Apparatus and method for predicting dementia by using polyhedral drawing

Abstract

An apparatus and a method for predicting dementia by using polyhedral drawing are disclosed. The disclosed method comprises: step (a) of acquiring a polyhedral drawing image drawn by a user; step (b) of converting the polyhedral drawing image into graph structure data having nodes and edges; step (c) of inputting the graph structure data into a graph neural network to obtain an embedding vector for the graph structure; step (d) of obtaining feature vectors for additional information including the user's age input from the user; step (e) of inputting the embedding vector and the feature vectors for the additional information into a self-attention neural network to calculate the importance of each of the vectors; and step (f) of determining whether the user has dementia through neural network calculation on the basis of a result of the importance calculated in step (e). According to the disclosed apparatus and method, it is advantageous in that the initial diagnosis accuracy of dementia can be increased by using an artificial neural network by appropriate pre-processing of a polyhedral drawing, and the initial diagnosis accuracy of dementia can be increased by integrally analyzing various types of data by using a complex artificial intelligence system.

Description

Dementia prediction device and method using polyhedral drawing

[0001] The present invention relates to a dementia prediction device and method, and more specifically, to a dementia prediction device and method using polyhedral drawing.

[0002] Dementia, a cognitive impairment of the brain, presents a problem in that it is difficult to diagnose early based solely on clinical evaluation, as initial symptoms are often mild or atypical. Previously, clinicians determined whether a patient had dementia by utilizing test data such as cube drawing and surveys.

[0003] However, as this method is an empirical diagnostic method that did not fully utilize numerous medical records and complex data sources, it had the problem of not being able to fully utilize diverse data.

[0004] In addition, existing dementia diagnostic procedures, such as neuropsychological testing, require a significant amount of time and resources, and there was also the problem that specialists had to dedicate a considerable amount of time to them.

[0005]

[0006] The present invention proposes a federated learning method and apparatus capable of minimizing the influence of malicious terminals among terminals participating in federated learning and minimizing data exchanged between a federated learning server and terminals during federated learning.

[0007] The present invention proposes an apparatus and method that can improve the accuracy of early diagnosis of dementia by using an artificial neural network through appropriate preprocessing of polyhedral drawings.

[0008] In addition, the present invention proposes a device and method capable of improving the accuracy of early dementia diagnosis by integrally analyzing various types of data using a complex artificial intelligence system.

[0009] According to one aspect of the present invention, a method for predicting dementia performed in a computing device including a processor and a memory is provided, comprising the steps of: acquiring a polyhedral drawing image drawn by a user (a); converting the polyhedral drawing image into graph structure data having nodes and edges (b); inputting the graph structure data into a graph neural network to acquire an embedding vector for the graph structure (c); acquiring feature vectors for additional information including the user's age input by the user (d); inputting the embedding vector and the feature vectors for the additional information into a self-attention neural network to calculate the importance of each vector (e); and determining whether the user has dementia through neural network calculation based on the importance result calculated in step (e) (f).

[0010] The above step (b) includes: a step of obtaining a binarized image by binarizing the polyhedral drawing image (b1); a step of obtaining a plurality of lines for the binarized image through a line vectorization operation on the binarized image (b2); a step of removing unnecessary points among the points included in each of the plurality of lines (b3); a step of defining the points remaining by the above step (b3) as nodes of graph data and defining the coordinates of the remaining points as attribute information of the nodes and obtaining graph structure data (b4); and a step of merging nodes that are at a distance less than or equal to a first threshold value set based on the distance between each node (b5).

[0011] The above step (b3) includes: a step of connecting the starting point and the last point of the line with a straight line (b3-1); a step of searching for the intermediate point furthest from the connected straight line (b3-2); a step of determining whether the distance between the straight line and the furthest intermediate point is greater than or equal to a preset second threshold (b3-3); and a step of removing the furthest intermediate point if the distance between the straight line and the furthest intermediate point is less than or equal to the preset second threshold (b3-4).

