Intracranial arterial stenosis stroke risk assessment method based on multi-modal data

Abstract

The application discloses an intracranial arterial stenosis stroke risk assessment method based on multi-modal data and belongs to the technical field of risk assessment. The method comprises the following steps: acquiring multi-modal data of a target object and pre-processing the multi-modal data; extracting multi-modal features of the multi-modal data and fusing the multi-modal features based on cross attention to obtain initial multi-modal fusion features; processing the initial multi-modal fusion features by principal component analysis to obtain final multi-modal fusion features; inputting the multi-modal fusion features into a constructed stroke risk prediction model to obtain stroke risk assessment results of the target object at different time periods, and determining a risk grade according to the stroke risk assessment results at different time periods; and screening different rehabilitation schemes according to the stroke risk grade and a pre-constructed rehabilitation strategy specificity prediction model to obtain an optimal decision scheme of the target object. The application can provide accurate stroke risk assessment and optimal treatment schemes for ICAS patients by using the above method.

Description

Technical Field

[0001] This invention relates to the field of risk assessment technology, and in particular to a method for assessing the risk of stroke in intracranial arterial stenosis based on multimodal data. Background Technology

[0002] Intracranial atherosclerotic stenosis (ICAS) is a major cause of ischemic stroke and its recurrence. Accurately assessing the stroke risk of patients with ICAS and developing individualized treatment strategies for patients at different risk levels are key to reducing ICAS-related mortality and disability rates.

[0003] Currently, some studies have attempted to use machine learning methods for ICAS stroke risk prediction, but the following limitations still exist: (1) Insufficient integration of multimodal data. Existing ICAS stroke prediction schemes are mainly based on a combination of single images and clinical features, which makes it difficult to fully reflect the patient's overall condition and the prediction accuracy needs to be improved.

[0004] (2) Disconnect between risk prediction and treatment strategy. Existing technologies can only achieve static prediction of stroke risk and lack a mechanism to correlate the risk prediction results with the effectiveness of different treatment strategies, making it difficult to obtain the best technical solution. Summary of the Invention

[0005] The purpose of this invention is to provide a method for assessing the risk of stroke due to intracranial arterial stenosis based on multimodal data, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, this invention provides a method for assessing the risk of stroke due to intracranial arterial stenosis based on multimodal data, comprising the following steps: S1. Acquire multimodal data of the target object and preprocess the multimodal data, which includes imaging data, clinical text data, detection data, biomarker data and the collection timestamps corresponding to each data. S2. Based on the preprocessed multimodal time series data, extract multimodal features and fuse the multimodal features based on cross-attention to obtain the initial multimodal fused features; S3. Principal component analysis is used to process the initial multimodal fusion features to obtain the final multimodal fusion features; S4. Input the multimodal fusion features from step S3 into the constructed stroke risk prediction model to obtain the stroke risk assessment results of the target object at different time periods, and determine the risk level based on the risk assessment results at different time periods. S5. Based on the stroke risk level and the pre-built rehabilitation strategy-specific prediction model, different rehabilitation programs are screened to obtain the optimal decision program for the target population.

[0007] Preferably, the imaging data includes high-resolution magnetic resonance vascular wall imaging, magnetic resonance angiography, arterial spin labeling perfusion imaging, diffusion-weighted imaging, and hemodynamic parameter maps based on computational fluid dynamics simulation; the clinical data includes the target subject's medical history, clinical manifestations, family history, and stroke etiology classification text data extracted from the discharge summary using natural language processing; the biomarker data includes peripheral blood mononuclear cell gene expression profiles and serum glial fibrillary acidic protein levels; and the detection data includes lipid profiles, glycated hemoglobin, high-sensitivity C-reactive protein, homocysteine, and lipoprotein-associated phospholipase A2.

[0008] Preferably, step S2 specifically includes: Parallel extraction was employed to extract radiomics features, habitat radiomics features, hemodynamic features, detection features, clinical features, and biomarker features from preprocessed imaging data, clinical text data, detection data, and biomarker data, respectively. Among these, habitat radiomics features reflect plaque regional heterogeneity. After Z-score standardization of detection features, clinical features, and biomarker features, feature selection based on the maximum information coefficient is adopted to retain the feature subset related to the time of stroke occurrence. The selected detection features, clinical features, and biomarker features are tensor-concatenated to obtain the first fusion feature. Feature transformations were performed on the first fusion feature, radiomics feature, habitat radiomics feature, and hemodynamic feature respectively, mapping the features of different modalities to a unified feature space; A multi-head cross-attention mechanism is used to fuse features from different modalities. The feature of any modality is used as the query, and the features of the other three modalities are used as the key and value. The attention-weighted representation of the modality relative to the other modalities is calculated. The cross-attention outputs of all modalities are concatenated and projected to obtain the fused multimodal feature vector.

