BBB repairing agent evaluation method and system based on ischemia reperfusion injury model

Abstract

The invention discloses a BBB repairing agent evaluation method and system based on an ischemia reperfusion injury model, and relates to the technical field of biologication.The method comprises the following specific steps that the ischemia reperfusion injury model is built, and a sample is collected, specifically, an animal ischemia reperfusion injury model is built, and experimental animals are divided into an experimental group and a control group; brain tissue and blood samples of the two groups of animals are collected at a plurality of preset time nodes, and behavioral evaluation is carried out at the same time to obtain score data; according to the method, the deep learning model fused with the dynamic attention mechanism is constructed, the weights of different time nodes and different modal data on BBB repair effect evaluation after ischemia reperfusion injury can be automatically identified, the problems that early signals are ignored and key time node evaluation is deviated due to fixed static data weights in a traditional evaluation method are solved, and the accuracy of BBB repair effect evaluation is improved. The model can flexibly adjust the weight of each modal data according to the dynamic change characteristics of the BBB repair process, and a more reliable basis is provided for research, development and application of a repair agent.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of biotechnology, in particular to a BBB repair agent evaluation method and system based on an ischemia-reperfusion injury model. BACKGROUND

[0002] The blood-brain barrier (BBB) is a key barrier structure between the central nervous system and peripheral blood circulation, and its integrity is important for maintaining the normal physiological function of the central nervous system. In clinical practice, ischemia-reperfusion injury is a relatively common pathological process, which can easily lead to damage to the BBB and cause a series of secondary injuries, seriously affecting the prognosis of patients. Therefore, scientifically and accurately evaluating the repair effect of BBB repair agents has become an important direction in this field.

[0003] Currently, existing BBB repair agent evaluation methods mainly rely on detection and analysis of static data. These methods usually detect static data of a single or a small number of modalities at a specific time point and conduct comprehensive evaluation based on fixed weights. However, the repair process of the BBB after ischemia-reperfusion injury presents significant dynamic characteristics. At different time nodes, the damage degree, repair process, and expression level of each biomarker of the BBB all have obvious differences. The fixed weight mode of static data used in existing evaluation methods cannot adapt to this dynamic change. On the one hand, this mode easily ignores early signals in the repair process, which are crucial for timely intervention and adjustment of treatment plans. On the other hand, fixed weights cannot truly reflect the actual contribution of each modality data to the evaluation at different time nodes, leading to biased evaluation results and providing unreliable basis for the research and application of BBB repair agents. SUMMARY

[0004] The present application aims to overcome the shortcomings of the prior art and provides a BBB repair agent evaluation method and system based on an ischemia-reperfusion injury model. It can effectively fuse and dynamically allocate weights to multi-modal time series data by constructing a deep learning model with a dynamic attention mechanism. By automatically identifying the weights of different time nodes and different modality data for BBB repair effect evaluation after ischemia-reperfusion injury, it solves the problems of early signal being ignored and evaluation bias at key time nodes caused by fixed static data weights in traditional evaluation methods. By obtaining multi-modal time series data at multiple time nodes, the deep learning model extracts and fuses time series features, dynamically tracks the action process of BBB repair agents, and predicts the subsequent repair trend based on existing time series data. This provides strong support for the research, optimization, and clinical application of BBB repair agents.

[0005] The application provides the following technical solutions to solve the above technical problems: On the one hand, a BBB repair agent evaluation method based on an ischemia-reperfusion injury model, the method comprising the following specific steps:

[0006] Ischemia-reperfusion injury model construction and sample collection: an animal ischemia-reperfusion injury model is constructed, and experimental animals are divided into an experimental group and a control group; brain tissue and blood samples of the two groups of animals are collected at preset multiple time nodes, and behavioral evaluation is performed to obtain score data;

[0007] Multi-modal time series data acquisition: brain tissue samples are processed and scanned to obtain pathological section image data, blood samples are detected to obtain molecular biology index data, and behavioral scores are recorded; the three types of data are standardized, respectively, to generate corresponding time series data sets;

