DIKWP semantic map generation and intention recognition interface method for multi-modal health data

Abstract

The invention provides a DIKWP semantic map generation and intention recognition interface method for multi-modal health data, which is used for active medical intelligent diagnosis and treatment. According to the method, data such as texts, images and voices are mapped to a data-information-knowledge-intelligence-intention five-layer semantic map, a patient health target is extracted in combination with a Transform and knowledge distillation model, and intention layer nodes are generated. The semantic map realizes cross-layer association through a multi-hop relationship, provides a standardized DIKWP-API interface, supports AI model query, explanation and dynamic updating, and enables the decision process to be explained and aligned with patient targets. The system further comprises a graph updating and version management mechanism, and knowledge closed-loop correction and intention evolution tracking are achieved through clinical feedback. The method improves the intelligent level of multi-modal medical data fusion analysis, ensures that diagnosis and treatment reasoning is transparent and controllable and conforms to the willingness of a patient, and has important clinical application value.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of medical artificial intelligence and knowledge graph, and specifically relates to a method for constructing a semantic graph of multi-modal health data and an intent recognition interface. In particular, the method maps structured and unstructured medical data (including electronic medical record texts, medical images, voice interactions, etc.) into a five-layer semantic graph of data-information-knowledge-wisdom-purpose (DIKWP), and provides a standardized interface for AI models to call, to support explainable reasoning of medical decisions and alignment with patient health goals. BACKGROUND

[0002] With the exponential growth of medical data, its forms are increasingly diverse, including clinical text records, test values, medical images, and voice exchanges between patients and doctors, among other multi-modal information. However, existing medical AI systems mostly use end-to-end deep learning models, and there are still limitations in the fusion and understanding of multi-modal data, often lacking transparent semantic understanding and explainability, like a "black box". This makes it difficult for AI to give clinical trust in diagnosis and treatment recommendations, and also difficult to adjust according to the individual preferences of patients.

[0003] Knowledge graph technology provides a new way to solve the above problems. Knowledge graph organizes and links entities and concepts in the medical field in the form of a semantic network, enabling AI to use structured medical knowledge for reasoning, thereby improving the accuracy and explainability of recommendations. In recent years, multi-modal knowledge graphs have gradually attracted attention, by representing data in different modalities such as text, images, and voice as graph nodes, they can integrate multi-source information to provide a more comprehensive medical context. For example, some studies have constructed multi-modal medical knowledge graphs containing text and images to enrich the basis for clinical decision support. However, traditional knowledge graphs generally focus on the data-information-knowledge-wisdom (DIKW) hierarchy, and lack direct representation of patients' subjective intentions or goals. In other words, existing technologies rarely include patient intent in the knowledge representation of medical AI, making it difficult for AI systems to understand and respond to patients' health goal preferences in a timely manner.

[0004] To address this issue, the DIKWP semantic model adds "Purpose" as the fifth layer based on the classic DIKW cognitive hierarchy. The DIKWP model emphasizes the representation of target intent on top of data perception, information processing, knowledge reasoning, and wisdom decision-making, to drive the entire cognitive process with purpose. In the medical context, the introduction of the "intent layer" means that the AI system will not only consider medical knowledge and objective indicators, but also understand the goals and preferences behind patient or clinical decision-making, thereby achieving truly patient-centered intelligent diagnosis and treatment. Integrating patient intent into the semantic graph is expected to address the misalignment between current medical AI decision-making and patient needs, and improve the acceptability and effectiveness of AI recommendations.