[0009] Preferably, the hemodynamic characteristics include the cross-lesion pressure ratio, wall shear stress, oscillatory shear index, and vessel diameter stenosis rate, wherein the cross-lesion pressure ratio is the ratio of pressure distal to proximal to the stenosis.

[0010] Preferably, the stroke risk prediction model is constructed using the DeepSurv model, which contains four fully connected hidden layers, employs the SELU activation function and Dropout regularization, and uses the negative log-partial likelihood function as the loss function.

[0011] Preferably, the target object includes different time periods, including the predicted probability of stroke occurrence in the first, second, and third years of the target object's future; Risk levels are categorized into high risk, medium risk, and low risk, and the classification method is as follows: 1-year stroke probability ≤10%, 2-year stroke probability ≤15%, and 3-year stroke probability ≤20%: dynamic low risk; If the probability exceeds the corresponding threshold at any given time but does not meet the high-risk standard, it is considered a dynamic medium-risk situation. A stroke probability of >25% within 1 year or an increase of >50% in the probability of any period compared to the previous period: dynamic high risk, indicating unstable plaque progression.

[0012] Preferably, step S5 specifically includes: Based on the target group's stroke risk level, age, and comorbidity index, candidate decision-making schemes are generated from a pre-set rehabilitation strategy library, and candidate decision-making schemes for different rehabilitation strategies are generated from the pre-set rehabilitation strategy library. Based on a pre-built rehabilitation strategy-specific prediction model, the residual stroke risk of the target subjects after receiving candidate decision-making options is predicted. Calculate the net rehabilitation benefit for the target population based on residual stroke risk and different candidate decision-making options. ; Net recovery benefit of different candidate decision-making options Sort according to net recovery benefit Generate the optimal decision-making scheme.

[0013] Preferably, the machine learning models used to construct the rehabilitation strategy-specific prediction model include DeepSurv model and TCN-LSTM time-series hybrid model.

[0014] The present invention also provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and when the processor calls the computer program in the memory, it implements the content of the intracranial artery stenosis stroke risk assessment method based on multimodal data.

[0015] The present invention also provides a storage medium storing computer-executable instructions, which, when loaded and executed by a processor, implement the content of a method for assessing the risk of intracranial arterial stenosis stroke based on multimodal data.

[0016] Therefore, the present invention employs the above-mentioned method for assessing intracranial arterial stenosis stroke risk based on multimodal data, which has the following beneficial effects: (1) This application uses a cross-attention mechanism to deeply fuse multimodal features, which can adaptively learn the correlation and complementarity between different modal features, overcoming the feature redundancy and dimensionality explosion problems caused by simple splicing in traditional methods. At the same time, through multi-image feature extraction and multimodal feature fusion, the prediction accuracy is improved. (2) By constructing a treatment strategy-specific prediction model, the individualized residual stroke risk of patients under each strategy can be directly predicted, and the optimal decision-making scheme can be obtained based on the residual stroke risk, which can directly guide clinical decision-making.

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating an embodiment of the intracranial arterial stenosis stroke risk assessment method based on multimodal data of the present invention. Figure 2 This is an embodiment of the intracranial arterial stenosis stroke risk assessment method based on multimodal data of the present invention to obtain the optimal decision scheme for the target object. Detailed Implementation

[0019] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0021] Example like Figure 1-2 As shown, this invention provides a method for assessing the risk of stroke due to intracranial arterial stenosis based on multimodal data, comprising the following steps: S1. Acquire multimodal data of the target object and preprocess the multimodal data, which includes imaging data, clinical text data, detection data, biomarker data, and the collection timestamps corresponding to each data.

[0022] In this embodiment, imaging data includes high-resolution magnetic resonance vascular wall imaging, magnetic resonance angiography, arterial spin labeling perfusion imaging, diffusion-weighted imaging, and hemodynamic parameter maps based on computational fluid dynamics simulation; clinical data includes the target subject's medical history, clinical manifestations, family history, and stroke etiology classification text data extracted from the discharge summary using natural language processing; biomarker data includes peripheral blood mononuclear cell gene expression profiles and serum glial fibrillary acidic protein levels, reflecting atherosclerotic plaque activity and neurological injury markers; detection data includes lipid profiles, glycated hemoglobin, high-sensitivity C-reactive protein, homocysteine, and lipoprotein-associated phospholipase A2, which are independent of biomarker data and reflect metabolic and biochemical indicators rather than inflammatory states.

