A knowledge graph-based rehabilitation diagnosis and treatment system and method for cerebrovascular diseases

Abstract

This application provides a knowledge graph-based rehabilitation diagnosis and treatment system and method for cerebrovascular diseases. It maps multimodal rehabilitation data onto a pre-defined knowledge graph, generating a feature subgraph. Starting with the core functional impairment node in the patient's feature subgraph, a multi-hop relationship expansion search is performed in the knowledge graph to obtain a set of candidate rehabilitation intervention nodes. Based on the empirical validity data of each candidate rehabilitation intervention node in a historical patient group with similar characteristics to the current patient, the topological association strength of the knowledge graph, and the feasibility constraints of the target patient's current medical scenario, the personalized fit of each candidate rehabilitation intervention node is determined. The results of ranking each candidate rehabilitation intervention node according to all personalized fits form a draft recommendation of a structured rehabilitation plan for the target patient. Using the scheme of this application, multimodal rehabilitation data and dynamic medical knowledge of patients can be integrated to conduct intelligent rehabilitation plan recommendations.

Description

Technical Field

[0001] This application relates to the field of medical information technology, and in particular to a knowledge graph-based rehabilitation diagnosis and treatment system and method for cerebrovascular diseases. Background Technology

[0002] Cerebrovascular disease is one of the leading causes of disability, often leaving patients with varying degrees of motor, cognitive, and speech impairments. Rehabilitation is a long-term, complex, and highly individualized process. Precise and efficient rehabilitation programs are key to improving patients' functional outcomes and quality of life. However, currently, the development of clinical rehabilitation programs mainly relies on the experience of rehabilitation physicians and therapists. This approach faces many challenges, such as insufficient integration and utilization of information for clinical decision-making, barriers and delays in the application of medical knowledge, and limited personalization and precision.

[0003] In recent years, knowledge graph technology has provided a powerful tool for the structured representation of medical knowledge systems. Some studies have attempted to construct disease knowledge graphs to assist in diagnosis or recommend treatment plans. However, in the sub-field of cerebrovascular disease rehabilitation, existing technical solutions typically suffer from the following fundamental defects or deficiencies: directly mapping multimodal patient data to concept nodes in the knowledge graph ignores the necessary transformation and reasoning process from underlying data to high-level clinical semantic information, resulting in flawed technical logic; relying primarily on static knowledge graphs for logical reasoning or simple retrieval leads to overly theoretical recommendations lacking validation and calibration from large-scale historical efficacy data from the real world, resulting in insufficient practicality and credibility; and the systems are mostly open-loop designs, unable to continuously learn and optimize from clinician decision feedback and patient efficacy tracking, leading to stagnant intelligence levels. Therefore, how to integrate patients' multimodal rehabilitation data and dynamic medical knowledge to provide intelligent rehabilitation plan recommendations has become a challenge for the industry. Summary of the Invention

[0004] Based on this, this application provides a knowledge graph-based rehabilitation diagnosis and treatment system and method for cerebrovascular diseases that integrates patients' multimodal rehabilitation data and dynamic medical knowledge to make intelligent rehabilitation program recommendations.

[0005] Firstly, this application provides a rehabilitation information retrieval method for the rehabilitation diagnosis and treatment of cerebrovascular diseases, applied to a knowledge graph-based cerebrovascular disease rehabilitation diagnosis and treatment system. The method includes the following steps: Acquire multimodal rehabilitation data from target patients; The multimodal rehabilitation data is mapped to a preset cerebrovascular disease rehabilitation knowledge graph to generate a feature subgraph of the target patient. The cerebrovascular disease rehabilitation knowledge graph contains the medical logical relationships between disease, functional impairment, assessment indicators, rehabilitation intervention measures and expected efficacy nodes. Starting with the core functional impairment node in the patient feature subgraph, and combining it with the clinical diagnosis and treatment stage label, a multi-hop relationship expansion retrieval is performed in the cerebrovascular disease rehabilitation knowledge graph to obtain a set of candidate rehabilitation intervention nodes that are associated with the patient's current state. Based on empirical validity data from historical patient groups with similar characteristics to the current patient for each candidate rehabilitation intervention node, the topological association strength of the cerebrovascular disease rehabilitation knowledge graph, and the feasibility constraints of the target patient's current medical scenario, the personalized fit of each candidate rehabilitation intervention node is determined. Based on all individualized fits, each candidate rehabilitation intervention node is ranked, and a draft recommendation of a structured rehabilitation program for the target patient is formed based on the ranking results.

[0006] In some embodiments, mapping the multimodal rehabilitation data to a preset cerebrovascular disease rehabilitation knowledge graph to generate a feature subgraph of the target patient specifically includes: Obtain a pre-defined knowledge graph of cerebrovascular disease rehabilitation; Feature extraction and diagnostic reasoning are performed on the medical imaging data in the multimodal rehabilitation data, and the results are associated with the corresponding disease and pathological basic nodes in the cerebrovascular disease rehabilitation knowledge graph. The structured assessment scale data in the multimodal rehabilitation data are standardized and transformed, and the results are associated with the corresponding functional impairment and assessment indicator nodes in the cerebrovascular disease rehabilitation knowledge graph. Medical entity recognition and extraction are performed on the unstructured medical record text data in the multimodal rehabilitation data, and the results are associated with the corresponding symptoms, signs and comorbidity nodes in the cerebrovascular disease rehabilitation knowledge graph. Based on all associations, all associated nodes and predefined medical logical relationship edges between associated nodes are extracted from the cerebrovascular disease rehabilitation knowledge graph to form a feature subgraph representing the current clinical state of the target patient.

[0007] In some embodiments, based on all associations, extracting all associated nodes and predefined medical logical relationship edges between the associated nodes from the cerebrovascular disease rehabilitation knowledge graph to construct a feature subgraph representing the current clinical state of the target patient specifically includes: Create and initialize a central node representing the target patient, and link the processed association results of the medical imaging data, structured assessment scale data, and unstructured medical record text data to the central node respectively; Collect all the associated edges directly connected to the central node, and obtain all directly associated disease nodes, pathological basis nodes, functional impairment nodes, assessment indicator nodes, symptom nodes, sign nodes, and comorbidity nodes; Query the cerebrovascular disease rehabilitation knowledge graph and extract the predefined medical logical relationship edges between all directly related nodes; The central node, all directly associated nodes, and the extracted medical logical relationship edges are integrated into an independent graph data structure to obtain a feature subgraph of the target patient's current clinical state.

[0008] In some embodiments, starting with the core functional impairment node in the patient feature subgraph and combining it with clinical treatment stage tags, a multi-hop relationship expansion retrieval is performed in the cerebrovascular disease rehabilitation knowledge graph to obtain a set of candidate rehabilitation intervention nodes that are associated with the patient's current state. Specifically, this set includes: Based on the severity and influence weight of each functional impairment node in the patient feature subgraph, at least one core functional impairment node is identified. Determine the whitelist of relationship types, path evidence-based level thresholds, and extension hop limits corresponding to the tags of clinical diagnosis and treatment stages to form search constraint rules; Using the core functional impairment node as the initial node, and in accordance with the retrieval constraint rules, a constrained multi-hop graph traversal is performed in the cerebrovascular disease rehabilitation knowledge graph; Collect all nodes visited during the multi-hop graph traversal process that are of the type of rehabilitation intervention measures, perform deduplication and path confidence integration to form the candidate rehabilitation intervention node set.

[0009] In some embodiments, determining a whitelist of relationship types, a path evidence-based level threshold, and an extension hop limit corresponding to clinical diagnosis and treatment stage tags to form retrieval constraint rules specifically includes: Obtain the clinical treatment stage labels from the patient feature submap; Using the clinical diagnosis and treatment stage label as the key, query the preset stage-constraint rule mapping table to obtain the corresponding relationship type whitelist, path evidence-based level threshold, and extension hop limit; The whitelist of relation types, path evidence level thresholds, and extended hop limit obtained from the query are output and encapsulated into retrieval constraint rules.

[0010] In some embodiments, determining the personalized fit of each candidate rehabilitation intervention node based on empirical validity data from historical patient groups similar to the current patient's characteristics, the topological association strength of the cerebrovascular disease rehabilitation knowledge graph, and the feasibility constraints of the target patient's current medical scenario specifically includes: For each candidate rehabilitation intervention node, historical patient groups with similar characteristics to the target patient are retrieved from the historical medical database, and the empirical validity score of the candidate rehabilitation intervention node is determined based on the efficacy data corresponding to the candidate rehabilitation intervention node and the historical patient groups. From the multi-hop relation expansion retrieval results, the calculated association confidence of the candidate rehabilitation intervention node, which represents the theoretical association strength with the patient's core functional impairment node, is extracted as the topological association strength score of the candidate rehabilitation intervention node. Based on the resource allocation and constraint rule base of the target patient's current medical scenario, assess the feasibility of implementing candidate rehabilitation intervention nodes in the current medical scenario, and calculate the feasibility constraint score of candidate rehabilitation intervention nodes; The empirical validity score, the topological correlation strength score, and the scenario feasibility constraint score are calculated using a linear weighted fusion algorithm with preset weights to obtain the personalized fit of the selected rehabilitation intervention node, thereby obtaining the personalized fit of each candidate rehabilitation intervention node.