[0012] The above step (b2) further comprises: a step (b3-5) of maintaining the farthest intermediate point and dividing the line based on the farthest intermediate point when the distance between the line and the farthest intermediate point is greater than or equal to a preset second threshold; and a step (b3-6) of repeating steps (b3-1) through (b3-5) for the divided line until there are no longer any intermediate points whose distance from the line is greater than or equal to the second threshold.

[0013] In the above step (b4), the edge information of the graph structure data is defined such that nodes belonging to the same line are connected and nodes belonging to non-identical lines are not connected.

[0014] In the above step (b5), the attribute information of the merged node is the average coordinate information of the two nodes to be merged.

[0015] After the above step (b6), the number of nodes for the merged nodes is further reduced in the same way as in the above step (b3).

[0016] After the above step (b6), the method further includes adding information reflecting whether each node is similar to a preset number of grapplets as node attribute information.

[0017] According to another aspect of the present invention, a dementia prediction device is provided comprising: a processor; and a memory connected to the processor, wherein the processor performs the steps of: acquiring a polyhedral drawing image drawn by a user (a); converting the polyhedral drawing image into graph structure data having nodes and edges (b); inputting the graph structure data into a graph neural network to acquire an embedding vector for the graph structure (c); acquiring feature vectors for additional information including the user's age input by the user (d); inputting the embedding vector and the feature vectors for the additional information into a self-attention neural network to calculate the importance of each vector (e); and determining whether the user has dementia through neural network calculation based on the importance result calculated in step (e) (f).

[0018] The present invention has the advantage of being able to increase the accuracy of early diagnosis of dementia by using an artificial neural network through appropriate preprocessing of polyhedral drawings.

[0019] In addition, the present invention has the advantage of improving the accuracy of early dementia diagnosis by using a complex artificial intelligence system to integrally analyze various types of data.

[0020]

[0021] FIG. 1 is a drawing illustrating the overall structure of a dementia prediction device using polyhedral drawing according to one embodiment of the present invention.

[0022] FIG. 2 is a block diagram showing the detailed structure of a preprocessing module according to one embodiment of the present invention.

[0023] FIG. 3 is a diagram showing an example of the result of performing line vectorization on an image drawn by a user according to an embodiment of the present invention.

[0024] FIG. 4 is a flowchart showing the operation of a line simplification module according to an embodiment of the present invention.

[0025] FIG. 5 is a drawing showing a cube image drawn according to the result of node merging according to an embodiment of the present invention.

[0026] FIG. 6 is a diagram showing an example of grapplets that can be used in the grapplet attribute information addition module.

[0027] FIG. 7 is a flowchart illustrating the overall flow of a dementia prediction method using polyhedral drawing according to an embodiment of the present invention.

[0028]

[0029] Hereinafter, specific embodiments according to embodiments of the present invention will be described with reference to the drawings. The following detailed description is provided to facilitate a comprehensive understanding of the methods, devices, and / or systems described herein. However, this is merely illustrative and the present invention is not limited thereto.

[0030] In describing the embodiments of the present invention, if it is determined that a detailed description of known technology related to the present invention may unnecessarily obscure the essence of the embodiments, such detailed description will be omitted. Furthermore, the terms described below are defined in consideration of their functions in the present invention, and these may vary depending on the intentions or practices of the user or operator. Therefore, such definitions should be based on the content throughout this specification. Terms used in the detailed description are intended merely to describe specific embodiments and should not be limiting. Unless explicitly stated otherwise, expressions in the singular form include the meaning of the plural form. In this description, expressions such as “include” or “comprising” are intended to refer to certain characteristics, numbers, steps, actions, elements, parts thereof, or combinations thereof, and should not be interpreted to exclude the existence or possibility of one or more other characteristics, numbers, steps, actions, elements, parts thereof, or combinations thereof other than those described.

[0031] FIG. 1 is a drawing illustrating the overall structure of a dementia prediction device using polyhedral drawing according to one embodiment of the present invention.

[0032]

[0033] Referring to FIG. 1, a dementia prediction device using polyhedral drawing according to one embodiment of the present invention may include a preprocessing module (100), a graph neural network (200), a multi-layer neural network (300), a cell attention neural network (400), and a diagnostic neural network (500).

