KR102395312B1 - Dementia prediction apparatus based on artiticial intelligence learning using gait motion - Google Patents

Abstract

An artificial intelligence learning-based dementia prediction device using walking motion is provided. According to an embodiment of the present invention, there is provided an apparatus for predicting dementia based on artificial intelligence using gait motion, comprising: a data preprocessing unit for preprocessing time series acceleration data and time series angular acceleration data respectively obtained from left and right feet according to a gait motion into data of a certain size; a feature extraction unit for extracting signal features using the preprocessed acceleration data and angular acceleration data; a learning unit for generating a plurality of learning models for predicting dementia by learning the extracted signal features; a cross-validation unit that evaluates and cross-validates the plurality of learning models; and a dementia prediction unit that selects an optimal learning model from among the plurality of learning models through the cross-validation and predicts dementia using the optimal learning model.

Description

Dementia prediction device based on artificial intelligence learning using gait motion {DEMENTIA PREDICTION APPARATUS BASED ON ARTITICIAL INTELLIGENCE LEARNING USING GAIT MOTION}

The present invention relates to an apparatus for predicting dementia based on artificial intelligence using a walking motion.

With advances in medicine, the proportion of the elderly population is increasing worldwide. In particular, Korea is aging very rapidly compared to other developed countries, and it is expected that the number of dementia patients will increase rapidly with the rapid aging of the population.

Although various drugs are used to treat dementia patients and the like, they do not fundamentally cure dementia, but merely slow the progression of dementia. However, it is reported to have a relatively high therapeutic effect when treatment is carried out with a prescription in the early stage of dementia.

Therefore, early prediction and diagnosis of dementia is very important in alleviating the symptoms of dementia. Accordingly, it is possible to alleviate symptoms through prediction of dementia, and to reduce social loss and economic cost due to dementia.

Accordingly, several studies for predicting dementia are being conducted in various countries around the world. Recently, it has been reported in clinical research results supported by the United Nations and national health authorities that gait speed decreases with disease or poor health due to old age. Walking speed refers to the movement speed of the trunk, and this human dynamic characteristic is considered to be a characteristic that distinguishes between healthy and unhealthy or diseased people. In order to extract human dynamic characteristics more precise than the moving speed of the trunk, the speed concept, which represents the average value by dividing the moving distance by time, is difficult to express detailed characteristics. Therefore, most of the methods are used to measure and analyze the movements of the left and right legs and feet that touch the ground and repeat floating.

Conventional plantar pressure distribution analysis only measures the time the foot touches the ground and the time the foot touches the ground, so it is difficult to precisely analyze the interaction between the trunk and lower extremity moving in a three-dimensional space. This is because, in time, the foot is attached to the ground for about 60% of the time.

In addition, the global standard for 3D motion inspection is a method using a set of 6 or more infrared cameras. After acquiring the location tracking data of the marker according to time by attaching a marker that a camera can track to the upper and lower body of the examiner, it is possible to extract inertial information such as speed and acceleration by performing a differential operation. However, this method does not utilize inertial data such as speed and acceleration in most medical institutions and extracts only simple joint angle (Range of Motion, range of motion).

Therefore, technology for using inertial data such as acceleration data or angular acceleration data obtained according to a gait motion for accurate early diagnosis and prediction of dementia is insufficient.

Republic of Korea Patent Publication No. 2019-0123186 (published on October 31, 2019) Republic of Korea Patent Publication No. 2019-0123187 (published on October 31, 2019) Republic of Korea Patent No. 1929965 (Registered on Dec. 11, 2018)

The present invention is to solve the above problem, and by extracting features from time-series inertial data obtained according to gait motion and using a model generated through artificial intelligence learning, artificial intelligence learning using gait motion that can predict dementia at an early stage A device for predicting dementia based on the present invention is provided.

The problems to be solved by the present invention are not limited to the problems mentioned above, and other problems not mentioned will be clearly understood by those skilled in the art from the following description.

In order to achieve the above object, the apparatus for predicting dementia based on artificial intelligence using walking motion according to an embodiment of the present invention for achieving the above object is to convert time series acceleration data and time series angular acceleration data respectively obtained from left and right feet according to the walking motion to data of a certain size. data pre-processing unit for pre-processing; a feature extraction unit for extracting signal features using the preprocessed acceleration data and angular acceleration data; a learning unit for generating a plurality of learning models for predicting dementia by learning the extracted signal features; a cross-validation unit that evaluates and cross-validates the plurality of learning models; and a dementia prediction unit that selects an optimal learning model from among the plurality of learning models through the cross-validation and predicts dementia using the optimal learning model.