[0011] In some embodiments, the multimodal rehabilitation data includes medical imaging data, structured assessment scale data, and unstructured medical record text data.

[0012] Secondly, this application provides a knowledge graph-based rehabilitation and treatment system for cerebrovascular diseases, which includes a rehabilitation information retrieval unit, the rehabilitation information retrieval unit comprising: The acquisition module is used to acquire multimodal rehabilitation data of the target patient; The processing module is used to map the multimodal rehabilitation data to a preset cerebrovascular disease rehabilitation knowledge graph to generate a feature subgraph of the target patient. The cerebrovascular disease rehabilitation knowledge graph contains medical logical relationships between disease, functional impairment, assessment indicators, rehabilitation intervention measures and expected efficacy nodes. The processing module is also used to perform multi-hop relationship expansion retrieval in the cerebrovascular disease rehabilitation knowledge graph, starting from the core functional impairment node in the patient feature subgraph and combining the clinical diagnosis and treatment stage label, to obtain a set of candidate rehabilitation intervention nodes that are associated with the patient's current state. The processing module is also used to determine the personalized fit of each candidate rehabilitation intervention node based on empirical validity data from historical patient groups with similar characteristics to the current patient, the topological association strength of the cerebrovascular disease rehabilitation knowledge graph, and the feasibility constraints of the target patient's current medical scenario. The execution module is used to sort each candidate rehabilitation intervention node according to all personalized adaptations, and to form a recommended draft of a structured rehabilitation plan for the target patient based on the sorting results.

[0013] Thirdly, this application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described rehabilitation information retrieval method for cerebrovascular disease rehabilitation diagnosis and treatment.

[0014] Fourthly, this application provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described rehabilitation information retrieval method for cerebrovascular disease rehabilitation diagnosis and treatment.

[0015] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects: The knowledge graph-based rehabilitation diagnosis and treatment system and method for cerebrovascular diseases provided in this application firstly acquires multimodal rehabilitation data of the target patient, maps the multimodal rehabilitation data to a preset cerebrovascular disease rehabilitation knowledge graph, and generates a feature subgraph of the target patient. The cerebrovascular disease rehabilitation knowledge graph contains medical logical relationships between disease, functional impairment, assessment indicators, rehabilitation interventions, and expected efficacy nodes. This step can accurately map and integrate heterogeneous and multi-source patient data into a unified medical knowledge framework (knowledge graph) through standardized processing and pre-trained models, thereby associating abstract clinical concepts with specific patient information and generating a comprehensive and structured feature subgraph, which provides a basis for subsequent precise diagnosis and treatment. The retrieval and reasoning process provides a reliable and semantically rich foundation of personalized context. Secondly, starting with the core functional impairment nodes in the patient feature subgraph and combining them with clinical treatment stage tags, a multi-hop relation expansion retrieval is performed in the cerebrovascular disease rehabilitation knowledge graph to obtain a set of candidate rehabilitation intervention nodes related to the patient's current state. This step enables controlled and medically logical graph path exploration, using the patient's most core functional impairment as a precise starting point and combining it with the clinical rules of their current rehabilitation stage. This systematically discovers a set of potential rehabilitation interventions that are directly and indirectly theoretically related to the patient's current state, ensuring the clinical relevance, stage appropriateness, and logical coherence of the retrieval results, and avoiding retrieval errors. Blindness and irrelevance; then, based on empirical validity data from historical patient groups similar to the current patient's characteristics, the topological association strength of the cerebrovascular disease rehabilitation knowledge graph, and the feasibility constraints of the target patient's current medical scenario, the personalized fit of each candidate rehabilitation intervention node is determined. This step can calculate the empirical validity score based on practical evidence from similar patient groups, the topological association strength score based on knowledge graph theory, and the feasibility constraint score based on real-world scenario conditions for each candidate intervention measure. This yields a quantitative score that comprehensively reflects the measure's fit in terms of evidence-based, logical, and operable aspects, expanding the recommendation basis from a single dimension to a multi-dimensional fusion, greatly improving the efficiency of the recommendation. This improves the scientific rigor, personalization, and practical feasibility of the recommendations. Finally, based on all personalized fits, each candidate rehabilitation intervention node is ranked, and a draft recommendation of a structured rehabilitation plan for the target patient is formed based on the ranking results. This step objectively ranks all candidate measures according to their comprehensive fit and automatically extracts and assembles detailed plan components associated with the top-ranked measures, generating a clearly structured, complete draft recommendation with indicated recommendation strength and evidence sources. This provides rehabilitation physicians with an efficient, intuitive, and traceable decision support blueprint. In summary, the proposed solution integrates patients' multimodal rehabilitation data and dynamic medical knowledge to provide intelligent rehabilitation plan recommendations. Attached Figure Description

[0016] Figure 1 This is an exemplary flowchart of a rehabilitation information retrieval method for cerebrovascular disease rehabilitation diagnosis and treatment, as shown in some embodiments of this application. Figure 2 This is a schematic diagram illustrating an application scenario of a rehabilitation program recommendation data processing system according to some embodiments of this application; Figure 3 This is a flowchart illustrating the process of determining a set of candidate rehabilitation intervention nodes according to some embodiments of this application; Figure 4 This is a schematic diagram of the structure of a rehabilitation information retrieval unit according to some embodiments of this application; Figure 5 This is a schematic diagram of the structure of a computer device that implements a rehabilitation information retrieval method for the diagnosis and treatment of cerebrovascular diseases, according to some embodiments of this application. Detailed Implementation

[0017] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0018] refer to Figure 1 The figure is an exemplary flowchart of a rehabilitation information retrieval method for cerebrovascular disease rehabilitation diagnosis and treatment, according to some embodiments of this application. The rehabilitation information retrieval method for cerebrovascular disease rehabilitation diagnosis and treatment mainly includes the following steps: In step 101, multimodal rehabilitation data of the target patient is acquired.

[0019] In practice, obtaining multimodal rehabilitation data for the target patient can be achieved in the following ways: Multimodal rehabilitation data can be obtained from the medical information system of the target patient. This multimodal rehabilitation data includes medical imaging data, structured assessment scale data, and unstructured medical record text data. For example, firstly, the original medical imaging data of the patient's most recent cranial computed tomography (CT) scan or magnetic resonance imaging (MRI) examination can be retrieved from the integrated medical imaging archive. This data is stored in DICOM standard format and contains continuous tomographic image sequences and their metadata, such as scan parameters, patient identification, and examination date. Secondly, by connecting to the hospital's information system or an independent rehabilitation assessment database, the structured assessment scale data that has been completed for the target patient can be queried and exported. This data is stored in key-value pairs or tables, including but not limited to the total score and sub-item scores of the National Institutes of Health Stroke Scale, the limb motor function score of the Fugl-Meyer Assessment (FMA), and the Modified Barthel Index. The activity of daily living (ADL) score is obtained by accessing the text database of electronic medical records. Finally, unstructured medical record text data related to the current rehabilitation treatment is extracted, mainly including plain text fields in recent medical records, rehabilitation assessment reports, and discharge summaries. These texts contain free descriptions by doctors or therapists of the patient's symptoms, signs, functional performance, and treatment response. Other methods can also be used in other embodiments, and this application does not limit them.

[0020] It should be noted that, in order to ensure the timeliness and quality of the above-mentioned multimodal rehabilitation data, the completeness and freshness of the data can be verified when performing this step. For example, check whether the key scales are complete and prioritize the assessment records within the last two weeks. Also, perform preliminary time alignment on data from different sources, such as ensuring that the timelines of images, scales, and text records correspond to the same treatment stage, thereby forming a complete, consistent, and chronologically clear set of multimodal rehabilitation data. Other methods can also be used in other embodiments, and this application does not limit them.

[0021] In some embodiments, reference Figure 2 As shown in the figure, this figure is a schematic diagram of the application scenario of the rehabilitation program recommendation data processing system shown in some embodiments of this application. The figure includes three main components: a data acquisition device, a server, and a data storage device. The data acquisition device is responsible for collecting multimodal rehabilitation data of the target patient and sending the collected multimodal rehabilitation data to the server through a communication network. The rehabilitation program recommendation data processing system runs on the server, and the server stores the processing results in the data storage device and visualizes them.