[0023] Preprocessing of multimodal data includes: For image data in radiographic data, the N4 bias field correction algorithm is used to eliminate intensity inhomogeneity in the image, and images of different sequences are registered to the same spatial coordinate system and then standardized.

[0024] For both test data and clinical data, each numerical data point is standardized, and categorical clinical data, such as gender, is uniquely encoded.

[0025] Logarithmic transformation and standardization were performed on biomarker data to reduce the impact of variation.

[0026] S2. Based on the preprocessed multimodal time-series data, extract multimodal features and fuse these features using cross-attention to obtain initial multimodal fused features. Specifically, this includes: Parallel extraction was employed to extract radiomics features, habitat radiomics features, hemodynamic features, detection features, clinical features, and biomarker features from preprocessed imaging data, clinical text data, detection data, and biomarker data, respectively. Among these, habitat radiomics features reflect plaque regional heterogeneity.

[0027] After Z-score standardization of detection features, clinical features, and biomarker features, feature selection based on the maximum information coefficient is adopted to retain the feature subset related to the time of stroke occurrence.

[0028] The selected detection features, clinical features, and biomarker features are spliced ​​using tensors to obtain the first fusion feature.

[0029] Feature transformations were performed on the first fusion feature, radiomics feature, habitat radiomics feature, and hemodynamic feature respectively, mapping the features of different modalities to a unified feature space.

[0030] A multi-head cross-attention mechanism is used to fuse features from different modalities. The feature of any one modality is used as the query, and the features of the other three modalities are used as the key and value. The attention-weighted representation of the modality relative to the other modalities is calculated.

[0031] The cross-attention outputs of all modalities are concatenated after layer normalization and projected through a fully connected layer to obtain the initial multimodal fusion features.

[0032] Specifically, radiomics features, habitat radiomics features, and hemodynamic features are all extracted from imaging data. Radiomics features include morphological features, texture features, and gray-level histogram features. Morphological features include vascular cross-sectional area stenosis rate, plaque eccentricity index, average fibrous cap thickness, high-dimensional brain perfusion features, and basic morphological features such as grid volume and voxel volume. Gray-level histogram features include, but are not limited to, maximum gray value, mean gray value, median gray value, 10th percentile, 90th percentile, and interquartile range. Texture features include gray-level co-occurrence matrix, gray-level run length matrix, gray-level size region matrix, neighborhood gray-level difference matrix, and gray-level dependency matrix.

[0033] Habitat radiomics features are extracted from high-resolution magnetic resonance vascular wall imaging data to reflect plaque regional heterogeneity. The extraction process includes: Extract the region of interest (ROI) and construct a four-dimensional signal intensity vector for each voxel at the same spatial location in the four high-resolution magnetic resonance imaging (MRI) sequences (T1WI, T2WI, PDWI, and T1CE). ,in, Traverse all voxels within the ROI of the plaque.

[0034] The K-means clustering algorithm was used to divide all voxels within the patch into groups. Each habitat subregion. The clustering objective function is: ; in, For the first The signal intensity vector of an individual element. For the first The center vectors of the clusters, To belong to the A set of voxels in a cluster, For Euclidean distance, The value is determined using the elbow rule: Calculate different The sum of squared errors within clusters under the given value is selected based on the inflection point of the curve. The value K is usually 3 or 4.

[0035] Each habitat subregion is delineated individually for its own ROI, and radiomics features are extracted from each subregion. The volume fraction and spatial eccentricity of each subregion are calculated to construct a habitat radiomics feature matrix, which serves as the habitat radiomics feature. The spatial eccentricity is the distance from the centroid of the subregion to the geometric center of the patch.

[0036] Hemodynamic features are extracted from hemodynamic parameter maps derived from computational fluid dynamics simulations, including cross-lesion pressure ratio, wall shear stress, oscillatory shear index, and vessel diameter stenosis rate, among which: The cross-lesion pressure ratio is defined as the ratio of the pressure distal to the stenosis to the pressure proximal to the stenosis, and the formula is: ; In the formula, The average pressure at the vascular cross-section 5 mm distal to the stenosis. The average pressure at the vascular cross-section 5 mm proximal to the stenosis. The lower the value, the more severe the hemodynamic impairment caused by the stenosis. The normal value is about 1.0, and PR < 0.85 indicates significant hemodynamic impairment.