[0008] Construction of a deep learning model with a dynamic attention mechanism: a deep learning model containing a data input layer, a modal feature extraction layer, a dynamic attention weight distribution layer, a feature fusion layer, and a repair effect evaluation output layer is constructed to realize multi-modal data reception, feature extraction, dynamic weight distribution, feature fusion, and repair effect evaluation result output;

[0009] Model training and verification: collect multi-modal time series data with BBB repair effect labels to construct training and verification data sets, input the model, and optimize the minimum prediction loss through parameter iteration, verify the model performance to the standard, and obtain a stable BBB repair agent evaluation model;

[0010] BBB repair agent evaluation: input the multi-modal time series data of the two groups of animals at each time node into the trained model, compare the output repair effect evaluation results, analyze the intervention of the repair agent on BBB repair, and evaluate its long-term repair effect in combination with the repair trend prediction.

[0011] Further, in the multi-modal time series data acquisition step, the collected brain tissue samples are sequentially fixed, dehydrated, embedded, sliced, and stained, a pathological slice scanner is used to scan the stained pathological slices, pathological slice image data reflecting the integrity of vascular endothelial cell tight junction and the activation state of astrocytes at different time nodes are acquired, and the image data is preprocessed by denoising, normalization, and image segmentation to obtain a standardized pathological slice image time series dataset; the collected blood samples are processed, molecular biology detection techniques such as enzyme-linked immunosorbent assay, real-time fluorescent quantitative polymerase chain reaction, and Western blotting are used to detect molecular biology indicators such as the concentration of inflammation factors related to BBB repair, the expression level of vascular endothelial cell tight junction related proteins, and the expression level of astrocyte activation related proteins in the blood, to acquire a molecular biology detection time series dataset at different time nodes, and the data is standardized by removing outliers and data normalization; at each preset time node, a standard neurobehavioral rating scale is used to evaluate the behavior of animals in the experimental group and the control group, and the scores related to animal motor function, balance ability, and sensory function are recorded to form a behavior score time series dataset, and the score data is subjected to consistency test and standardization processing.

[0012] Further, in the deep learning model construction step with dynamic attention fusion mechanism, a deep learning model containing a data input layer, a modal feature extraction layer, a dynamic attention weight distribution layer, a feature fusion layer, and a repair effect evaluation output layer is constructed, wherein the data input layer receives the three standardized time series datasets to provide a basis for subsequent processing; the modal feature extraction layer processes the pathological slice image time series data, the molecular biology detection time series data, and the behavior score time series data to output time series feature vectors of each modality; the dynamic attention weight distribution layer introduces a dynamic attention mechanism, takes the time series feature vectors of each modality as input, determines the BBB repair effect evaluation target feature vector trained by the sample data with known BBB repair effect, calculates the similarity between the feature vectors of each modality and the evaluation target feature vector at each time node using a similarity calculation function, obtains the attention score of each modality at each time node through exponential processing and normalization operation according to the similarity result, and determines the weight of each modality at the corresponding time node according to the attention score; the feature fusion layer weights and fuses the time series feature vectors of each modality according to the weight of each modality at different time nodes obtained by the dynamic attention weight distribution layer, obtains the weighted feature vectors of each modality, and sums all the weighted feature vectors of all modalities to obtain a multi-modal time series comprehensive feature vector at the time node; the repair effect evaluation output layer adopts a fully connected network structure to process the fused multi-modal time series comprehensive feature vector and outputs the repair effect evaluation result of the BBB at each time node, including repair grade, repair index, and repair trend prediction.

[0013] Furthermore, in the deep learning model construction step that integrates the dynamic attention mechanism, the attention score of each modality at each time point is obtained through exponential processing and normalization based on the similarity results. The calculation formula is as follows: ,in, It is the first At the _ time point Attention score for each modality It is the first At the _ time point Temporal feature vectors of each modality It is the target feature vector for evaluating the BBB repair effect. This represents the total number of modalities, corresponding to three modalities: pathological images, molecular biology, and behavioral science. It is the Euclidean norm of a vector. It is the first At the _ time point Temporal feature vectors of each modality.

