Apparatus and method for diagnosing mild cognitive impairment subtype

Abstract

A dementia subtype diagnosis apparatus according to an embodiment of the present invention comprises: a data acquisition unit for collecting first protein data for effective protein candidates from a blood sample; a data processing unit for generating second protein data from the first protein data on the basis of an artificial intelligence analysis module, and generating combined protein data by combining the first protein data and the second protein data; and a diagnosis unit for determining a dementia subtype by using the combined protein data.

Description

Device and method for diagnosing mild cognitive impairment subtypes

[0001] The present invention relates to an apparatus and method for diagnosing a subtype of mild cognitive impairment. More specifically, the invention relates to an apparatus and method for diagnosing a subtype of mild cognitive impairment by obtaining biomarkers from a blood sample and using a graph-based neural network model that reflects the independent and interaction effects of the biomarkers.

[0002] The present invention is derived from research conducted as part of the Ministry of Science and ICT’s Individual Basic Research (MSIT) (Project Unique Number: 2710085913, Project Number: 2021R1A2C2003474, Research Project Name: Development of an Artificial Intelligence Model for Precision Medicine and Diagnosis Diversification of Dementia, Project Performing Institution Name: Ajou University).

[0003] The present invention was derived from research conducted as part of the Ministry of Science and ICT’s Artificial Intelligence Convergence Innovation Talent Development (R&D) (Project Unique Number: 2710033934, Project Number: RS-2023-00255968, Research Project Name: Artificial Intelligence Convergence Innovation Talent Development (Ajou University), Project Performing Organization Name: Ajou University Industry-Academic Cooperation Foundation).

[0004] The present invention is derived from research conducted as part of the Ministry of Education's establishment of a research infrastructure for science and engineering (Project No.: 2340032323, Project No.: 2022R1A6A3A01086784, Research Project Title: Development of a Multimodal-Multi-Domain Based Machine Learning Algorithm for Predicting Dementia Progression, Project Performing Institution: Ajou University).

[0005] The present invention was derived from research conducted as part of the Ministry of Health and Welfare's Global Training Support for Medical Scientists (Project No.: 2460002543, Project No.: RS-2024-00407544, Research Project Title: Development of a New Anticancer Treatment Strategy Based on Liver Cancer-TME Connectome, Project Performing Organization: Ajou University Industry-Academic Cooperation Foundation).

[0006] The present invention is derived from research conducted as part of the Ministry of Science and ICT’s Individual Basic Research (MSIT) (R&D) (Project No.: 2710078560, Project No.: RS-2022-00165386, Research Project Title: Discovery of Epidrivers to Control Cancer Progression and Development of Diagnostic and Therapeutic Candidates, Project Performing Organization: Ajou University Industry-Academic Cooperation Foundation).

[0007] The present invention was derived from research conducted as part of the Ministry of Health and Welfare's project to foster convergence talents specialized in medical artificial intelligence (Project No.: 2460004446, Project No.: RS-2025-02310331, Research Project Name: Ajou University Medical Artificial Intelligence Specialized Convergence Talent Development Project Group, Project Performing Organization Name: Ajou University Industry-Academic Cooperation Foundation).

[0008] The present invention was derived from research conducted as part of the Ministry of Health and Welfare's Research-Oriented Hospital Development (R&D) (Project No.: 2460004015, Project No.: RS-2021-KH113821, Research Project Title: Establishment of Human-Environmental Interaction Beyond Target Platform, Project Performing Organization: Ajou University Industry-Academic Cooperation Foundation).

[0009] Meanwhile, the Korean government, the provider of the problem, has no property interest in all aspects of the present invention.

[0010] Due to the aging of modern society, the incidence of dementia-related neurodegenerative diseases is increasing. Among these, Mild Cognitive Impairment (MCI) serves as a pre-dementia stage, making early diagnosis and appropriate treatment crucial clinical challenges. In particular, MCI can be classified into various subtypes, with Alzheimer's Disease Cognitive Impairment (ADCI) and Subcortical Vascular Cognitive Impairment (SVCI) being representative examples. Since these subtypes differ in their causes and treatment strategies, accurate classification and early diagnosis are essential. Existing diagnostic methods for MCI rely on surveys and cognitive function tests to assess the presence of MCI, which has been criticized for being subjective and having low accuracy in early diagnosis.

