Biphasic affective disorder prediction system based on cognitive pathology association

Abstract

The invention discloses a dual-phase affective disorder prediction system based on cognitive pathology association, and relates to the technical field of biological signal processing, and the system comprises a data preprocessing module which is used for preprocessing data in a data set; the cognitive perception attention module is used for extracting brain region features related to the bipolar affective disorder from the preprocessed data; the cooperative multi-expert network module is used for simulating cooperative interaction between brain regions according to the output of the cognitive perception attention module to obtain brain region features including the cooperative interaction relationship between the brain regions; and the causal normalization classifier is used for carrying out classification processing on the output of the cooperative multi-expert network module and generating a dual-phase affective disorder prediction result of the to-be-tested person. According to the method, the prediction accuracy, sensitivity and interpretability of the bipolar affective disorder are remarkably improved, meanwhile, the calculation complexity is reduced, efficient prediction of the bipolar affective disorder is achieved, and important technical support is provided for early intervention of mental and psychological diseases.

Description

Technical Field

[0001] The present invention relates to the technical field of biological signal processing, and in particular to a bipolar disorder prediction system based on cognitive pathology association. Background Art

[0002] The early prediction of bipolar disorder (BD) faces many challenges due to the complex and hidden pathological characteristics. It is characterized by the alternation of manic and depressive states, but in the early stages, the symptoms are similar to the mood swings of healthy individuals, making clinical identification difficult. Traditional diagnostic methods based on clinical evaluation rely on the subjective judgment of doctors, lack objective basis, and the prediction results have high uncertainty. Structural magnetic resonance imaging (sMRI) has become an important tool for exploring the pathological characteristics of bipolar disorder with its high spatial resolution and stable image quality. In recent years, the combined use of sMRI and functional magnetic resonance imaging (fMRI) has revealed the complex pathological mechanisms of patients with bipolar disorder: fMRI captures dynamic functional activity of the brain, and sMRI provides static anatomical images, and the two complement each other.

[0003] However, the high cost and complex data processing of fMRI limit its clinical popularity. In contrast, sMRI is more suitable for large-scale screening and disease prediction due to its low cost, high stability and easy operation. Therefore, the development of automated prediction methods based on sMRI data can significantly improve the efficiency of early identification and management of bipolar disorder.

[0004] The occurrence of bipolar disorder is closely related to structural and functional abnormalities in specific brain regions. Existing prediction methods based on sMRI are divided into three categories according to the granularity of feature extraction: slice level, whole brain level, and brain region level. The slice level method decomposes three-dimensional sMRI images into two-dimensional slice sequences and extracts local features. This method can capture local structural changes, but due to the lack of global modeling capabilities, it is difficult to effectively identify complex pathological relationships across brain regions, which limits its application in pathological causal association analysis.

[0005] The whole-brain method takes the three-dimensional image of the entire brain as input, retains the global spatial information, and implements end-to-end predictive analysis based on the deep learning model. Although the whole-brain method provides a holistic perspective and facilitates the analysis of complex pathological changes, the pathological manifestations of different brain regions vary greatly, and the subtle pathological features of specific brain regions are easily masked by global information, resulting in limited prediction accuracy.

[0006] Brain region-level methods use brain maps (such as the AAL template) to standardize the brain and extract statistical features of each brain region, which can more accurately focus on key brain regions related to bipolar disorder, especially in pathological studies involving local brain region abnormalities. However, existing brain region-level methods rely on predefined statistical features and fail to fully mine the original information of sMRI data and its complex high-dimensional associations, limiting the in-depth analysis of pathological causal relationships between brain regions.

[0007] In summary, the existing technologies in the prediction of bipolar disorder generally fail to fully capture the pathological associations of specific brain regions, and it is difficult to establish an effective prediction model based on cognitive pathological mechanisms. Therefore, there is an urgent need for a bipolar disorder prediction method with brain region-level cognitive pathological associations as the core, which can improve the accuracy and reliability of mental and psychological disease prediction from the perspective of pathological causality. Summary of the invention

[0008] In view of the above-mentioned deficiencies in the prior art, the present invention provides a bipolar disorder prediction system based on cognitive pathology association to solve the problems of low accuracy and insufficient reliability of the prior art.

