Memory curve-based cognitive memory dysfunction auxiliary diagnosis device

Abstract

The application discloses an auxiliary diagnosis device for cognitive memory dysfunction based on a memory curve and belongs to the technical field of memory cognition auxiliary diagnosis. The auxiliary diagnosis device comprises a data preprocessing unit, a brain network feature acquisition unit, an activity signal feature extraction unit of a specified brain region, a memory model generation unit based on a memory curve and a cognitive memory dysfunction detection unit. The brain network feature acquisition unit is used for constructing a corresponding brain network based on preprocessed magnetic resonance imaging data and performing brain network feature extraction, and the extracted activity signal features are combined to serve as the input of the cognitive memory dysfunction detection unit, so that the auxiliary diagnosis result is output based on a preset detection model. The auxiliary diagnosis combines the auxiliary diagnosis result and a brain network topology map to generate a corresponding memory model and visually output and display. The application can be used for improving the accuracy of cognitive memory dysfunction auxiliary diagnosis and realizing the visual display of the memory forgetting process of a detection object.

Description

Technical Field

[0001] This invention belongs to the field of memory and cognition auxiliary diagnostic technology, specifically relating to an auxiliary diagnostic device for cognitive memory dysfunction based on memory curves. Background Technology

[0002] Functional magnetic resonance imaging (fMRI) is a representative non-invasive imaging technique used to study brain cognition and function in different populations. It is generally used to create images of brain physiological activity. This imaging technique has extremely high temporal and spatial resolution, and a major advantage is that it does not expose subjects to ionizing radiation, thus avoiding secondary harm to the subjects.

[0003] Since its invention, functional magnetic resonance imaging (fMRI) has made many contributions to the field of brain science. Many studies rely on fMRI for experimental support. fMRI helps researchers obtain physiological activity signals of the brain and can reveal abnormal lesions and physiological characteristics of the brain through images. FMRI is also frequently used in clinical practice for tumor observation.

[0004] From a modeling perspective, brain simulation can generally be approached at three scales: microscale, mesoscale, and macroscale. The microscale involves the brain composed of neurons and synapses; from a microscopic perspective, it's a network topology where nodes are tiny neurons and edges are synapses. The mesoscale involves clusters of neurons forming nodes within the network, with connections between these clusters forming edges. The macroscale divides the brain into regions based on physiological and structural properties, using these regions as nodes and connections (functional and physical) between them as edges.

[0005] Traditional memory theories support the idea that initial memories are dynamic, but once consolidated, they remain unchanged. Modern psychology and neuroscience, however, suggest that memory is always dynamic; even stable memories can become unstable after activation. In other words, the human brain's memory system exhibits high plasticity throughout the process of encoding, consolidation, and re-consolidation. Memory is categorizable. After information flows through the brain and is encoded, it can be divided into short-term memory, long-term memory, and sensory memory based on the duration and frequency of brain stimulation. Sensory memory is the most basic form, receiving information only from the five senses. Sensory information is immediate and time-sensitive, and the sheer volume of sensory information means much is unconsciously discarded by the brain. After a large amount of information becomes sensory memory, it undergoes brain attention (processing) to become short-term memory. The brain remembers this information for a short time (usually only a few seconds), but without any further processing, short-term memory is forgotten very quickly. Existing memory models each have their advantages and can be used as a reference. However, most models provide very limited descriptions of information and are not applicable to people with impaired cognitive and memory functions. Summary of the Invention

[0006] This invention provides a cognitive memory dysfunction auxiliary diagnostic device based on memory curves, which can be used to improve the accuracy of auxiliary diagnosis of cognitive memory dysfunction, and at the same time realize the visualization of the memory forgetting process of the test subject.

[0007] The technical solution adopted in this invention is as follows:

[0008] A memory curve-based auxiliary diagnostic device for cognitive memory dysfunction includes: a data preprocessing unit, a brain network feature acquisition unit, a designated brain region activity signal feature extraction unit, a memory curve-based memory model generation unit, and a cognitive memory dysfunction detection unit.

[0009] The data preprocessing unit is used to preprocess the input magnetic resonance imaging data and send the preprocessed data to the brain network feature acquisition unit and the activity signal feature extraction unit of the specified brain region, respectively.

[0010] The brain network feature acquisition unit constructs a brain functional network, or a brain functional network and a brain structural network, based on the Pearson correlation coefficient between brain regions.

[0011] The brain functional network is as follows: Based on a pre-set brain region template, the brain is divided into regions, and the time series of pre-processed functional magnetic resonance imaging data of each brain region is obtained. The Pearson correlation coefficient between the time series of any two brain regions is calculated. If the correlation coefficient is greater than a specified threshold, the two brain regions are connected; otherwise, they are not connected. The brain functional network is constructed and saved.

[0012] The brain structure network is constructed and saved as follows: Based on a pre-set brain region template, the brain is divided into regions, gray matter images from the structural magnetic resonance imaging data of each brain region are obtained, the Pearson correlation coefficient between the gray matter images of any two brain regions is calculated, if the correlation coefficient is greater than a specified threshold, the two brain regions are connected, otherwise they are not connected.

[0013] The brain network feature acquisition unit extracts network features of brain functional networks and brain structural networks based on specified network attribute indicators, obtains brain network features, and sends them to the cognitive memory dysfunction detection unit.

[0014] The activity signal feature extraction unit for a specified brain region is used to extract activity signal features of a specified brain region related to cognitive memory and send them to the cognitive memory dysfunction detection unit.

[0015] The cognitive memory dysfunction detection unit is pre-set with a trained cognitive memory dysfunction detection model. It uses brain network features and activity signal features as inputs to the cognitive memory dysfunction detection model, obtains the cognitive memory dysfunction detection result of the current detection object based on its output, and sends the cognitive memory dysfunction detection result to the memory model generation unit based on the memory curve. The cognitive memory dysfunction detection result includes two categories: normal or abnormal.