[0005] However, there are several technical challenges in achieving the above goals: (1) how to perform hierarchical semantic parsing of health data of different modalities and map them to a unified multi-layer knowledge graph structure; (2) how to automatically identify health goals and intentions from patients' natural language descriptions or interactions and integrate them into the highest layer of the knowledge graph; (3) how to fuse nodes at each layer in the graph to achieve multi-hop semantic association from data layer to intention layer to support complex reasoning; (4) how to design a standardized interface for various medical AI models to query and update the graph, enabling embedded and interpretable semantic feedback; and (5) how should the graph support dynamic updating and version management of nodes to adapt to closed-loop feedback and evolution of patient intentions in medical practice. Therefore, it is necessary to provide a new method to construct a multi-modal medical semantic graph containing an intention layer and design a corresponding interface to ensure that the medical AI system can not only understand multi-modal medical information but also "understand" the needs and goals of patients. SUMMARY

[0006] To overcome the lack of semantic explanation and intention alignment in existing medical AI, the present application provides a DIKWP semantic graph generation and intention recognition interface method for multi-modal health data. The present application aims to automatically parse and map multi-modal health data such as text, images, and voice to nodes at each layer of the DIKWP semantic graph, and extract the health goal preferences of patients from their related data through an intention recognition algorithm to generate semantic nodes in the intention layer. Further, the present application provides a standardized DIKWP-API interface for medical AI models to query, embed, or update semantic graph node information, thereby achieving interpretable reasoning and dynamic alignment with patient goals during the diagnosis and treatment process.

[0007] The method of the present application comprises the following steps and modules:

[0008] Multi-modal data semantic parsing and graph mapping: For different types of health data such as electronic health records (EHR), medical images, and voice conversations, design semantic parsing mechanisms respectively: Text data processing: Perform natural language processing on electronic medical record texts, doctor's notes, and consultation conversations, use named entity recognition and relation extraction in the medical field, extract the symptoms, signs, diagnoses, test indicators, and drugs mentioned by patients as concept entities, and determine their relationships according to the context, and map them to the corresponding nodes of the graph (such as symptom nodes, disease nodes, etc.); Image data processing: Radiological images, pathological pictures, etc. are analyzed using deep learning models (such as CNN convolutional neural networks or visual Transformers) to identify key lesions or image features (such as tumor size, location, and abnormal types shown in the image). These information is converted into structured descriptions and mapped into information layer or knowledge layer nodes (such as "lung nodule: 5mm" as an information node, linked to the "lung nodule" knowledge node). Voice data processing: Voice recognition (ASR) is used to transcribe medical voice interactions into text, and the above text processing method is used to extract medical entities and relationships. Voice tone or emotion-related information can be marked as supplementary attributes on related nodes (such as marking the intensity of the patient's statement of pain). Data hierarchical representation: After analyzing each modality data, the DIKWP model is used to represent information in layers: the data layer saves the original data and simple facts (such as sensor readings, original image fragments); the information layer saves meaningful information extracted from data (such as "body temperature = 39℃" or "X-ray shows shadow"); the knowledge layer represents general medical knowledge entities (such as disease name, anatomical site, treatment plan, etc.); the wisdom layer represents decisions or plans generated by knowledge application (such as "suggest antibiotic treatment for 10 days"); the intent layer represents the patient's health goals or the purpose behind the clinical decision (such as "patient wants to return to work as soon as possible" or "prefer to use conservative treatment to avoid surgery").

[0009] Intention recognition based on Transformer: An intention recognition model is provided to extract health goal preferences from patient interaction data and generate intent layer nodes. The model uses a deep learning Transformer architecture, pre-trained on large-scale Chinese medical corpus (such as medical records, patient feedback) and knowledge distillation optimized, which can understand the implicit intention expressed by patients in natural language: Semantic understanding and term alignment: The model supports mapping patients' daily language to professional medical terms. For example, the patient says "I don't want to take too many drugs", the model can identify the intention as "hope to reduce drug intake", and align it with the standard medical preference concept to form the intent node; Transformer + knowledge distillation: First, use a large medical language model (containing rich medical knowledge and context understanding ability) to train the intention recognition task, and then use knowledge distillation technology to transfer its knowledge to a lightweight model to improve reasoning speed and deployability. The model combines rule base (such as common patient goals in clinical guidelines) and deep learning to achieve high accuracy of intent classification and key point extraction.