[0014] Furthermore, in the deep learning model construction step that integrates the dynamic attention mechanism, the temporal feature vectors of each modality are weighted and fused, and the fusion formula is as follows: ,in, It is the first The fused multimodal temporal comprehensive feature vector at each time point It is the first At the 1st time point The weights of each modality It is the first At the 1st time point Temporal feature vectors of each modality It represents the total number of modes.

[0015] Furthermore, in the model training and validation steps, multimodal time-series data of ischemia-reperfusion injury model samples with known BBB repair effects and corresponding BBB repair effect annotation information are collected to construct model training datasets and validation datasets. The training dataset is input into the constructed deep learning model that integrates a dynamic attention mechanism, and the model parameters are adjusted using an iterative optimization algorithm to minimize the loss between the model's predicted repair effect evaluation results and the actual annotation information. The optimization iteration is performed using a gradient descent algorithm.

[0016] Furthermore, in the model training and validation steps, the gradient descent algorithm is used for optimization iteration, and its algorithm formula is as follows: ,in, For the updated model parameters, the model parameters before the update, the learning rate, the gradient of the loss function on In the model training process, the performance of the model is verified in real time by using the validation data set, the model performance is optimized by adjusting the model structure parameters, and the model is trained until the evaluation index of the model on the validation data set reaches the preset standard, and the BBB repair agent evaluation model with stable performance is obtained.

[0017] Further, in the BBB repair agent evaluation step, the long-term repair effect is evaluated by combining repair trend prediction, and the prediction formula is: wherein, is the repair effect evaluation result of the predicted time node, is the multi-modal time series comprehensive feature vector of the time node, is the weight matrix of the prediction layer, is the bias term of the prediction layer.

[0018] On the other hand, the BBB repair agent evaluation system based on the ischemia-reperfusion injury model comprises:

[0019] The model and sample collection module collects brain tissue samples and blood samples from experimental and control animals, and is equipped with a sample preservation device to ensure sample quality. At the same time, the experimental personnel can be reminded to conduct behavioral evaluation on the experimental animals at the sample collection time node to obtain behavioral score data.

[0020] The multi-modal time series data processing module processes pathological section images, molecular biology data and behavioral scores, completes the acquisition, processing and standardization of the three types of modal data, and generates a time series data set.

[0021] The deep learning model construction module provides a visual interface for users to select parameters, builds a deep learning model with multiple layers, realizes multi-modal data processing and repair effect evaluation result output, and realizes multi-modal data processing and repair effect evaluation result output.

[0022] The model training and verification module receives externally imported training sample data, combines the time series data set generated by the multi-modal time series data processing module, performs format conversion and division, and obtains the model training data set and the validation data set. According to the user's set training parameters, the constructed deep learning model is trained and the parameters are optimized, and the loss change and model parameters in the training process are recorded in real time. The performance of the model in the training is verified by using the validation data set, and the training is automatically stopped and the model is saved when the model performance meets the standard.

[0023] The repair agent evaluation application module imports the multi-modal time series data of the experimental group and the control group at each time node, performs format checking and standardization processing on the data, calls the qualified model saved by the model training and verification module, inputs the imported data into the model, generates the evaluation results of each time node, and compares and analyzes the evaluation results of the experimental group and the control group, displays the repair effect time series change, the modal weight change and the difference between the two groups, and generates a detailed evaluation report for the user to view and export

[0024] Compared with the prior art, the BBB repair agent evaluation method and system based on the ischemia-reperfusion injury model have the following beneficial effects:

[0025] Firstly, the deep learning model fused with a dynamic attention mechanism can automatically identify the weight of different time nodes and different modal data for BBB repair effect evaluation after ischemia-reperfusion injury, solve the problem of early signal being ignored and key time node evaluation deviation caused by fixed static data weight in the traditional evaluation method, and make the model flexibly adjust the weight of each modal data according to the dynamic change characteristics of the BBB repair process, thereby providing a more reliable basis for the development and application of repair agents.