[0022] In step 102, the multimodal rehabilitation data is mapped to a preset cerebrovascular disease rehabilitation knowledge graph to generate a feature subgraph of the target patient. The cerebrovascular disease rehabilitation knowledge graph contains the medical logical relationships between disease, functional impairment, assessment indicators, rehabilitation intervention measures and expected efficacy nodes.

[0023] In some embodiments, mapping the multimodal rehabilitation data to a preset cerebrovascular disease rehabilitation knowledge graph to generate a feature subgraph of the target patient can be achieved through the following steps: Obtain a pre-defined knowledge graph of cerebrovascular disease rehabilitation; Feature extraction and diagnostic reasoning are performed on the medical imaging data in the multimodal rehabilitation data, and the results are associated with the corresponding disease and pathological basic nodes in the cerebrovascular disease rehabilitation knowledge graph. The structured assessment scale data in the multimodal rehabilitation data are standardized and transformed, and the results are associated with the corresponding functional impairment and assessment indicator nodes in the cerebrovascular disease rehabilitation knowledge graph. Medical entity recognition and extraction are performed on the unstructured medical record text data in the multimodal rehabilitation data, and the results are associated with the corresponding symptoms, signs and comorbidity nodes in the cerebrovascular disease rehabilitation knowledge graph. Based on all associations, all associated nodes and predefined medical logical relationship edges between associated nodes are extracted from the cerebrovascular disease rehabilitation knowledge graph to form a feature subgraph representing the current clinical state of the target patient.

[0024] It should be noted that the cerebrovascular disease rehabilitation knowledge graph in this application is a structured semantic network. Its core function is to systematically integrate, represent, and store the medical knowledge system in the field of cerebrovascular disease rehabilitation in a form that can be understood and processed by computers. Through a graph structure of "nodes" and "edges," it models and associates discrete medical concepts, such as diseases, functional disorders, intervention measures, and their complex logical relationships, such as causal relationships, applicability relationships, and evaluation relationships, to form a knowledge base that supports automated retrieval and reasoning. The medical logical relationship edge is a directed or undirected edge connecting two concept nodes in the cerebrovascular disease rehabilitation knowledge graph. Its core function is to formally define and carry the clinically significant logical relationships between medical concepts. The feature subgraph is a substructure network dynamically derived from the cerebrovascular disease rehabilitation knowledge graph. Its core function is to concisely and structurally represent the individualized, multi-dimensional clinical state of the target patient in the form of a graph. It serves as the starting point for subsequent retrieval and is isolated from the pre-stored rehabilitation intervention measures and efficacy expectation nodes in the knowledge graph, but can be connected through relationship edges.

[0025] In specific implementation, obtaining the pre-defined knowledge graph for cerebrovascular disease rehabilitation can be achieved in the following way: Load the pre-constructed and stored knowledge graph for cerebrovascular disease rehabilitation. This knowledge graph is stored in a graph database as graph data or in a relational database after serialization. Its node types include disease, pathological basis, functional impairment, assessment indicators, symptoms, signs, comorbidities, rehabilitation interventions, and expected efficacy. Each node has a unique identifier, name, detailed definition, level of evidence, and possible related attribute values. The edges between nodes represent predefined medical logical relationships, such as "leads to," "manifests as," "assessed as," "applicable to," "contraindicated in," and "expected improvement." Each edge can store the relationship strength weight and the source of evidence (such as clinical evidence). The knowledge graph for cerebrovascular disease rehabilitation includes attributes such as the guideline name and chapters, and the last update time. The construction of this knowledge graph is completed through the following process: First, natural language processing technology is used to automatically extract entities such as diseases, functional impairments, and interventions, along with their relationships such as 'cause,' 'applicable to,' and 'contraindicated', from structured medical literature such as the "Chinese Guidelines for Cerebrovascular Disease Rehabilitation" and Cochrane systematic reviews, forming an initial graph. Then, at least two rehabilitation medicine experts review, correct, and supplement the extraction results to ensure the accuracy of the knowledge. During operation, the cerebrovascular disease rehabilitation knowledge graph can be incrementally updated through the administrator interface to incorporate new medical evidence or correct existing knowledge. Other methods can also be used in other embodiments, and this application does not limit this approach.

[0026] In specific implementation, feature extraction and diagnostic reasoning are performed on the medical imaging data in the multimodal rehabilitation data, and the results are associated with the corresponding disease and pathological basic nodes in the cerebrovascular disease rehabilitation knowledge graph. This can be achieved in the following way: First, the acquired medical imaging data is preprocessed and then input into a pre-trained deep learning model, which is used as a lesion segmentation and classification model. The preprocessing aims to convert the original medical imaging data into a standardized format suitable for input to the deep learning model. The lesion segmentation and classification model is built on the U-Net architecture and trained on a dataset containing 5,000 labeled stroke images. Its lesion segmentation Dice coefficient reaches above 0.85, and the classification accuracy reaches above 90%. It can automatically output the semantic segmentation mask of the lesion, the classification result (such as ischemic, hemorrhagic) and key features, including the precise anatomical location of the lesion (such as the left basal ganglia region, right frontal lobe), volume, and affected Brønsted lesion. The model first identifies the lesion segmentation and classification area or cerebrovascular region (e.g., the blood supply area of ​​the middle cerebral artery). Then, based on the structured features output by the lesion segmentation and classification model, combined with pre-set image-diagnosis mapping rules, such as mapping "left basal ganglia ischemic lesion" to the disease concept "ischemic stroke" and the pathological location concept "left basal ganglia," a standardized diagnostic description is generated. Subsequently, in the cerebrovascular disease rehabilitation knowledge graph, precise or fuzzy matching is performed through node names, aliases, and concept codes to find the "disease" node (e.g., "cerebral infarction") and "pathological basis" node (e.g., "left basal ganglia lesion") corresponding to the diagnostic description. Finally, a temporary "patient node" created for the current target patient is established with the matched disease node and pathological basis node, respectively, with "diagnosed by image" and "lesion located" types of relationship edges, thereby completing the association. This association is recorded in the temporary graph structure. Other methods can also be used in other embodiments, and this application does not limit this.

[0027] In specific implementation, the standardized transformation of the structured assessment scale data in the multimodal rehabilitation data and the association of the results with the corresponding functional impairment and assessment indicator nodes in the cerebrovascular disease rehabilitation knowledge graph can be achieved in the following way: First, read the structured assessment scale data. For each scale, such as the National Institutes of Health Stroke Scale, the Fugl-Meyer Motor Function Rating Scale, and the Modified Barthel Index, access the pre-set scale-level conversion rule base. This rule base defines the mapping relationship between the original score range and the standardized functional impairment level label. For example, the rule "If the upper limb score of the Fugl-Meyer Motor Function Rating Scale is between 0 and 35, then the functional impairment level is 'severe upper limb motor dysfunction'"; Second, based on the patient's specific score, from... The system dynamically determines the functional impairment level label to which the original score belongs, while retaining the specific numerical value of the assessment indicator. Subsequently, in the cerebrovascular disease rehabilitation knowledge graph, the system uses a query function to find a "functional impairment" node that is completely consistent with the name of the functional impairment level label, as well as an "assessment indicator" node corresponding to the scale name, such as "Fugl-Meyer upper limb score". Then, after finding the corresponding node, a "manifestation as" relationship edge is established between the temporary "patient node" and the functional impairment node, and the severity level can be stored on this edge. Finally, a "assessment value" relationship edge is established between the temporary "patient node" and the assessment indicator node, and the specific score value and assessment date are stored on this edge, thereby completing the dual association. Other methods can also be used in other embodiments, and this application does not limit them.

[0028] In specific implementation, the medical entity recognition and extraction of unstructured medical record text data in the multimodal rehabilitation data, and the association of the results with the corresponding symptoms, signs, and comorbidity nodes in the cerebrovascular disease rehabilitation knowledge graph, can be achieved in the following way: First, the cleaned unstructured medical record text data is input into a named entity recognition and relation extraction model trained on professional medical corpora, such as Chinese medical literature and electronic medical records. This named entity recognition and relation extraction model is fine-tuned on a corpus containing 10,000 labeled electronic medical records based on a RoBERTa pre-trained language model. Its F1 score on symptom, sign, and disease entity recognition tasks is no less than 0.92, and it can recognize clinical entities in the text and normalize them to standard medical terms. The identified entity types include symptoms (such as dizziness, slurred speech), signs (such as shallow left nasolabial fold, grade III muscle strength), medical history (such as history of hypertension, atrial fibrillation), and treatment history. Secondly, for each identified entity, the model outputs its standardized name and entity type, such as normalizing "speech disorder" to "diarrhea"; then, based on the entity type, the model queries the corresponding node category in the cerebrovascular disease rehabilitation knowledge graph, for example, associating the symptom entity with the "symptom" node, the sign entity with the "sign" node, and the medical history entity with the "comorbidity" node. The query uses a combination of precise matching based on term strings and concept encoding matching. If a direct match is not possible, the model uses semantic similarity calculation based on word vectors or ontology hierarchy to find the closest node in the knowledge graph; finally, between the temporary "patient node" and these found nodes, a corresponding semantic relationship edge is established based on the entity type, such as "chief complaint is", "examination findings", "past condition", etc., to complete the association. Other methods can also be used in other embodiments, and this application does not limit this.