[0034] A polyhedral drawing image created by a user is input to the preprocessing module (100), and the preprocessing module (100) functions to generate graph structure data for the polyhedral drawing image. The preprocessing module (100) is a core component of the present invention and receives a polyhedral image drawn by a patient and outputs graph structure data having node attribute information and edge information, which is connection information between nodes.

[0035] There has been a method for clinicians to empirically diagnose dementia using cube drawing, and in this embodiment, the case where the polyhedron is a cube will be explained as an example. However, those skilled in the art will understand that the object drawn by a dementia patient is not limited to a cube, and that cases where various polyhedra are drawn are also included within the scope of the present invention.

[0036] A detailed method for outputting a cube image drawn by a user as graph structure data in a preprocessing module (100) is explained with reference to a separate drawing.

[0037] The graph neural network (200) receives graph structure data output from the preprocessing module (100) and outputs an embedding vector for the input graph structure data through graph neural network operations. The graph neural network (200) is pre-trained and performs neural network operations on the input graph structure data based on the learned neural network weights. According to the research of the inventor of the present invention, when a cube drawn by a user is represented by the embedding vector of the graph neural network, the features of the drawn cube can be expressed more clearly.

[0038] The multi-layer neural network (300) receives additional information. Here, the additional information is information about the patient who performed the cued drawing, and may include the patient's age, the patient's education period, the patient's neuropsychological test data, etc. It will be obvious to those skilled in the art that the additional information can be varied and the number of additional information items can be set differently as needed. The multi-layer neural network (300) outputs a feature vector for each additional information item through neural network operations on the input additional information.

[0039] Although a single multi-layer neural network (300) is conceptually illustrated in FIG. 1, independent multi-layer neural networks may be used for each type of additional information, and each multi-layer neural network operates to output a feature vector for the input additional information. Since the operation of outputting a feature vector through neural network operations on the input information of the multi-layer neural network (300) is a widely known operation, a detailed description thereof is omitted.

[0040] The embedding vector for the drawn cube output from the graph neural network and the feature vector output from the multilayer neural network are input to the self-attention neural network (400). The self-attention neural network (400) performs self-attention operations on the embedding vector for the graph structure of the cube drawn by the user and the feature vector for each additional information. The self-attention neural network (400) ensures that each multimodal data is integrated with different weights through self-attention operations on the input multimodal data (embedding vector of the graph neural network and feature vector for each additional information). By assigning different weights to each multimodal data through the self-attention neural network (400), information with high diagnostic value can be reflected with higher weight in determining whether or not there is dementia.

[0041] The diagnostic neural network (500) receives the output vector of the self-attention neural network (400) as input and outputs final dementia diagnosis prediction information regarding whether or not there is dementia. The diagnostic neural network (500) can effectively predict dementia through neural network operations of multimodal data in which the importance of diagnostic value is adjusted through the self-attention neural network (400).

[0042] The dementia prediction device of the present invention comprises a plurality of neural networks (graph neural network, multi-layer neural network, self-attention neural network, diagnostic neural network), and the learning of each neural network may be performed using pre-prepared labels. The diagnostic neural network (500) is trained by backpropagating the loss to the diagnostic neural network (500) by the labels, the self-attention neural network (400) is trained by the loss backpropagated from the diagnostic neural network (500), and the graph neural network (200) and multi-layer neural network (300) can be trained by the loss backpropagated from the self-attention neural network (400). Of course, the above-described learning method using labels is an exemplary learning method, and it will be obvious to those skilled in the art that the learning of the dementia prediction device of the present invention can be performed through various known learning methods.

[0043] FIG. 2 is a block diagram showing the detailed structure of a preprocessing module according to one embodiment of the present invention.

[0044] Referring to FIG. 2, a preprocessing module (100) according to one embodiment of the present invention includes a drawing image binarization module (110), a line vectorization module (120), a line simplification module (130), a graph structure conversion module (140), a node merging module (150), a node simplification module (160), and a grapplet attribute addition module (170).