[0026] Secondly, by obtaining multi-modal time series data at multiple time nodes and using a deep learning model for time series feature extraction and fusion, the BBB repair agent action process can be dynamically tracked, the subsequent repair trend can be predicted based on the existing time series data, effective technical means are provided for in-depth understanding of the action mechanism and effect of the repair agent, and the model related parameters can be adjusted to adapt to different types of ischemia-reperfusion injury models and different evaluation requirements, thereby effectively supporting the development, optimization and clinical application of BBB repair agents.

[0027] Other advantages, objects and features of the present application will be set forth in part in the description which follows, and in part will become apparent to those skilled in the art upon examination of the following or can be learned from practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without any creative effort.

[0029] Figure 1 FIG. 1 is a structural schematic diagram of the BBB repair agent evaluation system based on the ischemia-reperfusion injury model;

[0030] Figure 2This is a flowchart of the evaluation method for BBB repair agents based on an ischemia-reperfusion injury model. Detailed Implementation

[0031] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0032] This invention provides a method and system for evaluating BBB repair agents based on an ischemia-reperfusion injury model, such as... Figure 1 As shown, the system includes a model and sample acquisition module, a multimodal time-series data processing module, a deep learning model construction module, a model training and validation module, and a repair agent evaluation and application module. By constructing a deep learning model that integrates a dynamic attention mechanism, it achieves effective fusion and dynamic weight allocation of multimodal time-series data. By automatically identifying the weights of different time points and different modalities of data on the evaluation of BBB repair effects after ischemia-reperfusion injury, it solves the problems of neglecting early signals and evaluation bias at key time points caused by fixed weights of static data in traditional evaluation methods. By acquiring multimodal time-series data at multiple time points, the system uses a deep learning model to extract and fuse temporal features of the data, enabling dynamic tracking of the action process of BBB repair agents. Based on existing time-series data, it predicts subsequent repair trends, providing strong support for the research, optimization, and clinical application of BBB repair agents.

[0033] Example 1

[0034] like Figure 2 As shown, an ischemia-reperfusion injury model was constructed using SPF-grade male rats of a specific strain. The rats were randomly divided into an experimental group and a control group. The experimental group was given a repair agent via intraperitoneal injection at the beginning of reperfusion, while the control group was injected with an equal amount of physiological control agent. Multiple sample collection time points were preset. At each time point, a certain number of rats were taken from both groups to collect brain tissue samples and venous blood samples. At the same time, the Longa Rating Scale was used to conduct behavioral assessments on the rats before sample collection, and the scoring results were recorded.

[0035] The collected brain tissue samples were first fixed with a fixative, then subjected to gradient dehydration, embedding, and sectioning. HE staining and immunofluorescence staining for specific tight junction proteins were then performed. The sections were scanned using a pathological slide scanner to obtain pathological slide images. Image preprocessing software was used to remove image noise, perform grayscale normalization, and segment the target region, generating a time-series dataset of pathological slide images (denoted as...). ,in (As the image feature set corresponding to the corresponding time node), for the collected venous blood samples, plasma was separated by centrifugation, and the concentration of specific inflammatory factors in the plasma was detected by ELISA. The expression level of specific tight junction proteins in the plasma was detected by Western blotting. After removing outliers from the detection data, normalization was performed to form a molecular biology detection time series dataset (denoted as ). The Longa scores at each time point are compiled, and the consistency of the scores is verified by the Kappa consistency test (Kappa values ​​meet the preset standard). Then, the scores are standardized to generate a behavioral score time series dataset (denoted as ). .