[0016] The memory model generation unit based on memory curves obtains the brain network of the current detection object from the brain network feature extraction unit, extracts the network topology of a specified brain region to obtain the initial network topology of the memory model, treats each specified brain region as a memory attribute node, and obtains the correlation coefficient between memory attribute nodes based on the Pearson correlation coefficient between brain regions; and divides each memory attribute node into two categories: level A and level B, with level A being more important than level B; then, based on the stimulus information of the memory attribute nodes input by the user, and according to the memory curves of the pre-set healthy population and the memory curves of the cognitive memory disorder population, it calculates the total correlation coefficient of each memory attribute node at different elapsed times. Based on the currently calculated total correlation coefficient, the network topology of the memory model is updated: if the value of the currently calculated total correlation coefficient is greater than 1, it is reset to 1, and memory attribute nodes with a total correlation coefficient lower than the fuzzy threshold are regarded as fuzzy nodes and labeled as fuzzy nodes, and level B fuzzy nodes are deleted from the current memory model network topology, thus obtaining the network topology of the memory model at the corresponding elapsed time and visualizing it.

[0017] The technical solution provided by this invention brings at least the following beneficial effects:

[0018] This invention preprocesses the acquired MRI data to construct a brain structural network and a brain functional network to extract specified brain network features for use by other units. It also extracts brain activity signals (such as ALFF, fALFF, and Reho signals) from abnormal brain regions (specific brain regions related to cognitive memory). Then, based on the extracted brain network features and brain region activity signal features of the subjects, the machine learning model for assisting in the diagnosis of cognitive memory disorders is trained, resulting in a trained machine learning model. This model is then used to assist in the diagnosis of cognitive memory disorders in the test subjects, by inputting the brain network features of the test subjects and the brain region activity signal features of the specified brain regions into the trained machine learning model, and obtaining the auxiliary diagnostic results based on its output. Simultaneously, this invention combines abnormal brain regions, network topology attributes, and attribute topology to propose a memory forgetting curve (memory curve) suitable for people with cognitive memory disorders and a memory model based on attribute topology for people with cognitive memory disorders. Finally, it provides a visualized simulation of the memory forgetting process in this population. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the structure of the cognitive memory dysfunction auxiliary diagnostic device based on the memory curve provided in an embodiment of the present invention.

[0021] Figure 2 This is a topological diagram of the graded "biological and water" attributes in an embodiment of the present invention.

[0022] Figure 3 In a specific embodiment of the present invention, the AD forgetting curve and the HC forgetting curve are compared.

[0023] Figure 4 This is a topological diagram of the attributes of "organisms and water" two hours later, as shown in a specific embodiment of the present invention. (4-a) is the attribute topology diagram of group AD; (4-b) is the attribute topology diagram of group HC. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0025] Most memory models provide very limited descriptions of information and are not applicable to people with impaired cognitive memory function. Furthermore, there is a lack of research and diagnosis of brain networks for cognitive memory disorders, which relies on doctors for diagnosis, which is highly subjective and not conducive to disease detection. In view of this, this invention proposes an auxiliary diagnostic device for cognitive memory dysfunction based on memory curves.

[0026] As one possible implementation, the memory curve-based auxiliary diagnostic device for cognitive memory dysfunction provided in this embodiment of the invention includes: a data preprocessing unit, a brain network feature acquisition unit, an activity signal feature extraction unit for a specified brain region, a memory curve-based memory model generation unit, and a cognitive memory dysfunction detection unit, such as... Figure 1 As shown.

[0027] The data preprocessing unit is used to preprocess the input MRI data (including functional and structural images), and then send the preprocessed data to the brain network feature acquisition unit and the activity signal feature extraction unit for a specified brain region, respectively. The brain network feature acquisition unit is used to acquire the time series of the preprocessed functional MRI data for each brain region based on a preset brain region template (e.g., ALL116), calculate the Pearson correlation coefficient between any two brain regions, and determine connectivity between the two brain regions if the correlation coefficient is greater than a specified threshold. If the brain regions are not connected, a brain functional network (also known as a brain-inspired functional network) is constructed and saved. Alternatively, structural MRI data of each brain region (extracting gray matter images) can be obtained. Based on the Pearson correlation coefficient between the gray matter images of brain regions, if the correlation coefficient is greater than a specified threshold, the two brain regions are connected; otherwise, they are not connected. A brain structural network (also known as a brain-inspired decoupling network) is constructed and saved. Then, based on specified network attribute indicators (such as network topology attribute indicators, the upper triangular matrix of the adjacency matrix of the brain network, etc.), brain network features of the brain functional network and brain structural network are extracted. The activity signal feature extraction unit is used to extract activity signal features of specified brain regions related to cognitive memory, such as ALFF signals, fALFF signals, and Reho signals. The cognitive memory dysfunction detection unit has a pre-trained cognitive memory dysfunction detection model. It uses the brain network features output by the brain network feature acquisition unit and the activity signal features output by the activity signal feature extraction unit of the specified brain region as input to the cognitive memory dysfunction detection model. Based on its output, it obtains the cognitive memory dysfunction detection result for the current test subject and sends the result to the memory model generation unit based on the memory curve. The cognitive memory dysfunction detection result is a binary classification result: normal or abnormal. The memory model generation unit based on the memory curve obtains the brain network (brain functional network or brain structural network) of the current test subject from the brain network feature extraction unit, extracts the network topology of specified brain regions to obtain the initial network topology of the memory model, treats each specified brain region as a memory attribute node, and obtains the correlation coefficient between memory attribute nodes based on the Pearson correlation coefficient between brain regions. It also classifies each memory attribute node into two categories: Grade A and Grade B, with Grade A being more important than Grade B.Based on the stimulus information of the memory attribute nodes input by the user, and according to the preset memory curves of healthy individuals and cognitive memory impairment individuals, the total correlation coefficient of each memory attribute node at different elapsed times is calculated. Based on the currently calculated total correlation coefficient, the network topology of the memory model is updated: if the currently calculated total correlation coefficient is greater than 1, it is reset to 1, and B-level memory attribute nodes with total correlation coefficients below the fuzzy threshold are regarded as forgotten nodes and deleted from the current memory model's network topology. The network topology of the memory model at the corresponding elapsed time is obtained and visualized.