[0045] The drawing image binarization module (110) functions to binarize a cube image drawn by a patient. Binarization functions to binarize the pixels of the image so that the area where the user drew a line has data of 1 and the area where no line was drawn has data of 0. The 0 and 1 of the pixel data are exemplary, and since the binarization of the image is a widely known method, a detailed description thereof is omitted.

[0046] The line vectorization module (120) represents the binarized image as multiple lines having multiple points based on the binarized image. The line vectorization module (120) divides the image into multiple lines based on the kinematic structure of the input image, and each divided line includes multiple points. Line vectorization, which divides an image into multiple lines based on its kinematic structure, is a known technique, and line vectorization can be performed using a known line vectorization module method. For example, the line vectorization proposed by Bessmeltsev and Solomon in 2019 can be used.

[0047] If the patient has drawn an accurate cube (rectangular prism), the number of lines for the image drawn by the line vectorization module (!20) is fixed. However, ordinary people, including dementia patients, cannot draw an accurate cube, and the number of lines output by the line vectorization module (120) may differ from the fixed number, and in the case of a severe dementia patient, the drawing image may be expressed with a very large number of lines.

[0048] FIG. 3 is a diagram showing an example of the result of performing line vectorization on an image drawn by a user according to an embodiment of the present invention.

[0049] Referring to Fig. 3, it can be seen that the image drawn by the user is represented by a total of 4 lines through line vectorization.

[0050] The line simplification module (130) functions to reduce the number of points included in the lines generated by the line vectorization module (120). Each line generated by the line vectorization module (120) includes multiple points, and the line is formed by multiple points. The ultimate goal of the preprocessing module (100) of the present invention is to turn a cube drawn by a user into a graph structure, and the line simplification module (130) is the most critical module in the process of turning the drawn cube into a graph structure.

[0051] The line simplification module (130) performs line simplification by removing unnecessary points while keeping only the key points of the line.

[0052] FIG. 4 is a flowchart illustrating the operation of a line simplification module according to one embodiment of the present invention.

[0053] Line simplification is performed independently for each of the multiple lines, and Figure 4 illustrates a line simplification operation performed for a single line.

[0054] Referring to Fig. 4, first, the starting point and the ending point of each line are selected (step 400). The line consists of several points (P1, P2, , Pn), and P1 and Pn are selected when P1 is the starting point and Pn is the ending point. The points can be represented by two-dimensional coordinates.

[0055] When the start and end points of the line are selected, a straight line connecting the start point and the end point is created (step 410).

[0056] When a straight line is generated, the distances between the straight line and the intermediate points are each calculated, and the point among the multiple intermediate points that is farthest from the straight line is searched (step 420).

[0057] The distance to the point (Pmax) farthest from the straight line is compared with a preset threshold value to determine whether the distance to the point (Pmax) farthest from the straight line exceeds the preset threshold value (step 430).

[0058] If the distance between the line and Pmax is greater than a preset threshold, the corresponding point (Pmax) is retained (step 450). If the distance between the line and Pmax is less than a preset threshold, the corresponding point (Pmax) is removed (step 440).

[0059] If the distance between the line and Pmax is greater than a preset threshold, the line is divided based on the corresponding point (Pmax) rule (step 460). This divides the line, where P1 and Pn are each end points, into the P1-Pmax line and the Pmax-Pn line.

[0060] When line splitting is performed, steps 420 through 460 above are repeated for the split lines (step 470). Line simplification is performed repeatedly until there are no longer any points exceeding the threshold preset in step 430.

[0061] Through iterative point removal, only important points remain on the line, resulting in a line having only simplified points. As previously explained, the line simplification in Fig. 4 is performed for each line, and if there are four lines as in Fig. 3, the line simplification shown in Fig. 4 is performed for each of the first to fourth lines.

[0062] Referring again to FIG. 2, the graph structure conversion module (140) converts a cube image drawn by a user into graph structure data using points left by line simplification. The graph structure data is data containing attribute information of nodes and edge information connecting nodes. The graph structure conversion module (140) sets the points left by the line simplification module (130) as nodes. Meanwhile, the attribute information of each node is set as the (x,y) coordinates of the node.

