A method and system for acute large vessel occlusion identification and reperfusion decision making

Abstract

The application discloses an acute large blood vessel occlusion identification and reperfusion decision method and system, and the method comprises the following steps: S1, obtaining medical record text and head multi-modal CT image data of a patient to be diagnosed; S2, performing natural language processing on the medical record text to obtain text features; S3, processing the multi-modal CT image data, calculating the ASPECTS score of non-enhanced CT, and positioning the blood vessel occlusion position and evaluating the collateral circulation from the CTA image to obtain image features; S4, performing weighted fusion on the text features and the image features to generate fusion features; and S5, inputting the fusion features into a pre-constructed guideline knowledge graph and a decision model to generate an individualized acute large blood vessel occlusion identification result and a reperfusion treatment suggestion. The application can automatically, quickly and accurately generate an individualized identification result and a treatment suggestion by fusing the features of the medical record text and the multi-modal CT image and combining a real-time updated guideline knowledge graph.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence technology, specifically to a method and system for identifying and reperfusion decision-making in acute large vessel occlusion. Background Technology

[0002] Acute ischemic stroke with large vessel occlusion (AIS-LVO) is a severe type of acute ischemic stroke. Patient prognosis is closely related to the time window from onset to receiving effective reperfusion therapy. Therefore, timely and accurate identification and treatment decision-making are crucial.

[0003] Currently, in clinical practice, the identification and treatment decision-making process for AIS-LVO has the following limitations: 1. The diagnostic process is highly dependent on the physician's clinical experience. Physicians need to manually interpret the patient's computed tomography (CT) images (such as non-contrast CT and CT angiography) and analyze medical record text information (such as chief complaint and present illness). This process is time-consuming and easily affected by subjective factors in image interpretation and symptom assessment, resulting in diagnostic efficiency and consistency that need to be improved.

[0004] 2. Existing auxiliary analysis tools have limited capabilities. Traditional image analysis systems typically automate the processing of CT images in a single modality, such as calculating infarct areas or detecting vascular occlusion. They lack the ability to effectively integrate and jointly analyze clinical symptoms and medical history information in text form with multimodal image features, thus failing to fully utilize multi-source information for comprehensive judgment.

[0005] 3. There is a lag between clinical decision support and the latest medical evidence. Many decision support systems operate based on fixed rule bases, making it difficult to keep pace with rapidly updating clinical guidelines. As a result, the recommendations they generate may not reflect the current best treatment standards and are not timely enough.

[0006] 4. Existing technical solutions are mostly isolated functional modules. While independent medical image analysis devices or thrombectomy surgical equipment exist on the market, there is a lack of an integrated, end-to-end solution that can seamlessly connect front-end rapid identification, mid-end multimodal analysis, and back-end guideline-based personalized treatment decision recommendations.

[0007] Existing technologies mostly focus on the single analysis of medical images or the improvement of treatment devices. There is no complete system solution that effectively integrates natural language processing (for medical record text analysis), deep learning (for multimodal CT image processing), and dynamic knowledge graphs (for integrating and reasoning clinical guidelines) to achieve AIS-LVO identification and automatic generation of reperfusion treatment decisions.

[0008] Therefore, there is an urgent need to develop an AIS-LVO identification and treatment decision support system that can overcome the above limitations and achieve rapid, accurate, automated, and up-to-date clinical practice. Summary of the Invention

[0009] To overcome the shortcomings of the prior art, one of the objectives of this invention is to provide a method for identifying and reperfusion decision-making in acute large vessel occlusion, which can automatically, quickly and accurately generate personalized identification results and treatment recommendations by fusing features from medical record text and multimodal CT images and combining them with a real-time updated guideline knowledge graph.

[0010] The second objective of this invention is to provide an acute large vessel occlusion identification and reperfusion decision system, which can automatically, quickly and accurately generate personalized identification results and treatment suggestions by fusing features from medical record text and multimodal CT images and combining them with a real-time updated guideline knowledge graph.

