjacent in the graph G.

The probability of a given label sequence y given from the observation sequence x is defined as a normalized product of the potential function as follows.

&Quot; (13) "

는 전체 관측 시퀀스와 위치 i와 i-1에서의 레이블의 쌍 함수(pairwise function);

는

에서의 레이블과 관측 시퀀스의 데이터 함수 (data function); λ_k와 μ_k는 학습데이터로부터 추정되는 상수 파라미터다.here

Is a pairwise function of the entire observation sequence and the label at positions i and i-1;

The

The data function of the label and observational sequence at; λ _k and μ _k are constant parameters estimated from the training data.

로 치환하여 위의 식을 다음과 같이 변환할 수 있다.

, The above equation can be transformed as follows.

&Quot; (14) "

에 대한 확실한 값의 확률을 결정한다. Where Z (x) denotes a normalized factor. The above equation can be regarded as a measure of the input sequence that partially determines the likelihood of each possible value for Y _i . The model assigns weights to each feature and fuses

Lt; RTI ID = 0.0 > a < / RTI >

시퀀스 (

)로부터, 파라미터

의 학습은 최대 우도 학습법 (maximum likelihood learning)

를 통해 수행되며, 기울기 하강 (gradient descent)를 통해 풀린다.Of the given learning data

sequence (

),

Learning is the maximum likelihood learning method.

, And is solved through a gradient descent.

&Quot; (15) "

2. Intermodel (EC)

에 속하는 다른 사용자의 플레이 특징 데이터를 사용한다. 도 5는 인터 모델(EC)의 모델을 보여준다. 인터 모델(EC)는 입력 생성부분과 인공 신경망 (Artificial nueral network, 이하 ANN)으로 나뉘며,

는 사용자

의 각 인지능력 측정 콘텐츠에서 추출된 특징 벡터;

는 사용자

의 인구학적 데이터 (나이(

), 교육 수준

년, 성별 (

));

는 사용자

의 인지 상태 레이블;

는 추론돼야할 사용자;

은 특징 사이의 비유사도 함수 (본 발명에서는

norm을 사용한다);

는 인구학적 데이터 사이의 비유사도 함수 (본 발명에서는

norm을 사용한다)를 의미한다. On the other hand, the inter-model (EC) has user's play feature data and demographic data similar to the user in order to infer the perceived state of the user

The play characteristic data of the other user belonging to the play function is used. Figure 5 shows a model of the intermodel (EC). The inter-model (EC) is divided into an input generation part and an artificial neural network (ANN)

User

A feature vector extracted from each cognitive capability measurement content of the feature vector;

User

Demographic data (age (

), Education level

Year, gender (

));

User

The cognitive status label of;

The user to be inferred;

Is a non-linear function between features

norm);

Is a non-linear function between demographic data

norm is used).

)을 G1으로부터 임의로 선택한다.

에 속하는 다섯 사용자의 데이터는 추론되어야 할 사용자와 짝이 되어 사용된다. 이 때,

에 속하는 사용자는

를 만족하는 사용자 u만 속하며

로 설정될 때 최적의 성능을 보여준다. ANN의 입력은

와

사이의 각 특징

간의

과

를 사용한다.In order to generate the input, five persons

) Is arbitrarily selected from G1.

The data of the five users belonging to the group are used in combination with the user to be inferred. At this time,

Users belonging to

Belongs only to user u satisfying

, It shows the best performance. The input of ANN is

Wow

Each feature between

Liver

and

Lt; / RTI >

&Quot; (16) "

That is, a vector of 30 dimensions is used as input.

중 최소한 3명의 사용자는 나머지

과 겹치지 않는 사용자가 선택되어야 더욱 다양한 입력을 생성할 수 있다. In the present invention, the input is generated by repeating the above process five times in order to learn the ANN considering the number of input cases among users belonging to similar demographic data. In other words, a total of five inputs are generated for one x. At this time,

At least three of the users

A non-overlapping user must be selected to generate more variety of inputs.

가 사용된다.The ANN model of the present invention uses a back propagation neural network with 100 hidden units. For the hidden layer, the sigmoid function

Is used.

을 5번 생성하고 분류하여 5번 중 가장 높은 평균 질량 함수 (인지상태 레이블에 대한)값을 갖는 인지상태를 최종 x의 인지상태로 판단한다.In the inference phase, as in the learning phase,

Is generated and classified 5 times, and the recognition state having the highest average mass function (for the recognition state label) among the 5 is judged as the recognition state of the final x.