In addition, the data preprocessor may obtain and preprocess the time series acceleration data and time series angular acceleration data from a 3D acceleration sensor and a 3D angular acceleration sensor respectively attached to the left and right feet of the pedestrian.

Also, the data preprocessor may divide the time series acceleration data and the time series angular acceleration data to have the same size, and normalize the divided data by overlapping the divided data with a predetermined size.

Also, the feature extractor may extract signal features including an average, standard deviation, energy, entropy, a first cross-correlation coefficient, and a second cross-correlation coefficient of each of the preprocessed acceleration data and the angular acceleration data.

Also, the feature extractor may calculate the sub-signal features of the average, standard deviation, energy, entropy, first cross-correlation coefficient, and second cross-correlation coefficient, respectively.

Also, the learning unit may learn the signal features by using at least one of k-NN, SVM, Random Forest, Extra Tree, and XGBoost algorithms.

In addition, the cross-validation unit may perform cross-validation using a learning data set change cross-validation method, a K-fold cross-validation method, and a Leave-P-Out cross-validation method.

The dementia prediction unit may predict dementia by applying the time series acceleration data obtained from the examiner and signal features extracted from the time series angular acceleration data to the optimal learning model.

Other specific details of the invention are included in the detailed description and drawings.

According to the present invention, it is possible to predict and diagnose dementia early by using a model generated through artificial intelligence learning by extracting features from time-series inertial data obtained according to a gait motion.

1 is a diagram illustrating the concept of an entire system including an artificial intelligence learning-based dementia prediction device using a walking motion according to an embodiment of the present invention.
2 is a diagram illustrating the configuration of an apparatus for predicting dementia based on artificial intelligence using a walking motion according to an embodiment of the present invention.
3 is a diagram for explaining orthogonality of inertia data.
4 and 5 are diagrams for explaining a preprocessing process of time series inertia data.
6 is a diagram illustrating a pattern arranged by quantifying signal features derived from time series inertia data.
7 is a view for explaining a method of distinguishing a healthy person from a dementia patient.
8 is a diagram illustrating a sequence of a method for predicting dementia based on artificial intelligence learning using a walking motion according to an embodiment of the present invention.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. Advantages and features of the present invention and methods of achieving them will become apparent with reference to the embodiments described below in detail in conjunction with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, but will be implemented in a variety of different forms, and only these embodiments allow the disclosure of the present invention to be complete, and common knowledge in the technical field to which the present invention belongs It is provided to fully inform the possessor of the scope of the invention, and the present invention is only defined by the scope of the claims. Like reference numerals refer to like elements throughout.

It should be understood that although first, second, etc. are used to describe various elements, components, and/or sections, these elements, components, and/or sections are not limited by these terms. These terms are only used to distinguish one element, component, or sections from another. Accordingly, it goes without saying that the first element, the first element, or the first section mentioned below may be the second element, the second element, or the second section within the spirit of the present invention.

The terminology used herein is for the purpose of describing the embodiments and is not intended to limit the present invention. In this specification, the singular also includes the plural unless specifically stated otherwise in the phrase. As used herein, “comprises” and/or “made of” refers to a referenced component, step, operation and/or element of one or more other components, steps, operations and/or elements. The presence or addition is not excluded.

Unless otherwise defined, all terms (including technical and scientific terms) used herein may be used with the meaning commonly understood by those of ordinary skill in the art to which the present invention belongs. In addition, terms defined in a commonly used dictionary are not to be interpreted ideally or excessively unless clearly defined in particular.

In this case, the same reference numerals refer to the same components throughout the specification, and it will be understood that each configuration of the process flowchart drawings and combinations of the flowchart drawings may be performed by computer program instructions. These computer program instructions may be embodied in a processor of a general purpose computer, special purpose computer, or other programmable data processing equipment, such that the instructions performed by the processor of the computer or other programmable data processing equipment are not described in the flowchart configuration(s). It creates a means to perform functions.

It should also be noted that in some alternative embodiments it is also possible for the functions recited in the configurations to occur out of order. For example, it is possible that two components shown one after another may in fact be performed substantially simultaneously, or that the components may sometimes be performed in the reverse order according to the corresponding function.

Hereinafter, the present invention will be described in more detail with reference to the accompanying drawings.

1 is a diagram illustrating the concept of the entire system including the apparatus for predicting dementia based on artificial intelligence using a walking motion according to an embodiment of the present invention.

Referring to FIG. 1 , an artificial intelligence learning-based dementia prediction apparatus 100 (hereinafter referred to as a 'dementia prediction apparatus') using a walking motion according to an embodiment of the present invention is transmitted from an examination institution server 20 through a network. Data and health care information stored in the health care database 10 may be obtained. Examples of the network include the Internet.