[0011] To achieve one of the objectives of this invention, the following solution is adopted: A method for identifying and reperfusion decision-making in acute large vessel occlusion includes the following steps: S1. Obtain the medical record text and head multimodal CT image data of the patient to be diagnosed; S2. Perform natural language processing on the medical record text to obtain text features, including clinical symptom keyword embedding and onset time window features, and calculate the clinical severity score based on the text features; S3. Process the multimodal CT image data, calculate the ASPECTS score of non-contrast CT, locate the vascular occlusion location and assess collateral circulation from the CTA images, and obtain image features. S4. Based on a cross-modal attention mechanism and a learnable gating network, the text features and the image features are weighted and fused to generate fused features; S5. Input the fused features into the pre-built guide knowledge graph and decision model to generate personalized acute large vessel occlusion identification results and reperfusion treatment suggestions.

[0012] Furthermore, step S2 specifically includes: A medical BERT model adapted for acute stroke was used to extract clinical symptom keyword embeddings and onset time description features from the medical record text. A rule engine is used to extract structured time information from the medical record text, including the onset time, arrival time, and completion time of imaging examinations. The clinical symptom keywords, onset time description features, and structured time information are jointly input into the risk prediction model to predict the NIHSS score and the likelihood of acute large vessel occlusion. The clinical severity score includes the NIHSS score.

[0013] Furthermore, step S3 specifically includes: A deep learning model based on the U-Net architecture is used to analyze non-enhanced CT images and automatically calculate the ASPECTS score. A 3D CNN model based on ResNet3D-50 and a self-attention module was used to process CTA images to locate the vascular occlusion and assess collateral circulation from CTA images. A deep learning segmentation model combined with HU value threshold segmentation and morphological processing was used to detect bleeding in CT images and exclude contraindications for treatment.

[0014] Furthermore, in step S4, the weighted fusion specifically involves: The cross-modal attention mechanism is used to align text features with image features; The contribution weights of text features and image features are dynamically calculated through the learnable gating network. The aligned features are weighted and concatenated according to the contribution weights to form fused features.

[0015] Furthermore, in step S5, the pre-built guideline knowledge graph stores treatment rules including clinical treatment guidelines and their levels of evidence; the decision model is built based on rule reasoning and machine learning, and is used to reason based on the fused features and the guideline knowledge graph to generate treatment recommendations.

[0016] Furthermore, it also includes a real-time update step for the guideline knowledge graph: regularly and automatically capturing the latest clinical guidelines and literature, and updating them into the guideline knowledge graph after review.

[0017] Furthermore, after step S5, the following steps are also included: S6. When the confidence level of the identification result or treatment suggestion output by the decision model is lower than the preset threshold, the manual review process is triggered, and the relevant data is pushed to the review terminal. S7. Receive confirmation or correction instructions for the identification result or treatment suggestion returned by the verification terminal; S8. Based on the confirmation or correction instruction, generate optimized training data for incremental training and optimization of the decision model.

[0018] Furthermore, prior to step S1, a data preprocessing step is also included: The medical record text was standardized using a medical dictionary and rule base. The multimodal CT image data is subjected to grayscale normalization, noise reduction, and registration in a standard space.

[0019] Furthermore, the multimodal CT image data includes at least non-contrast CT and CTA images; the reperfusion treatment recommendations include at least one of intravenous thrombolysis, mechanical thrombectomy, or bridging therapy.

[0020] To achieve the second objective of this invention, the following solution is adopted: An acute large vessel occlusion identification and reperfusion decision-making system includes: The data acquisition module is used to acquire the medical record text and head multimodal CT image data of the patient to be diagnosed; The medical record text analysis module is used to perform natural language processing on the medical record text to obtain text features, including clinical symptom keyword embedding and onset time window features, and to calculate a clinical severity score based on the text features. The CT image analysis module is used to process the multimodal CT image data, calculate the ASPECTS score of non-contrast CT, locate the vascular occlusion location and assess collateral circulation from CTA images, and obtain image features. The multimodal fusion module is used to perform weighted fusion of the text features and the image features based on a cross-modal attention mechanism and a learnable gating network to generate fused features; The decision generation module is used to input the fused features into a pre-built guideline knowledge graph and decision model to generate personalized acute large vessel occlusion identification results and reperfusion treatment suggestions.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention significantly improves diagnostic efficiency and objectivity, saving valuable time for reperfusion therapy. By automating natural language processing and image analysis, this invention replaces the cumbersome process of relying entirely on doctors manually interpreting medical records and CT images. It drastically reduces the identification and assessment time for acute large vessel occlusion from the usual 30-60 minutes to within minutes, greatly reducing diagnostic delays and avoiding subjective judgment biases introduced by individual experience differences.