[0063] Meanwhile, edge information is determined by the line to which each node (remaining point) belongs. Nodes belonging to the same line are connected, while nodes belonging to different lines are configured not to be connected.

[0064] Below, the graph structure transformation of the present invention is explained with one example.

[0065] For example, assume that there are 4 lines and a total of 10 points remain.

[0066] At this time, the points remaining on each line are as follows.

[0067] Line 1 (A): V(1)=(20,10), V(2)=(40,8)

[0068] Line 2 (B): V(3)=(40,8), V(4)=(43,32), V(5)=(38,40)

[0069] Line 3 (C): V(6)=(25,20), V(7)=(30,18), V(8)=(40,8)

[0070] Line 4 (D): V(9)=(30,18), V(10)=(37,40)

[0071] The points remaining on each line are defined as nodes, and the attribute information of a node becomes its (x,y) coordinates.

[0072] As explained earlier, edge information is configured to connect nodes belonging to the same line, and when nodes are configured as in the example above, the following edges are defined. In the example below, E(a,b) represents the edge connecting node a and node b.

[0073] Line 1 (A): E(12)=(V(1), V(2))

[0074] Line 2 (B): E(34)=(V(3), V(4)), E(45)=(V(4), V(5))

[0075] Line 3 (C): E(67)=(V(6), V(7)), E(78)=(V(7), V(8))

[0076] Line 4 (D): E(910)=(V(9), V(10))

[0077] The node merging module (150) selectively merges each node of the graph structure data. According to a preferred embodiment of the present invention, node merging is determined by the distance between nodes to be merged. If the distance between nodes to be merged is less than or equal to a preset threshold, the nodes to be merged are merged. If the distance between nodes to be merged is greater than or equal to a preset threshold, the nodes to be merged are not merged. The attribute value ((x,y) coordinate) of the merged node may be set as the average value of the nodes to be merged. Of course, any one of the original coordinate values ​​of the nodes to be merged may be selected and set as the attribute value of the merged node. As a result of the node merging, if the distance to another node becomes less than or equal to a preset threshold due to a change in the coordinate value of the node that was merged first, the merged node may be merged again.

[0078] The decision regarding whether to merge nodes is made for all existing node pairs.

[0079] In the example described above, nodes V(2) and V(3) are merged, and the attribute information of the merged node is defined as (40,8). At this time, (40,8) is the average coordinate value of the coordinates of V(2) and V(3). Also, nodes V(5) and V(10) are merged, and the attribute information of the merged node is defined as (37.5,40). Nodes V(7) and V(9)) are merged, and the attribute information of the merged node is defined as (30,18). Meanwhile, the node formed by the merger of V(2) and V(3) is at a distance less than or equal to a preset threshold from node V(8), and thus the merged node V(23) and node V(8) are merged.

[0080] The node merging module (150) assigns new node numbers to the merged nodes and corrects edge information during node merging. The node merging module (150) corrects edge information in a way that maintains the edges that existed in the nodes prior to merging, without generating new edges due to node merging.

[0081] FIG. 5 is a drawing showing a cube image drawn according to the result of node merging according to one embodiment of the present invention.

[0082] Referring to Fig. 5, it can be seen that the number of points (number of nodes) has decreased from 10 to 6 compared to Fig. 3.

[0083] The node simplification module (160) functions to reduce the number of nodes in the same way as the line simplification module (130). Although the node merging module (150) reduced the number of nodes through the distance between nodes, noise nodes may still exist, and noise nodes may have an adverse effect on dementia diagnosis. Since the node removal task of the node simplification module (160) is the same as the point removal task of the line simplification module (130), a detailed explanation thereof is omitted. Node simplification by the node simplification module (160) may not be performed as necessary.

[0084] The final graph structure data is completed by the node merging module (150) and the node simplification module (160), and the grapplet attribute information addition module (170) functions to reflect the grapplet information in the attributes of each node. Specifically, the grapplet attribute information addition module (170) is a module that adds information to the attribute information of a node regarding which of the multiple preset grapplet structures the graph structure centered on the node is similar to. For example, if four grapplet structures are prepared in advance and the graph structure information centered on the first node is similar to the first grapplet among the four grapplets, the grapplet attribute information addition module adds information that the graph structure centered on the first node is similar to the first grapplet to the attribute information of the first node.