[0037] Wall shear stress is defined as the tangential force exerted by blood flow on a unit area of ​​the vessel wall, and the formula is: ; In the formula, For blood dynamic viscosity, This represents the velocity gradient at the vessel wall.

[0038] The oscillatory shear index quantifies the degree of oscillation in the direction of blood flow within a cardiac cycle. The formula is: ; In the formula, T represents one cardiac cycle. This represents the instantaneous wall shear stress.

[0039] The stenosis rate of blood vessel diameter, extracted from digital subtraction angiography images, is calculated using the following formula: ; In the formula, The diameter of the normal blood vessel proximal to the stenosis. This is the minimum lumen diameter at the narrowest point.

[0040] This step involves K-means clustering of multiple sequences from high-resolution magnetic resonance imaging of the blood vessel wall to segment the plaque into multiple sub-regions with different signal enhancement patterns. This quantifies the biological heterogeneity within the plaque and provides richer risk information compared to traditional whole plaque analysis methods. In particular, it can capture the spatial distribution characteristics of high-risk components within the plaque, further improving prediction accuracy.

[0041] S3. Principal Component Analysis (PCA) is used to process the initial multimodal fusion features to obtain the final multimodal fusion features. Specifically, PCA is used to reduce the dimensionality of the initial multimodal fusion features, retaining principal components that explain more than 95% of the variance, thus obtaining the final multimodal fusion features. PCA effectively mitigates the overfitting risk caused by high-dimensional features; simultaneously, it transforms the original highly correlated multimodal features into mutually orthogonal principal components, eliminating the interference of multicollinearity among features on the model weights and filtering out confounding noise, thereby improving prediction stability and inference efficiency.

[0042] S4. Input the multimodal fusion features from step S3 into the constructed stroke risk prediction model to obtain the probability of stroke occurrence in the target object in the 1st, 2nd, and 3rd years, and determine the risk level based on the risk assessment results for different time periods.

[0043] Risk levels are categorized into high risk, medium risk, and low risk, and the classification method is as follows: 1-year stroke probability ≤10%, 2-year stroke probability ≤15%, and 3-year stroke probability ≤20%: dynamic low risk; If the probability exceeds the corresponding threshold at any given time period but does not meet the high-risk standard: dynamic medium risk; A stroke probability of >25% within 1 year or an increase of >50% in the probability of any period compared to the previous period: dynamic high risk, indicating unstable plaque progression.

[0044] In this embodiment, the stroke risk dynamic prediction model is constructed using a DeepSurv network based on a competitive risk framework. The network contains four fully connected hidden layers, each with 256 nodes. It employs the SELU activation function and Dropout regularization, and uses the negative log-partial likelihood function as the loss function, as shown in the formula: ; In the formula, For the first The event indicator variable for each sample, where 1 indicates a stroke event has occurred and 0 indicates no stroke event has occurred. For the first The observation time for each sample Indicates time The sample index set that is still in the risk set at that time A deep neural network is represented as: ;in, , , This is the weight matrix. , , For bias vectors, This is the activation function.

[0045] For the target object, the model outputs its stroke probability within a certain time window, using the following formula: ; In the formula, Let be the survival function, representing the patient's survival over time. The probability of not having had a stroke previously.

[0046] During the training of the stroke risk prediction model, multimodal data from multiple subjects were acquired. After preprocessing, feature extraction, feature fusion, and principal component analysis of the multimodal data, a dataset was obtained. The dataset was then divided into training, testing, and validation sets in a 7:2:1 ratio to train the stroke risk prediction model.

[0047] S5. Based on stroke risk level and a pre-built rehabilitation strategy-specific prediction model, different rehabilitation plans are screened to obtain the optimal decision plan for the target population. The machine learning models used to construct the rehabilitation strategy-specific prediction model include the DeepSurv model and the TCN-LSTM time-series hybrid model.

[0048] In this embodiment, the DeepSurv model is used, and step S5 is as follows: Based on the target individuals' stroke risk level, age, and comorbidity index, candidate decision-making schemes for different rehabilitation strategies are generated from a pre-set rehabilitation strategy library. The candidate decision-making schemes include: intensified drug therapy, endovascular stent placement, intracranial and extracranial vascular bypass surgery, and standard drug therapy.