[0022] 2. This invention achieves more accurate lesion identification and assessment through multimodal deep information fusion. Instead of analyzing text or images in isolation, this invention innovatively utilizes cross-modal attention mechanisms and learnable gating networks to adaptively weight and deeply fuse clinical manifestations and imaging features. This fusion mechanism overcomes the limitations of single-modal analysis, significantly improving the accuracy of comprehensive judgment (up to over 92%), which is superior to traditional single-modal analysis methods.

[0023] 3. This invention generates highly compliant, individualized treatment decisions based on structured guideline knowledge. By combining deeply fused features with a pre-built structured guideline knowledge graph for reasoning, it can generate personalized reperfusion treatment recommendations that conform to authoritative clinical guideline standards. This solves the problem of inconsistent guideline compliance caused by reliance on personal memory or static rules in traditional decision-making, achieving a guideline compliance rate of over 98%.

[0024] 4. This invention provides a seamless end-to-end clinical decision support closed loop. Starting with data input, the invention proceeds through feature extraction, fusion analysis, and intelligent decision-making, forming a complete, coherent, and automated process. It effectively solves the problem of disconnect between the identification and decision-making stages in existing technologies, achieving integrated output from rapid identification to treatment suggestion generation, providing clinicians with an efficient and reliable one-stop auxiliary tool. Attached Figure Description

[0025] Figure 1 This is a block diagram of the acute large vessel occlusion identification and reperfusion decision system in an embodiment of the present invention; Figure 2 This is a block diagram of the overall architecture of the acute large vessel occlusion identification and reperfusion decision system in an embodiment of the present invention; Figure 3 This is a flowchart of multimodal data fusion in an embodiment of the present invention; Figure 4 This is a network structure diagram of the CT image analysis module in an embodiment of the present invention; Figure 5 This is a schematic diagram of the decision generation logic in an embodiment of the present invention; Figure 6 This is a flowchart of the acute large vessel occlusion identification and reperfusion decision method in an embodiment of the present invention. Detailed Implementation

[0026] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0027] Example 1 like Figures 1 to 5 As shown, this embodiment of the invention provides an acute large vessel occlusion identification and reperfusion decision-making system, comprising: The data acquisition module is used to acquire the medical record text and head multimodal CT image data of the patient to be diagnosed.

[0028] The medical record text analysis module is used to perform natural language processing on the medical record text to obtain text features, including clinical symptom keyword embedding and onset time window features, and to calculate a clinical severity score based on the text features.

[0029] The CT image analysis module is used to process the multimodal CT image data, calculate the ASPECTS score of non-contrast CT, locate the vascular occlusion location and assess collateral circulation from CTA images, and obtain image features.

[0030] The multimodal fusion module is used to perform weighted fusion of the text features and the image features based on a cross-modal attention mechanism and a learnable gating network to generate fused features.

[0031] The decision generation module is used to input the fused features into a pre-built guideline knowledge graph and decision model to generate personalized acute large vessel occlusion identification results and reperfusion treatment suggestions.

[0032] The acute large vessel occlusion identification and reperfusion decision system of the present invention will be further described in detail below.

[0033] The acute large vessel occlusion identification and reperfusion decision system of this invention adopts a multimodal data fusion architecture.

[0034] In this embodiment, to implement the medical record text analysis function, the system has a dedicated medical record text analysis module. The specific contents of the medical record text analysis module are as follows: This module employs a BERT-based natural language processing model to extract key clinical symptoms (such as hemiplegia, aphasia, and limb weakness) from medical record texts. The specific configuration and training method are as follows: Model selection and training: A pre-trained Chinese medical BERT model was used as the foundation. Based on this, domain-adaptive training was performed using medical literature, clinical guidelines and electronic medical record data in the field of acute stroke to enhance the model's understanding of professional terminology and context.