Ⅴ. Detection and prediction of the final brain disease (cognitive disorder, dementia)

와

)의 외적을 통해 최종 인지상태를 판단한다.In order to perform the detection and prediction of the final cognitive impairment and dementia of the user, an intra-model (RC) and an inter-model (EC) are applied to the respective inputs, and two probability score vectors

Wow

) To determine the final perceived state.

&Quot; (17) "

As described above, the brain disease detection and prediction system 10 according to the present invention can detect and predict brain diseases of each user by deriving the final cognitive state c (x) through the controller 400.

In addition, by detecting and predicting the cognitive impairment and dementia of a user using the cognitive-strength measurement contents and the log-based machine learning model measured from the contents, problems of subjectivity or data management of these inspection methods can be solved , It can lower the cost compared to existing methods.

It is to be understood that the present invention is not limited to the above-described embodiment, but may be modified and changed without departing from the scope of the present invention. It will be understood by those of ordinary skill in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined in the appended claims.

10; Brain disease detection and prediction system 100; display
200; An input unit 300; The storage unit
400; The control unit

Claims (16)

A method for detecting and predicting brain diseases including cognitive disorders and dementia based on a user's cognitive ability,
Obtaining a user's response to questions related to memory, attention, computation, linguistic and perceptual ability through an input unit;
Extracting data related to memory, attention, computation power, linguistic ability, and perception ability from the user's response;
Extracting feature data related to a cognitive power of the user from the extracted data;
Learning the brain disease detection and prediction model using the feature data; And
And the control unit deducing the perceived state of the user from the learned brain disease detection and prediction model. The method according to claim 1,
Wherein the brain disease detection and prediction model includes an intra model that uses past and current play feature data and perceptual state of the user as inputs and an inter-model that uses play data of a user community with demographic data similar to the user Brain diseases and prediction methods. 3. The method of claim 2,
Wherein the intra-model and the inter-model are configured to include probabilities of a user's cognitive state class, each of the cognitive state normal, cognitive disorder, and dementia probability. 3. The method of claim 2,
Wherein learning of the brain disease detection and prediction model includes simultaneous learning through partial teacher learning of the intra-model and inter-model. 3. The method of claim 2,
Wherein the intra-model is constructed using a linear-chain conditional random field in which a user's previous feature vector set and a current feature vector set are input. 3. The method of claim 2,
Wherein the inter-model comprises an input generating part and an artificial neural network. 3. The method of claim 2,
And a probability score vector generated through the intra-model and the inter-model is externally determined to determine a cognitive state. The method according to claim 1,
The learning of the brain disease detection and prediction model may include a step of learning whether or not a condition in which there is no user who does not know the perceived state or a condition in which the number of iterations for learning becomes equal to a predetermined threshold value, A method for detecting and predicting repeated brain diseases until first achieved. 1. A system for detecting and predicting brain diseases, including cognitive disorders and dementia, based on cognitive abilities of a user,
A display that presents the user with questions related to memory, attention, compute power, linguistic ability and perception;
An input unit for inputting a user's response to questions related to memory, attention, calculation, linguistic ability, and perception;
And a controller for extracting data related to memory, attention, calculation, language, and perception from the user's response,
Wherein the controller extracts feature data related to a cognitive power of the user from the extracted data, learns brain disease detection and prediction models using the feature data, and inferences the user's perception state from the learned brain disease detection and prediction model A brain disease detection and prediction system. 10. The method of claim 9,
Wherein the brain disease detection and prediction model includes an intra model that uses past and current play feature data and perceptual state of the user as inputs and an inter-model that uses play data of a user community with demographic data similar to the user Brain diseases and prediction systems. 11. The method of claim 10,
Wherein the intra-model and the inter-model are configured to include probabilities of a user's cognitive state class, each of the cognitive state normal, cognitive disorder, and dementia. 11. The method of claim 10,
The brain disease detection and prediction system learns the brain disease detection and prediction model by performing simultaneous learning through partial teacher learning of the intra-model and inter-model. 11. The method of claim 10,
Wherein the intra-model is constructed using a linear-chain conditional random field that receives a user's previous feature vector set and a current feature vector set as inputs. 11. The method of claim 10,
Wherein the inter-model comprises an input generating part and an artificial neural network. 11. The method of claim 10,
A brain disease detection and prediction system for determining a perception state by externalizing a probability score vector generated through the intra-model and the inter-model. 10. The method of claim 9,
The learning of the brain disease detection and prediction model is performed by learning a condition in which there is no user who does not know the recognition state or a condition in which the number of iterations for learning becomes equal to a predetermined threshold value Repeated brain disease detection and prediction system until achieved.

Images