[0028] In this embodiment of the invention, MRI data from Alzheimer's patients and healthy controls were collected from an open-source database. After preprocessing, brain-inspired structural and functional networks were constructed for the AD group (patient group) and the HC group (healthy group). Graph theory was used to perform topological analysis on the two networks, and brain activity signals (ALFF signal (low-frequency amplitude signal), fALFF signal (low-frequency fractional amplitude, i.e., ALFF divided by the root mean square of the power spectrum over the entire frequency range), and Reho signal (local consistency signal)) were extracted. By comparison, abnormal brain regions in individuals with cognitive memory impairment were identified. By combining abnormal brain regions, network topological attributes, and attribute topology, a memory forgetting curve suitable for individuals with cognitive memory impairment and a memory model based on attribute topology for individuals with cognitive memory impairment were proposed to demonstrate the process of attribute-based memory forgetting, showing that attribute topology can intuitively simulate memory forgetting. From another perspective, the memory model of people with cognitive memory disorders is visualized more intuitively; based on the selected feature information, the corresponding feature information of the training data is extracted to train a machine learning model for assisting in the diagnosis of cognitive memory disorders, providing a theoretical basis for the early diagnosis of the disease and realizing the auxiliary diagnosis of the disease.

[0029] Data preprocessing is used to preprocess the acquired functional magnetic resonance imaging (fMRI) data, including: removal of interfering data (removing the first few time points of the data), time correction, head motion correction, orientation adjustment, image segmentation, voxel registration, spatial normalization, smoothing, linear drift removal, and noise filtering. Specifically, it removes several time points at the beginning of the functional MRI scan, arranges the corresponding scan layers in chronological order based on the machine's scanning sequence, performs head motion correction using a preset method, image segmentation and voxel registration establish the spatial correspondence between functional and structural images (to facilitate spatial normalization of functional images), and performs spatial normalization (to increase comparability, transforming all subjects' image data from the original space to a unified standard space), spatial smoothing, linear drift removal, and filtering (to preserve the BOLD (brain oxygenation level dependent) signal), thus obtaining the preprocessed functional MRI image sequence.

[0030] In this embodiment, the data preprocessing is described in detail as follows:

[0031] (1) Remove the first few time points of the functional nuclear magnetic resonance images: Since the scanning equipment has just started, the unstable magnetic field will cause the early data to be distorted. In order to ensure the stability and authenticity of the data, in this embodiment, the first 10 time points are removed.

[0032] (2) Slicing Time: Continuous scanning of adjacent layers or short intervals between scans will cause distortion of the scanning results of the MRI equipment. In order to eliminate this effect, staff generally do not use layer-by-layer scanning but scan every other layer. However, this also has certain drawbacks: the scanning time of different layers is quite different. If the data is not at the same time, the influence of time factors can only be eliminated as much as possible through mathematical fitting and other technical means.

[0033] (3) Head Movement Realignment: Subjects need to remain still in the instrument for a period of time. Even slight head movements during this process can cause image ghosting, which is very detrimental to researchers' observation of image data. Therefore, to solve this problem, devices are usually used to forcibly fix the subject's head. However, in many cases, the subject's head movements are unconscious, so there is still a chance of head movement. Researchers need to process the data for head movement. After all the data is processed, based on the average head movement data of the AD and HC groups, data with a large head movement influence (head movement data greater than a specified value) can be labeled as having a large head movement influence and manually removed, further improving the accuracy of the experiment.

[0034] (4) Reorientation: Adjust the image orientation to make the image data easier to observe and use. Resetting the image center improves the usability of the image. It can effectively improve the orientation problem for data with incorrect initial orientation. At the same time, the image can be scored, and unqualified images will be eliminated in subsequent simulation experiments.

[0035] (5) Image segmentation: Image segmentation is performed using the subject's structural image data, and then spatial standardization is performed using the corresponding transformation matrix.

[0036] (6) Voxel registration: Matching voxels in the subject's functional image and structural image to make the functional image clearer.

[0037] (7) Spatial Normalization: Due to the inconsistent head shapes of different subjects, the brain data collected are also different. To overcome this problem, a standardized template is established for the segmented gray matter, white matter, cerebrospinal fluid, and other parts of the data. In this way, all image data can be unified into the same specific space: the MNI (Montreal Neurological Institute) template space, which is beneficial to subsequent data processing.

[0038] (8) Smoothing: The equipment used to acquire MRI image data has internal noise, which more or less interferes with the scanning accuracy and makes the image appear noisy in certain brain regions. In order to reduce noise interference, noise reduction algorithms are used to improve the signal-to-noise ratio of the image data, such as Gaussian noise matrix.

[0039] (9) Detrend: During head scanning, the device will experience temperature rise, which will cause data to drift linearly. Appropriate algorithms are needed to counteract the resulting linear drift.

[0040] (10) Filtering: During data acquisition, noise such as the subject's heartbeat, breathing, and eyelid twitching, as well as some low-frequency and high-frequency noise, can interfere with the accuracy of the experiment. To remove these interferences, time filtering can be used to filter out meaningless frequency bands. The inverse Fourier transform is used to sum the frequencies, and then the noise is filtered out by the filter. In this embodiment, by removing low-frequency and high-frequency data, only the data in the 0.01Hz to 0.1Hz frequency band is left.

[0041] In the research process of this invention, the constructed brain network includes a brain functional network and a brain structural network. The acquired magnetic resonance structural images are preprocessed (the images are segmented according to different tissue structures, resulting in gray matter images). Then, the gray matter images are resized to match the resolution of the brain region partitioning template (e.g., the ALL116 template). Based on the set brain region partitioning template, gray matter image blocks of each brain region are obtained. The Pearson correlation coefficient between each brain region is calculated based on the gray matter image blocks. When the correlation coefficient is greater than a specified threshold, it is considered that there is an edge between the two current brain regions, thereby constructing a brain structural network (also known as a brain-inspired structural network).