[0035] Training data: The data comes from anonymized electronic medical records of multiple collaborating secondary and tertiary hospitals, totaling approximately 3,000 emergency department text records, covering chief complaints, present illness history, imaging reports, etc. Annotation was completed by three neurologists according to standard guidelines, including symptom keywords, time windows, and clinical severity, with an annotation consistency Kappa value greater than 0.85.

[0036] Feature extraction: This module extracts text features including symptom keyword embeddings (such as "sudden onset of hemiplegia" and "slurred speech"), time window features (such as the time from onset to consultation) and clinical score-related features (such as NIHSS score keywords). The BERT model outputs 768-dimensional features and extracts key features through an attention pooling mechanism.

[0037] Supplementary processing: Integrating regular expressions and a rule engine, structured time information is accurately extracted, including onset time, arrival time, consultation time, and completion time of imaging examinations. Based on the extracted features, a risk prediction model is constructed to calculate the NIHSS score and the probability of acute large vessel occlusion.

[0038] In this embodiment, multimodal CT images are processed through a CT image analysis module. The specific contents of the CT image analysis module are as follows: This module is responsible for processing multimodal CT images and specifically includes three functional parts: NCCT Image Analysis: A deep learning model based on the U-Net architecture analyzes non-contrast CT images and automatically calculates the ASPECTS score. This U-Net employs a 5-layer encoder-decoder structure, with each layer having a 3×3 kernel size and ReLU activation function. A combination of Dice Loss and Focal Loss is used as the loss function to address the imbalance between ischemic areas and normal tissue in the training samples.

[0039] CTA image analysis: A 3D CNN network based on ResNet3D-50 was used to process CTA images, with input data dimensions of 128×128×64×1. A self-attention module was added to the network to enhance the model's ability to focus on key locations of vascular occlusion, thereby achieving accurate localization of vascular occlusion and assessment of collateral circulation status.

[0040] Bleeding Detection: An integrated bleeding detection algorithm is used to exclude contraindications for treatment. This algorithm employs a deep learning model (such as U-Net++) for hematoma segmentation, supplemented by threshold segmentation (HU value > 45) and morphological processing (such as opening operations to remove noise), and finally outputs a bleeding probability map and hematoma volume estimation.

[0041] In this embodiment, the multimodal fusion module and the decision generation module are comprised of the multimodal fusion decision module, the specific contents of which are as follows: This module aims to deeply integrate text and image information and generate decisions based on clinical knowledge. Feature fusion: Fusion method: Feature-level fusion is adopted. Text features (768 dimensions) and image features (1024 dimensions) are aligned through a cross-modal attention mechanism, and then weighted and concatenated to form a 1536-dimensional fused feature.

[0042] Weighted learning: Dynamically adjust the contribution weights of text features and image features in the fusion process through a learnable gating network.

[0043] Feature optimization: Principal component analysis was used to reduce the 1536-dimensional fused features to 512-dimensional features, and then the nonlinear expressive power of the features was further enhanced through a fully connected layer.

[0044] Decision Generation: A knowledge graph incorporating the latest clinical guidelines is established. A feature-level fusion network is designed to integrate textual and image features, and personalized treatment recommendations are generated based on rule-based reasoning and machine learning methods, combined with the guideline knowledge graph.

[0045] The system architecture of this invention follows a clear hierarchical data processing flow, namely: data input layer → preprocessing layer → feature extraction layer → fusion decision layer → output layer, realizing end-to-end automated processing from raw data to treatment decisions.

[0046] The following section provides a detailed explanation of the entire process from data preparation to system deployment, using specific data as examples.