[0036] A deep learning model incorporating a dynamic attention mechanism is constructed. The data input layer has three data interfaces to receive data from each interface. , , In the modal feature extraction layer, spatial features of the temporal data of pathological slide images are extracted using a network with multiple convolutional layers, and temporal features are extracted using a multi-layer long short-term memory network, outputting an image modal temporal feature vector. Molecular biology detection time-series data are mapped using a multi-layer fully connected network, then fed into a multi-layer LSTM to extract temporal variation features, outputting a molecular modality temporal feature vector. Behavioral scoring time-series data is directly input into a multi-layer LSTM to extract temporal features, and outputs behavioral modality temporal feature vectors. Dynamic attention weight allocation layer As input, first load the evaluation target feature vector obtained by training with multiple rat samples known to have BBB repair effects. Then, cosine similarity is used to calculate the modal feature vectors at each time point and... The similarity is used to obtain the attention score after exponentialization and normalization operations. The calculation formula is as follows: (in the formula, Representing three modes, (For modality indexing), and the score is directly used as the weight of each modality at the corresponding time node. The feature fusion layer, according to the modal weights, weights and sums the feature vectors of the three modalities at the same time point to obtain the multimodal temporal comprehensive feature vector at that time point. The fusion formula is as follows: The repair effect evaluation output layer adopts a multi-layer fully connected network. Process the data and output the BBB repair level (Excellent, Good, Medium, Poor) and repair index (0-100 points, denoted as...) at each time point. The results include the predicted repair trend over the future preset duration.

[0037] Collect 200 sets of ischemia-reperfusion injury rat sample data with clear BBB repair effect (including pathological section images, molecular biology indicators, behavior score and corresponding repair effect annotation), divide them into training data set and validation data set according to the ratio of 7:3, input the training data set into the constructed deep learning model, adjust the model parameters by using gradient descent algorithm, take the difference between the model prediction and the actual annotation as the loss, iterate training for 50 rounds, test the model performance after each round of training with the validation data set, calculate the accuracy rate of the model prediction repair level (target > 90%) and the mean square error of the repair index (target < 5), when the training reaches the 35th round, the model validation accuracy rate reaches 92%, and the mean square error is 4.2, which meets the preset standard, stop training and save the model.

[0038] Input the multi-modal time series data of each time node of the experimental group and the control group into the trained model, and the model outputs the repair index of each time node of the two groups (experimental group) and (control group), for future repair trend prediction, use linear extrapolation logic based on comprehensive feature vectors, the prediction formula is: , (in the formula, is the future prediction time interval, is the prediction layer weight matrix, is the bias term), compare the repair effect evaluation results of the two groups at the same time node, analyze the intervention of the repair agent on the BBB repair after ischemia-reperfusion injury, and evaluate the influence of the repair agent A on the long-term repair effect of BBB combined with the repair trend prediction result.

[0039] Example two

[0040] SPF level male rats of a specific strain were selected to construct an ischemia-reperfusion injury model, and were randomly divided into an experimental group and a control group. The experimental group was given a repair agent by tail vein injection, and the control group was injected with an equal amount of phosphate buffer as a control. Multiple sample collection time nodes were preset, and a certain number of rats were selected from the two groups at each time node. Brain tissue samples and heart blood samples were collected. At the same time, before each sample collection time node, the neurological behavior function of the rats was evaluated using the Bederson score scale, focusing on the motor coordination and limb movement of the rats.

[0041] For the collected brain tissue samples, first fixed with fixing solution, then sequentially carried out sucrose gradient dehydration, OCT embedding operation, and then carried out frozen section according to the preset specification, and the section was respectively subjected to GFAP protein immunofluorescence staining (used to mark astrocytes and observe the activation state) and HE staining (used to observe the histopathological morphology); using a pathological section scanner to scan the stained section according to the preset clarity standard, obtain pathological section images containing astrocyte morphology and histopathological characteristics, remove the background of the scanned image through image preprocessing software to eliminate irrelevant interference information, then perform fluorescence intensity normalization processing to unify the image brightness standard, and segment the region where the astrocyte is located, and finally generate a standardized pathological section image time series dataset; for the collected heart blood samples, centrifugal treatment was carried out according to the preset condition, and serum samples were separated, qRT-PCR technology was used to detect the expression level of GFAP mRNA in serum, and ELISA technology was used to detect the concentration of specific inflammatory factors in serum; the molecular biology index data detected is subjected to outlier rejection, removing the data deviating from the normal range due to detection error, and then standardized to make the data in the same dimension, forming a molecular biology detection time series dataset, collating the Bederson score results recorded at each time node, using Kappa consistency test method to verify the consistency of the score results of different experimenters, to ensure the reliability of the score, and then the score data is standardized to generate a behavior score time series dataset.