[0085] For example, if the attribute information of the first node is (x1, y1) and it is determined to be similar to the first grapple when compared with four grapples, the node's attribute information can be transformed to (x1, y1, 1, 0, 0, 0). If it is determined to be similar to the second grapple, the node's attribute information can be transformed to (x1, y1, 0, 1, 0, 0). Additionally, counting information regarding similarity can also be reflected.

[0086] Figure 6 is a diagram showing an example of grapplets that can be used in the grapplet attribute information addition module.

[0087] It will be obvious to those skilled in the art that the grapplets shown in FIG. 6 are merely exemplary and that the shape and number of grapplets used may be changed as needed.

[0088] Node attribute information and edge information of graph structure data are obtained as described above, and such obtained graph structure data is input into a graph neural network (200), and the embedding vector output from the graph neural network (200) acts as one of the element data for determining dementia together with feature vectors of multiple additional information.

[0089] FIG. 7 is a flowchart illustrating the overall flow of a dementia prediction method using polyhedral drawing according to one embodiment of the present invention.

[0090] Referring to FIG. 7, first, a cube drawing image drawn by the user is obtained (step 700).

[0091] The preprocessing module (100) converts the cube drawing image into graph structure data having nodes and edges (step 702).

[0092] Graph structure data is input into a graph neural network to obtain embedding vectors for the graph structure (step 704).

[0093] Meanwhile, a feature vector for additional information including the user's age, which is input by the user, is obtained (step 706).

[0094] The embedding vectors output from the graph neural network and feature vectors for additional information are input into the self-attention neural network to calculate the importance of each vector (step 708).

[0095] Based on the results of the self-attention neural network operation, the diagnostic neural network is used to finally determine whether or not it is dementia (step 710).

[0096] The dementia prediction device of the present invention may be a computing device including a processor and memory, and the method of the present invention may be performed on a computing device including a processor and memory.

[0097] The present invention has been described with reference to embodiments illustrated in the drawings, but this is merely illustrative, and those skilled in the art will understand that various modifications and equivalent alternative embodiments are possible therefrom. Accordingly, the true technical scope of protection of the present invention should be determined by the technical spirit of the appended claims.

Claims

1. A method for predicting dementia performed on a computing device including a processor and memory, wherein A step of acquiring a polyhedral drawing image drawn by a user (a); Step (b) of converting the above polyhedron drawing image into graph structure data having nodes and edges; Step (c) of inputting the graph structure data into a graph neural network to obtain an embedding vector for the graph structure; Step (d) of obtaining feature vectors for additional information including the user's age input by the user; A step (e) of inputting the feature vectors for the above embedding vectors and the above additional information into a self-attention neural network to calculate the importance of each vector; A dementia prediction method comprising a step (f) of determining whether the user has dementia through neural network computation based on the importance result computed in step (e) above.

2. In Paragraph 1, The above step (b) is, Step (b1) of binarizing the polyhedron drawing image to obtain a binarized image; A step (b2) of obtaining a plurality of lines for the binarized image through a line vectorization operation for the binarized image; Step (b3) of removing unnecessary points among the points included in each of the above multiple lines; A step (b4) of defining the remaining points as nodes of the graph data and defining the coordinates of the remaining points as attribute information of the nodes, and obtaining graph structure data by the above step (b3); A dementia prediction method comprising the step (b5) of merging nodes that are at a distance less than or equal to a first threshold value set based on the distance between each node.

3. In Paragraph 2, The above step (b3) Step (b3-1) connecting the starting point and the ending point of the line with a straight line; Step (b3-2) of searching for the intermediate point furthest from the above connected straight line; A step (b3-3) of determining whether the distance between the above straight line and the above farthest intermediate point is greater than or equal to a preset second threshold; A dementia prediction method comprising the step (b3-4) of removing the farthest intermediate point when the distance between the above straight line and the above farthest intermediate point is less than or equal to a preset second threshold.