[0029] In some embodiments, the extraction of all associated nodes and predefined medical logical relationship edges between the associated nodes from the cerebrovascular disease rehabilitation knowledge graph, based on all associations, to construct a feature subgraph representing the current clinical state of the target patient can be achieved through the following steps: Create and initialize a central node representing the target patient, and link the processed association results of the medical imaging data, structured assessment scale data, and unstructured medical record text data to the central node respectively; Collect all the associated edges directly connected to the central node, and obtain all directly associated disease nodes, pathological basis nodes, functional impairment nodes, assessment indicator nodes, symptom nodes, sign nodes, and comorbidity nodes; Query the cerebrovascular disease rehabilitation knowledge graph and extract the predefined medical logical relationship edges between all directly related nodes; The central node, all directly associated nodes, and the extracted medical logical relationship edges are integrated into an independent graph data structure to obtain a feature subgraph of the target patient's current clinical state.

[0030] In specific implementation, creating and initializing a central node representing the target patient, and linking the processed medical image data, structured assessment scale data, and unstructured medical record text data to the central node can be achieved in the following way: First, create a new graph node in memory or temporary storage, mark its node type as "patient instance," and assign it a globally unique identifier, which is bound to the target patient's medical record number or the ID of the current treatment session in the medical information system; at the same time, initialize a set of basic attributes for the central node, including patient identifier, creation timestamp, current treatment stage label, and data source record; then, integrate all the association relationships pointing to various types of nodes in the knowledge graph generated in the previous steps, i.e., the edges recorded in the "temporary graph structure." Specifically, for the "diagnosed as by image" and "lesion location" relationships established after processing medical image data, link the "disease" node and the "pathological basis" node pointed to by the relationship by creating a path from this central node. New edges are created pointing to these nodes, with the relationship type "diagnosed as by imaging" and "lesion location". The association is transferred and solidified into the graph structure with this central node as the core. For the "performance" and "assessment value" relationship established after processing structured assessment scale data, new edges are also created starting from this central node and pointing to the corresponding "functional impairment" node and "assessment index" node. The original scores, grades and other attributes are stored as attributes of the edges. For the "chief complaint", "examination findings", "past condition" and other relationships established after processing unstructured medical record text data, new edges are also created starting from this central node and pointing to the corresponding "symptom", "sign" and "comorbidity" nodes. It should be noted that through the above operations, all clinical features extracted from multimodal data are organized and linked to this unique central node representing the target patient, forming a star-shaped association network with this central node as the root and various clinical feature nodes as leaves. Other methods can also be used in other embodiments, and this application does not limit them.

[0031] In specific implementation, collecting all related edges directly connected to the central node and obtaining all directly related disease nodes, pathological basis nodes, functional impairment nodes, assessment indicator nodes, symptom nodes, sign nodes, and comorbidity nodes can be achieved in the following way: Perform a graph traversal query on the star-shaped relational network constructed in the aforementioned steps, with the central node as the core. Specifically, using a graph query language or graph database API, using the unique identifier of the central node as the query starting point, perform a first-degree neighbor query, that is, find all nodes directly connected to the central node through an edge, and simultaneously return the type and attributes of these edges; the query result will return a set containing all relational edges originating from the central node such as "diagnosed by imaging", "lesion located", "manifestation", "assessment value", "chief complaint", "examination findings", "past history", etc. These edges point to the terminal nodes; then, based on the node type attributes of the terminal nodes, the set is filtered and classified: nodes with the node type "disease" are classified into the disease node set, nodes with the node type "pathological basis" are classified into the pathological basis node set, nodes with the node type "functional impairment" are classified into the functional impairment node set, nodes with the node type "assessment index" are classified into the assessment index node set, nodes with the node type "symptom" are classified into the symptom node set, nodes with the node type "sign" are classified into the sign node set, and nodes with the node type "comorbidity" are classified into the comorbidity node set. Through this step, a structured and categorized list of all directly related clinical feature nodes of the target patient can be obtained. These nodes are the core elements constituting the patient feature subgraph. Other methods can also be used in other embodiments, and this application does not limit them.

[0032] In specific implementation, querying the cerebrovascular disease rehabilitation knowledge graph and extracting predefined medical logical relationship edges between all directly related nodes can be achieved in the following way: Using the set of all directly related disease nodes, pathological basis nodes, functional impairment nodes, assessment indicator nodes, symptom nodes, sign nodes, and comorbidity nodes collected and classified in the previous step as input, a batch graph query is constructed for these node sets. The aim is to retrieve all predefined relationship edges between any two nodes from the predefined cerebrovascular disease rehabilitation knowledge graph. Specifically, each node in the above node set is used as the starting or ending point of the query. The batch path query function of the graph database is used to query whether there are predefined relationships of type "causing," "manifesting as," "assessed as," "co-occurring," or "related" between any two nodes. The medical logic relationship edges are extracted, ignoring paths through the central node of the "patient instance". For example, a query might find a "cause" relationship edge between "left basal ganglia lesion" (pathological basis node) and "severe upper limb motor dysfunction" (functional dysfunction node), or a "risk factor" relationship edge between "history of hypertension" (comorbidity node) and "cerebral infarction" (disease node). All these discovered internal relationship edges existing between directly related nodes are extracted, and the type, direction, strength weight, and evidence source of each edge are recorded. This step is essentially about mining the inherent, medical knowledge-based connections between various clinical characteristics of patients, thereby enriching the star-shaped patient feature network into a denser network that reflects pathophysiological connections. Other methods can also be used in other embodiments, and this application does not limit them.

[0033] In specific implementation, integrating the central node, all directly related nodes, and extracted medical logical relationship edges into an independent graph data structure to obtain the feature subgraph of the target patient's current clinical state can be achieved in the following way: Create a new, empty graph data structure. First, add the central node representing the target patient to this new graph. Then, add all directly related disease nodes, pathological basis nodes, functional impairment nodes, assessment indicator nodes, symptom nodes, sign nodes, and comorbidity nodes collected in the previous step, as well as all predefined medical logical relationship edges extracted from the global knowledge graph that exist between these nodes, to this new graph. When adding nodes, the key attributes of these nodes in the original knowledge graph can be copied, such as... Names, definitions, and codes are used to maintain information integrity. When adding edges, their type, direction, and weight attributes are also copied. Then, the edges initially established from the central node to each directly related node, such as "diagnosed by imaging" and "lesion location", are also added to the new graph, thereby completely reconstructing the association between the central node and all feature nodes. Finally, this new graph contains three types of elements: a unique central node, all directly related clinical feature nodes, and all edges connecting these nodes, including "central node-feature node" edges and "feature node-feature node" edges. This graph data structure is given an independent identifier and stored in the database as a patient feature subgraph. Other methods can also be used in other embodiments, and this application does not limit them.

[0034] It should be noted that the above steps can accurately map and integrate heterogeneous and multi-source patient data into a unified medical knowledge framework (knowledge graph) through standardized processing and pre-trained models. This will associate abstract clinical concepts with specific patient information, generating a comprehensive and structured feature subgraph, which provides a reliable and semantically rich individualized contextual basis for subsequent accurate retrieval and reasoning.

[0035] In step 103, starting from the core functional impairment node in the patient feature subgraph, and combining the clinical diagnosis and treatment stage label, a multi-hop relationship expansion search is performed in the cerebrovascular disease rehabilitation knowledge graph to obtain a set of candidate rehabilitation intervention nodes that are associated with the patient's current state.

[0036] In some embodiments, reference Figure 3 As shown in the figure, this is a flowchart illustrating the process of determining a set of candidate rehabilitation intervention nodes in some embodiments of this application. In this embodiment, starting with the core functional impairment node in the patient feature subgraph and combining it with the clinical diagnosis and treatment stage label, a multi-hop relationship expansion retrieval is performed in the cerebrovascular disease rehabilitation knowledge graph to obtain a set of candidate rehabilitation intervention nodes that are associated with the patient's current state. This can be achieved through the following steps: In step 1031, at least one core functional impairment node is determined based on the severity and influence weight of each functional impairment node in the patient feature subgraph. In step 1032, the whitelist of relationship types, path evidence-based level thresholds, and extension hop limits corresponding to the clinical diagnosis and treatment stage tags are determined to form retrieval constraint rules; In step 1033, taking the core functional impairment node as the initial node, a constrained multi-hop graph traversal is performed in the cerebrovascular disease rehabilitation knowledge graph according to the retrieval constraint rules; In step 1034, all nodes visited during the multi-hop graph traversal that are of the type of rehabilitation intervention measures are collected, deduplicated, and integrated with path confidence to form the candidate rehabilitation intervention node set.