[0010] DIKWP semantic graph construction and fusion: The above obtained semantic information of each layer is fused to construct five-layer semantic graph: Node generation: For each medical entity or concept parsed, a node is created in the corresponding DIKWP layer. For example, a data node is created for a certain test value in the data layer, a knowledge node is created for a disease diagnosis in the knowledge layer, and an intent node is created for a patient goal in the intent layer; Multi-hop semantic linking: Semantic relationship links are established between nodes, spanning the five-layer structure. For example, a data layer node "fasting blood glucose = 8.0 mmol / L" is associated with a knowledge layer node "diabetes" diagnosis through an information layer node "hyperglycemia", and the diagnosis node is connected to a wisdom layer node "suggest lifestyle intervention and metformin treatment" according to clinical guidelines, while the patient's intent layer node "hope not to take medication for life" is associated with the wisdom layer node (indicating the patient's preference for conservative management). Through such cross-layer multi-hop linking, a semantic path is established from data to information to knowledge to wisdom to intent, allowing the AI system to trace the decision-making basis and evaluate whether the patient's goals are met; Node fusion and disambiguation: For nodes with different sources but the same or related semantics, disambiguation and merging are performed through a fusion mechanism. For example, multiple modalities refer to the same entity (CT report and doctor's notes both point to "pneumonia"), so a unified "pneumonia" node is retained in the knowledge layer, and relevant attributes and evidence sources are combined. During the fusion process, medical ontologies and alias dictionaries are combined to ensure that synonymous concepts are mapped to the same node, improving the consistency and completeness of the graph.

[0011] Plug-in semantic explanation interface (DIKWP-API): The present invention provides a standardized DIKWP-API interface to facilitate various medical AI applications to access and update the semantic graph, and to integrate explainable AI and semantic feedback: Query service: AI models can query nodes and relationships in the graph through the API. For example, a diagnosis model can query the "symptom to possible disease" association in a patient's graph to assist in reasoning, and a question and answer system can retrieve patient intent layer nodes to obtain patient preferences that should be considered when answering. The API supports multiple query modes, including retrieving all semantic nodes by patient ID, filtering specific types of nodes by conditions (such as querying all unsatisfied patient goals), and multi-hop queries (such as retrieving associated intent nodes from a data node to explain the impact of the data on patient goals); Embedded explanation: AI models can call the API to obtain semantic paths as explanation basis during reasoning. For example, a treatment recommendation algorithm can obtain the path information "the drug → treatment of the disease → improvement of symptoms → satisfaction of patient intent" when recommending a drug, and provide it as an explanation to the doctor or patient. This embedded semantic feedback makes AI decision-making transparent and traceable to knowledge and intent nodes in the graph, enhancing user trust; Update service: The interface allows AI components or users to write or update new information to the graph nodes. For example, when a doctor corrects a diagnosis, the system can update the corresponding knowledge node through the API; when a patient expresses new preferences in follow-up, new or modified intent layer nodes are added. API design follows standard RESTful / Web Service specifications, using a unified data format (such as JSON or FHIR standard-based extensions) to represent graph nodes and relationships, ensuring interoperability between different systems. The interface is pluggable, meaning it can be easily integrated into any medical AI platform without significantly modifying existing models to utilize semantic graph functions.

[0012] Graph node update and version control: To adapt to the closed-loop process of medical diagnosis and treatment, the invention introduces a dynamic maintenance mechanism for the semantic graph: Closed-loop feedback correction: When the clinical results or patient feedback do not match the AI recommendations, the feedback is collected through the interface and applied to the graph. For example, if a treatment plan does not achieve the expected results, the system records the results and updates the status of the relevant wisdom layer nodes (such as marking "not up to standard"), while triggering a re-evaluation of the patient's intent nodes (the patient may have adjusted the goal). This closed-loop feedback enables the graph to "learn" the effects of each decision, continuously optimizing subsequent reasoning; Version management and intent evolution tracking: Graph node updates use version control, and each node's changes (such as numerical changes, status updates, and intent changes) are recorded with a history version and timestamp. This allows tracking of the entire process of patient state and intent evolution. For example, a patient's initial intent node "quickly alleviate symptoms" evolves into "avoid recurrence" in long-term management, and the system clearly records this change through versioning and can reference past intent to balance the current plan when making decisions. In addition, version control can support semantic rollback, allowing the restoration of previous correct states in case of incorrect updates, and providing a temporal dimension of semantic reference for models to help AI identify trends (such as changes in semantic chains indicating disease deterioration or improvement).