[0011] In addition, although there were diagnostic methods based on specific proteins or genetic biomarkers, diagnostic methods based on a single biomarker had the problem of lower diagnostic accuracy and reliability because they could not reflect complex biological interactions.

[0012] Therefore, there is a need for a diagnostic method for mild cognitive impairment that offers improved accuracy and reliability compared to diagnostic methods based on a single biomarker.

[0013] The present disclosure aims to provide a diagnostic method for mild cognitive impairment subtypes that offers improved accuracy and reliability compared to existing single-biomarker-based diagnostic methods by considering the independent and interaction effects of biomarkers.

[0014] In addition, the present disclosure aims to improve the interpretability of diagnostic results compared to general neural network model-based diagnostic methods by diagnosing mild cognitive impairment subtypes using an explainable graph propagation network (XGPN).

[0015] A diagnostic device for mild cognitive impairment subtypes according to one embodiment of the present disclosure may include: a biomarker acquisition unit that acquires biomarkers from a blood sample; a neural network model processing unit that generates a graph-based neural network model reflecting independent and interaction effects of the biomarkers and trains the graph-based neural network model by adjusting parameters of the graph-based neural network model; and a diagnostic unit that diagnoses a mild cognitive impairment subtype using the trained graph-based neural network model.

[0016] In some embodiments, the graph-based neural network model may include an explainable graph propagation network (XGPN).

[0017] In some embodiments, subtypes of mild cognitive impairment may include Alzheimer's-associated cognitive impairment (ADCI) and subcortical vascular cognitive impairment (SVCI).

[0018] In some embodiments, the neural network model processing unit can adjust the inter-node propagation parameters of the graph-based neural network model and the classification parameters for classifying mild cognitive impairment subtypes during the learning process of the graph-based neural network model.

[0019] In some embodiments, if the value of the inter-node propagation parameter is low, it weakens the connection strength between nodes to prevent overfitting, and if the value is high, it strengthens information propagation to enhance the interaction effect between proteins.

[0020] In some embodiments, classification parameters may be used to calculate probability values ​​for ADCI and SVCI subtypes based on the learned inter-node interaction effect matrix.

[0021] In some embodiments, the biomarker acquisition unit may acquire a diagnostic marker value from a blood sample, acquire effective protein candidates based on the diagnostic marker value, and acquire biomarkers by performing a p-value test on the effective protein candidates.

[0022] In some embodiments, the biomarker acquisition unit may acquire SUVR (Standardized Uptake Value Ratio), an indicator representing how much radioisotope has accumulated in a specific brain region in a PET (Positron Emission Tomography) brain image, and PSMD (Peak Width of Skeletonized Mean Diffusivity), an indicator for evaluating microstructural damage of white matter in a DTI (Diffusion Tensor Imaging) brain image, as diagnostic marker values.

[0023] In some embodiments, the neural network model processing unit updates the graph-based neural network model by reflecting the interaction effects between biomarkers in the graph-based neural network model, obtains a loss function value to evaluate the performance of the graph-based neural network model, and terminates the training of the graph-based neural network model when the amount of change of the loss function value is below a threshold value.

[0024] A method for diagnosing a subtype of mild cognitive impairment according to one embodiment of the present disclosure may include: obtaining biomarkers from a blood sample; generating a graph-based neural network model that reflects independent and interaction effects of the biomarkers; training a graph-based neural network model by adjusting parameters of the graph-based neural network model; and diagnosing a subtype of mild cognitive impairment using the trained graph-based neural network model.

[0025] In some embodiments, the graph-based neural network model may include an explainable graph propagation network (XGPN).

[0026] In some embodiments, subtypes of mild cognitive impairment may include Alzheimer's-associated cognitive impairment and subcortical vascular cognitive impairment.

[0027] In some embodiments, the step of training a graph-based neural network model may include the step of adjusting inter-node propagation parameters of the graph-based neural network model; and the step of adjusting classification parameters for classifying mild cognitive impairment subtypes.

[0028] In some embodiments, the step of obtaining biomarkers may include: obtaining diagnostic marker values ​​from a blood sample; obtaining effective protein candidates based on diagnostic marker values; and obtaining biomarkers by performing a p-value test on the effective protein candidates.

[0029] In some embodiments, the step of training a graph-based neural network model may include: a step of updating the graph-based neural network model by reflecting the interaction effects between biomarkers in the graph-based neural network model; a step of obtaining a loss function value to evaluate the performance of the graph-based neural network model; and a step of terminating the training of the graph-based neural network model when the amount of change in the loss function value is below a threshold value.