[0009] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is: a bipolar disorder prediction system based on cognitive pathology association, comprising: A data preprocessing module is used to preprocess the data in the data set; The cognitive perception attention module is used to extract brain region features related to bipolar disorder from the preprocessed data; The collaborative multi-expert network module is used to simulate the collaborative interaction between brain regions according to the output of the cognitive perception attention module, and obtain the brain region characteristics including the collaborative interaction relationship between brain regions; The causal normalization classifier is used to classify the output of the collaborative multi-expert network module and generate a prediction result of bipolar disorder for the person to be tested.

[0010] Furthermore, the data in the dataset includes T1-weighted MRI data and AAL template.

[0011] Furthermore, the specific method of preprocessing is: using the N4 algorithm to correct the intensity inhomogeneity of the T1-weighted MRI data; using the HD-BET network to perform skull stripping on the corrected T1-weighted MRI data; aligning the skull-stripped T1-weighted MRI data to the standard MNI space and adjusting it to a uniform size; performing zero-mean unit variance normalization on the resized T1-weighted MRI data to obtain the preprocessed T1-weighted MRI data; directly mapping the AAL template to the MNI space and aligning it according to the brain region annotations so that it is aligned with the preprocessed T1-weighted MRI data in the anatomical space; analyzing the cerebellum region in the AAL template, merging the brain regions with a volume less than 1000 into the adjacent brain regions to obtain the preprocessed AAL template; the preprocessed T1-weighted MRI data and the preprocessed AAL template are both in a unified MNI space; the preprocessed T1-weighted MRI data and the preprocessed AAL template are both preprocessed data.

[0012] Furthermore, the cognitive perception attention module includes a three-dimensional convolutional encoding unit, a brain region feature extraction unit, and a brain region self-attention weighting unit which are connected in sequence.

[0013] Furthermore, the specific method of the cognitive perception attention module extracting brain region features related to bipolar disorder from the preprocessed data includes the following steps: A1. The whole brain features in the preprocessed T1-weighted MRI data are extracted through a three-dimensional convolutional coding unit, and the expression is:

[0014] in, is the preprocessed T1-weighted MRI data, Represents the three-dimensional spatial structure feature extraction process, Indicates whole-brain characteristics; A2. Extract local features of each brain region through the brain region feature extraction unit combined with the preprocessed AAL template and the whole brain features , whose expression is:

[0015] in, is the preprocessed AAL template; represents feature extraction processing; A3: The local features of each brain region are analyzed by combining the brain region self-attention weighted unit with the preprocessed AAL template. Attention weighting was performed to obtain brain region features related to bipolar disorder .

[0016] Furthermore, the specific method of feature extraction processing in step A2 includes the following sub-steps: A2-1. Three-dimensional rectangular frame of the brain area As the cropping range, for the whole brain features Perform cropping to obtain the characteristic blocks of each brain region; The expression is:

[0017] in, Indicates i The specific boundaries of each brain region in three-dimensional space, Indicates that the brain area x The minimum and maximum values ​​on the axis, Indicates that the brain area y The minimum and maximum values ​​on the axis, Indicates that the brain area z Minimum and maximum values ​​on the axis; N is the total number of brain regions; A2-2, extracting effective regional features of each brain region from the feature blocks of each brain region using the brain region mask; A2-3. Adaptively average pool the effective regional features of each brain region to obtain local features of uniform size .

[0018] Furthermore, the collaborative multi-expert network module includes an expert hybrid encoder, which includes multiple expert networks, and the multiple expert networks are used to process brain areas related to different cognitive functions.

[0019] Furthermore, the collaborative multi-expert network module simulates the collaborative interaction between brain regions according to the output of the cognitive perception attention module, and the specific method for obtaining brain region features including the collaborative interaction relationship between brain regions includes the following steps: B1. Perform linear transformation on the input brain region features related to bipolar disorder to obtain features with unified dimensions , whose expression is:

[0020] in, is the weight matrix, is the bias term, Indicates j brain regions associated with bipolar disorder M is the total number of brain region features associated with bipolar disorder; B2. Interact the unified dimension features with the CLS token representing global information to form a sequence , whose expression is:

[0021] in, Represents global information, Indicates K The features obtained by interacting with the CLS token after the unified dimension; B3. Input the sequence obtained in step B2 into the expert hybrid encoder for optimization, and obtain brain region features including the collaborative interaction relationship between brain regions after optimization. , whose expression is:

[0022] in, Indicates optimization processing.