[0013] The invention constructs a multi-modal health semantic graph covering the "data, information, knowledge, wisdom, and intent" five layers, and provides complete algorithm and interface solutions, enabling medical AI systems to connect patient data, medical knowledge, and individual intent through semantics. Based on this, AI models can achieve: (a) comprehensive cognition of multi-modal information, improving the accuracy of diagnosis and treatment decisions; (b) explainable reasoning based on knowledge graphs, providing transparent display of decision-making basis; (c) personalized decision-making based on patient goals and preferences, ensuring that the diagnosis and treatment plan meets the patient's wishes; and (d) seamless integration of semantic feedback mechanisms through standard interfaces, continuously improving AI behavior. These technological innovations will effectively promote the implementation of DIKWP semantic closed-loop proactive healthcare, forming a new paradigm of patient-centered, goal-driven medical AI.

[0014] Compared with the prior art, the present application has the following beneficial effects: - Multi-modal fusion and semantic consistency: Support for unified semantic analysis and graph mapping of multi-source health data such as text, image, voice, etc. Solving the problem of heterogeneous medical information combination, realizing the lossless fusion of knowledge. AI can therefore make decisions based on complete patient graphs, reducing cognitive blind spots; - Patient intention alignment: By introducing an intention layer node in the semantic graph, the AI decision-making process takes into account the patient's goals. Diagnosis and treatment plans can therefore better meet patient preferences and ethical requirements, embodying the human-oriented direction of "proactive medical care", and improving plan acceptance and compliance; - Explainability and traceability: With the help of DIKWP-API interface, the reasoning process of any medical AI model can be linked to the causal chain in the graph, realizing white-box explanation. Doctors and patients can trace the data, knowledge and intentions on which AI recommendations are based, greatly improving system transparency and trustworthiness; - Standard interface and ecological compatibility: The interface provided by the present application complies with standard design and has good compatibility and scalability. AI modules from different manufacturers or research teams can access the unified semantic graph through DIKWP-API, realizing synergistic effect. This lays the foundation for building a medical semantic AI ecosystem and promotes the sharing and utilization of data and knowledge; - Dynamic evolution and continuous optimization: The update and version management mechanism of the semantic graph enables the system to evolve itself. Through closed-loop feedback accumulation, the graph will be continuously improved with clinical use, and the AI model can also be continuously learned and adjusted, avoiding rigid decisions that remain unchanged, and gradually approaching better diagnosis and treatment strategies.

[0015] The present application provides a complete method and interface, which realizes the semantic graph generation and intention recognition integration of multi-modal health data, significantly enhances the explainability, personalization and intelligent decision-making level of medical AI system, and has high practical application value and popularization prospect. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 Flowchart for mapping multi-modal health data to DIKWP semantic graph.

[0017] Figure 2 Integration architecture diagram of DIKWP-API interface in AI diagnosis and treatment system. DETAILED DESCRIPTION

[0018] The specific implementation of the present application will be further described below in conjunction with the drawings and examples. It should be understood that the following examples are intended to explain the present application rather than limit the scope of the present application, and those skilled in the art can make several modifications or improvements without departing from the spirit of the present application.

[0019] Example 1: Semantic graph generation and application for multi-modal diabetes management.

[0020] This example illustrates how to construct a patient semantic graph and support personalized decision making using the method of the present application, taking the management of a diabetes patient as an example. Figure 1 The data processing and graph generation process under this scenario is shown.

[0021] (1) Data collection and preprocessing: Patient Zhang, male, 45 years old, diagnosed with type 2 diabetes. The system obtains his structured data (recent blood glucose values, blood pressure, medication records, etc.), unstructured text (outpatient follow-up records, including chief complaints and doctor's recommendations), medical images (fundus photographs for retinopathy screening), and follow-up telephone recordings (patient feedback) from the hospital electronic medical record. These multi-modal data are preprocessed, including transcribing voice recordings into text and standardizing image formats.