[0030] The apparatus and method for diagnosing mild cognitive impairment subtypes according to an embodiment of the present disclosure can provide a method for diagnosing mild cognitive impairment subtypes with improved accuracy and reliability compared to existing single biomarker-based diagnostic methods by considering the independent and interaction effects of biomarkers.

[0031] In addition, the device and method for diagnosing mild cognitive impairment subtypes according to the embodiments of the present disclosure can increase the interpretability of the diagnosis results compared to general neural network model-based diagnosis methods by diagnosing mild cognitive impairment subtypes using an explainable graph propagation network.

[0032] FIG. 1 is a drawing showing a diagnostic device for mild cognitive impairment subtypes according to an embodiment of the present disclosure.

[0033] FIG. 2 is a drawing showing a neural network model processing unit and a diagnostic unit according to an embodiment of the present disclosure.

[0034] FIG. 3 is a diagram illustrating a method for learning a graph-based neural network model according to an embodiment of the present disclosure.

[0035] FIG. 4 is a flowchart illustrating a method for diagnosing a mild cognitive impairment subtype according to an embodiment of the present disclosure.

[0036] FIG. 5 is a flowchart illustrating a method for obtaining a biomarker according to an embodiment of the present disclosure.

[0037] FIG. 6 is a diagram showing a p-value test according to an embodiment of the present disclosure.

[0038] FIG. 7 is a flowchart illustrating a graph-based neural network model learning method according to an embodiment of the present disclosure.

[0039] Hereinafter, preferred embodiments of the present invention will be described as follows with reference to the attached drawings.

[0040] In the following, terms such as "upper," "middle," and "lower" may be replaced with other terms, such as "first," "second," and "third," to describe the components of the specification. While terms such as "first," "second," and "third" may be used to describe various components, they are not limited by these terms, and "first component" may be named "second component."

[0041]

[0042] FIG. 1 is a drawing showing a diagnostic device for mild cognitive impairment subtypes according to an embodiment of the present disclosure.

[0043] Referring to FIG. 1, a mild cognitive impairment subtype diagnosis device (100) according to an embodiment of the present disclosure may include a biomarker acquisition unit (110), a neural network model processing unit (120), and a diagnosis unit (130).

[0044] A mild cognitive impairment subtype diagnostic device (100) can collect data regarding blood proteins from a blood sample and can diagnose a subtype of mild cognitive impairment by analyzing the blood protein data.

[0045] The biomarker acquisition unit (110) can acquire diagnostic marker values ​​from a blood sample. For example, the biomarker acquisition unit (110) can acquire SUVR (Standardized Uptake Value Ratio), which is an indicator of how much radioisotope has accumulated in a specific brain region in a PET (Positron Emission Tomography) brain image, and PSMD (Peak Width of Skeletonized Mean Diffusivity), which is an indicator for evaluating microstructural damage of white matter in a DTI (Diffusion Tensor Imaging) brain image, a type of MRI (Magnetic Resonance Imaging) technique, as diagnostic marker values.

[0046] The biomarker acquisition unit (110) can select effective protein candidates through data processing and filtering of diagnostic marker values, and can acquire effective protein candidates with high association with mild cognitive impairment subtypes as biomarkers by performing a p-value test to confirm statistically significant data.

[0047] The neural network model processing unit (120) can perform the role of generating and training a graph-based neural network model (121) based on the biomarkers acquired by the biomarker acquisition unit (110).

[0048] A graph-based neural network model (121) can construct a protein-protein interaction (PPI) network, and each node (biomarker) in the network can learn the independent effects of each node and the interaction effects between nodes to update the protein-protein interaction network. For example, the graph-based neural network model (121) can be designed as an XGPN (eXplainable Graph Propagation Network) that processes data based on a graph structure and provides the function of explaining the model's output results in a way that humans can understand. In this case, the graph-based neural network model (121) can be more reliable than existing neural network models by providing interpretability of the diagnostic results.

[0049] The neural network model processing unit (120) can improve the accuracy of the graph-based neural network model (121) by adjusting the inter-node propagation parameter (Steadiness) that controls the degree of propagation for each node during the learning process of the graph-based neural network model (121) and the classification parameter for classifying the final mild cognitive impairment subtype.

[0050] If the value of the inter-node propagation parameter is low, it weakens the connection strength between nodes to prevent overfitting, and if the value is high, it strengthens information propagation to enhance the effect of interactions between proteins.