[0049] Based on a pre-built rehabilitation strategy-specific prediction model, the residual stroke risk of the target subject at time t after receiving candidate decision options is predicted: ; In the formula, For the main effect parameters of the strategy, This is the feature-policy interaction effect vector. This is an indicator function that takes the value 1 when the patient actually receives the strategy k, and 0 otherwise. , To enhance the drug treatment regimen, This refers to the intravascular stent placement protocol. This refers to the intracranial and extracranial vascular bypass surgery plan, with "med" indicating the standard drug treatment plan.

[0050] Calculate the net rehabilitation benefit for the target population based on residual stroke risk and different candidate decision-making options. : ; In the formula, The residual stroke risk for the target subjects after receiving standard drug treatment is determined by the output of the rehabilitation strategy-specific prediction model when k=med; The residual risk after accepting candidate strategy k; The perioperative complication risk for candidate strategy k was estimated by propensity score matching based on the incidence of perioperative events in patients with similar baseline characteristics. This is a risk weighting coefficient, calibrated according to clinical severity, with a value ranging from 0.5 to 0.9; when k=med, =0, =0 is used as the reference standard.

[0051] Net recovery benefit of different candidate decision-making options Sort in descending order and select the largest. The corresponding strategy is the optimal decision-making scheme, when the top two When the difference is less than 5%, a less invasive strategy should be preferred.

[0052] During the training of the rehabilitation strategy-specific prediction model, clinical and imaging data of subjects with different risk levels using different rehabilitation strategies are acquired. After preprocessing, feature extraction, and feature fusion of the acquired data, a dataset is formed. The dataset is divided into training set, test set, and validation set in a 7:2:1 ratio to train, test, and validate the rehabilitation strategy-specific prediction model.

[0053] S6. Employ an incremental learning strategy to gradually improve the predictive capabilities of the stroke risk prediction model and the rehabilitation strategy-specific prediction model.

[0054] It should be noted that all subject data used in this embodiment were used with the consent of the subjects.

[0055] To verify the effectiveness of the method of the present invention, multimodal data of multiple subjects from multiple medical centers were obtained for comparison. The results of the comparative experiment are shown in Table 1.

[0056] Table 1. Comparison of different technologies

[0057] In contrast group 1, predictions based solely on radiomics and clinical features failed to comprehensively reflect the patient's overall condition, with a C-index of only 0.71. In clinical practice, the predictive accuracy of models constructed by different researchers using plaque features from imaging data varies significantly, limiting their applicability and reproducibility.

[0058] Compared to Group 2, directly concatenating multimodal features and inputting them into the DeepSurv model improved the C-index to 0.81. However, simple concatenation leads to feature redundancy and dimensionality explosion, and it fails to learn the correlation and complementarity between different modal features.

[0059] The proposed scheme achieves deep fusion of multimodal features through a multi-head cross-attention mechanism, which can adaptively learn the correlation and complementarity between different modalities. After combining multimodal features and reducing dimensionality through principal component analysis, the scheme is input into the DeepSurv model, achieving a C-index of 0.87. The AUC at each time period is significantly better than the previous two comparative schemes.

[0060] The present invention also provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and when the processor calls the computer program in the memory, it implements the content of the intracranial artery stenosis stroke risk assessment method based on multimodal data.

[0061] The present invention also provides a storage medium storing computer-executable instructions, which, when loaded and executed by a processor, implement the content of a method for assessing the risk of intracranial arterial stenosis stroke based on multimodal data.

[0062] Therefore, the present invention employs the above-mentioned intracranial artery stenosis stroke risk assessment method based on multimodal data, which can provide ICAS patients with accurate stroke risk assessment and optimal treatment plan.

[0063] 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for assessing the risk of stroke due to intracranial arterial stenosis based on multimodal data, characterized in that, Includes the following steps: S1. Acquire multimodal data of the target object and preprocess the multimodal data, which includes imaging data, clinical text data, detection data, biomarker data and the collection timestamps corresponding to each data. S2. Based on the preprocessed multimodal time series data, extract multimodal features and fuse the multimodal features based on cross-attention to obtain the initial multimodal fused features; S3. Principal component analysis is used to process the initial multimodal fusion features to obtain the final multimodal fusion features; S4. Input the multimodal fusion features from step S3 into the constructed stroke risk prediction model to obtain the stroke risk assessment results of the target object at different time periods, and determine the risk level based on the risk assessment results at different time periods. S5. Based on the stroke risk level and the pre-built rehabilitation strategy-specific prediction model, different rehabilitation programs are screened to obtain the optimal decision program for the target population.