The dementia prediction device 100 designs a model that can predict dementia from time-series inertial data obtained from each of the left and right feet according to gait motion, and uses this to determine dementia, mild cognitive impairment, normality, etc. of a new examiner to prevent dementia predictable. A detailed configuration and function of the dementia prediction device 100 will be described later.

First, the health care database 10 includes a medical institution DB 11 operated by a medical institution, a public institution DB 12 operated by a public institution, and the like. The health care database 10 may store health care data open to the public, health care data related to a patient visiting a hospital, and the like. Data of subjects diagnosed with dementia disease, data of subjects diagnosed with mild cognitive impairment (MCI), and data of healthy people may be obtained from the health care database 10 .

Next, the inspection institution server 20 includes a plurality of sub-inspection inspector servers 21 and 22 operated by various inspection institutions such as hospitals, public institutions, research institutes, and corporations.

Each of the sub-inspection agency servers 21 and 22 may acquire inertial data from the inertial sensor 5 attached to both feet of the examiner. The inertial data includes acceleration data, angular acceleration data, and the like. Specifically, the inertial sensor 5 may include an acceleration sensor and an angular acceleration sensor, and may represent a single sensor unit rather than a concept that distinguishes the acceleration sensor or the angular acceleration sensor. The acceleration sensor is a sensor that measures acceleration in a three-axis direction in three dimensions, and the angular acceleration sensor is a sensor that measures angular acceleration in a three-axis direction in three dimensions.

In addition, the sub-inspection agency servers 21 and 22 are communicatively connected with the inertial sensor 5 to acquire data. Specifically, each of the subordinate inspection agency servers 21 and 22 may perform short-range wireless communication from each inertial sensor 5 installed in the inspector. For example, short-range wireless communication may be performed using Bluetooth, Zigbee, Near Field Communication (NFC), Wibree, or the like. For example, it connects wirelessly within a radius of 10 to 100m by Bluetooth method, uses a 2.45Ghz frequency, and uses a frequency of 2.4Ghz band in a Wibri method, and connects wirelessly for a distance of up to 10m at a communication speed of 1Mbps. can

Therefore, by receiving reliable inertial data from the inertial sensor 5 using a short-distance communication module, each sub-inspection agency server 21, 22?? (100) and so on. Of course, it will be apparent to those skilled in the art that the inertial sensor 5 is not limited to the short-range wireless communication described above, and other wireless communication methods may be adopted.

In addition, although not shown in the drawing, inertial data obtained by each individual, not an inspection institution, by using the inertial sensor 5 is transferred to the health care database 10, dementia prediction device ( 100) and so on.

Here, the user terminal may have a web browser (Netscape, Internet Explorer, Chrome, etc.) capable of displaying the contents of a web page such as HTML and XML. In particular, it serves to output dementia prediction information transmitted from the dementia prediction apparatus 100 . Such user terminals include general mobile communication terminals such as cellular phone, PCS phone (Personal Communications Services phone), synchronous/asynchronous IMT-2000 (International Mobile Telecommunication-2000), 2G/3G/4G/5G , WiBro wireless network serviceable terminal, Palm PC (Palm Personal Computer), personal digital assistant (PDA: Personal Digital Assistant), smart phone (Smart phone), WAP phone (Wireless application protocol phone), etc. It may comprehensively mean all wired and wireless home appliances/communication devices having a user interface for this purpose, and may be a device equipped with an interface for wireless LAN connection, such as an IEEE 802.11 wireless LAN network card. In addition, the user terminal may be an information communication device such as a computer or a notebook computer or a device including the same in addition to the mobile communication terminal.

In particular, when the user terminal accesses the health care database 10, the dementia prediction device 100, etc., various sensitive information related to dementia prediction, for example, user personal information, user medical information, user biometric information, etc. Since there is a risk of information being exposed, authentication of the user terminal is essential. At least one authentication protocol among text file authentication module (mod_auth), DBM authentication module (mod_auth_dbm), Berkeley DB authentication module (mod_auth_db), Anonymous authentication module (mod_auth_anon), PostgreSQL authentication module, and XNS authentication service in the authentication step of the user terminal It will be apparent to those skilled in the art that the user terminal is authenticated using In particular, when the user terminal accesses for the first time, it is preferable to request a member registration and perform authentication of the user terminal after the membership registration is completed.

Referring back to FIG. 1 , the dementia prediction apparatus 100 preprocesses the time series acceleration data and time series angular acceleration data respectively obtained from the left and right feet according to the walking motion, and then extracts signal features, and the extracted signal It learns the characteristics and predicts dementia based on it.