[0023] Furthermore, the optimization process in step B3 is specifically as follows: After being input into the expert hybrid encoder, the gating mechanism in the expert hybrid encoder is based on the sequence The fit between the features in the network and the expert network dynamically selects the most relevant P Expert network, through P Expert Network Based on Sequence The features in the model are used to model the pathological correlation characteristics of different brain regions in specific cognitive functions; the results obtained by each expert network are weighted and aggregated to obtain the optimized brain region features including the collaborative interaction relationship between brain regions. .

[0024] Furthermore, the system is a system obtained after training, and a loss function FocalLoss is introduced during the training process, and its expression is:

[0025] in, is the category weight factor, Indicates that the prediction is a category The probability of is the focus parameter; represents the loss value, is the natural logarithm; The parameters involved in this system are updated in a cyclic iterative manner, and the AdamW optimizer is used to optimize this system until the loss value no longer decreases, and the training is terminated.

[0026] The beneficial effects of the present invention are: 1. The cognitive perception attention module involved in the present invention performs three-dimensional spatial structure feature extraction, brain region spatial structure extraction and brain region attention weighting in sequence, and combines the brain region attention mechanism to dynamically strengthen the feature expression of disease-related brain regions, thereby significantly enhancing the system's ability to capture features associated with cognitive pathology and improving the prediction sensitivity and reliability of bipolar disorder.

[0027] 2. Each expert network in the collaborative multi-expert network module focuses on a subset of brain regions related to a specific cognitive task, simulates the collaborative process of brain regions under heavy cognitive tasks, deeply explores the pathological changes related to specific cognitive tasks, and further improves the accuracy of prediction.

[0028] 3. The system proposed in the present invention accurately captures the pathological associations of bipolar disorder by combining brain region-level structural imaging features with analysis of cognitive functions, significantly improving the efficiency of the system and reducing computational complexity. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a schematic diagram of the structure of the system proposed by the present invention. DETAILED DESCRIPTION

[0030] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0031] like Figure 1 As shown, a bipolar disorder prediction system based on cognitive pathology association includes: The data preprocessing module is used to preprocess the data in the data set. The data in the data set includes T1-weighted MRI data and AAL template. The N4 algorithm was used to correct the intensity inhomogeneity of T1-weighted MRI data. The HD-BET network was used to perform skull stripping on the corrected T1-weighted MRI data. The skull-stripped T1-weighted MRI data were registered to the standard MNI space and resized to a uniform size (128×128×128). The resized T1-weighted MRI data were normalized to zero mean and unit variance to obtain the preprocessed T1-weighted MRI data. The AAL template was directly mapped to the MNI space and registered according to the brain region annotations to align it with the preprocessed T1-weighted MRI data in the anatomical space, omitting the intensity correction and stripping steps. The cerebellum in the AAL template was analyzed, and the brain regions with a volume less than 1000 were merged into the adjacent brain regions to enhance the information expression and analysis stability of the cerebellum, and finally the preprocessed AAL template was obtained. The preprocessed T1-weighted MRI data and the preprocessed AAL template were both in a unified MNI space. The preprocessed T1-weighted MRI data and the preprocessed AAL template were both preprocessed data.

[0032] The cognitive perception attention module (CAAM) includes a three-dimensional convolutional encoding unit, a brain region feature extraction unit, and a brain region self-attention weighting unit connected in sequence, which is used to extract brain region features related to bipolar disorder from the preprocessed data, and specifically includes the following steps: A1. The whole brain features in the preprocessed T1-weighted MRI data are extracted by a three-dimensional convolutional coding unit. The three-dimensional convolutional coding unit includes a three-dimensional convolutional neural network, which consists of two convolutional layers. The size of each convolutional kernel is 3×3×3. First, the dimension is increased to and then restore back By expanding the receptive field layer by layer, it is possible to capture spatial information in a wider area. This design provides effective support for subsequent learning within a specific brain region, ensuring that a wider range of spatial information can be combined when focusing on specific features of the brain region, thereby improving the recognition accuracy of pathological changes in bipolar disorder. The entire three-dimensional spatial structure feature extraction process is expressed as:

[0033] in, is the preprocessed T1-weighted MRI data, represents the three-dimensional spatial structure feature extraction processing performed by the three-dimensional convolutional neural network, Indicates whole-brain characteristics; A2. Extract local features of each brain region through the brain region feature extraction unit combined with the preprocessed AAL template and the whole brain features , whose expression is:

[0034] in, is the preprocessed AAL template; represents feature extraction processing; The specific method of the feature extraction process includes the following sub-steps: A2-1. Three-dimensional rectangular frame of the brain area As the cropping range, for the whole brain features Perform cropping to obtain the characteristic blocks of each brain region; The expression is:

[0035] in, Indicates i The specific boundaries of each brain region in three-dimensional space, Indicates that the brain area x The minimum and maximum values ​​on the axis, Indicates that the brain area y The minimum and maximum values ​​on the axis, Indicates that the brain area z Minimum and maximum values ​​on the axis; N is the total number of brain regions; A2-2. Using the brain region mask for each brain region’s feature block Extract the effective regional features of each brain region; The position and shape of the brain region in three-dimensional space are defined, and the specific spatial area of ​​the brain region can be identified. Through masking operations, the cognitive perception attention module (CAAM) can ensure that only structural information related to the brain region is extracted, thereby avoiding interference from other brain regions.

[0036] A2-3. Adaptively average pool the effective regional features of each brain region so that the features of all brain regions are integrated into regional features of uniform size to ensure the spatial consistency of the features of all brain regions and finally obtain local features of uniform size. .

[0037] A3: The local features of each brain region are analyzed by combining the brain region self-attention weighted unit with the preprocessed AAL template. Attention weighting was performed to obtain brain region features related to bipolar disorder .

[0038] The collaborative multi-expert network module (CMEN) is used to simulate the collaborative interaction between brain regions according to the output of the cognitive perception attention module, and obtain brain region features including the collaborative interaction relationship between brain regions. The collaborative multi-expert network module (CMEN) includes an expert mixed encoder, which includes multiple expert networks. The expert networks are used to process brain regions related to different cognitive functions. Each expert network focuses on modeling brain region features related to different cognitive functions. For example, some expert networks focus on brain regions related to memory functions (such as the hippocampus and prefrontal cortex), mainly analyzing the neural activity patterns and feature changes of these brain regions during memory encoding, storage and retrieval, and performing feature modeling; some expert networks focus on brain regions related to emotion regulation (such as the amygdala and insula), focusing on modeling brain region features involved in emotion generation, recognition and regulation; in addition, some expert networks focus on brain regions related to decision-making (such as the prefrontal cortex and basal ganglia), focusing on exploring the collaborative mechanisms and feature representations of these brain regions in the decision-making process. This modular structure enables the entire system to capture the complex interactions between different cognitive functions more comprehensively. Different cognitive functions are closely related to specific brain regions, and pathological changes in these brain regions may reflect potential disease mechanisms. Therefore, accurately capturing the interactions between these brain regions is crucial for understanding the pathology of psychiatric disorders.

[0039] The specific method of obtaining brain region features including collaborative interaction relationships between brain regions through a collaborative multi-expert network module (CMEN) includes the following steps: B1. Characteristics of brain regions related to bipolar disorder in response to input Perform linear transformation to obtain features with unified dimensions , whose expression is:

[0040] in, is the weight matrix, is the bias term, Indicates j brain regions associated with bipolar disorder M is the total number of brain region features associated with bipolar disorder; B2. Interact the unified dimension features with the CLS token representing global information to form a sequence , whose expression is:

[0041] in, Represents global information, Indicates K The features obtained by interacting with the CLS token after the unified dimension; B3. Input the sequence obtained in step B2 into the expert hybrid encoder for optimization, and obtain brain region features including the collaborative interaction relationship between brain regions after optimization. , whose expression is:

[0042] in, Indicates optimization processing.

[0043] The optimization process is as follows: After being input into the expert hybrid encoder, the gating mechanism in the expert hybrid encoder is based on the sequence The most relevant P expert networks are dynamically selected based on the fit between the features in the dataset and the expert network. This selection process is dynamic and is adjusted in real time according to the requirements of different cognitive tasks. For example, in memory tasks, the gating mechanism will automatically identify the features of brain regions that are closely related to memory functions and activate the expert network specifically responsible for memory. In emotion perception tasks, the gating mechanism will select an expert network that focuses on emotion regulation for analysis. The selected P expert networks model the pathological association features of different brain regions in specific cognitive functions based on the input features. The results obtained from each expert network are weighted and aggregated to obtain the optimized brain region features including the collaborative interaction relationship between brain regions. .

[0044] The causal normalization classifier is used to classify the output of the collaborative multi-expert network module and generate a prediction result of bipolar disorder for the person to be tested.