[0042] In order to extract effective feature information from MRI data that can be effectively used to identify cognitive and memory dysfunction, this invention conducts a specific analysis of brain networks (brain structural networks and brain functional networks) based on graph theory and topological network knowledge during the research process.

[0043] In this invention, the metrics used to measure network performance when performing topological network analysis on brain networks include: characteristic path length, clustering coefficient, optimal community structure, degree centrality, betweenness centrality, global efficiency, local efficiency, isomatch coefficient, small-world property, degree distribution, enriched community coefficient, and participation coefficient. These are explained in detail below:

[0044] (1) Characteristic path length: the average of the shortest paths of all connected node pairs in the network;

[0045] (2) Clustering coefficient: It can describe the phenomenon of nodes "clustering", which is equivalent to the ratio of the number of interconnected node pairs among the nodes adjacent to the node to the total number of node pairs.

[0046] (3) Optimal community structure: The network can be divided into multiple modules. The modularity attribute can represent the degree of module division and can determine the degree of network aggregation. The community structure metric is: calculate the sum of all edge weights of each module, then divide it by the sum of all edge weights in the network to get the metric of each module. Accumulate the difference between the metric of each module and the score of the sum of all edge weights of that module, and then divide it by the sum of all edge weights of all modules.

[0047] (4) Degree Centrality: The degree is defined as the number of other nodes a given node is connected to. In a directed graph, degree can be further divided into out-degree and in-degree. The higher the degree centrality of a given node, the more important the node is in the network. It is used to characterize the degree to which a node is connected to all other nodes. This includes proximity centrality, which reflects the proximity of a node to other nodes in the network; betweenness centrality, which is an indicator of node importance based on the number of shortest paths passing through a node; and eigenvector centrality, which determines the importance of a node based on both the number of its neighbors (i.e., the degree of the node) and the importance of its neighbors.

[0048] (5) Betweenness centrality: This metric measures the connectivity between different nodes connected to a given node.

[0049] (6) Global efficiency: The global efficiency of a network is defined as the inverse of the average characteristic path length of the network, and the network’s information transmission capability depends on this value.

[0050] (7) Local efficiency: There are many relatively independent subnetworks in the network. After calculating the global efficiency of all subnetworks, we can calculate their average value to obtain the local efficiency.

[0051] (8) Homomatch coefficient: A node with a higher degree is more likely to connect to other nodes with a higher degree, which is called homomatch; otherwise, it is called dismatch.

[0052] (9) Small-world property: Calculate the ratio of the clustering coefficient of the empirical network to that of the random network, and calculate the ratio of the average path length of the empirical network to that of the random network. Then, use the ratio of the clustering coefficient to the average path length as a metric. If the metric is greater than 1, it indicates that the network has the characteristics of a small-world network.

[0053] (10) Degree distribution: Count the degree of each node and calculate the proportion of nodes with the same degree to the total number of nodes.

[0054] (11) Enriched community coefficient: This coefficient indicates that the connections between nodes with high degree in the network are more compact than the connections between other nodes. The members of the enriched community can be found by using the k-degree kernel method to find nodes with high degree. This coefficient can be calculated by dividing the number of edges between nodes with degree k by the total number of possible edges between these nodes. The larger the value, the more compact the connections between the members of the enriched community.

[0055] (12) Participation coefficient: The participation coefficient refers to the ratio of the degree of a given node in the local module to its degree in the entire network. The larger the participation coefficient, the greater the role and the more connections the node has in this local module.

[0056] To identify brain regions associated with cognitive memory function, this invention utilizes DPABI software (a toolkit for data processing and analysis in brain imaging) to collect ALFF, fALFF, and Reho signals. Fisher-Z transformation is then applied to these signals to identify brain regions with significant signal intensity differences between the experimental and control groups. These regions are identified as those involved in cognitive memory function: hippocampus, prefrontal cortex, diencephalon, striatum, cerebellum, neocortex, amygdala, and reflex pathways—a total of eight brain regions. These eight regions can then be abstracted into attribute nodes (attribute node 1-attribute node 8), and memory attributes can be categorized (level A and level B attributes). Attribute topology is then used to simulate memory-related processes. To simulate memory loss, this invention also combines the Ebbinghaus forgetting curve with brain-inspired functional and structural networks.

[0057] In this invention, the specific analysis of brain networks is as follows:

[0058] Based on the range of connectivity density values ​​obtained from the experiment: the minimum connectivity density parameter is 0.12354 and the maximum density is 0.48. The topological properties of the brain-inspired structural networks of the AD group and the HC group at various densities (connectivity densities) were compared with a step size of 0.02.

[0059] Without the permutation test, the AD group had a density between 0.14 and 0.36, and the characteristic path length of the structural network was higher than that of the normal healthy subjects. After the permutation test, although there were differences between the two groups, they were not statistically significant.

[0060] Experiments revealed that the transitivity of the brain-inspired network in the AD group was higher than that in the HC group, exceeding the confidence interval within the connectivity density range of 0.2–0.24, indicating a significant difference between the AD and HC groups within this density range. Modular analysis of the brain-inspired network showed that within a density range of 0.12354–0.26, the HC group's modularity was higher than the AD group's, but within a density range of 0.26–0.48, the AD group's modularity was higher than the HC group's. Therefore, within a density range of 0.12354–0.36, excluding the first outlier, the average node betweenness number of the AD group's brain-inspired network was generally higher than that of the HC group, while they were almost the same within the 0.36–0.48 density range. Within a density range of 0.12354–0.18, the average edge betweenness number of the HC group's brain-inspired network was generally higher than that of the AD group. Within a density range of 0.2–0.36, the average edge betweenness number of the AD group's brain-inspired network was generally higher than that of the HC group, while they were almost the same within the 0.36–0.48 density range.