[0037] It should be noted that the multi-hop relation expansion retrieval in this application is a search strategy performed on a graph data structure. Its core function is to start from a specific node in the patient feature subgraph, such as the core functional impairment node, and perform multi-step, constrained graph traversal along the medical logical relation edges in the knowledge graph. The aim is to discover other concept nodes that are indirectly related to the starting point but logically related, especially rehabilitation intervention nodes.

[0038] In specific implementation, determining at least one core functional impairment node based on the severity and impact weight of each functional impairment node in the patient feature subgraph can be achieved in the following way: First, from the patient feature subgraph, all nodes with the node type attribute of "functional impairment" are selected; for each functional impairment node, the "severity level" field stored in its node attribute is read. The value of this field may be a predefined level label such as "mild," "moderate," or "severe." If it is not directly stored, it is obtained by converting it through a pre-defined score-level mapping table based on the specific score on the associated "assessment index" node; simultaneously, a predefined "functional impairment impact weight table" is accessed, based on the name or code of the functional impairment, such as upper limb motor dysfunction, balance dysfunction, etc. The system queries the preset influence weight coefficient of the speech impairment under the current rehabilitation goal. This weight coefficient reflects the relative importance of the functional impairment to the patient's overall living ability and rehabilitation priority. Then, a comprehensive priority score is calculated for each functional impairment node. The calculation method can be to quantify the severity level into a numerical value (e.g., severe = 3, moderate = 2, mild = 1) and then multiply it by the influence weight coefficient. Finally, all functional impairment nodes are sorted in descending order according to this comprehensive priority score, and the node ranked first is selected as the core functional impairment node. Alternatively, when the comprehensive scores are similar or the rehabilitation plan needs to cover multiple goals, the top N nodes (N is usually 1 to 3) are selected to form the core functional impairment node set. Other methods can also be used in other embodiments, and this application does not limit them.

[0039] In specific implementation, taking the core functional impairment node as the initial node and performing a constrained multi-hop graph traversal in the cerebrovascular disease rehabilitation knowledge graph according to the retrieval constraint rules, can be implemented in the following way: First, initialize a priority queue, add the core functional impairment node as the starting element to the queue, and record that its current path is empty and the path confidence is 1.0; at the same time, initialize an access record list to record all nodes visited during the traversal and their related information; then, enter a loop, retrieve the current node and its corresponding path and path confidence from the queue; if the current path has hops... If the number of hops has reached the "maximum number of hops for extension" specified in the retrieval constraint rule, then the path will not be extended further. Otherwise, the cerebrovascular disease rehabilitation knowledge graph is queried to obtain all direct relation edges originating from the current node. For each relation edge, it is checked whether the type of the edge is within the "relation type whitelist" of the retrieval constraint rule, and whether the "evidence level" attribute value carried by the edge is not lower than the "path evidence level threshold" in the rule. Only relation edges that pass both of these checks will be retained. For each retained edge, the confidence of the extended new path is calculated as follows: New path confidence = Current path confidence × Predefined relation strength weight of the edge × Preset path confidence decay coefficient. The relationship strength weight is obtained by mapping the medical evidence level associated with the edge. Evidence levels A, B, and C correspond to weight values ​​of 1.0, 0.7, and 0.3, respectively. The path confidence decay coefficient is set to 0.8, meaning that the path confidence decreases by 20% due to the increase in indirectness for each extended hop. If the new path confidence is still higher than a preset global minimum threshold, such as 0.1, the neighbor node pointed to by this edge, as well as the updated path and path confidence, are added to the priority queue as a new exploration item. Each time a node is taken from the queue for processing, the node's identifier, node type, discovered path, and current path confidence are recorded in the access record list. This loop continues until the queue is empty or the preset maximum number of exploration nodes is reached. It should be noted that this step not only completes the constrained graph traversal but also fully records all visited nodes and their detailed information, providing a data foundation for subsequent filtering steps. Other methods can also be used in other embodiments, and this application does not limit them.

[0040] In specific implementation, collecting all nodes visited during the multi-hop graph traversal that are of the rehabilitation intervention type, performing deduplication and path confidence integration to form the candidate rehabilitation intervention node set can be achieved in the following way: Read the access record list recorded in the previous steps. This access record list contains all nodes visited during the traversal, their node types, discovered paths, and path confidence. First, filter out all node records with the node type attribute "rehabilitation intervention" from this list. Each such node record may correspond to one or more different access paths, i.e., reaching the node through different relational paths. Next, perform deduplication using the node's unique identifier as the key to ensure that each unique rehabilitation intervention node appears only once in the subsequent list. For each deduplicated rehabilitation intervention node, aggregate... All different path records leading to the node are collected, and the path with the highest final path confidence is selected. This path confidence value is used as the "association confidence" between the node and the initial core functional impairment node. At the same time, the detailed relationship sequence of this highest confidence path is retained as the "theoretical basis chain" for recommending the intervention. Finally, all deduplicated rehabilitation intervention nodes, the "association confidence" of each node, and the optional "theoretical basis chain" are encapsulated together as the "candidate rehabilitation intervention node set". This set can be sorted in descending order according to the "association confidence" to form an ordered list, which serves as the direct input for the next step of calculating personalized fit. Other methods can also be used in other embodiments, and this application does not limit them.

[0041] In some embodiments, determining the whitelist of relationship types, path evidence-based level thresholds, and extension hop limits corresponding to clinical diagnosis and treatment stage tags, and forming retrieval constraint rules, can be achieved through the following steps: Obtain the clinical treatment stage labels from the patient feature submap; Using the clinical diagnosis and treatment stage label as the key, query the preset stage-constraint rule mapping table to obtain the corresponding relationship type whitelist, path evidence-based level threshold, and extension hop limit; The whitelist of relation types, path evidence level thresholds, and extended hop limit obtained from the query are output and encapsulated into retrieval constraint rules.

[0042] It should be noted that the retrieval constraint rules in this application are a set of parameters and conditions used to control and guide the multi-hop relation expansion retrieval process. Their core function is to encode the characteristics of the clinical diagnosis and treatment stage and the principles of evidence-based medicine into computable and executable rules, and to filter and constrain the type, quality and depth of the retrieval path in real time, so as to ensure that the retrieval results are not only based on the graph topology, but also conform to the clinical practice norms and evidence requirements of the specific rehabilitation stage.

[0043] In specific implementation, obtaining the clinical treatment stage label in the patient feature subgraph can be achieved in the following way: read the attribute set of the central node of the "patient instance" in the patient feature subgraph and directly obtain its pre-stored "clinical treatment stage label" attribute value; if the attribute value does not exist or is empty, then automatically calculate and assign a stage label based on the difference between the target patient's onset date and the current date, combined with predefined stage division rules. The stage division rules are, for example: 0-2 weeks after onset is the acute phase, 2 weeks-6 months is the early recovery phase, 6 months-1 year is the late recovery phase, and more than 1 year is the sequelae phase; after the calculation is completed, the automatically determined label is updated to the attribute of the central node to ensure data consistency; finally, whether by direct reading or automatic calculation, a clear clinical treatment stage label string is obtained, such as the early recovery phase, which serves as the input for subsequent queries. Other methods can also be used in other embodiments, and this application does not limit this.

[0044] It should be noted that the stage-constraint rule mapping table can be preset in the following way: during the initialization stage, the administrator or medical expert presets the stage-constraint rule mapping table through the configuration management interface; the construction of the mapping table is based on authoritative clinical guidelines for cerebrovascular disease rehabilitation, such as the "Chinese Guidelines for Cerebrovascular Disease Rehabilitation", expert consensus and common paths in clinical practice. Medical experts define the following for each clinical diagnosis and treatment stage, such as the acute phase, early recovery phase, late recovery phase and sequelae phase: (1) the set of relation types that can be considered in this stage, i.e., the relation type whitelist; (2) the rigor requirements for the evidence adopted in this stage, i.e., the path evidence level threshold; (3) the reasonable step limit from functional impairment to intervention measures, i.e., the extension jump limit; the mapping table is persistently stored in the system's configuration database or file, can be read multiple times during runtime and supports dynamic updates during system maintenance to adapt to the evolution of medical knowledge. Other methods can also be used in other embodiments, and this application does not limit this.