[0047] First, during the data collection and preprocessing phase, a multicenter retrospective dataset from five hospitals (a total of 3,200 acute stroke patients) was used, and all data collection was ethically approved. To enhance the model's generalization ability, publicly available datasets such as ATLAS and ISLES were supplemented for data augmentation. Key metrics, such as ASPECTS scores, were independently labeled by two physicians with associate chief physician titles or above. Cases with discrepancies were arbitrated by a third expert to ensure labeling quality, with an intraclass correlation coefficient (ICC) greater than 0.9. The ratio of positive samples from large vessel occlusion to negative samples from non-large vessel occlusion in the dataset was approximately 1:3. To address class imbalance, oversampling techniques and data augmentation methods such as elastic deformation were employed to balance the sample distribution.

[0048] After data acquisition, preprocessing was performed: patient medical records and DICOM format head CT images were obtained from the Hospital Information System (HIS) and Picture Archiving and Communication System (PACS). For the medical record text data, medical terminology was standardized using a Chinese medical dictionary (such as CMeDict), and key time windows and symptom keywords were extracted using rule bases (such as regular expressions). For the CT image data, the following standardization process was performed: grayscale normalization was performed, truncating the Henlein Unit (HU) values ​​to the [-100, 400] range, and then linearly normalizing them to the [0, 1] range; a nonlocal mean denoising algorithm was applied to reduce image noise; and based on the Elastix toolkit, affine transformations were used to register the images to the standard brain template (MNI) space to eliminate inter-individual anatomical differences.

[0049] Next, in the model training and optimization phase, a dataset containing approximately 5,000 patients diagnosed with acute large vessel occlusion was used for model training. A transfer learning strategy was employed, with domain-specific fine-tuning performed on a pre-trained model on large public datasets such as ImageNet. Simultaneously, a multi-task learning framework was used to optimize both vessel occlusion identification and precise localization tasks. The hyperparameter settings were as follows: the optimizer was AdamW (learning rate set to 2e-5, weight decay to 0.01), the batch size was 16, the number of training epochs was 100, the learning rate scheduling used a cosine annealing strategy, and a 10% warm-up training step was performed in the initial stage.

[0050] For model validation and evaluation, 5-fold cross-validation was used, and 20% of the total data was independently reserved as a test set. The main evaluation metrics included the area under the curve (AUC, with a target value greater than 0.95), F1 score, sensitivity, and specificity. The 95% confidence intervals for each metric were calculated to assess the stability of the model performance.

[0051] Furthermore, regarding system construction and deployment, a guideline knowledge graph is first constructed: the Neo4j graph database is used to store knowledge such as symptom-disease associations, rules on contraindications for reperfusion therapy, and levels of evidence. The system has a knowledge update mechanism: it automatically retrieves the latest literature from authoritative guideline publishing websites monthly; it has an API monitoring function that automatically triggers update alerts when new guidelines are released by institutions such as the CDC; all automatically retrieved content must be manually reviewed and confirmed before being updated to the knowledge graph. In knowledge reasoning, the Drools rule engine is integrated. An example reasoning path is: if the symptom is "sudden hemiplegia" and imaging indicates "internal carotid artery (ICA) occlusion," the system automatically recommends a "mechanical thrombectomy" option (with Class I evidence level).

[0052] This invention ensures real-time, personalized treatment decisions and high adherence to the latest clinical guidelines. It combines fusion features obtained from deep analysis with a pre-built and updatable guideline knowledge graph to generate treatment recommendations through a decision model. This guarantees that the output plan instantly aligns with the latest authoritative clinical guidelines, solving the problem of lagging updates in traditional decision support systems, achieving a guideline compliance rate of over 98%, and providing personalized reperfusion strategy recommendations for each patient.

[0053] For system integration and interface design, Docker containerization technology is used for deployment to ensure environmental consistency and facilitate integration with the hospital's existing information systems. A RESTful API interface is provided for clinical systems to call, with the main endpoint being POST / api / stroke / prediction. The request body is in JSON format and includes the patient ID, text data, and image storage link. The interface response is also in JSON format, for example: {"decision": "mechanical_thrombectomy", "confidence":0.92, "contraindications": ["hemorrhagic transformation"]}.