The dementia prediction apparatus 100 integrates and controls the characteristic information DB 200 . The characteristic information DB 200 stores and manages signal features and learning models generated using time-series acceleration and angular acceleration data obtainable from a walking motion in the dementia prediction apparatus 100 .

At this time, in order to facilitate the search and management of various information and data stored in the characteristic information DB 200, the characteristic information DB 200 is a DB for a dementia prediction device and a DB for a health care database according to the source and characteristics of the information. , and may include a DB for inspection agency servers. The characteristic information DB 200 means a logical or physical storage server for storing information, for example, Oracle DBMS of Oracle, MS-SQL DBMS of Microsoft, Sybase ) may be in the form of SYBASE DBMS, etc., but it will be apparent to those skilled in the art that it is not limited thereto.

2 is a diagram illustrating the configuration of an apparatus for predicting dementia based on artificial intelligence using a walking motion according to an embodiment of the present invention. 3 is a diagram for explaining orthogonality of inertia data. 4 and 5 are diagrams for explaining a preprocessing process of time series inertia data.

Referring to FIG. 2 , the dementia prediction device 100 includes a device communication unit 105 , a data preprocessing unit 110 , a feature extraction unit 120 , a learning unit 130 , a cross validation unit 140 , and a dementia prediction unit ( 150 ), a device input unit 160 , a device storage unit 170 , a device output unit 180 , a device control unit 190 , and the like.

The communication unit 105 communicates with the health care database 10 , the inspection institution server 20 , and the like through a network. Data stored in the health care database 10, data obtained from the inertial sensor 5 in the inspection institution server 20, etc. are acquired through the communication unit 105, and dementia prediction information is transmitted through the communication unit 105 for health care It can be transmitted to the database 10, the inspection agency server 20, and the like.

The data preprocessor 110 preprocesses the time-series acceleration data and the time-series angular acceleration data respectively obtained from the left and right feet according to the walking motion into data of a predetermined size. Here, the preprocessing means setting the time series acceleration data and the time series angular acceleration data as values to be used for dementia prediction, or processing the time series acceleration data and time series angular acceleration data as data in a format required by the learning model to be used for dementia prediction.

Here, as shown in FIG. 1, the data preprocessing unit 110 obtains time series acceleration data and time series angular acceleration data from the three-dimensional acceleration sensor and the three-dimensional angular acceleration sensor respectively attached to the left and right feet of the pedestrian and can be pre-processed. .

The data for predicting dementia is time series data for each left and right sensors (acceleration sensor and angular acceleration sensor) according to the examiner's gait motion. As shown in FIG. 3 , the time series data includes time series acceleration data Ax, Ay, and Az and time series angular acceleration data Gx, Gy, and Gz. In addition, Ax, Ay, and Az are orthogonal to each other, and Gx, Gy, and Gz are orthogonal to each other.

In this case, Ax denotes acceleration data in the front-rear direction, Ay denotes acceleration data in the inner and outer directions, and Az denotes acceleration data in the vertical direction. In addition, Gx denotes angular acceleration data that rotates in the front-back direction as an axis, Gy denotes angular acceleration data that rotates in the inner and outer directions as an axis, and Gz denotes angular acceleration data that rotates in the vertical direction as an axis. In addition, each of the acceleration data and the angular acceleration data is divided into time series data obtained from the left foot and time series data obtained from the right foot.

In particular, the data preprocessor 110 may divide the time-series acceleration data and the time-series angular acceleration data to have the same size, and normalize the divided data by overlapping the divided data with a predetermined size. Time series data is continuous data, and if the quantitative range is too large, it may be judged as noise or cause overfitting. normalization is performed.

For example, as shown in FIG. 4 , in the gait motion of the examiner, the data size is defined as n, and the data is defined as a one-time trial, and the data is overlapped by placing a gap by the data size m. can be divided

And, as shown in FIG. 5 , the first normalized (Normalized) data size n signals for each sensor are called a first trial (1st trial), and the last normalized (Normalized) data size n signals for each sensor are N It can be called a round trial (Nth trial).

Specifically, when sampling is set to 100 Hz, a trial data set of about 80 times may be generated in one minute inspection. When one examiner wears the sensor for 1 minute and performs the examination, 12 pieces of data are obtained by adding the left and right sides.

The feature extraction unit 120 extracts signal features by using the preprocessed acceleration data and the angular acceleration data.

For example, the feature extraction unit 120 may extract signal features including the mean, standard deviation, energy, entropy, first cross-correlation coefficient, and second cross-correlation coefficient of each of the preprocessed acceleration data and angular acceleration data. . The mean, standard deviation, energy, entropy, first cross-correlation coefficient, and second cross-correlation coefficient have different values depending on the healthy person and the dementia patient.

Here, the average (average, Mij) can be calculated by Equation 1 below.