[0042] The deep learning model integrating the dynamic attention mechanism is built, and the data input layer maintains three data interfaces to receive the generated pathological section image time series dataset, molecular biology detection time series dataset and behavior score time series dataset. In the modal feature extraction layer, for the pathological section image time series data, the number of convolution layers is adjusted to adapt to the specification of capturing astrocyte morphological features. The spatial features in the image are extracted through the convolution layer, and then combined with the multi-layer network structure with time series feature capturing capability to extract the time series change features, and output the time series feature vector of the pathological section image modal. For the molecular biology detection time series data, the number of fully connected network layers is adjusted to adapt to the feature dimension of mRNA and protein indicators. After feature mapping of the data through the fully connected network, the multi-layer network structure with time series feature capturing capability is used to extract the time series change features, and the time series feature vector of the molecular biology detection modal is output. For the behavior score time series data, the multi-layer network structure with time series feature capturing capability is directly input to extract the time series features, and the time series feature vector of the behavior score modal is output. The dynamic attention weight distribution layer loads the BBB repair effect evaluation target feature vector optimized for the polypeptide repair agent evaluation scene, and takes the time series feature vectors of the three modes as input. By calculating the similarity of each modal feature and the evaluation target feature vector at different time nodes, and combining the exponential processing and normalization operation, the attention score of each modal at the corresponding time node is obtained. The attention score is directly used as the weight of each modal. The feature fusion layer weights the time series feature vectors of the three modes at the same time node according to the weight of each modal, and then sums the weighted feature vectors to obtain the fused multi-modal time series comprehensive feature vector at the time node. The repair effect evaluation output layer adopts a multi-layer fully connected network structure to process the multi-modal time series comprehensive feature vector, and outputs the repair level (divided into four levels of excellent, good, medium and poor), repair index (an index for quantifying repair effect) and repair trend prediction result within a preset time length of the BBB at each time node.

[0043] The sample data of the polypeptide repair agent acting on the ischemia-reperfusion injury rat samples verified by experiments to repair the BBB effect are collected, the data including corresponding pathological section image data, molecular biology detection data, behavior score data and professional determined BBB repair effect annotation information, the collected sample data is divided into model training data set and model verification data set according to a preset ratio, the training data set is used for model parameter learning, and the verification data set is used for evaluating the model performance, the training data set is input into the constructed deep learning model, and the gradient descent algorithm is used to iteratively optimize and train the model parameters, and the training process lasts for a preset number of rounds; the difference between the model predicted BBB repair effect and the actual labeled repair effect is taken as the loss, the model parameters are continuously adjusted to reduce the loss, after each round of training is completed, the verification data set is input into the model, the accuracy of the model predicted BBB repair grade and the deviation degree of the predicted repair index and the actual labeled value are calculated, and the model performance is evaluated, when the performance of the model on the verification data set reaches the preset standard (such as the repair grade prediction accuracy meets the preset requirement, and the repair index deviation is controlled within the preset range), the model training is stopped, and the trained and stable model is saved.

[0044] The pathological section image time series data, molecular biology detection time series data and behavior score time series data of the experimental group and the control group at each preset time node are input into the trained deep learning model, the model performs feature extraction, dynamic weight distribution, multi-modal feature fusion and repair effect calculation on the input data, and outputs the BBB repair grade, repair index and repair trend prediction result of the experimental group and the control group at each time node, by comparing the repair effect evaluation results of the experimental group and the control group at the same time node, the influence of the repair agent on the activation state of the astrocytes (such as the morphological change and distribution density change of the astrocytes) and the intervention effect on the BBB repair process (such as the repair starting time and repair progress speed) are analyzed, combined with the future repair trend prediction result output by the model, the maintenance ability of the repair agent B on the BBB repair effect in a long time period is evaluated, so that the effect of the repair agent B on the BBB repair after ischemia-reperfusion injury is comprehensively judged.