[0051] Specifically, inter-node propagation parameters ( ) can be expressed according to the following mathematical formula 1.

[0052]

[0053] step, is an interaction effect matrix representing the interaction effects between nodes, is a diagonal matrix representing the inter-node propagation parameters of individual nodes, is the normalized graph Laplacian matrix represented according to the following mathematical formula 2, is an independent effect matrix representing the independent effects of the nodes.

[0054]

[0055] step, is the pxp identity matrix, is a matrix of order indicating how many times each node is connected to other nodes, is an adjacency matrix representing the connection strength between nodes.

[0056] Additionally, the classification parameter for classifying mild cognitive impairment subtypes may be a variable used to classify ADCI and SVCI subtypes, and may be a parameter used to calculate probability values ​​for each subtype based on the learned inter-node interaction effect matrix.

[0057] Specifically, classification parameters for classifying subtypes of mild cognitive impairment ( , ) can be expressed according to the following mathematical formulas 3 and 4.

[0058]

[0059] step, is the interaction effect matrix, is the natural constant, is a classification parameter for classifying ADCI subtypes, is the probability of being ADCI.

[0060]

[0061] step, is the interaction effect matrix, is the natural constant, is a classification parameter for classifying SVCI subtypes, is the probability of being an SVCI.

[0062] The diagnosis unit (130) can classify subtypes of mild cognitive impairment using a graph-based neural network model (121) learned in the neural network model processing unit (120) and provide a diagnosis result. Specifically, the diagnosis unit (130) can calculate probability values ​​for ADCI and SVCI subtypes and diagnose the subtype corresponding to the higher probability value as a subtype of mild cognitive impairment.

[0063] A diagnostic device (100) for mild cognitive impairment subtypes according to an embodiment of the present disclosure can improve accuracy and reliability compared to existing single biomarker-based diagnostic methods by considering the independent effects and interaction effects of biomarkers.

[0064] In addition, the mild cognitive impairment subtype diagnosis device (100) according to an embodiment of the present disclosure can increase the interpretability of the diagnosis results compared to a general neural network model-based diagnosis method by diagnosing mild cognitive impairment subtypes using an explainable graph propagation network. For example, the mild cognitive impairment subtype diagnosis device (100) according to an embodiment of the present disclosure can be designed in an XGPN (eXplainable Graph Propagation Network) manner in which the relationships between nodes and how each node influenced the final result can be interpreted.

[0065]

[0066] FIG. 2 is a drawing showing a neural network model processing unit and a diagnostic unit according to an embodiment of the present disclosure.

[0067] FIG. 2 can be described with reference to FIG. 1 described above. Referring to FIG. 2, a neural network model processing unit (220) and a diagnostic unit (230) are shown. The neural network model processing unit (220) and the diagnostic unit (230) shown in FIG. 2 may correspond to the neural network model processing unit (120) and the diagnostic unit (130) of FIG. 1, respectively.

[0068] The neural network model processing unit (220) may, for example, receive information regarding biomarkers and data regarding the independent effects of biomarkers from the biomarker acquisition unit (110). The neural network model processing unit (220) may generate a graph-based neural network model (221) using the biomarkers and may acquire data regarding the interaction effects of biomarkers based on the output of the graph-based neural network model (221) according to the inter-node propagation parameters. The neural network model processing unit (220) may provide data regarding the independent effects of biomarkers and data regarding the interaction effects to the diagnosis unit (230).

[0069] The diagnostic unit (230) can diagnose a subtype of mild cognitive impairment based on data regarding the independent effects of biomarkers and data regarding the interaction effects of biomarkers from the neural network model processing unit (220). Specifically, the diagnostic unit (230) can calculate probability values ​​for ADCI and SVCI respectively, and can finally diagnose the subtype based on the higher of the two calculated probability values.

[0070]

[0071] FIG. 3 is a diagram illustrating a method for learning a graph-based neural network model according to an embodiment of the present disclosure.

[0072] FIG. 3 may be described with reference to FIG. 1 and FIG. 2 described above. Referring to FIG. 3, a graph-based neural network model (221) according to an embodiment of the present disclosure includes a propagation parameter (Steadiness) between nodes of the graph-based neural network model (221) during the learning process, an ADCI classification parameter for classifying mild cognitive impairment subtypes ( ) and SVCI classification parameters( ) can be adjusted.