이 된다. 상기 k는 데이터의 크기 단위에 따라 bit 단위로 설정되어 임의의 자연수가 된다. 예를 들어, k의 상한인 n은 데이터 크기가 되므로, 32, 64, 128, 256, 512, 1024 등으로 적절하게 설정될 수 있다.In Equation 1, n is the size of data by data preprocessing, and j is 12 pieces of data that are left and right acceleration and angular acceleration data. In addition, i is a trial number, which may be determined according to the examination time of the examiner. And, k is determined according to the size of the data,

becomes this The k is set in bit units according to the data size unit and becomes an arbitrary natural number. For example, since n, which is the upper limit of k, becomes the data size, it may be appropriately set to 32, 64, 128, 256, 512, 1024, or the like.

In addition, standard deviation (Stdij) can be calculated by Equation 2 below.

In addition, energy (Eij) can be calculated by Equation 3 below.

In Equation 3, FFT means fast Fourier transform.

In addition, entropy (Epij) can be calculated by the following Equation (4).

Specifically, in the mean (Mij) according to Equation 1, 12 data have different values, but in the data set of a healthy person, Ax/Ay, Az/Ay, left and right Gx, Gy, Gz among Ax, Ay, and Az on the left and right Among them, both Gy/Gx and Gy/Gz are greater than 1. Upon examination of the gait motion for 1 minute, a trial data set of about 80 rounds is generated, and in a healthy person, the average Ax/Ay, Az/Ay, Gy/Gx, Gy/Gz of 80 times is greater than 1. Dementia patients have an average of 80 Ax/Ay, Az/Ay, Gy/Gx, and Gy/Gz less than 1.

In addition, the CV (Coefficient of Variation) can be calculated by dividing the standard deviation (Stdij) according to Equation 2 by the mean (Mij) according to Equation 1, and the CV value of a healthy person is Ax, Ay, The average of 80 CVs of Az, Gx, Gy, and Gz has 1 digit, whereas in dementia patients, the average of 80 CVs of Ax, Ay, Az, Gx, Gy, and Gz on the left and right has 2 digits.

In addition, the energy (Eij) according to Equation 3 has different values for the 12 data sets, but when the motion test is performed for 1 minute, the result of the data set is about 80 times. For a healthy person, the energy value of the average Ax of 80 times is Ay It is greater than the energy value and the energy value of Gy is 80 times greater than the energy value of Gx and the energy value of Gz. However, in the case of dementia patients, the energy value of Ax is smaller than the energy value of Ay, and the energy value of Gy is 80 times smaller than the energy value of Gx and the energy value of Gz.

In addition, although the entropy (Epij) according to Equation 4 has different values for 12 data, the sum of entropy of Ax, Ay, Az, Gx, Gy, and Gz of the left and right sides of a healthy person is larger than that of a dementia patient. In addition, the entropy values of left and right Ay are greater in healthy people than in dementia patients, and the entropy values in left and right Gx are greater in healthy people than in dementia patients. In this case, the high entropy corresponds to the total average of 80 times.

The first cross-correlation coefficient and the second cross-correlation coefficient refer to respective correlations of preprocessed acceleration data and angular acceleration data.

Specifically, the first cross-correlation coefficient is calculated by pairing the preprocessed acceleration data by two, and three signal features are extracted. In addition, the first cross-correlation coefficient is calculated by pairing the preprocessed angular acceleration data by two, and three signal features are extracted. Accordingly, as for the first cross-correlation coefficient, 12 signal features may be extracted by combining both left and right feet.

Specifically, the second cross-correlation coefficient calculates the correlation between the left foot and the right foot based on the X-axis, Y-axis, and Z-axis of the preprocessed acceleration data, respectively, and nine signal features are extracted. In addition, the second cross-correlation coefficient calculates the correlation between the left foot and the right foot based on the preprocessed angular acceleration data on the X-axis, Y-axis, and Z-axis, respectively, and nine signal features are extracted. Accordingly, as for the second cross-correlation coefficient, 18 signal features may be extracted.

As described above, the feature extraction unit 120 extracts dozens of signal features including the mean, standard deviation, energy, entropy, the first cross-correlation coefficient, and the second cross-correlation coefficient, thereby accurately determining the criterion for classifying dementia. You can learn and set it up.

Also, the feature extractor 120 may calculate sub-signal features of the average, standard deviation, energy, entropy, the first cross-correlation coefficient, and the second cross-correlation coefficient, respectively. In this case, the sub-signal features refer to signal features calculated by the feature extracting unit 120 based on respective values of the mean, standard deviation, energy, entropy, the first cross-correlation coefficient, and the second cross-correlation coefficient.