[0045] The above is only a preferred embodiment of the present application, and does not limit the present application in any form. Although the preferred embodiment of the present application has been disclosed as above, it is not intended to limit the present application. Any person skilled in the art can make some changes or modifications to the above disclosed technical content without departing from the scope of the technical solution of the present application, and any brief introduction, modification, equivalent change and modification of the above embodiment according to the technical essence of the present application are still within the scope of the technical solution of the present application.

Claims

1. A method for evaluating a BBB repair agent based on an ischemia-reperfusion injury model, characterized by, The method comprises the following specific steps: Ischemia-reperfusion injury model construction and sample collection: construct an animal ischemia-reperfusion injury model and divide the experimental animals into an experimental group and a control group, collect brain tissue and blood samples of the two groups of animals at preset multiple time nodes, and perform behavioral assessment to obtain score data; Multi-modal time series data acquisition: process the brain tissue samples and scan to obtain pathological section image data, detect the blood samples to obtain molecular biology index data, and record behavioral scores, and standardize the three types of data respectively to generate corresponding time series data sets; Deep learning model construction with fusion of dynamic attention mechanism: construct a deep learning model comprising a data input layer, a modal feature extraction layer, a dynamic attention weight distribution layer, a feature fusion layer, and a repair effect evaluation output layer, to realize multi-modal data reception, feature extraction, dynamic weight distribution, feature fusion, and repair effect evaluation result output; Model training and verification: collect multi-modal time series data with BBB repair effect labels to construct training and verification data sets, input the model, and optimize the parameters to minimize the prediction loss, verify the model performance to reach the standard, and obtain a stable BBB repair agent evaluation model; BBB repair agent evaluation: input the multi-modal time series data of the two groups of animals at each time node into the trained model, compare the output repair effect evaluation results, analyze the intervention of the repair agent on BBB repair, and evaluate its long-term repair effect in combination with the repair trend prediction.

2. The method for evaluating a BBB repair agent based on an ischemia-reperfusion injury model according to claim 1, characterized by, In the multi-modal time series data acquisition step, the collected brain tissue samples are sequentially fixed, dehydrated, embedded, sectioned, and stained, the stained pathological sections are scanned using a pathological section scanner, pathological section image data reflecting the integrity of vascular endothelial cell tight junction and the activation state of astrocytes at different time nodes are obtained, and the image data are preprocessed by denoising, normalization, and image segmentation to obtain standardized pathological section image time series data sets; the collected blood samples are processed, molecular biology detection techniques such as enzyme-linked immunosorbent assay, real-time fluorescent quantitative polymerase chain reaction, and Western blotting are used to detect molecular biology indexes such as the concentration of inflammatory factors related to BBB repair, the expression level of vascular endothelial cell tight junction related proteins, and the expression level of astrocyte activation related proteins in the blood, to obtain molecular biology detection time series data sets at different time nodes, and the data are standardized by removing outliers and data normalization; At each preset time node, the experimental group and the control group of animals are assessed by behavioral assessment using a standard neurological scoring scale, the scores related to the motor function, balance ability, and sensory function of the animals are recorded, a behavioral score time series data set is formed, and the score data are subjected to consistency test and standardization.

3. The method for evaluating a BBB repair agent based on an ischemia-reperfusion injury model according to claim 1, characterized by, The deep learning model with the fusion dynamic attention mechanism comprises a data input layer, a modal feature extraction layer, a dynamic attention weight distribution layer, a feature fusion layer, and a repair effect evaluation output layer.