[0073] The internode propagation parameter (Steadiness) can play a role in controlling the strength of information propagation between each node (biomarker) of a graph-based neural network model (221). The internode propagation parameter can prevent overfitting that may occur due to excessive information propagation and maintain the balance of information between nodes. The internode propagation parameter can play a role in assigning weights in the process of combining independent effects of each node and information propagated from neighboring nodes at each learning step.

[0074] ADCI classification parameters ( ) may be a parameter for the diagnostic unit (230) to classify ADCI subtypes, and may be used to extract activation values ​​of nodes (biomarkers) related to ADCI from independent effects and interaction effects between nodes transmitted through the graph propagation process, and to calculate the final ADCI probability.

[0075] SVCI classification parameters ( ) can be a parameter for the diagnostic unit (230) to classify SVCI subtypes, and can be used to extract activation values ​​of nodes (biomarkers) associated with SVCI from independent effects and interaction effects between nodes transmitted through the graph propagation process, and to calculate the final SVCI probability.

[0076]

[0077] FIG. 4 is a flowchart illustrating a method for diagnosing a subtype of mild cognitive impairment according to an embodiment of the present disclosure. FIG. 4 can be described with reference to FIG. 1 to 3 described above.

[0078] A method for diagnosing a mild cognitive impairment subtype (S100) according to an embodiment of the present disclosure is,

[0079] The method may include the steps of obtaining biomarkers from a blood sample (S110), generating a graph-based neural network model that reflects the independent and interaction effects of the biomarkers (S120), training the graph-based neural network model by adjusting the parameters of the graph-based neural network model (S130), and diagnosing a subtype of mild cognitive impairment using the trained graph-based neural network model (S140).

[0080] FIG. 4 illustrates steps S110 to S140 being performed sequentially, but is not limited thereto; some steps may be merged and performed simultaneously, some steps may be omitted, or new steps may be added.

[0081] In step S110, biomarkers can be obtained from blood samples. For example, diagnostic marker values ​​(SUVR, PSMD) extracted from PET and DTI brain images can be obtained from blood samples, and effective protein candidates can be selected based on these diagnostic marker values. For instance, effective protein candidates can be selected by extracting SUVR, an indicator of how much radioisotope has accumulated in a specific brain region from PET brain images, and PSMD, an indicator for evaluating microstructural damage in white matter from DTI brain images, and selecting proteins that exceed a specific threshold or have statistically significant values. Subsequently, statistically significant effective protein candidates can be obtained as biomarkers by performing a p-value test on the selected effective protein candidates.

[0082] In step S120, a graph-based neural network model can be generated that reflects the independent and interaction effects of the biomarkers. For example, the neural network model processing unit (220) can construct a protein-protein interaction (PPI) network and generate a graph-based neural network model by representing the relationships between each biomarker (node) as a graph structure.

[0083] Specifically, proteins selected as biomarkers can each be set as nodes in the graph, and each node may include characteristic values ​​representing the independent effects of the corresponding protein. Additionally, edges representing the connection relationships between nodes may include characteristic values ​​representing the interaction effects between proteins.

[0084] In step S130, the graph-based neural network model can be trained by adjusting the parameters of the graph-based neural network model. For example, the neural network model processing unit (220) can train the graph-based neural network model by adjusting the inter-node propagation parameters of the graph-based neural network model and the classification parameters for classifying mild cognitive impairment subtypes. For example, the neural network model processing unit (220) can define a loss function to quantify and measure the difference between the predicted value and the actual value, and can terminate training when the loss function value is below a threshold.

[0085] In step S140, a subtype of mild cognitive impairment can be diagnosed using a learned graph-based neural network model. For example, the diagnostic unit (230) may receive data regarding independent effects and interaction effects between nodes from the neural network model processing unit (220), and may calculate probability values ​​for each subtype of mild cognitive impairment based on the data received from the neural network model processing unit (220). The diagnostic unit (230) may diagnose a subtype of mild cognitive impairment (ADCI or SVCI) based on the higher probability value among the probability values ​​for each subtype.

[0086]

[0087] FIG. 5 is a flowchart illustrating a method for obtaining a biomarker according to an embodiment of the present disclosure.