[0061] Experiments revealed that within a density range of 0.18–0.36, the average node betweenness and average edge betweenness of the AD group were slightly higher than those of the HC group. This indirectly reflects that the control ability of nodes in the brain-inspired structural network of the AD group is higher than that of nodes in the HC group. Furthermore, the interaction between nodes in the brain-inspired structural network of the AD group and other nodes within the module is also higher than that of the HC group.

[0062] Experiments using small-world properties of brain-inspired network structures revealed that within a density range of 0.15–0.44, the Lambda value in the AD group was higher than that in the HC group. However, overall, the Sigma and Gamma values ​​in the HC group were higher than those in the AD group. Under permutation testing, the Sigma value exceeded the confidence interval within a density range of 0.22–0.24, indicating a significant difference in Sigma values ​​between the AD and HC groups within this range. Here, Lambda represents the feature path length in small-world properties, Gamma represents the clustering coefficient, and Sigma represents the small-world properties considering both feature path length and clustering coefficient. Comparing the Sigma values ​​of the two experimental groups, it was found that the small-world properties of the brain-inspired network in the AD group showed a significant degradation. These values ​​indicate that the average feature path length of the brain-inspired network in the AD group was longer than that in the HC group, fully demonstrating the severe damage that cognitive memory dysfunction inflicts on brain structure.

[0063] In the clustering coefficient experiment, it was found that the clustering coefficient of the AD group's brain-inspired structural network was higher than that of the HC group overall. However, under the permutation test, there was no significant difference in the clustering coefficient. The higher clustering coefficient of the AD group indicates that the "clustering phenomenon" among the nodes of the AD group's brain-inspired structural network is more severe. This suggests that the transmission efficiency of the AD group's brain-inspired structural network nodes may be higher in local subnetworks, but from a global perspective, communication between different subnetworks is more difficult, meaning that the global network efficiency may be lower.

[0064] Experiments on global and local network efficiency revealed that the AD group generally had higher local efficiency than the HC group. Within the density range of 0.36–0.48, the difference in local efficiency between the AD and HC groups after permutation tests was significant. The HC group had slightly higher global efficiency than the AD group, but the overall difference was not significant. The comparison of local and global network efficiency is consistent with the analysis from the clustering coefficient experiment: the AD group's higher clustering coefficient and longer characteristic path length led to higher local network efficiency, but the reduced transmission efficiency between different sub-networks resulted in lower global network efficiency.

[0065] The preprocessed functional image data was imported into Gretna, and the Gretna software was used to construct and analyze the brain-inspired functional network. After constructing the brain-inspired functional networks for healthy subjects and subjects with cognitive and memory dysfunction, experiments on hierarchical attributes revealed that, excluding the first two outliers, within a connectivity density range of 0.06–0.46, the hierarchical attributes of the AD group's brain-inspired functional network were significantly lower than those of the HC group. The lower the hierarchical attributes, the more unstable and less maintainable the functional network.

[0066] Synchronization experiments revealed that within a density range of 0.04–0.1, the synchronicity of the brain-inspired functional network in the AD group was almost identical to that in the HC group. However, within a density range of 0.1–0.46, the synchronicity of the brain-inspired functional network in the AD group was higher than that in the HC group.

[0067] For brain-inspired functional networks, the relevant attributes of nodes are also crucial. Experiments have shown that, given a node, the average shortest path length from other nodes to that node in the AD group brain-inspired functional network is generally longer than that in the HC group. However, the node clustering coefficients of the AD group brain-inspired functional network and the HC group are similar, with little overall difference.

[0068] Experiments on the small-world properties of the network revealed that, in the brain-inspired functional network, the Cp (clustering coefficient) value was higher in the HC group than in the AD group within a density range of 0.04–0.16, while in the AD group it was higher than in the HC group within a density range of 0.16–0.46. For the Gamma and Lp (shortest path length) values ​​of the small-world properties, the AD group was generally slightly higher than the HC group. For the Sigma value of the small-world properties, within a density range of 0.04–0.18, after excluding the first outlier, the HC group was higher than the AD group; within a density range of 0.18–0.46, the AD group was higher than the HC group.

[0069] Efficiency is one of the core indicators for evaluating networks. For brain-inspired functional networks, network efficiency tests revealed that the global efficiency of the AD group's brain-inspired functional network was slightly higher than that of the HC group. Within the density range of 0.04 to 0.26, the local efficiency of the HC group's brain-inspired functional network was higher than that of the AD group. Within the density range of 0.26 to 0.46, the local efficiency of the AD group's brain-inspired functional network was the same as that of the HC group overall.

[0070] Based on the above analysis, the feature information extracted by this invention for the identification of cognitive memory dysfunction includes: gray matter volume (reflecting feature information related to brain structural networks), ALFF signals, fALFF signals, and Reho signals of brain regions related to cognitive memory dysfunction (hippocampus, prefrontal cortex, diencephalon, striatum, cerebellum, neocortex, amygdala, and reflex pathways), as well as the topological properties of brain functional networks and the adjacency matrix of brain functional networks.

[0071] Attribute topology graphs, as a way to express formal background, can intuitively show the attributes an object possesses and the relationships between them. Using attribute topology graphs, this invention employs graph theory and other related methods to conduct theoretical research on the topology graphs and analyze the intrinsic relationships between different attributes. To describe attributes more accurately, this invention classifies attributes into levels A and B, where level A attributes are important attributes of the object, and level B attributes are secondary attributes. Classifying attributes also reflects reality; for a complete object, the attributes it contains are often not all equal. Humans are sophisticated beings, and different attributes have different stimulating effects on different brains; even the same attribute can have different stimulating effects on different brains. Research has found that important attributes are more likely to stimulate brain memory, and the forgetting time is longer; some attribute features can even be permanently remembered. However, in the memory model based on attribute topology, the degree of attribute memory becomes blurred over time, specifically manifested in attribute nodes m. i The overall correlation coefficient gradually decreases from 1 until it reaches the fuzzy threshold, eventually returning to zero. Once the memory level of a Class B attribute drops from 100% to the fuzzy threshold, that Class B attribute, along with its associated edges, is removed from the attribute topology graph—a process known as forgetting. Class A attributes, however, are not removed from the original attribute topology graph even if they reach the fuzzy threshold. The practical basis for this principle is that certain attributes of an object that have been memorized (Class B attributes) will be forgotten after a period of time without further stimulation, but important specific attributes of that object (Class A attributes) are difficult to forget.