4. The method for evaluating a BBB repair agent based on an ischemia-reperfusion injury model according to claim 3, characterized by, In the deep learning model construction steps that integrate the dynamic attention mechanism, the attention score of each modality at each time point is obtained through exponential processing and normalization based on the similarity results. The calculation formula is as follows: ,in, It is the first At the 1st time point Attention score for each modality It is the first At the 1st time point Temporal feature vectors of each modality It is the target feature vector for evaluating the BBB repair effect. This represents the total number of modalities, corresponding to three modalities: pathological images, molecular biology, and behavioral science. It is the Euclidean norm of a vector. It is the first At the 1st time point Temporal feature vectors of each modality.

5. The method for evaluating a BBB repair agent based on an ischemia-reperfusion injury model according to claim 3, characterized by, In the deep learning model construction step of the fusion dynamic attention mechanism, the time sequence feature vectors of each modality are weighted and fused, and the fusion formula is: wherein, is the fusion multi-modal time sequence comprehensive feature vector at the th time node, is the weight of the th modality at the th time node, is the time sequence feature vector of the th modality at the th time node, and is the total number of modalities.

6. The method for evaluating a BBB repair agent based on an ischemia-reperfusion injury model according to claim 1, characterized by, In the model training and verification step, the multi-modal time series data and the corresponding BBB repair effect label information of the ischemia-reperfusion injury model samples with a clear BBB repair effect are collected to construct a model training data set and a verification data set.

7. The method for evaluating a BBB repair agent based on an ischemia-reperfusion injury model according to claim 6, characterized by, In the model training and verification step, optimization iteration is performed by using a gradient descent algorithm, and the algorithm formula is: wherein, is the updated model parameter, is the model parameter before updating, is the learning rate, is the gradient of the loss function on In the model training process, the performance of the model is verified in real time by using the validation dataset, the model performance is optimized by adjusting the model structure parameters, and the evaluation index of the model on the validation dataset reaches the preset standard until the BBB repair agent evaluation model is obtained.

8. The method for evaluating a BBB repair agent based on an ischemia-reperfusion injury model according to claim 1, characterized by, In the BBB repair agent evaluation step, the long-term repair effect is evaluated in combination with the repair trend prediction, and the prediction formula is: wherein, is the repair effect evaluation result of the predicted time node, is the multi-modal time sequence comprehensive feature vector of the time node, is the weight matrix of the prediction layer, is the bias term of the prediction layer.

9. A BBB repair agent evaluation system based on an ischemia-reperfusion injury model, the system being suitable for use in the BBB repair agent evaluation method based on an ischemia-reperfusion injury model according to any one of claims 1 to 8, characterized by, The method comprises: A model and sample collection module: brain tissue samples and blood samples are collected from experimental and control animals, and sample preservation devices are provided to ensure sample quality; at the same time, experimental personnel are reminded to conduct behavioral evaluation on experimental animals at the sample collection time node to obtain behavioral score data; A multi-modal time series data processing module: pathological section images, molecular biology data, and behavioral scores are processed to obtain, process, and standardize three types of modal data, and generate time series data sets; A deep learning model construction module: provides a visual interface for users to select parameters, builds a deep learning model with a multi-layer structure, and realizes multi-modal data processing and repair effect evaluation result output; The model training and verification module receives externally imported training sample data, combines the time series data set generated by the multi-modal time series data processing module, performs format conversion and division, obtains a model training data set and a verification data set, optimizes and trains parameters of the constructed deep learning model according to user-set training parameters, and records loss changes and model parameters in the training process in real time, verifies the performance of the model in the training process by using the verification data set, automatically stops the training when the model performance meets the standard, and saves the model; The repair agent evaluation application module imports multi-modal time series data of the experimental group and the control group at each time node, performs format checking and standardization processing on the data, calls the standard model saved by the model training and verification module, inputs the imported data into the model, generates evaluation results at each time node, compares and analyzes the evaluation results of the experimental group and the control group, displays the repair effect time series change, the modal weight change and the difference between the two groups, and generates a detailed evaluation report for the user to view and export.

Images