4. In Paragraph 3, The above step (b2) Step (b3-5) of maintaining the farthest intermediate point and dividing the line based on the farthest intermediate point when the distance between the above straight line and the farthest intermediate point is greater than or equal to a preset second threshold; and A dementia prediction method further comprising the step (b3-6) of repeating steps (b3-1) through (b3-5) for the divided line until there are no longer any intermediate points whose distance from the straight line is greater than or equal to the second threshold.

5. In Paragraph 2, A dementia prediction method in which, in step (b4) above, the edge information of the graph structure data is defined such that nodes belonging to the same line are connected and nodes belonging to non-identical lines are not connected.

6. In Paragraph 2, A dementia prediction method in which the attribute information of the merged node in step (b5) is the average coordinate information of the two target nodes to be merged.

7. In Paragraph 2, A dementia prediction method comprising further a step of reducing the number of nodes for the merged nodes after the above step (b6) in the same way as the above step (b3).

8. In Paragraph 2, A dementia prediction method that further includes the step of adding information reflecting whether each node is similar to a preset number of grapples as node attribute information after the above step (b6).

9. Processor; It includes memory connected to the above processor, The above processor is, A step of acquiring a polyhedral drawing image drawn by a user (a); Step (b) of converting the above polyhedron drawing image into graph structure data having nodes and edges; Step (c) of inputting the graph structure data into a graph neural network to obtain an embedding vector for the graph structure; Step (d) of obtaining feature vectors for additional information including the user's age input by the user; A step (e) of inputting the feature vectors for the above embedding vectors and the above additional information into a self-attention neural network to calculate the importance of each vector; A dementia prediction device that executes step (f) of determining whether the user has dementia through neural network computation based on the importance result computed in step (e) above.

10. In Paragraph 9, The above step (b) is, Step (b1) of binarizing the polyhedron drawing image to obtain a binarized image; A step (b2) of obtaining a plurality of lines for the binarized image through a line vectorization operation for the binarized image; Step (b3) of removing unnecessary points among the points included in each of the above multiple lines; A step (b4) of defining the remaining points as nodes of the graph data and defining the coordinates of the remaining points as attribute information of the nodes, and obtaining graph structure data by the above step (b3); A dementia prediction device comprising the step (b5) of merging nodes that are at a distance less than or equal to a preset first threshold based on the distance between each node.

11. In Paragraph 10, The above step (b3) Step (b3-1) connecting the starting point and the ending point of the line with a straight line; Step (b3-2) of searching for the intermediate point furthest from the above connected straight line; A step (b3-3) of determining whether the distance between the above straight line and the above farthest intermediate point is greater than or equal to a preset second threshold; A dementia prediction device comprising the step (b3-4) of removing the farthest intermediate point when the distance between the above straight line and the above farthest intermediate point is less than or equal to a preset second threshold.

12. In Paragraph 11, The above step (b2) Step (b3-5) of maintaining the farthest intermediate point and dividing the line based on the farthest intermediate point when the distance between the above straight line and the farthest intermediate point is greater than or equal to a preset second threshold; and A dementia prediction device further comprising the step (b3-6) of repeating steps (b3-1) through (b3-5) for the divided line until there are no longer any intermediate points whose distance from the straight line is greater than or equal to the second threshold value.

13. In Paragraph 10, A dementia prediction device in which, in the above step (b4), the edge information of the graph structure data is defined such that nodes belonging to the same line are connected and nodes belonging to non-identical lines are not connected.

14. In Paragraph 10, In the above step (b5), the attribute information of the merged node is the average coordinate information of the two target nodes for merging, a dementia prediction device.

15. In Paragraph 10, A dementia prediction device comprising a step of further reducing the number of nodes for the merged nodes after the above step (b6) in the same way as the above step (b3).

16. In Paragraph 10, A dementia prediction device further comprising the step of adding information reflecting whether each node is similar to a preset number of grapples as node attribute information after the above step (b6).

Images