For example, the feature extraction unit 120 may include an average, standard deviation, energy, entropy, average, standard deviation, coefficient of variation (CV), and root mean square (RMS) of each of the first and second cross-correlation coefficients. ) can be computed as sub-signal features. Accordingly, the feature extraction unit 120 may more precisely learn and set a criterion for classifying dementia by extracting hundreds to thousands of signal features.

The learning unit 130 generates a plurality of learning models for predicting dementia by learning the extracted signal features. At this time, learning is to learn how to classify data that is not classified as dementia into dementia or the like (eg, dementia, mild cognitive impairment, the general public) through a learning process. In addition, the learning unit 130 may classify according to the learned criteria when the inertia data of a new test subject is input.

For example, the learning unit 130 may learn signal features by using at least one of k-NN, SVM, Random Forest, Extra Tree, and XGBoost algorithms.

The Random Forest algorithm is a learning algorithm that consists of a forest of numerous decision trees and averages each prediction result into one outcome variable. The k-NN (k-Nearest Neighbors) algorithm and SVM algorithm have the largest width in the data distribution space. It is a non-stochastic learning algorithm that determines the classification to which data belongs by classifying the boundaries of XGBoost Algorithm It is a learning algorithm that expands the tree structure of a random forest to near-infinite cases.

The cross-validation unit 140 evaluates and cross-validates a plurality of learning models. The efficiency of the learning algorithm performed by the learning unit 130 is verified by the cross validation unit 140 .

Here, the cross-validation unit 140 may perform cross-validation using the training data set change cross-validation method, the K-fold cross-validation method, and the Leave-P-Out cross-validation method.

For example, in the known K-fold cross-validation method, the cross-validation unit 140 may start with K=10 and analyze the change in precision and consistency by changing the K value. At this time, even the label distribution of stratified k-fold data is considered.

Specifically, the cross-validation unit 140 randomly divides N data into equal K groups, puts one group as a test set, and sets the remaining K-1 groups as a training set. to be repeated K times.

In addition, the cross-validation unit 140 may analyze a change in precision and consistency according to a change in the 80/20 training set in a known 70/30 cross-validation method.

In addition, the cross-validation unit 140 may analyze the variation in precision and consistency from the well-known leave one out (LOO) cross-validation method to the leave-P-out cross-validation method.

Specifically, the cross-validation unit 140 sets the P data samples from the N data samples as a test set and the remaining N-P after subtracting the P numbers as a training set to verify the learning model. there is.

Thus, the cross-validation unit 140 may evaluate the precision and consistency of the learning algorithm through the learning data set change cross-validation method, the K-fold cross-validation method, and the Leave-P-Out cross-validation method.

The dementia prediction unit 150 selects an optimal learning model from among a plurality of learning models through cross-validation, and predicts dementia using the optimal learning model. At this time, the optimal learning model is a learning algorithm with maximum precision and consistency through the cross-validation cross-validation method, the K-fold cross-validation method, and the Leave-P-Out cross-validation method by the cross-validation unit 140 . it means.

Here, the dementia prediction unit 150 may predict dementia by applying the time series acceleration data obtained from the examiner and signal features extracted from the time series angular acceleration data to the optimal learning model.

A detailed model for predicting dementia will be described later.

6 is a diagram illustrating a pattern arranged by quantifying signal features derived from time series inertia data. And, FIG. 7 is a view for explaining a method of distinguishing a healthy person from a dementia patient.

The dementia prediction model may output a relative distance result based on the data of a new test subject by inputting the data of the dementia patient and the data of the normal person. The infinity concept of the Hilbert space can be introduced through artificial intelligence-based learning in the existing prediction method that generates discontinuities in the prediction section due to the limited composition of the number of dementia patients.

For example, the dementia prediction model may output the dementia status result of a new test subject based on the distance (similarity) between the dementia characteristic variable and the data of normal persons. The dementia similarity level and the health level may be determined based on the threshold value set for the distance (similarity). The closer to the dementia similarity, the lower the health level, and the farther from the dementia similarity, the higher the health level.

Here, the similarity can be iteratively subdivided into K clusters based on the average value (similarity) of distances using a traditional classification technique such as the K-means clustering algorithm. It will be apparent to those skilled in the art that other useful algorithms other than the K-means clustering algorithm can be used.

In particular, it is possible to improve the accuracy of the dementia prediction model by repeatedly performing the derivation of the output by the data input of the new test subject.

The dementia prediction unit 150 may determine whether there is an abnormality according to the signal characteristics of the new test subject based on the dementia prediction model. When data of a new subject is input, as described above, a quantitative level of the dementia state level is calculated based on the similarity, and a dementia prediction result may be output.