[0022] (2) Semantic analysis and node extraction: First, run the medical NLP model on the text data to identify Zhang's symptom entities such as "dry mouth, excessive drinking", drug entities such as "metformin", and detection indicators such as "glycated hemoglobin", and identify the relationships implied in the text (such as the association between symptoms and diseases, and the treatment relationship between drugs and diseases). Second, structure numerical values such as blood glucose readings are directly mapped to data layer nodes (fasting blood glucose = 8.0 mmol / L, with a timestamp); combined with reference value ranges, generate information layer nodes "high blood sugar". Third, image data is analyzed by image analysis models to find "mild retinopathy signs", and the system converts this result to an information layer node and links it to the "diabetic retinopathy" concept in the knowledge layer. Through the above process, the system generates a number of semantic nodes for Zhang, initially covering the DIKWP layers: data layer nodes record original measurement values and observations, information layer nodes summarize clinically meaningful information (such as "high blood sugar" and "signs of complications"), knowledge layer nodes include disease diagnoses ("type 2 diabetes"), related complications and drugs, wisdom layer nodes represent the current treatment plan ("control diet + oral hypoglycemic drugs"), and intent layer nodes are currently empty or pending.

[0023] (3) Intention identification and intention layer construction: The system analyzes Zhang's previous medical records and follow-up conversations to identify his preferences for treatment goals. The Transformer intention model identifies that Zhang has mentioned "concerns about long-term medication side effects and hopes to try to control blood sugar through exercise and diet." Based on this, the system generates an intention layer node: "Patient prefers lifestyle intervention over medication." This node is represented in a structured form (intention category = treatment preference: non-drug first), and the supporting evidence (summary of Zhang's original words) is stored as a node attribute. The model also combines Zhang's demographic information and cultural background to infer that his implied intentions may include "wants to avoid complications as much as possible and maintain normal quality of life." These inferences are also recorded as intention layer nodes after being confirmed by the doctor. At this point, Zhang's semantic graph has five layers of content, with the intention layer clearly identifying his health goals and value orientation.

[0024] (4) Cross-layer semantic linking and graph completion: Next, the system integrates the above nodes into the global DIKWP graph and establishes connections: - The data layer node "empty stomach blood sugar = 8.0" is connected to the information layer node "high blood sugar" (indicating that the health information is derived from the original data). - The "high blood sugar" information node is associated with the knowledge layer node "type 2 diabetes" (indicating that this information supports the diagnosis). - The "type 2 diabetes" knowledge node is connected to the wisdom layer node "current treatment plan: drug + lifestyle intervention," which is further associated with a wisdom layer node "target blood sugar control <7.0 mmol / L" (representing the clinical target set by the medical team). - The wisdom layer treatment plan node is also linked to the intention layer nodes "prefer lifestyle intervention" and "avoid long-term medication side effects." This link indicates that the treatment plan needs to balance the patient's wishes: achieving blood sugar control goals while using the patient's preferred methods as much as possible. - In addition, the knowledge layer node "diabetic retinopathy" is connected to the wisdom layer node "regular eye examination plan," which is associated with the intention layer node "avoid complications and deterioration," indicating that this monitoring measure serves the patient's intention to avoid blindness and other serious complications.

[0025] Through the above multi-hop linking, Zhang's DIKWP semantic graph is constructed into a coherent network. Figure 1 The example flowchart shows the process of node generation and association: how the original data is elevated to information and knowledge, and through the wisdom decision node, it corresponds to the patient's intention node, forming a semantic closed loop.