2. The method for assessing the risk of stroke due to intracranial arterial stenosis based on multimodal data according to claim 1, characterized in that: Imaging data included high-resolution magnetic resonance vascular wall imaging, magnetic resonance angiography, arterial spin labeling perfusion imaging, diffusion-weighted imaging, and hemodynamic parameter maps based on computational fluid dynamics simulations; clinical data included the target subjects' medical history, clinical manifestations, family history, and stroke etiology classification text data extracted from discharge summaries using natural language processing; biomarker data included peripheral blood mononuclear cell gene expression profiles and serum glial fibrillary acidic protein levels; and detection data included lipid profiles, glycated hemoglobin, high-sensitivity C-reactive protein, homocysteine, and lipoprotein-associated phospholipase A2.

3. The method for assessing the risk of stroke due to intracranial arterial stenosis based on multimodal data according to claim 2, characterized in that: Step S2 specifically includes: Parallel extraction was employed to extract radiomics features, habitat radiomics features, hemodynamic features, detection features, clinical features, and biomarker features from preprocessed imaging data, clinical text data, detection data, and biomarker data, respectively. Among these, habitat radiomics features reflect plaque regional heterogeneity. After Z-score standardization of detection features, clinical features, and biomarker features, feature selection based on the maximum information coefficient is adopted to retain the feature subset related to the time of stroke occurrence. The selected detection features, clinical features, and biomarker features are tensor-concatenated to obtain the first fusion feature. Feature transformations were performed on the first fusion feature, radiomics feature, habitat radiomics feature, and hemodynamic feature respectively, mapping the features of different modalities to a unified feature space; A multi-head cross-attention mechanism is used to fuse features from different modalities. The feature of any modality is used as the query, and the features of the other three modalities are used as the key and value. The attention-weighted representation of the modality relative to the other modalities is calculated. The cross-attention outputs of all modalities are concatenated and projected to obtain the fused multimodal feature vector.

4. The method for assessing the risk of stroke due to intracranial arterial stenosis based on multimodal data according to claim 3, characterized in that: Hemodynamic characteristics include the cross-lesion pressure ratio, wall shear stress, oscillatory shear index, and vessel diameter stenosis rate, where the cross-lesion pressure ratio is the ratio of pressure distal to proximal to the stenosis.

5. The method for assessing the risk of stroke due to intracranial arterial stenosis based on multimodal data according to claim 1, characterized in that: The stroke risk prediction model is constructed using the DeepSurv model, which contains four fully connected hidden layers. It employs the SELU activation function and Dropout regularization, and uses the negative log-partial likelihood function as the loss function.

6. The method for assessing the risk of stroke due to intracranial arterial stenosis based on multimodal data according to claim 1, characterized in that: The target population is included in different time periods, including the predicted probability of stroke in the first, second, and third years of the target population. Risk levels are categorized into high risk, medium risk, and low risk, and the classification method is as follows: The probability of stroke in 1 year is ≤10%, the probability of stroke in 2 years is ≤15%, and the probability of stroke in 3 years is ≤20%, indicating a low dynamic risk. If the probability exceeds the corresponding threshold at any given time but does not meet the high-risk standard, it is considered a dynamic medium-risk situation. A stroke probability of >25% within 1 year or an increase of >50% in the probability of any period compared to the previous period indicates dynamic high risk and suggests unstable plaque progression.

7. The method for assessing the risk of stroke due to intracranial arterial stenosis based on multimodal data according to claim 1, characterized in that, Step S5 specifically includes: Based on the target group's stroke risk level, age, and comorbidity index, candidate decision-making schemes for different rehabilitation strategies are generated from a pre-set rehabilitation strategy library. Based on a pre-built rehabilitation strategy-specific prediction model, the residual stroke risk of the target subjects after receiving candidate decision-making options is predicted. Calculate the net rehabilitation benefit for the target population based on residual stroke risk and different candidate decision-making options. ; Net recovery benefit of different candidate decision-making options Sort according to net recovery benefit Generate the optimal decision-making scheme.

8. The method for assessing the risk of stroke due to intracranial arterial stenosis based on multimodal data according to claim 7, characterized in that: Machine learning models for constructing rehabilitation strategy-specific prediction models include DeepSurv model and TCN-LSTM time-series hybrid model.

Images