[0045] In specific implementation, using the clinical treatment stage label as the key to query the preset stage-constraint rule mapping table and obtain the corresponding relationship type whitelist, path evidence-based level threshold, and extension hop limit can be achieved in the following way: using the obtained clinical treatment stage label string as the query key to access the preloaded stage-constraint rule mapping table; this mapping table uses the clinical treatment stage as the primary key, and each record contains three fields: relationship type whitelist, path evidence-based level threshold, and extension hop limit; wherein, the relationship type whitelist field stores a list of relationship type names, for example, for early recovery, its list content may be [functional impairment - applicable intervention (early)]. [Intervention - Comorbidity Contraindications, Intervention - Device Dependence]; The Path Evidence Level Threshold field stores a string or numerical code representing the minimum required level of evidence, such as Level B; The Extended Jump Limit field stores an integer value, such as 3, which is an average value derived from the analysis of the steps of typical rehabilitation pathways in the guidelines; Perform an exact match query. If a record corresponding to the key value is found, the values ​​of these three fields are extracted as output; if not found, a predefined default rule is used, which usually corresponds to the most stringent constraint, as output to ensure the safety and robustness of the retrieval process. Other methods can also be used in other embodiments, and this application does not limit them.

[0046] In practice, the retrieved relation type whitelist, path evidence level threshold, and extended hop limit are output and encapsulated into retrieval constraint rules. This can be achieved as follows: the relation type whitelist, path evidence level threshold, and extended hop limit obtained from the previous steps are combined with globally preset fixed parameters, such as the path confidence decay coefficient (default value 0.7) and the minimum path confidence threshold (default value 0.1), to instantiate a retrieval constraint rule object. Internally, this retrieval constraint rule object is an instance of a structure or class, and its attribute fields correspond one-to-one with the aforementioned parameters. For example, in object-oriented programming, this object could contain `allowed_rel`. The object contains attributes such as `ation_types` (list type), `min_evidence_level` (string type), `max_hops` (integer type), `confidence_decay_factor` (floating-point type), and `min_confidence_threshold` (floating-point type). After encapsulation, the retrieval constraint rule object is output as a complete and independent set of parameters and directly passed to the subsequent constrained multi-hop graph traversal. This serves as the sole basis for path expansion, filtering, and evaluation during its execution. Other methods can also be used in other embodiments, and this application does not limit them.

[0047] It should be noted that the above steps can take the patient's most core functional impairment as a precise starting point, and combine it with the clinical rules of the current rehabilitation stage to explore a controlled and medically logical path map. This systematically discovers a set of potential rehabilitation interventions that are directly and indirectly theoretically related to the patient's current state, ensuring the clinical relevance, stage appropriateness and logical coherence of the search results, and avoiding the blindness and irrelevance of the search.

[0048] In step 104, based on the empirical validity data of each candidate rehabilitation intervention node in a historical patient group with similar characteristics to the current patient, the topological association strength of the cerebrovascular disease rehabilitation knowledge graph, and the feasibility constraints of the target patient's current medical scenario, the personalized fit of each candidate rehabilitation intervention node is determined.

[0049] In some embodiments, determining the personalized fit of each candidate rehabilitation intervention node based on empirical validity data from historical patient groups similar to the current patient's characteristics, the topological association strength of the cerebrovascular disease rehabilitation knowledge graph, and the feasibility constraints of the target patient's current medical scenario can be achieved through the following steps: For each candidate rehabilitation intervention node, historical patient groups with similar characteristics to the target patient are retrieved from the historical medical database, and the empirical validity score of the candidate rehabilitation intervention node is determined based on the efficacy data corresponding to the candidate rehabilitation intervention node and the historical patient groups. From the multi-hop relation expansion retrieval results, the calculated association confidence of the candidate rehabilitation intervention node, which represents the theoretical association strength with the patient's core functional impairment node, is extracted as the topological association strength score of the candidate rehabilitation intervention node. Based on the resource allocation and constraint rule base of the target patient's current medical scenario, assess the feasibility of implementing candidate rehabilitation intervention nodes in the current medical scenario, and calculate the feasibility constraint score of candidate rehabilitation intervention nodes; The empirical validity score, the topological correlation strength score, and the scenario feasibility constraint score are calculated using a linear weighted fusion algorithm with preset weights to obtain the personalized fit of the selected rehabilitation intervention node, thereby obtaining the personalized fit of each candidate rehabilitation intervention node.

[0050] It should be noted that the personalized fit in this application is a final quantitative score of the overall suitability and priority of each candidate rehabilitation intervention relative to the current specific target patient; the empirical validity score is based on real-world historical medical data and measures the efficacy level or probability of effectiveness of a candidate rehabilitation intervention when actually applied in a group with similar characteristics to the current patient; the topological association strength score measures the theoretical correlation and logical support strength between a candidate rehabilitation intervention node and the core node representing the current state of the patient in the semantic network structure of the cerebrovascular disease rehabilitation knowledge graph; and the feasibility constraint score assesses the degree of matching between the resources and conditions required for the actual implementation of a candidate rehabilitation intervention in the specific medical scenario in which the target patient is currently located, such as a specific medical institution or home environment, and the inherent resources and limitations of the scenario.

[0051] In specific implementation, retrieving historical patient groups with similar characteristics to the target patient from the historical medical database and determining the empirical validity score of the candidate rehabilitation intervention nodes based on the efficacy data corresponding to the candidate rehabilitation intervention nodes can be achieved in the following way: First, access the associated historical medical database, which stores desensitized medical data of more than 3,000 cerebrovascular disease rehabilitation patients from the rehabilitation departments of cooperating tertiary hospitals over the past five years. The data table structure includes a patient basic information table, a diagnosis and assessment record table, a rehabilitation intervention implementation table, and an efficacy follow-up table; Second, extract a set of key feature dimensions for defining similarity from the patient feature subgraph and perform similarity retrieval in the historical medical database. These dimensions include at least the core disease diagnosis, main pathological location, type and severity level of core functional impairment, and patient age group. These dimensions and their values ​​are defined, such as disease: "cerebral infarction", pathological location: "left basal ganglia", core functional impairment: "severe upper limb motor dysfunction", and age group: "50-60 years old". A feature vector is constructed. Then, in the historical medical database, this feature vector is compared with the corresponding feature vector of each historical patient in the database for similarity calculation. The calculation method can involve one-hot encoding the categorical features and calculating the cosine similarity, thereby filtering out all historical patients with similarity higher than a preset threshold, such as 0.75, to form a similar historical patient group. Then, within this similar group, a subset of patients who have actually applied the specific measures corresponding to the current candidate rehabilitation intervention node are further filtered out. Finally, based on the outcome data of patients in this subset, such as the average improvement value of the Fugl-Meyer score before and after treatment, or the proportion of patients achieving clinically significant improvement, normalization is performed. For example, the average improvement value is divided by the theoretical maximum improvement value of the scale to obtain a value between 0 and 1, which is used as the empirical validity score of the candidate rehabilitation intervention node. If the subset is empty or the sample size is too small, such as less than 5 cases, a baseline score indicating insufficient evidence, such as 0.5, is assigned. Other methods can also be used in other embodiments, and this application does not limit this.

[0052] In specific implementation, the calculated association confidence score, which characterizes the theoretical association strength with the patient's core functional impairment node, is extracted from the multi-hop relationship expansion retrieval results and used as the topological association strength score of the candidate rehabilitation intervention node. This can be achieved in the following way: directly accessing all nodes of rehabilitation intervention type that are visited during the multi-hop graph traversal process collected in step 103, performing deduplication and path confidence integration to form the candidate rehabilitation intervention node set, the association confidence attribute generated and stored for each candidate rehabilitation intervention node is the highest path confidence value calculated when traversing from the core functional impairment node through a path that conforms to the clinical diagnosis and treatment stage constraints to the intervention node in the cerebrovascular disease rehabilitation knowledge graph. It is itself an indicator between 0 and 1 that quantifies the theoretical association strength. This pre-stored value is directly read out as the topological association strength score corresponding to the candidate rehabilitation intervention node. Other methods can also be used in other embodiments, and this application does not limit this.

[0053] In specific implementation, based on the resource allocation and constraint rule base of the target patient's current medical scenario, the feasibility of implementing candidate rehabilitation intervention nodes in the current medical scenario is evaluated, and the feasibility constraint score of the candidate rehabilitation intervention nodes is calculated. This can be achieved in the following way: Maintaining a medical scenario resource allocation and constraint rule base, which records information such as available equipment lists, qualified therapist skill tags, rehabilitation project catalogs covered by medical insurance policies, and typical single treatment durations under different medical scenario identifiers, such as rehabilitation departments of tertiary hospitals, community rehabilitation stations, and home rehabilitation; simultaneously, obtaining the medical scenario identifier of the target patient's current location, and for each node in the candidate rehabilitation intervention node set, extracting the list of conditions required for its implementation from the attributes of its associated rehabilitation intervention measure node or its linked solution component nodes. This list typically includes necessary conditions. The system includes items such as equipment, therapist skills, environmental requirements, single session duration, and cost level. Then, each requirement is matched and its compliance is judged against the corresponding resources in the medical scenario resource configuration and constraint rule base: a fully matched requirement receives 1 point, a partially matched requirement or one that can be substituted by an equivalent method receives 0.5 points, and a completely mismatched requirement or one lacking resources receives 0 points. If there is an absolute contraindication, such as an intervention requiring standing balance training but the patient's scenario is bedridden home care, a veto factor is triggered, significantly reducing the total score. Finally, the scores of all requirement items are weighted and averaged. The weights can be set according to the necessity of the requirement item and adjusted in conjunction with the veto factor, outputting a value between 0 and 1 as the scenario feasibility constraint score for the candidate rehabilitation intervention node. Other methods can also be used in other embodiments, and this application does not limit this.