[0054] In terms of security mechanisms, HTTPS protocol (TLS 1.3) is used for end-to-end encrypted transmission to ensure data security; and a role-based three-level access control is implemented: physicians have full access and operation permissions, nurses have read-only permissions for results, and system administrators are responsible for user management and configuration permissions.

[0055] To illustrate the system workflow, a clinical application example is provided below: Enter the medical record text: "Male, 68 years old, sudden onset of right limb weakness for 2 hours, history of hypertension"; Enter the imaging data: the patient's non-contrast CT (NCCT) and CT angiography (CTA) images.

[0056] The system outputs the following results after processing: "Occlusion of the M1 segment of the left middle cerebral artery, ASPECTS score of 8, estimated NIHSS score of 12"; based on the recognition results and the guideline knowledge graph, the treatment recommendation is "Meets the criteria for intravenous thrombolysis combined with mechanical thrombectomy bridging therapy, and it is recommended to start immediately".

[0057] To verify the actual effectiveness of the acute large vessel occlusion identification and reperfusion decision system described in this invention, a clinical validation study was conducted at the stroke center of a tertiary hospital.

[0058] This invention included 200 patients clinically suspected of having acute large vessel occlusion (AIS-LVO). The identification results and treatment recommendations generated by the system of this invention were compared and statistically analyzed in a blinded manner with the "gold standard" (including final imaging diagnosis and treatment decisions based on the latest guidelines) composed of a team of senior stroke experts.

[0059] The system of this invention exhibits excellent performance, and the verification results are as follows: Sensitivity: 94.3%; Specificity: 91.2%; Average decision time: 4.7 minutes; Guideline compliance rate: 98.5%.

[0060] The above results demonstrate that the system of the present invention can quickly and accurately identify acute large vessel occlusion, and the reperfusion treatment decisions it generates have high clinical guideline compliance, significantly improving the efficiency and consistency of diagnostic decisions.

[0061] The following table compares the acute large vessel occlusion identification and reperfusion decision-making system of this invention with existing technologies: Experimental example: During the validation process, rigorous statistical methods were employed: 95% confidence intervals were calculated for all performance indicators. To quantify the improvement of this system compared to existing technologies, the McNemar test was used to compare its performance with baseline models (such as RAPID software), with a p-value < 0.05 defined as statistically significant.

[0062] The experimental setup mainly consisted of two comparisons: first, comparing the system's automatically generated diagnostic results and treatment recommendations with the diagnostic and recommendation decisions of an expert group composed of multiple senior stroke specialists, using this as the clinical "gold standard"; second, conducting generalization tests to verify the system's stability in different scenarios. Generalization tests included: cross-device testing on data acquired from three different brands of CT equipment (GE, Siemens, and Philips), with results showing that the area under the curve (AUC) for the key indicator fluctuated within a range of <0.03; and cross-center testing on clinical data from two hospitals that did not participate in model training, where the system performance remained stable with an F1 score >0.88.

[0063] Failure Case Analysis and Handling Mechanism: A dedicated analysis and handling process has been established to address potential challenges the system may encounter in practical applications. First, typical scenarios prone to system misjudgment are summarized, including: atypical patient clinical symptoms (e.g., posterior circulation infarction presenting only as dizziness), severe motion artifacts in CT images, and unclear records of the patient's onset time.

[0064] To address these uncertainties, the system employs a tiered processing mechanism: when the confidence level of the system's output identification results or treatment recommendations falls below a preset threshold (0.7), a manual review flag is automatically triggered, and all relevant information for the case is pushed to the senior physician's terminal for review. The system provides an embedded review interface where physicians can view the system's analysis results and have the right to confirm or override the system's decisions based on their professional judgment. The corrected results are automatically fed back into the model optimization loop as feedback data. Furthermore, a failure case database is established to continuously collect and analyze erroneous cases. These collected erroneous cases are used as new training data for incremental training and optimization of the model. Optimization measures include, but are not limited to, adding a posterior cycle symptom feature module.

[0065] Example 2 like Figure 6 As shown, this embodiment of the invention provides a method for identifying and reperfusion decision-making in cases of acute large vessel occlusion, comprising the following steps: S1. Obtain the patient's medical record text and head multimodal CT image data.