[0026] (5) DIKWP-API interface application: Figure 2The semantic graph is shown in the interface application scenario of the AI diagnosis and treatment system: - Query explanation: when the AI assistant provides the next treatment suggestion (such as whether to need insulin intensive treatment) to Zhang, it queries the relevant nodes in Zhang's graph through the DIKWP-API. The assistant first retrieves all connections of the knowledge layer "type 2 diabetes" node and finds the existing wisdom layer scheme and patient intention nodes. The information returned by the API lets the AI know: "the patient's current medication control, and the intention preference is to use less medication". Therefore, when generating the suggestion, the AI will tend to increase non-drug means (such as exercise dosage) first instead of directly adding new drugs. This process ensures that the model output aligns with the patient's goal. - Embedded feedback: at the same time, the AI assistant uses the API to obtain a semantic path for explaining its suggestion. For example, the assistant recommends "increasing daily brisk walking from 30 minutes to 60 minutes", it obtains the knowledge chain from "exercise increase" to "blood sugar improvement", and the whole path from "blood sugar improvement" to "reducing drug dependence" to "meeting the patient's intention" through the interface, forming the explanation: "it is recommended to increase exercise, because "exercise increase" helps "lower blood sugar", which in turn "reduces the need for medication", which meets your goal of preferring to control blood sugar through lifestyle". These explanation contents are provided in real time by the interface, ensuring that the information is accurate, reliable and easy to understand. - Update the graph: after a period of time, if Zhang's blood sugar control improves and no new drugs are added, the system will update the relevant nodes through the API: the data layer adds a recent blood glucose monitoring value node (reduced to 6.8 mmol / L), the information layer generates a node "blood glucose target reached", the wisdom layer node "current scheme" updates the status to "effective", and an attribute label "partially satisfied" is added to the intention layer node "hope to use less medication". Conversely, if the therapeutic effect is not ideal and medication is needed, the system also updates the nodes and may add intentions (such as the patient may express "can accept insulin treatment" at this time, then add the intention node). All update operations are completed through the API with secure authorization, and new version records are generated for subsequent audit and analysis.

[0027] (6) Closed-loop optimization: in the long-term follow-up process, Zhang's semantic graph is constantly improved with each interaction, recording the development of the disease and the change of intention. When the AI system formulates an annual management scheme for Zhang, the graph provides a consistent knowledge and intention context throughout the process. For example, if the patient cannot adhere to exercise due to busy work at a certain stage, the intention node may temporarily change to "hope to simplify the treatment scheme", and the system adjusts the scheme to focus on simplicity accordingly. After that, if the situation improves, the intention node version may return to "emphasis on lifestyle". Throughout the process, the AI obtains the latest graph status through the interface at each decision point, truly forming a new semantic closed-loop medical process of data -> knowledge -> decision -> feedback.

[0028] Through the above embodiments, it can be clearly seen that the application effect of the application in the actual medical scene: the multi-modal data is fused into a unified semantic graph, the AI can obtain an explanatory knowledge chain to support decision-making, and the scheme can be adjusted at any time according to the patient's intention. With the help of the DIKWP model framework and interface, the application enables the AI diagnosis and treatment system to have the ability similar to that of a doctor, that is, to know the phenomenon and the reason, and to respect the patient's intention. This ability has a significant advantage in improving the quality of medical decision-making, communication efficiency and patient satisfaction.

[0029] It should be noted that the embodiments of the application are not limited to the above examples. In different medical fields (such as personalized tumor treatment, integration of traditional Chinese medicine diagnosis and treatment knowledge, etc.), corresponding DIKWP semantic graphs and interfaces can be constructed according to the idea of the application. For example, in tumor treatment, the intention layer can represent the patient's preference for the trade-off between survival time and quality of life; the system can integrate multi-modal data such as lesion changes in images and gene test results into the graph to provide more comprehensive background knowledge and target guidance for tumor AI. For another example, in the traditional Chinese medicine scene, the knowledge layer can include traditional Chinese medicine syndrome knowledge, the wisdom layer can include syndrome differentiation and treatment scheme, and the intention layer can also cover the patient's willingness for efficacy and conditioning method. All these variations fall within the protection scope of the application.

[0030] In summary, the multi-modal health data DIKWP semantic graph generation and intention recognition interface method provided by the application can be widely applied to various medical AI systems, significantly enhancing the semantic understanding depth of intelligent decision-making and the harmony of human-computer interaction. The technical scheme of the application is novel, feasible and has industrial application value.