Referring to FIG. 6 , an arrangement pattern may be quantified according to the frequency of extraction of signal features, and a healthy person and a dementia patient may be determined according to the frequency of signal features disposed at an extreme distance. In this case, the weakened part of the detailed body part may be learned by extracting signal features according to the inertia data.

Referring to FIG. 7 , a dementia patient may be identified in red, and a healthy person may be identified in green. On the other hand, signal features having a frequency similar to dementia or health may be classified by a geometrical distance d.

If the location of the new test subject appears the same on both sides of the center of dementia and the center of health, it can be explained that the similarity between dementia and health status is 50:50. At this time, when various diseases are combined rather than a simple pattern that determines the similarity between dementia and health only with the distance concept, as shown in FIG. part can be learned.

8 is a diagram illustrating a sequence of a method for predicting dementia based on artificial intelligence using a walking motion according to an embodiment of the present invention.

Referring to FIG. 8 , in the method for predicting dementia based on artificial intelligence using walking motion according to an embodiment of the present invention, time series acceleration data and time series angular acceleration data are respectively obtained from left and right feet according to the walking motion (S110), the Pre-processing time series acceleration data and time series angular acceleration data into data of a certain size (S120), extracting signal features using the preprocessed acceleration data and angular acceleration data (S130), and plural learning to predict dementia Learning the extracted signal features to generate a model (S140), evaluating and cross-validating the plurality of learning models (S150), selecting an optimal learning model among the plurality of learning models through the cross-validation, Dementia is predicted using the optimal learning model (S160).

Hereinafter, a detailed embodiment of a method for predicting dementia based on artificial intelligence using a walking motion according to an embodiment of the present invention will be described.

First, the data for detecting dementia characteristics is time series data for each left and right inertial sensors of the test subject. The time series data sampled N times per second is averaged data for each sensor, standard deviation data for each sensor, energy data for each sensor, entropy data for each sensor, first cross-correlation coefficient for each sensor, second cross-correlation every 512 through the normalization process. is the coefficient If the first 512 normalized signals for each sensor are called the 1st trial, and the sampling is 100Hz, about 80 trials of data are generated for 1 minute of operation. When one test subject performed a motion test for 1 minute, there were 12 left and right 6 sensors, including mean, standard deviation, energy, entropy, and first and second cross-correlation coefficients for a total of 78 data columns. this is created

The database obtained based on 200 people diagnosed with dementia and 400 people with mild cognitive impairment (MCI) was referred to as a reference. Healthy people were based on a database of 3,000 men and women under the age of 40 who had no hospital visits for 6 months and had no congenital diseases.

Here, in a healthy person, there is no obstruction for the central nervous system and peripheral nervous system, as well as muscles and joints to transmit motor commands. However, in the case of central nervous system degenerative diseases such as dementia, the function of the central nervous system to generate motor commands is deteriorated, making it difficult to generate normal motor commands. These results can be measured as variability. At this time, the motor command is a motor command signal transmitted from the central nervous system (CNS) to the peripheral nervous system.

Specifically, the variability of healthy people is smaller than that of dementia patients, and the variability is smaller in the young group than in the older group.

Signal features can be extracted through the trial data of 80 times. Here, it is possible to differentiate the standard of a healthy person. A general health standard is taken based on the total average of 80 times, and the healthiest person from 1 to 80 individually is all greater than 1 and the CV average is all 1 digits It can be said that health is the most ideal case when all are large. If all 80 trials are satisfied in a total of 80 trials, it can be said to be 100 points, and the health score can be quantified using a proportional formula based on this. Similarly, the severity of dementia can be quantified inversely to that of a healthy person.

Since we found the intergroup characteristics of healthy people, dementia-diagnosed patients, and elderly people with mild cognitive impairment (developed into dementia), the results of applying learning algorithms such as SVM and k-NN have very high accuracy, sensitivity, and specificity.

Based on the dementia prediction model to which the best learning algorithm is applied from various learning algorithms, it is possible to determine whether a new test subject has dementia-related abnormalities. That is, when data of a new subject is input, the quantitative level of the dementia state level is calculated based on the distance (similarity), and the dementia prediction result can be output.

This artificial intelligence learning-based dementia prediction apparatus and method using walking motion according to an embodiment of the present invention, unlike the existing artificial intelligence learning method that performs image patterns only with brain image data of dementia patients, extracting signal features Do it before learning. Dementia patients are diagnosed by examining cerebrospinal fluid, blood, and imaging images, starting with mild cognitive impairment showing cognitive decline. Although the characteristics of dementia are based on the clinical characteristics of cognitive decline and memory loss, it is an unknown new technology that detects abnormal elements occurring in the walking motion of dementia patients and performs AI-based learning. Currently, there is no perfect treatment for dementia, and the only alternative is to slow the progression of dementia through exercise. Since dementia characteristics can be detected and predicted at an early stage based on artificial intelligence learning using gait motion, social losses and economic costs associated with dementia can be reduced.