[0088] FIG. 5 can be described with reference to FIG. 1 to FIG. 4 described above. Referring to FIG. 5, a method for obtaining a biomarker (S210) according to an embodiment of the present disclosure may include the step of obtaining a diagnostic marker value from a blood sample (S211), the step of obtaining effective protein candidates based on the diagnostic marker value (S212), and the step of obtaining biomarkers by performing a p-value test on the effective protein candidates (S213). The method for obtaining a biomarker (S210) illustrated in FIG. 4 may, for example, correspond to step S110 of FIG. 4.

[0089] FIG. 5 illustrates steps S211 to S213 being performed sequentially, but is not limited thereto; some steps may be merged and performed simultaneously, some steps may be omitted, or new steps may be added.

[0090] In step S211, diagnostic marker values ​​can be obtained from a blood sample. For example, the biomarker acquisition unit (110) can obtain diagnostic marker values ​​(SUVR, PSMD) from a blood sample through brain imaging techniques such as PET and DTI.

[0091] In step S212, effective protein candidates can be obtained based on diagnostic marker values. For example, if the diagnostic marker value exceeds a specific threshold, the corresponding proteins can be obtained as effective protein candidates.

[0092] In step S213, biomarkers can be obtained by performing a p-value test on the effective protein candidates. For example, biomarkers can be obtained based on a pre-set significance level (e.g., p < 0.05) by evaluating the association between each effective protein candidate and a subtype of mild cognitive impairment.

[0093]

[0094] FIG. 6 is a diagram showing a p-value test according to an embodiment of the present disclosure.

[0095] Figure 6 can be explained with reference to the aforementioned Figures 1 to 5. Referring to Figure 6, the X-axis may represent the change amount of each biomarker, and the Y-axis may represent statistical significance. For each of the effective protein candidates, statistically significant effective protein candidates whose Y-axis value exceeds the p-value based on p=0.05 may be determined as biomarkers of ADCI and SVCI.

[0096]

[0097] FIG. 7 is a flowchart illustrating a graph-based neural network model learning method according to an embodiment of the present disclosure. FIG. 7 can be described with reference to FIG. 1 through 6 described above. Referring to FIG. 7, a graph-based neural network model learning method (S330) according to an embodiment of the present disclosure may include a step of initializing a graph-based neural network model (S331), a step of updating a graph-based neural network model by reflecting the interaction effect between biomarkers in the graph-based neural network model (S332), a step of obtaining a loss function value for evaluating the performance of the graph-based neural network model (S333), and a step of determining whether the amount of change of the loss function value is below a threshold value (S334). The graph-based neural network model learning method (S330) illustrated in FIG. 7 may, for example, correspond to step S130 of FIG. 4.

[0098] FIG. 7 illustrates steps S331 to S334 being performed sequentially, but is not limited thereto; some steps may be merged and performed simultaneously, some steps may be omitted, or new steps may be added.

[0099] In step S331, a graph-based neural network model can be initialized. For example, the initial state of the initial nodes and edges (interactions between biomarkers) of the graph-based neural network model can be set.

[0100] In step S332, the graph-based neural network model can be updated by reflecting the interaction effects between biomarkers in the graph-based neural network model. For example, the interaction effects between nodes can be calculated based on the inter-node propagation parameters, and the model can be updated by reflecting the inter-node interaction effects in the edges.

[0101] In step S333, a loss function value can be obtained to evaluate the performance of the graph-based neural network model. For example, known loss functions such as Cross-Entropy Loss or Mean Squared Error (MSE) may be used.

[0102] In step S334, it can be determined whether the amount of change in the loss function value is below a threshold value. For example, the amount of change can be calculated by comparing the difference between the loss function value calculated in the current training step and the loss function value from the previous step, and if the amount of change in the loss function value is below a threshold value (e.g., 0.001 or less), the training of the graph-based neural network model can be terminated. If the amount of change in the loss function value exceeds the threshold value, steps S332 through S334 can be repeated.

[0103]

[0104] The present invention is not limited by the embodiments described above and the attached drawings, but is intended to be limited by the appended claims. Accordingly, various substitutions, modifications, changes, and combinations of embodiments may be made by those skilled in the art without departing from the technical spirit of the invention as described in the claims, and such are also to be considered to fall within the scope of the present invention.

Claims

1. A biomarker acquisition unit for acquiring biomarkers from a blood sample; A neural network model processing unit that generates a graph-based neural network model reflecting the independent and interaction effects of the above biomarkers, and trains the graph-based neural network model by adjusting the parameters of the graph-based neural network model; and A mild cognitive impairment subtype diagnosis device comprising a diagnosis unit that diagnoses mild cognitive impairment subtypes using a learned graph-based neural network model.