[0066] S2. Perform natural language processing on the medical record text to obtain text features, including clinical symptom keyword embedding and onset time window features, and calculate the clinical severity score based on the text features.

[0067] S3. Process the multimodal CT image data, calculate the ASPECTS score of non-contrast CT, locate the vascular occlusion location and assess collateral circulation from the CTA images, and obtain image features.

[0068] S4. Based on cross-modal attention mechanism and learnable gating network, the text features and the image features are weighted and fused to generate fused features.

[0069] S5. Input the fused features into the pre-built guide knowledge graph and decision model to generate personalized acute large vessel occlusion identification results and reperfusion treatment suggestions.

[0070] Furthermore, step S2 specifically includes: A medical BERT model adapted for acute stroke was used to extract clinical symptom keyword embeddings and onset time description features from the medical record text. A rule engine is used to extract structured time information from the medical record text, including the onset time, arrival time, and completion time of imaging examinations. The clinical symptom keywords, onset time description features, and structured time information are jointly input into the risk prediction model to predict the NIHSS score and the likelihood of acute large vessel occlusion. The clinical severity score includes the NIHSS score.

[0071] Furthermore, step S3 specifically includes: A deep learning model based on the U-Net architecture is used to analyze non-enhanced CT images and automatically calculate the ASPECTS score. A 3D CNN model based on ResNet3D-50 and a self-attention module was used to process CTA images to locate the vascular occlusion and assess collateral circulation from CTA images. A deep learning segmentation model combined with HU value threshold segmentation and morphological processing was used to detect bleeding in CT images and exclude contraindications for treatment.

[0072] Furthermore, in step S4, the weighted fusion specifically involves: The cross-modal attention mechanism is used to align text features with image features; The contribution weights of text features and image features are dynamically calculated through the learnable gating network. The aligned features are weighted and concatenated according to the contribution weights to form fused features.

[0073] Furthermore, in step S5, the pre-built guideline knowledge graph stores treatment rules including clinical treatment guidelines and their levels of evidence; the decision model is built based on rule reasoning and machine learning, and is used to reason based on the fused features and the guideline knowledge graph to generate treatment recommendations.

[0074] Furthermore, it also includes a real-time update step for the guideline knowledge graph: regularly and automatically capturing the latest clinical guidelines and literature, and updating them into the guideline knowledge graph after review.

[0075] Furthermore, after step S5, the following steps are also included: S6. When the confidence level of the identification result or treatment suggestion output by the decision model is lower than the preset threshold, the manual review process is triggered, and the relevant data is pushed to the review terminal. S7. Receive confirmation or correction instructions for the identification result or treatment suggestion returned by the verification terminal; S8. Based on the confirmation or correction instruction, generate optimized training data for incremental training and optimization of the decision model.

[0076] Furthermore, prior to step S1, a data preprocessing step is also included: The medical record text was standardized using a medical dictionary and rule base. The multimodal CT image data is subjected to grayscale normalization, noise reduction, and registration in a standard space.

[0077] Furthermore, the multimodal CT image data includes at least non-contrast CT and CTA images; the reperfusion treatment recommendations include at least one of intravenous thrombolysis, mechanical thrombectomy, or bridging therapy.

[0078] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A method for identifying and reperfusion decision-making in cases of acute large vessel occlusion, characterized in that, Includes the following steps: S1. Obtain the medical record text and head multimodal CT image data of the patient to be diagnosed; S2. Perform natural language processing on the medical record text to obtain text features, including clinical symptom keyword embedding and onset time window features, and calculate the clinical severity score based on the text features; S3. Process the multimodal CT image data, calculate the ASPECTS score of non-contrast CT, locate the vascular occlusion location and assess collateral circulation from the CTA images, and obtain image features. S4. Based on a cross-modal attention mechanism and a learnable gating network, the text features and the image features are weighted and fused to generate fused features; S5. Input the fused features into the pre-built guide knowledge graph and decision model to generate personalized acute large vessel occlusion identification results and reperfusion treatment suggestions.