[0072] In this embodiment of the invention, the memory forgetting model for individuals with cognitive memory dysfunction based on attribute topology is specifically as follows:

[0073] Step (1): Classify the memory attributes.

[0074] In this embodiment, the grading of memory attributes is described using "biological and water" as an example.

[0075] In the section on "Biodiversity and Water," eight different types of plants and animals are listed, requiring specific elements. These eight organisms are: leeches, giant salamanders, frogs, dogs, aquatic plants, reeds, beans, and corn, numbered 1, 23, 4, 5, 6, 7, and 8 respectively. Memory attributes a through h represent: living in water, living on land, possessing chlorophyll, dicotyledonous, monocotyledonous, mobile, having four limbs, and mammalian, respectively. Memory attributes a and b are designated as Level A attributes, while attributes c, d, e, f, g, and h are designated as Level B attributes. Figure 2 As shown,

[0076] Label the A-level memory attributes and the edges connected to the A-level memory attribute nodes. At this time, the memory percentage of each memory attribute is set to 2. Figure 2In the brackets, the numbers in the curly braces {} of each memory attribute represent the element numbers included in the current memory attribute, and the numbers in the curly braces {} of the lines connecting memory attributes represent the common element numbers of the two memory attributes.

[0077] Step (2): Set the forgetting curve, attribute fuzziness threshold and attribute forgetting threshold.

[0078] In this embodiment, the memory activation and forgetting processes of healthy individuals and individuals with cognitive memory impairment are simulated. Therefore, it is necessary to set memory forgetting curves for healthy individuals and for individuals with cognitive memory impairment. Since efficiency is one of the core indicators for evaluating the human brain, previous experiments with brain-inspired functional networks and brain-inspired structural networks have shown that the global efficiency of the brain-inspired functional network for individuals with cognitive memory impairment is 80% of that for healthy individuals. Based on this, the forgetting curve formula for individuals with cognitive memory impairment is set as y = 0.8 - 0.44x. 0.06 The forgetting curve formula for healthy individuals is y = 1 - 0.56x. 0.06 ,like Figure 3 As shown, y represents the percentage of memory retention, x represents time (hours), and the fuzziness threshold is set to 25% and the forgetting threshold is set to 10%.

[0079] Step (3): According to the set forgetting curve, fuzzy threshold and forgetting threshold, the percentage of memory attributes in "biological and water" will decrease over time and be subjected to new object stimulation as shown in Table 1 after one hour.

[0080] Table 1. Formal Background of the New Object

[0081]

[0082] In the initial stage, all memory attributes are at 100%, that is...

[0083] θ a =θ b =θ c =θ d =θ e =θ f =θ g =θ h =1

[0084] According to the formula, the memory attributes a, c, and d were directly stimulated, i.e., θ aa =θ cc =θ dd =1. The remaining memory attributes are indirectly stimulated, among which memory attribute a is connected to g, f, e, b, and c.

[0085] Based on the correlation coefficient between any two attribute nodes (i, j) Where E(i,j) represents the number of elements shared by attributes i and j, and g(i) represents the number of elements possessed by attribute i, the following association coefficient values ​​can be obtained:

[0086]

[0087]

[0088] The memory attribute c is linked to attributes a, b, e, and d, and their corresponding association coefficients are as follows:

[0089]

[0090]

[0091] The memory attribute d is linked to attributes b and c, and their corresponding correlation coefficients are as follows:

[0092]

[0093] According to the established forgetting curve formula, the percentage of each memory attribute decreased to 0.352. Without stimulation, all memory attributes are at the edge of the fuzziness threshold. After restimulation, the total correlation coefficients of memory attributes a, b, c, d, e, f, g, and h are respectively:

[0094] θ a =θ aa +θ ca =1.286,θ b =θ ab +θ cb +θ db =0.95;

[0095] θ c =θ ac +θ cc +θ dc =1.536,θ d =θ dd +θ cd =1.25,θ e =θ ae +θ ce =1.083,θ f =θ af =0.5;

[0096] θ g =θ ag =0.333,θ h =0.

[0097] Add the remaining 0.352 to all memory attributes, and reset all percentages of total association coefficients greater than 1 to 1, i.e., θ.a =θ b =θ c =θ d =θ e =1,θ f =0.852,θ g =0.685,θ h =0.352. At this point, all memory attributes are still intact.

[0098] Step (4): After another hour, the total correlation coefficient of memory attributes a, b, c, d, e, f, g, h is θ. a =θ b =θ c =θ d =θ e =0.352,θ f =0.299,θ g =0.241,θ h = 0.123. At this point, the memory attribute θ g =0.241,θ h =0.123 is already below the fuzzy threshold of 0.25, and the attribute nodes and their associated edges should be removed from the attribute topology graph. For comparison, the memory percentages of each attribute in the healthy population are θ. a =θ b =θ c =θ d =θ e =0.44,θ f =0.41,θ g =0.34,θ h =0.193, such as Figure 4 As shown, (4-a) is the attribute topology diagram of group AD, and (4-b) is the attribute topology diagram of group HC.

[0099] Step (5): After 25 hours of the experiment, the total correlation coefficient of memory attributes a, b, c, d, e, f, g, and h is θ. a =θ b =θ c =θ d =θ e =0.258,θ f =0.219,θ g =0.177,θ h =0.091, indicating that memory attribute h is less than the forgetting threshold and memory attribute f is less than the fuzziness threshold. For comparison, the percentage of memory for each attribute in the healthy population is θ. a =θ b =θ c =θ d =θ e =0.322,θf =0.303,θ g =0.249,θ h =0.141. It can be observed that at this point, memory attributes g and f in people with cognitive memory dysfunction have entered the fuzzy threshold and need to be removed from the network graph. Memory attribute h has decreased from the fuzzy threshold to the forgetting threshold. However, in healthy individuals, only memory attributes g and h have decreased to the fuzzy threshold.