[0045] In one embodiment of the present invention, data in OpenfMRI and private data are used as data sets to train the system of the present invention. The data in OpenfMRI includes a large number of healthy individuals from a community in Los Angeles and individual samples with schizophrenia, bipolar disorder and ADHD. A healthy control group and a bipolar disorder patient group were selected to ensure data consistency. After screening, there were a total of 174 samples (including 125 healthy controls and 49 bipolar disorder patients). This embodiment only uses T1-weighted MRI data for training and testing. The private data comes from Wuchang Hospital in Wuhan, China, and includes 208 healthy controls and 82 bipolar disorder patients.

[0046] The loss function Focal Loss is introduced in the training process, and its expression is:

[0047] in, is the weight factor for the bipolar disorder category, which is used to adjust the contribution of different categories to the loss. Defined as the ratio of the number of categories: , n BD is the number of bipolar disorder categories, n HC is the number of healthy control categories, and the weight of the healthy control category is 1; Indicates that the prediction is a category The probability of is the focus parameter, which is used to control the degree of attention paid to difficult-to-classify samples; represents the loss value, is the natural logarithm.

[0048] During the training process, the initial learning rate is set to , the weight decays to , the batch size is 32, and the number of training rounds is 150. The cosine annealing learning rate scheduler is used, and the model is built with , and ,in represents the feature dimension, represents the pooling size, represents the hidden layer dimension.

[0049] The parameters involved in the system of the present invention are updated in a cyclic iterative manner, and the system is optimized using the AdamW optimizer until the loss value no longer decreases, and the training is terminated.

[0050] In order to verify the actual effect of the bipolar disorder prediction system based on cognitive pathology association provided by the present invention, a comparative experiment was carried out under the same training conditions. The experimental results are shown in Table 1.

[0051] Table 1

[0052] It can be seen from the experimental results that the sensitivity of the system involved in the present invention is as high as 0.918, which is significantly higher than other models, indicating that the system proposed in the present invention has excellent ability in capturing bipolar disorder patients (positive class). However, the specificity of the system designed by the present invention is 0.612, which is relatively low, which is also a common trade-off for high-sensitivity models. Nevertheless, this result is still acceptable because the relatively low specificity does not affect the accurate identification of patients. In contrast, models with higher specificity such as EfficientNet (specificity = 1.0) can accurately distinguish healthy controls (negative class), but their sensitivity is extremely low (sensitivity = 0), almost ignoring the identification of patients. Although this classification strategy that is extremely biased towards the negative class improves specificity, it leads to large-scale missed diagnoses, and its practicality is greatly reduced from the perspective of clinical application. Therefore, although the system designed by the present invention has a certain room for improvement in specificity, its excellent performance in sensitivity makes its advantages in early screening more prominent, and can effectively avoid missed patients from missing treatment opportunities.

[0053] In summary, the present invention can accurately predict whether a subject suffers from bipolar disorder, has high sensitivity and strong reliability, and has great implementation value.

Claims

1. A bipolar disorder prediction system based on cognitive pathology association, characterized in that: include: A data preprocessing module is used to preprocess the data in the data set; The cognitive perception attention module is used to extract brain region features related to bipolar disorder from the preprocessed data; The collaborative multi-expert network module is used to simulate the collaborative interaction between brain regions according to the output of the cognitive perception attention module, and obtain the brain region characteristics including the collaborative interaction relationship between brain regions; The causal normalization classifier is used to classify the output of the collaborative multi-expert network module and generate a prediction result of bipolar disorder for the person to be tested.

2. A bipolar disorder prediction system based on cognitive pathology association according to claim 1, characterized in that: The data in the dataset include T1-weighted MRI data and AAL template.

3. A bipolar disorder prediction system based on cognitive pathology association according to claim 2, characterized in that: The specific method of preprocessing is: The N4 algorithm was used to correct the intensity inhomogeneity of T1-weighted MRI data; the HD-BET network was used to perform skull stripping on the corrected T1-weighted MRI data; the skull-stripped T1-weighted MRI data was registered to the standard MNI space and resized to a uniform size; the resized T1-weighted MRI data was normalized to zero mean unit variance to obtain the preprocessed T1-weighted MRI data; the AAL template was directly mapped to the MNI space and registered according to the brain region annotations to align it with the preprocessed T1-weighted MRI data in the anatomical space; The cerebellum regions in the AAL template were analyzed, and the brain regions with a volume less than 1000 were merged into the adjacent brain regions to obtain the preprocessed AAL template; The preprocessed T1-weighted MRI data and preprocessed AAL templates were in the unified MNI space; The preprocessed T1-weighted MRI data and the preprocessed AAL template are both preprocessed data.