On the other hand, the apparatus and method for predicting dementia based on artificial intelligence using a walking motion according to an embodiment of the present invention can be implemented as one module by software and hardware, and the above-described embodiments of the present invention can be executed on a computer. It can be embodied in a general-purpose computer that operates the program using a computer-readable recording medium that can be written as a program. The computer-readable recording medium is implemented in the form of a magnetic medium such as a ROM, a floppy disk, a hard disk, an optical medium such as a CD or DVD, and a carrier wave such as transmission through the Internet. In addition, the computer-readable recording medium is distributed in a computer system connected to a network so that the computer-readable code can be stored and executed in a distributed manner.

In addition, a component or '~ part'/'~ module' used in an embodiment of the present invention is a task, class, subroutine, process, object, execution thread, or software such as a program performed in a predetermined area on a memory. ), may be implemented in hardware such as a field-programmable gate array (FPGA) or an application-specific integrated circuit (ASIC), or a combination of the software and hardware. The component or '~ part'/'~ module' may be included in a computer-readable storage medium, or a part thereof may be distributed and distributed among a plurality of computers.

Although the embodiments of the present invention have been described above with reference to the accompanying drawings, those of ordinary skill in the art to which the present invention pertains can realize that the present invention can be embodied in other specific forms without changing its technical spirit or essential features. you will be able to understand Therefore, it should be understood that the embodiments described above are illustrative in all respects and not restrictive.

10: Healthcare database
20: inspection agency server
100: dementia prediction device
110: data preprocessor 120: feature extraction unit
130: learning unit 140: cross-validation unit
150: dementia prediction unit

Claims (8)

a data preprocessor for preprocessing time-series acceleration data and time-series angular acceleration data respectively obtained from the left and right feet according to the gait motion into data of a predetermined size;
Extracting signal features including the average, standard deviation, energy, entropy, first cross-correlation coefficient and second cross-correlation coefficient of each of the pre-processed acceleration data and the pre-processed angular acceleration data, the average, the a feature extracting unit for calculating sub-signal features of standard deviation, the energy, the entropy, the first cross-correlation coefficient, and the second cross-correlation coefficient;
a learning unit for generating a plurality of learning models for predicting dementia by learning the extracted signal features;
a cross-validation unit that evaluates and cross-validates the plurality of learning models; and
and a dementia prediction unit that selects an optimal learning model from among the plurality of learning models through the cross-validation and predicts dementia using the optimal learning model,
As for the first cross-correlation coefficient, three signal features are extracted by pairing the preprocessed acceleration data two by two to calculate the correlation, and two signal features are paired by pairing the preprocessed angular acceleration data two by two to calculate the correlation 12 signal features are extracted by combining the left and right feet,
In the second cross-correlation coefficient, nine signal features are extracted by calculating the correlation of the left foot and the right foot based on the X-axis, Y-axis, and Z-axis of the preprocessed acceleration data, respectively, and the preprocessed angular acceleration data is X Dementia prediction based on artificial intelligence learning using gait motion, in which 9 signal features are extracted by calculating the correlation of the left and right feet based on the axis, Y-axis, and Z-axis, and 18 signal features are extracted by combining the left and right feet Device. The method of claim 1,
The data preprocessor,
A device for predicting dementia based on artificial intelligence learning using walking motion, which preprocesses the time series acceleration data and time series angular acceleration data from the 3D acceleration sensor and the 3D angular acceleration sensor respectively attached to the left and right feet of a pedestrian. The method of claim 1,
The data preprocessor,
An apparatus for predicting dementia based on artificial intelligence using a walking motion, which divides the time series acceleration data and the time series angular acceleration data to have the same size, and normalizes the divided data by overlapping the divided data with a constant size. delete delete The method of claim 1,
The learning unit,
An artificial intelligence learning-based dementia prediction device using gait motion, which learns the signal features using at least one of k-NN, SVM, Random Forest, Extra Tree, and XGBoost algorithms. The method of claim 1,
The cross-validation unit,
Dementia prediction device based on artificial intelligence learning using gait motion, which cross-validates using the learning data set change cross-validation method, the K-fold cross-validation method, and the Leave-P-Out cross-validation method. The method of claim 1,
The dementia prediction unit,
An artificial intelligence learning-based dementia prediction device using gait motion, which predicts dementia by applying the time series acceleration data obtained from the examiner and signal features extracted from the time series angular acceleration data to the optimal learning model.

Images