[0100] Step (6): After 100 hours of the experiment, the total correlation coefficient of memory attributes a, b, c, d, e, f, g, and h is θ. a =θ b =θ c =θ d =θ e =0.198,θ f =0.178,θ g =0.144,θ h =0.074, indicating that all attribute nodes are below the fuzzy threshold. For comparison, the memory percentages of each attribute in the healthy population are θ. a =θ b =θ c =θ d =θ e =0.262,θ f =0.246,θ g =0.202,θ h =0.115. At this point, for individuals with cognitive memory impairment, all attributes except memory attribute h have decreased to the fuzzy threshold. Memory attribute h has decreased to the forgetting threshold, but since memory attributes a and b are Class A attributes, they will not be removed from the network topology. In healthy individuals, only memory attributes g, h, and f have decreased to the fuzzy threshold and will be removed from the network topology; the remaining memory attributes are all above the fuzzy threshold.

[0101] Based on extensive experimental verification, in this embodiment of the invention, memory attributes stored in the neocortex, amygdala, cerebellum, striatum, and reflex pathways are classified as Class A attribute nodes, and memory attributes stored in the hippocampus, prefrontal cortex, and diencephalon are classified as Class B attribute nodes. Then, by combining the topological attributes of the current object's brain functional network (or brain structural network), an initial network topology graph of the memory model can be obtained. That is, the edge connections between specified brain regions (neocortex, amygdala, cerebellum, striatum, reflex pathways, hippocampus, prefrontal cortex, and diencephalon) related to memory are extracted from the brain functional network (or brain structural network) to obtain the initial network topology graph of the memory model (undirected). The correlation coefficient between memory attribute nodes with edges can be set as the correlation Pearson correlation coefficient between brain regions. Then, based on the stimulus information of the input memory attribute nodes, the total correlation coefficient of each memory attribute node at different elapsed times can be calculated. Based on the currently calculated total correlation coefficient, the network topology of the memory model is updated: after each calculation, if the value of the currently calculated total correlation coefficient is greater than 1, it is reset to 1. Then, the B-level memory attribute nodes with a total correlation coefficient lower than the fuzzy threshold will be regarded as fuzzy nodes. The currently determined forgotten nodes and their edges are deleted from the network topology of the current memory model to obtain the network topology of the memory model at the corresponding time.

[0102] The total correlation coefficient of memory attribute nodes is calculated as follows:

[0103] The total correlation coefficient of the current memory attribute node after descent is calculated based on the forgetting curve (memory curve), and the autocorrelation coefficient of the direct stimulus node is set to 1.

[0104] If the current memory attribute node is a direct stimulus node, then its total correlation coefficient is: total correlation coefficient after decrease + autocorrelation correlation coefficient + correlation coefficient between the current memory attribute node and its indirect stimulus node;

[0105] If the current memory attribute node is an indirect stimulus node, then its total correlation coefficient is: total correlation coefficient after decrease + correlation coefficient between the current memory attribute node and its indirect stimulus node;

[0106] If the current memory attribute node is neither a direct stimulus node nor an indirect stimulus node, then the total correlation coefficient is directly the decreased total correlation coefficient.

[0107] Furthermore, based on the set forgetting threshold (where the forgetting threshold is less than the fuzzy threshold), memory attribute nodes whose total correlation coefficient reaches the forgetting threshold are regarded as forgotten nodes and are labeled accordingly.

[0108] In this embodiment of the invention, the models used for memory dysfunction detection include SVM, LightGBM, and LogisticRegression. For the original data and the dimensionality-reduced data, the model performance, from smallest to largest, is SVM, LightGBM, and then LogisticRegression. To further improve the performance of the LogisticRegression model, this embodiment uses an ensemble learning model, employing SVM and LightGBM as base classifiers and the LogisticRegression model as a meta-classifier. The trained base classifiers are used to classify and predict the test and training sets, and the output values ​​are used as input values ​​for the next stage to train the meta-classifier. The resulting meta-classifier exhibits better robustness than a single classifier. As shown in Tables 2 and 3, the Stacking ensemble learning model outperforms the SVM, LightGBM, and LogisticRegression models in various performance metrics: accuracy, precision, recall, F1-Score, and specificity.

[0109] Table 2 Performance metrics of various models before dimensionality reduction

[0110]

[0111] Table 3 Performance metrics of various models after dimensionality reduction

[0112]

[0113]

[0114] This invention preprocesses the collected MRI data to construct brain-inspired structural and functional networks for AD and healthy subjects. Graph theory is then used to perform topological analysis on both networks, extracting brain activity signals (ALFF, fALFF, and Reho signals). Comparison reveals abnormal brain regions in individuals with cognitive memory impairment. By combining abnormal brain regions, network topological attributes, and attribute topology, a memory forgetting curve (memory curve) suitable for individuals with cognitive memory impairment and a memory model based on attribute topology are proposed. The memory forgetting process of this population is then simulated. Based on the feature set obtained from the above experiments, a machine learning model for assisting in the diagnosis of cognitive memory impairment is trained, providing a solid theoretical foundation for the early diagnosis of this disease.