2. The method for identifying and reperfusion decision-making in acute large vessel occlusion according to claim 1, characterized in that, Step S2 specifically includes: A medical BERT model adapted for acute stroke was used to extract clinical symptom keyword embeddings and onset time description features from the medical record text. A rule engine is used to extract structured time information from the medical record text, including the onset time, arrival time, and completion time of imaging examinations. The clinical symptom keywords, onset time description features, and structured time information are jointly input into the risk prediction model to predict the NIHSS score and the likelihood of acute large vessel occlusion. The clinical severity score includes the NIHSS score.

3. The method for identifying and reperfusion decision-making in acute large vessel occlusion according to claim 1, characterized in that, Step S3 specifically includes: A deep learning model based on the U-Net architecture is used to analyze non-enhanced CT images and automatically calculate the ASPECTS score. A 3D CNN model based on ResNet3D-50 and a self-attention module was used to process CTA images to locate the vascular occlusion and assess collateral circulation from CTA images. A deep learning segmentation model combined with HU value threshold segmentation and morphological processing was used to detect bleeding in CT images and exclude contraindications for treatment.

4. The method for identifying and reperfusion decision-making in acute large vessel occlusion according to claim 1, characterized in that, In step S4, the weighted fusion specifically involves: The cross-modal attention mechanism is used to align text features with image features; The contribution weights of text features and image features are dynamically calculated through the learnable gating network. The aligned features are weighted and concatenated according to the contribution weights to form fused features.

5. The method for identifying and reperfusion decision-making in acute large vessel occlusion according to claim 1, characterized in that, In step S5, the pre-built guideline knowledge graph stores treatment rules including clinical treatment guidelines and their levels of evidence; the decision model is built based on rule reasoning and machine learning, and is used to reason based on the fused features and the guideline knowledge graph to generate treatment recommendations.

6. The method for identifying and reperfusion decision-making in acute large vessel occlusion according to claim 1, characterized in that, It also includes a real-time update process for the guideline knowledge graph: regularly and automatically capturing the latest clinical guidelines and literature, and updating them into the guideline knowledge graph after review.

7. The method for identifying and reperfusion decision-making in acute large vessel occlusion according to claim 1, characterized in that, Following step S5, the following is also included: S6. When the confidence level of the identification result or treatment suggestion output by the decision model is lower than the preset threshold, the manual review process is triggered, and the relevant data is pushed to the review terminal. S7. Receive confirmation or correction instructions for the identification result or treatment suggestion returned by the verification terminal; S8. Based on the confirmation or correction instruction, generate optimized training data for incremental training and optimization of the decision model.

8. The method for identifying and reperfusion decision-making in acute large vessel occlusion according to claim 1, characterized in that, Before step S1, a data preprocessing step is also included: The medical record text was standardized using a medical dictionary and rule base. The multimodal CT image data is subjected to grayscale normalization, noise reduction, and registration in a standard space.

9. The method for identifying and reperfusion decision-making in acute large vessel occlusion according to claim 1, characterized in that, The multimodal CT imaging data includes at least non-contrast CT and CTA images; the reperfusion treatment recommendations include at least one of intravenous thrombolysis, mechanical thrombectomy, or bridging therapy.

10. An acute large vessel occlusion identification and reperfusion decision-making system, characterized in that, include: The data acquisition module is used to acquire the medical record text and head multimodal CT image data of the patient to be diagnosed; The medical record text analysis module is used to perform natural language processing on the medical record text to obtain text features, including clinical symptom keyword embedding and onset time window features, and to calculate a clinical severity score based on the text features. The CT image analysis module is used to process the multimodal CT image data, calculate the ASPECTS score of non-contrast CT, locate the vascular occlusion location and assess collateral circulation from CTA images, and obtain image features. The multimodal fusion module is used to perform weighted fusion of the text features and the image features based on a cross-modal attention mechanism and a learnable gating network to generate fused features; The decision generation module is used to input the fused features into a pre-built guideline knowledge graph and decision model to generate personalized acute large vessel occlusion identification results and reperfusion treatment suggestions.

Images