2. In Paragraph 1 A diagnostic device for mild cognitive impairment subtypes, comprising an explained graph propagation network (XGPN) based on the graph-based neural network model.

3. In Paragraph 1, A diagnostic device for mild cognitive impairment subtypes, wherein the above-mentioned mild cognitive impairment subtypes include Alzheimer's Disease Cognitive Impairment (ADCI) and Subcortical Vascular Cognitive Impairment (SVCI).

4. In Paragraph 1, A mild cognitive impairment subtype diagnosis device, wherein the neural network model processing unit adjusts the inter-node propagation parameters of the graph-based neural network model and the classification parameters for classifying the mild cognitive impairment subtype during the learning process of the graph-based neural network model.

5. In Paragraph 4, A diagnostic device for a subtype of mild cognitive impairment, wherein the above-mentioned inter-node propagation parameter weakens the connection strength between nodes to prevent overfitting when the value is low, and strengthens information propagation to enhance the effect of protein-protein interaction when the value is high.

6. In Paragraph 4, A mild cognitive impairment subtype diagnostic device, wherein the above classification parameters are used to calculate probability values ​​for ADCI and SVCI subtypes based on a learned inter-node interaction effect matrix.

7. In Paragraph 1, A diagnostic device for a mild cognitive impairment subtype, wherein the biomarker acquisition unit acquires a diagnostic marker value from the blood sample, acquires effective protein candidates based on the diagnostic marker value, and acquires the biomarkers by performing a p-value test on the effective protein candidates.

8. In Paragraph 7, A diagnostic device for a subtype of mild cognitive impairment, wherein the above-mentioned biomarker acquisition unit acquires SUVR (Standardized Uptake Value Ratio), an indicator representing how much radioisotope has accumulated in a specific brain region in a PET (Positron Emission Tomography) brain image, and PSMD (Peak Width of Skeletonized Mean Diffusivity), an indicator for evaluating microstructural damage of white matter in a DTI (Diffusion Tensor Imaging) brain image, as diagnostic marker values.

9. In Paragraph 1, A diagnostic device for mild cognitive impairment subtypes, wherein the neural network model processing unit updates the graph-based neural network model by reflecting the interaction effect between the biomarkers in the graph-based neural network model, obtains a loss function value for evaluating the performance of the graph-based neural network model, and terminates the training of the graph-based neural network model when the amount of change in the loss function value is below a threshold value.

10. A step of obtaining biomarkers from a blood sample; A step of generating a graph-based neural network model that reflects the independent and interaction effects of the above biomarkers; A step of training the graph-based neural network model by adjusting the parameters of the graph-based neural network model; and A method for diagnosing a subtype of mild cognitive impairment, comprising the step of diagnosing a subtype of mild cognitive impairment using a trained graph-based neural network model.

11. In Paragraph 10, A method for diagnosing a subtype of mild cognitive impairment, wherein the above graph-based neural network model includes an explainable graph propagation network (XGPN).

12. In Paragraph 10, A diagnostic method for subtypes of mild cognitive impairment, wherein the above-mentioned subtypes of mild cognitive impairment include Alzheimer's Disease Cognitive Impairment (ADCI) and Subcortical Vascular Cognitive Impairment (SVCI).

13. In Paragraph 10, The step of training the above graph-based neural network model is, A step of adjusting inter-node propagation parameters of the above graph-based neural network model; and A method for diagnosing mild cognitive impairment subtypes, comprising the step of adjusting classification parameters for classifying the above mild cognitive impairment subtypes.

14. In Paragraph 10, The step of acquiring the above biomarkers is, A step of obtaining a diagnostic marker value from the above blood sample; A step of obtaining effective protein candidates based on the above diagnostic marker values; and A method for diagnosing a subtype of mild cognitive impairment, comprising the step of obtaining the biomarkers by performing a p-value test on the effective protein candidates.

15. In Paragraph 10, The step of training the above graph-based neural network model is, A step of updating the graph-based neural network model by reflecting the interaction effects between the biomarkers in the graph-based neural network model; A step of obtaining a loss function value to evaluate the performance of the above graph-based neural network model; and A method for diagnosing a subtype of mild cognitive impairment, comprising the step of terminating the training of the graph-based neural network model when the amount of change in the loss function value is below a threshold value.

Images