[0054] In specific implementation, the empirical validity score, the topological correlation strength score, and the scenario feasibility constraint score are calculated using a linear weighted fusion algorithm with preset weights to obtain the personalized fit of the selected rehabilitation intervention node. This can be achieved in the following way: First, the empirical validity score, topological correlation strength score, and scenario feasibility constraint score already obtained by the candidate rehabilitation intervention node are denoted as S_e, S_t, and S_f, respectively. Second, the weight coefficients corresponding to the current medical scenario are read from a preset weight configuration table. The weight configuration table stores the expert consensus values ​​of the weights of each dimension under different scenarios. For example, in the scenario of a hospital rehabilitation department, the empirical validity score weight We = 0.5, the topological correlation strength score weight Wt = 0.3, and the scenario feasibility constraint score weight Wf = 0.2 are set, satisfying W_e + W_t + W_f = 1. The consensus value is determined by taking the average of three or more high-level rehabilitation medicine experts who independently score according to clinical decision preferences. Then, a linear weighted fusion model is used to calculate the final fit, i.e., Personalized Fit Score = W_e × S_e + W_t × S_t + W_f × S_f; Finally, this calculation is performed independently for each candidate rehabilitation intervention node in the candidate rehabilitation intervention node set to generate a final, comprehensive, personalized fit for each candidate rehabilitation intervention node. Other methods can also be used in other embodiments, and this application does not limit them.

[0055] It should be noted that the above steps can integrate and calculate the empirical validity score based on practical evidence from similar patient groups, the topological association strength score based on knowledge graph theory, and the feasibility constraint score based on real-world scenario conditions for each candidate intervention. This results in a quantitative score that comprehensively reflects the measure's suitability in terms of evidence-based practice, logical consistency, and operability. This expands the recommendation basis from a single dimension to a multi-dimensional fusion, greatly improving the scientific validity, personalization, and practical feasibility of the recommendation results.

[0056] In step 105, each candidate rehabilitation intervention node is ranked according to all individualized fit, and a draft recommendation of a structured rehabilitation plan for the target patient is formed based on the ranking results.

[0057] In some embodiments, ranking each candidate rehabilitation intervention node according to all personalized fits and forming a draft recommendation of a structured rehabilitation program for the target patient based on the ranking results can be achieved through the following steps: Based on the personalized fit calculated for each candidate rehabilitation intervention node, all candidate rehabilitation intervention nodes are sorted in descending order to generate an ordered candidate list; From the ordered candidate list, the top N candidate rehabilitation intervention nodes are selected to form the core recommended intervention node set; For each node in the core recommended intervention node set, the associated rehabilitation program component information is extracted from the cerebrovascular disease rehabilitation knowledge graph and organized according to a preset format; The organized structured rehabilitation program components are integrated with the personalized suitability and key evidence of the corresponding intervention measures to generate a recommended draft of the structured rehabilitation program for the target patient, which is then provided for manual review.

[0058] In specific implementation, N can be set in the following way: N is a preset positive integer, and the default value of N is 5. The basis for setting it is to balance the comprehensiveness and efficiency of clinical decision-making. Through a survey of 20 rehabilitation physicians, it was found that 5 recommendations can cover the core intervention options without causing information overload. At the same time, the system interface allows physicians to dynamically adjust the value of N in the range of 3 to 7 according to the complexity of the diagnosis and treatment. Other methods can also be used in other embodiments, and this application does not limit them.

[0059] In specific implementation, for each node in the core recommended intervention node set, extracting its associated rehabilitation program component information from the cerebrovascular disease rehabilitation knowledge graph and organizing it according to a preset format can be achieved in the following way: For each rehabilitation intervention measure node corresponding to each core recommended intervention node in the core recommended intervention node set, query all program component nodes directly connected to the node through relational edges containing component or detailed description types in the cerebrovascular disease rehabilitation knowledge graph. Extract predefined attribute fields from these program component nodes. These fields include at least a description of the training method, such as mirror therapy training, intensity parameter suggestions, such as 20 minutes each time, once a day, applicable stage, and a summary of key evidence-based evidence. Subsequently, organize the extracted information according to a predefined, structured data template, for example, using a JSON object format, where the top-level key is the intervention name and the subkeys are the method, parameters, and basis, etc. For each core recommended intervention node, generate a data structure object that conforms to the data template, which is filled with the specific content extracted from the knowledge graph. Other methods can also be used in other embodiments, and this application does not limit this.

[0060] In practice, the organized structured rehabilitation program component information is integrated with the personalized fit and key evidence of the corresponding intervention measures to generate a recommended draft of the structured rehabilitation program for the target patient. This draft, which is then submitted for manual review, can be implemented as follows: First, a new document or data structure is created as a container for the recommended draft. Then, based on the original order of the core recommended intervention node set (determined by the personalized fit score), the structured program component data objects generated for each node in the previous steps are sequentially inserted into the draft container. When inserting each component object, the personalized fit score corresponding to that node is simultaneously associated and embedded, along with more detailed key evidence extracted from the knowledge graph, such as guideline name, publication date, and recommendation information. The final generated draft recommendation is a structured list, where each item clearly corresponds to a recommended intervention and clearly displays its specific plan content, recommendation strength (the greater the personalization fit, the greater the recommendation strength), and theoretical / evidence support. This draft recommendation is presented to rehabilitation physicians through a graphical user interface. Each recommendation is accompanied by interactive controls such as "adopt," "exclude," and "adjust parameters" buttons and text boxes. Rehabilitation physicians can review the draft content and use these controls to confirm, reject, or personalize the recommendations. After review and adjustment, the physician submits the plan via the "confirm plan" button on the interface. The physician's final decision is recorded, forming an executable, personalized rehabilitation plan that integrates AI recommendations and human professional judgment. Simultaneously, the differences between the plan before and after review can be saved as feedback data. Other methods can also be used in other embodiments, and this application does not limit this.

[0061] It should be noted that the above steps can objectively rank all candidate measures according to their comprehensive suitability, and automatically extract and assemble the detailed plan components associated with the top-ranked measures to generate a well-structured, complete recommendation draft that is marked with the strength of recommendation and the source of evidence. This provides rehabilitation physicians with an efficient, intuitive and traceable decision support blueprint.

[0062] Furthermore, in another aspect of this application, in some embodiments, this application provides a knowledge graph-based rehabilitation and treatment system for cerebrovascular diseases, which includes a rehabilitation information retrieval unit, referencing... Figure 4 The figure is a schematic diagram of the structure of a rehabilitation information retrieval unit according to some embodiments of this application. The rehabilitation information retrieval unit includes: an acquisition module 401, a processing module 402, and an execution module 403, which are described below: The acquisition module 401 in this application is mainly used to acquire multimodal rehabilitation data of the target patient; Processing module 402, in this application, is mainly used to map the multimodal rehabilitation data to a preset cerebrovascular disease rehabilitation knowledge graph to generate a feature subgraph of the target patient. The cerebrovascular disease rehabilitation knowledge graph contains medical logical relationships between disease, functional impairment, assessment indicators, rehabilitation intervention measures and expected efficacy nodes. The processing module 402 described in this application is also used to perform multi-hop relationship expansion retrieval in the cerebrovascular disease rehabilitation knowledge graph, starting from the core functional impairment node in the patient feature subgraph and combining the clinical diagnosis and treatment stage label, to obtain a set of candidate rehabilitation intervention nodes that are associated with the patient's current state. The processing module 402 described in this application is also used to determine the personalized fit of each candidate rehabilitation intervention node based on empirical validity data in historical patient groups with similar characteristics to the current patient, the topological association strength of the cerebrovascular disease rehabilitation knowledge graph, and the feasibility constraints of the target patient's current medical scenario. The execution module 403 in this application is mainly used to sort each candidate rehabilitation intervention node according to all the personalized adaptations, and to form a recommended draft of the structured rehabilitation plan for the target patient based on the sorting results.