[0115] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0116] The above descriptions are merely some embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A diagnostic aid for cognitive memory dysfunction based on the forgetting curve, characterized in that, include: The system includes a data preprocessing unit, a brain network feature acquisition unit, a designated brain region activity signal feature extraction unit, a memory curve-based memory model generation unit, and a cognitive memory dysfunction detection unit. The data preprocessing unit is used to preprocess the input magnetic resonance imaging data and send the preprocessed data to the brain network feature acquisition unit and the activity signal feature extraction unit of the specified brain region, respectively. The brain network feature acquisition unit constructs a brain functional network, or a brain functional network and a brain structural network, based on the Pearson correlation coefficient between brain regions. The brain functional network is as follows: Based on a pre-set brain region template, the brain is divided into regions, and the time series of functional magnetic resonance imaging data of each brain region after data preprocessing is obtained. The Pearson correlation coefficient between the time series of any two brain regions is calculated. If the correlation coefficient is greater than a specified threshold, the two brain regions are connected; otherwise, they are not connected. The brain functional network is constructed and saved. The brain structure network is constructed and saved as follows: Based on a pre-set brain region template, the brain is divided into regions, gray matter images from the structural magnetic resonance imaging data of each brain region are obtained, the Pearson correlation coefficient between the gray matter images of any two brain regions is calculated, if the correlation coefficient is greater than a specified threshold, the two brain regions are connected, otherwise they are not connected. The brain network feature acquisition unit extracts network features of brain functional networks and brain structural networks based on specified network attribute indicators, obtains brain network features, and sends them to the cognitive memory dysfunction detection unit. The activity signal feature extraction unit for a specified brain region is used to extract activity signal features of a specified brain region related to cognitive memory and send them to the cognitive memory dysfunction detection unit. The cognitive memory dysfunction detection unit is pre-set with a trained cognitive memory dysfunction detection model. It uses brain network features and activity signal features as inputs to the cognitive memory dysfunction detection model, obtains the cognitive memory dysfunction detection result of the current detection object based on its output, and sends the cognitive memory dysfunction detection result to the memory model generation unit based on the memory curve. The cognitive memory dysfunction detection result includes two categories: normal or abnormal. The memory model generation unit based on memory curves obtains the brain network of the current detection object from the brain network feature extraction unit, extracts the network topology of a specified brain region to obtain the initial network topology of the memory model, treats each specified brain region as a memory attribute node, and obtains the correlation coefficient between memory attribute nodes based on the Pearson correlation coefficient between brain regions; and divides each memory attribute node into two categories: level A and level B, with level A being more important than level B; then, based on the stimulus information of the memory attribute nodes input by the user, and according to the memory curves of the pre-set healthy population and the memory curves of the cognitive memory disorder population, it calculates the total correlation coefficient of each memory attribute node at different elapsed times. Based on the currently calculated total correlation coefficient, the network topology of the memory model is updated: if the value of the currently calculated total correlation coefficient is greater than 1, it is reset to 1, and memory attribute nodes with a total correlation coefficient lower than the fuzzy threshold are regarded as fuzzy nodes and labeled as fuzzy nodes, and level B fuzzy nodes are deleted from the current memory model network topology, thus obtaining the network topology of the memory model at the corresponding elapsed time and visualizing it.

2. The auxiliary diagnostic device as described in claim 1, characterized in that, The designated brain regions include eight: hippocampus, prefrontal cortex, diencephalon, striatum, cerebellum, neocortex, amygdala, and reflex pathways.

3. The auxiliary diagnostic device as described in claim 2, characterized in that, The brain regions corresponding to Level A memory attribute nodes are: striatum, cerebellum, neocortex, amygdala, and reflex pathways; the brain regions corresponding to Level B memory attribute nodes are: hippocampus, prefrontal cortex, and diencephalon.

4. The auxiliary diagnostic device as described in claim 1, characterized in that, The specified network attribute metrics include: network topology attribute metrics and the upper triangular matrix of the adjacency matrix of brain functional networks.

5. The auxiliary diagnostic device as described in claim 3, characterized in that, Network topology attributes include: characteristic path length, clustering coefficient, optimal community structure, degree centrality, betweenness centrality, global efficiency, local efficiency, isomatch coefficient, small-world property, degree distribution, enriched community coefficient, and participation coefficient.

6. The auxiliary diagnostic device as described in claim 1, characterized in that, The memory model generation unit based on memory curves also includes: treating memory attribute nodes with a total correlation coefficient lower than the forgetting threshold as forgotten nodes and labeling them as forgotten nodes, wherein the forgetting threshold is less than the fuzzy threshold.

7. The auxiliary diagnostic device as described in claim 1, characterized in that, The memory curve for healthy individuals is set as follows: The memory curve for individuals with cognitive memory impairment is set as follows: Where x represents the elapsed time. , These represent the percentage of memory retention in healthy individuals and those with cognitive memory impairment, respectively.

8. The auxiliary diagnostic device as described in claim 1, characterized in that, The specific steps for calculating the total correlation coefficient of each memory attribute node at different elapsed times are as follows: If the cognitive memory impairment test result is normal, the total correlation coefficient of memory attribute nodes after the decrease is calculated based on the memory curve of healthy people; if the cognitive memory impairment test result is abnormal, the total correlation coefficient of memory attribute nodes after the decrease is calculated based on the memory curve of people with cognitive memory impairment. Based on the stimulus information of memory attribute nodes, the direct stimulus nodes and indirect stimulus nodes in the network topology of the current memory model are determined, and the autocorrelation coefficient of the direct stimulus nodes is set to 1; and the initial total correlation coefficient of each memory attribute node is set to 1. If the current memory attribute node is a direct stimulus node, then its current total correlation coefficient is the sum of the total correlation coefficient after the decrease, the autocorrelation correlation coefficient, and the correlation coefficient between the current memory attribute node and the indirect stimulus node with which it has an edge. If the current memory attribute node is an indirect stimulus node, then its current total correlation coefficient is the sum of the total correlation coefficient after the decrease and the correlation coefficient between the current memory attribute node and the indirect stimulus node with which it has an edge; If the current memory attribute node is neither a direct stimulus node nor an indirect stimulus node, then its current total correlation coefficient is directly the decreased total correlation coefficient.

9. The auxiliary diagnostic device according to any one of claims 1 to 8, characterized in that, The cognitive memory dysfunction detection model is a combination of a base classifier and a meta-classifier. The base classifier includes a support vector machine model and a LightGBM model; the meta-classifier adopts a logistic regression model; the trained base classifier is used to perform binary classification prediction of cognitive memory dysfunction detection on the test set and the training set, and the output value of the base classifier is used as the input value of the meta-classifier to train the meta-classifier.

Images