4. A bipolar disorder prediction system based on cognitive pathology association according to claim 1, characterized in that: The cognitive perception attention module includes a three-dimensional convolutional encoding unit, a brain region feature extraction unit, and a brain region self-attention weighting unit which are connected in sequence.

5. A bipolar disorder prediction system based on cognitive pathology association according to claim 4, characterized in that: The specific method of the cognitive perception attention module to extract the brain region features related to bipolar disorder from the preprocessed data includes the following steps: A1. The whole brain features in the preprocessed T1-weighted MRI data are extracted through a three-dimensional convolutional coding unit, and the expression is: in, is the preprocessed T1-weighted MRI data, Represents the three-dimensional spatial structure feature extraction process, Indicates whole-brain characteristics; A2. Extract local features of each brain region through the brain region feature extraction unit combined with the preprocessed AAL template and the whole brain features , whose expression is: in, is the preprocessed AAL template; represents feature extraction processing; A3: The local features of each brain region are analyzed by combining the brain region self-attention weighted unit with the preprocessed AAL template. Attention weighting was performed to obtain brain region features related to bipolar disorder .

6. A bipolar disorder prediction system based on cognitive pathology association according to claim 5, characterized in that: The specific method of feature extraction processing in step A2 includes the following sub-steps: A2-1. Three-dimensional rectangular frame of the brain area As the cropping range, for the whole brain features Perform cropping to obtain the characteristic blocks of each brain region; The expression is: in, Indicates i The specific boundaries of each brain region in three-dimensional space, Indicates that the brain area x The minimum and maximum values ​​on the axis, Indicates that the brain area y The minimum and maximum values ​​on the axis, Indicates that the brain area z Minimum and maximum values ​​on the axis; N is the total number of brain regions; A2-2, extracting effective regional features of each brain region from the feature blocks of each brain region using the brain region mask; A2-3. Adaptively average pool the effective regional features of each brain region to obtain local features of uniform size .

7. A bipolar disorder prediction system based on cognitive pathology association according to claim 1, characterized in that: The collaborative multi-expert network module includes an expert mixed encoder, which includes multiple expert networks. The multiple expert networks are used to process brain areas related to different cognitive functions.

8. A bipolar disorder prediction system based on cognitive pathology association according to claim 7, characterized in that: The collaborative multi-expert network module simulates the collaborative interaction between brain regions according to the output of the cognitive perception attention module, and the specific method of obtaining brain region features including the collaborative interaction relationship between brain regions includes the following steps: B1. Perform linear transformation on the input brain region features related to bipolar disorder to obtain features with unified dimensions , whose expression is: in, is the weight matrix, is the bias term, Indicates j brain regions associated with bipolar disorder M is the total number of brain region features associated with bipolar disorder; B2. Interact the unified dimension features with the CLS token representing global information to form a sequence , whose expression is: in, Represents global information, Indicates K The features obtained by interacting with the CLS token after the unified dimension; B3. Input the sequence obtained in step B2 into the expert hybrid encoder for optimization, and obtain brain region features including the collaborative interaction relationship between brain regions after optimization. , whose expression is: in, Indicates optimization processing.

9. A bipolar disorder prediction system based on cognitive pathology association according to claim 8, characterized in that: The optimization process in step B3 is specifically as follows: will sequence After being input into the expert hybrid encoder, the gating mechanism in the expert hybrid encoder is based on the sequence The fit between the features in the network and the expert network dynamically selects the most relevant P Expert network, through P Expert Network Based on Sequence The features in the model model the pathological association characteristics of different brain regions in specific cognitive functions; The results obtained by each expert network are weighted and aggregated to obtain the optimized brain region features including the collaborative interaction relationship between brain regions. .

10. A bipolar disorder prediction system based on cognitive pathology association according to claim 1, characterized in that: The system is a system obtained after training. The loss function Focal Loss is introduced during the training process, and its expression is: in, is the category weight factor, Indicates that the prediction is a category The probability of is the focus parameter; represents the loss value, is the natural logarithm; The parameters involved in this system are updated in a cyclic iterative manner, and the AdamW optimizer is used to optimize this system until the loss value no longer decreases, and the training is terminated.

Images