[0063] Each module in the aforementioned rehabilitation information retrieval unit can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0064] In another embodiment, this application provides a computer device, which may be a server, and its internal structure diagram may be as follows. Figure 5 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores rehabilitation information retrieval data for the diagnosis and treatment of cerebrovascular diseases. The network interface communicates with external terminals via a network connection. When executed by the processor, the computer program implements a rehabilitation information retrieval method for the diagnosis and treatment of cerebrovascular diseases.

[0065] Those skilled in the art will understand that Figure 5The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0066] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above embodiment of the rehabilitation information retrieval method for cerebrovascular disease rehabilitation diagnosis and treatment.

[0067] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps described in the embodiment of the rehabilitation information retrieval method for cerebrovascular disease rehabilitation diagnosis and treatment.

[0068] In one embodiment, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps described in the embodiment of the rehabilitation information retrieval method for cerebrovascular disease rehabilitation diagnosis and treatment.

[0069] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0070] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0071] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for recommending rehabilitation programs based on rehabilitation information retrieval, applied to a knowledge graph-based rehabilitation diagnosis and treatment system for cerebrovascular diseases, characterized in that, The method includes the following steps: Acquire multimodal rehabilitation data from target patients; The multimodal rehabilitation data is mapped to a preset cerebrovascular disease rehabilitation knowledge graph to generate a feature subgraph of the target patient. The cerebrovascular disease rehabilitation knowledge graph contains the medical logical relationships between disease, functional impairment, assessment indicators, rehabilitation intervention measures and expected efficacy nodes. Starting with the core functional impairment node in the patient feature subgraph, and combining it with the clinical diagnosis and treatment stage label, a multi-hop relationship expansion retrieval is performed in the cerebrovascular disease rehabilitation knowledge graph to obtain a set of candidate rehabilitation intervention nodes that are associated with the patient's current state. Based on empirical validity data from historical patient groups with similar characteristics to the current patient for each candidate rehabilitation intervention node, the topological association strength of the cerebrovascular disease rehabilitation knowledge graph, and the feasibility constraints of the target patient's current medical scenario, the personalized fit of each candidate rehabilitation intervention node is determined. Based on all individualized fits, each candidate rehabilitation intervention node is ranked, and a draft recommendation of a structured rehabilitation program for the target patient is formed based on the ranking results.

2. The method as described in claim 1, characterized in that, Mapping the multimodal rehabilitation data to a pre-defined cerebrovascular disease rehabilitation knowledge graph to generate a feature sub-graph for the target patient specifically includes: Obtain a pre-defined knowledge graph of cerebrovascular disease rehabilitation; Feature extraction and diagnostic reasoning are performed on the medical imaging data in the multimodal rehabilitation data, and the results are associated with the corresponding disease and pathological basic nodes in the cerebrovascular disease rehabilitation knowledge graph. The structured assessment scale data in the multimodal rehabilitation data are standardized and transformed, and the results are associated with the corresponding functional impairment and assessment indicator nodes in the cerebrovascular disease rehabilitation knowledge graph. Medical entity recognition and extraction are performed on the unstructured medical record text data in the multimodal rehabilitation data, and the results are associated with the corresponding symptoms, signs and comorbidity nodes in the cerebrovascular disease rehabilitation knowledge graph. Based on all associations, all associated nodes and predefined medical logical relationship edges between associated nodes are extracted from the cerebrovascular disease rehabilitation knowledge graph to form a feature subgraph representing the current clinical state of the target patient.

3. The method as described in claim 2, characterized in that, Based on all associations, all associated nodes and predefined medical logical relationship edges between the associated nodes are extracted from the cerebrovascular disease rehabilitation knowledge graph to form a feature subgraph representing the current clinical state of the target patient. Specifically, this includes: Create and initialize a central node representing the target patient, and link the processed association results of the medical imaging data, structured assessment scale data, and unstructured medical record text data to the central node respectively; Collect all the associated edges directly connected to the central node, and obtain all directly associated disease nodes, pathological basis nodes, functional impairment nodes, assessment indicator nodes, symptom nodes, sign nodes, and comorbidity nodes; Query the cerebrovascular disease rehabilitation knowledge graph and extract the predefined medical logical relationship edges between all directly related nodes; The central node, all directly associated nodes, and the extracted medical logical relationship edges are integrated into an independent graph data structure to obtain a feature subgraph of the target patient's current clinical state.

4. The method as described in claim 1, characterized in that, Starting with the core functional impairment node in the patient feature subgraph, and combining it with the clinical treatment stage label, a multi-hop relationship expansion retrieval is performed in the cerebrovascular disease rehabilitation knowledge graph to obtain a set of candidate rehabilitation intervention nodes that are associated with the patient's current state. Specifically, this set includes: Based on the severity and influence weight of each functional impairment node in the patient feature subgraph, at least one core functional impairment node is identified. Determine the whitelist of relationship types, path evidence-based level thresholds, and extension hop limits corresponding to the tags of clinical diagnosis and treatment stages to form search constraint rules; Using the core functional impairment node as the initial node, and in accordance with the retrieval constraint rules, a constrained multi-hop graph traversal is performed in the cerebrovascular disease rehabilitation knowledge graph; Collect all nodes visited during the multi-hop graph traversal process that are of the type of rehabilitation intervention measures, perform deduplication and path confidence integration to form the candidate rehabilitation intervention node set.

5. The method as described in claim 4, characterized in that, Determine the whitelist of relationship types, path evidence-based level thresholds, and extension hop limits corresponding to clinical diagnosis and treatment stage tags, and formulate retrieval constraint rules, specifically including: Obtain the clinical treatment stage labels from the patient feature submap; Using the clinical diagnosis and treatment stage label as the key, query the preset stage-constraint rule mapping table to obtain the corresponding relationship type whitelist, path evidence-based level threshold, and extension hop limit; The whitelist of relation types, path evidence level thresholds, and extended hop limit obtained from the query are output and encapsulated into retrieval constraint rules.

6. The method as described in claim 1, characterized in that, Based on empirical validity data from historical patient groups with similar characteristics to the current patient for each candidate rehabilitation intervention node, the topological association strength of the cerebrovascular disease rehabilitation knowledge graph, and the feasibility constraints of the target patient's current medical scenario, the personalized fit of each candidate rehabilitation intervention node is determined specifically including: For each candidate rehabilitation intervention node, historical patient groups with similar characteristics to the target patient are retrieved from the historical medical database, and the empirical validity score of the candidate rehabilitation intervention node is determined based on the efficacy data corresponding to the candidate rehabilitation intervention node and the historical patient groups. From the multi-hop relation expansion retrieval results, the calculated association confidence of the candidate rehabilitation intervention node, which represents the theoretical association strength with the patient's core functional impairment node, is extracted as the topological association strength score of the candidate rehabilitation intervention node. Based on the resource allocation and constraint rule base of the target patient's current medical scenario, assess the feasibility of implementing candidate rehabilitation intervention nodes in the current medical scenario, and calculate the feasibility constraint score of candidate rehabilitation intervention nodes; The empirical validity score, the topological correlation strength score, and the scenario feasibility constraint score are calculated using a linear weighted fusion algorithm with preset weights to obtain the personalized fit of the selected rehabilitation intervention node, thereby obtaining the personalized fit of each candidate rehabilitation intervention node.

7. The method as described in claim 1, characterized in that, The multimodal rehabilitation data includes medical imaging data, structured assessment scale data, and unstructured medical record text data.

8. A knowledge graph-based rehabilitation and treatment system for cerebrovascular diseases, comprising a rehabilitation information retrieval unit, characterized in that, The rehabilitation information retrieval unit includes: The acquisition module is used to acquire multimodal rehabilitation data of the target patient; The processing module is used to map the multimodal rehabilitation data to a preset cerebrovascular disease rehabilitation knowledge graph to generate a feature subgraph of the target patient. The cerebrovascular disease rehabilitation knowledge graph contains medical logical relationships between disease, functional impairment, assessment indicators, rehabilitation intervention measures and expected efficacy nodes. The processing module is also used to perform multi-hop relationship expansion retrieval in the cerebrovascular disease rehabilitation knowledge graph, starting from the core functional impairment node in the patient feature subgraph and combining the clinical diagnosis and treatment stage label, to obtain a set of candidate rehabilitation intervention nodes that are associated with the patient's current state. The processing module is also used to determine the personalized fit of each candidate rehabilitation intervention node based on empirical validity data from historical patient groups with similar characteristics to the current patient, the topological association strength of the cerebrovascular disease rehabilitation knowledge graph, and the feasibility constraints of the target patient's current medical scenario. The execution module is used to sort each candidate rehabilitation intervention node according to all personalized adaptations, and to form a recommended draft of a structured rehabilitation plan for the target patient based on the sorting results.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the rehabilitation information retrieval method for rehabilitation diagnosis and treatment of cerebrovascular diseases as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the rehabilitation information retrieval method for rehabilitation diagnosis and treatment of cerebrovascular diseases as described in any one of claims 1 to 7.

Images