Claims

1. A DIKWP semantic graph generation and intent recognition interface method for multimodal health data, characterized in that, Includes the following steps: a) Acquire multimodal health data from patients, including structured electronic health record data, medical imaging data, and voice interaction data; preprocess various types of data to meet the format requirements of subsequent semantic parsing; b) Perform natural language semantic parsing on the preprocessed text data to extract medical entities and relationships, conceptualize symptoms, signs, diagnoses, test indicators, treatment measures, etc., as semantic nodes, and identify the semantic relationships between them; perform image analysis on medical imaging data to obtain the clinical representation reflected by the images, and convert the representation into text or symbolic descriptions as semantic nodes; after speech recognition and transcription of voice data into text, apply the natural language semantic parsing to extract the corresponding entity and relationship nodes; c) Classify the semantic nodes extracted in step b) according to the DIKWP five-layer model: map the original observations or measurement data to data layer nodes, map the factual descriptions derived from the data to information layer nodes, map medical concepts and general knowledge to knowledge layer nodes, and map decision-making schemes for specific patients to wisdom layer nodes; d) Provide an intent recognition model to analyze text interaction information from patients, identify patients' health goals and preference intents, and represent the identified intent concepts as semantic nodes in the intent layer; wherein the intent recognition model adopts a Transformer-based deep learning architecture and integrates medical corpora through knowledge distillation to support the alignment of Chinese natural language with professional medical terminology. e) Construct the DIKWP semantic graph: Establish node sets in the data layer, information layer, knowledge layer, intelligence layer and intent layer respectively, and create cross-layer links based on the semantic relationships between nodes; This includes: connecting data layer nodes with their generated information layer nodes, connecting information layer nodes with corresponding knowledge layer medical concept nodes, connecting knowledge layer nodes with their driven wisdom layer decision nodes, and connecting wisdom layer nodes with related intent layer nodes, thereby forming a semantic network graph covering a multi-hop relationship of data → information → knowledge → wisdom → intent; performing node fusion and disambiguation for the same entity or concept from different sources to ensure that each entity in the graph has only one standardized representation and semantic consistency; f) providing a standardized DIKWP-API interface for external systems to access the semantic graph, including: a query interface for retrieving graph nodes and their relationships based on patient ID or semantic conditions, an update interface for receiving new data, new conclusions, or feedback from external input to dynamically update the corresponding semantic nodes or relationships, and an interpretation interface for returning semantic links for AI decision interpretation based on specified node paths; g) Establish a version control and closed-loop update mechanism for semantic graph nodes: assign a version identifier and timestamp to each node's content update and save historical change records; during patient diagnosis and treatment, when there is new medical feedback or changes in patient intentions, update the graph with the feedback information through the DIKWP-API interface and generate a new version node, thereby realizing closed-loop maintenance of the semantic graph and tracking of patient intention evolution.

2. The method according to claim 1, characterized in that, The text semantic parsing in step b) includes using a medical knowledge base and terminology ontology to improve the accuracy of entity recognition and disambiguation, and mapping synonyms or aliases to the same knowledge layer node when they are detected; the image data analysis uses a pre-trained medical image model and combines it with an image description report template to convert the model output into a structured description.

3. The method according to claim 1, characterized in that, The knowledge distillation process of the intent recognition model in step d) includes: using a large medical language model (pre-trained self-attention network) as the teacher model to train on a large amount of Chinese medical dialogue data labeled with patient intent, thereby obtaining high-performance intent classification and feature extraction capabilities; then training a lightweight student model to receive guidance from the output of the teacher model, so that it can significantly reduce the number of parameters while maintaining the accuracy of intent recognition, making it easy to embed into the medical AI system for real-time operation.

4. The method according to claim 1, characterized in that, Step e) The establishment of cross-layer links includes: to achieve multi-hop semantic reasoning, direct associations are allowed between any two layers of nodes in the DIKWP graph. When there is a reasoning relationship that crosses the intermediate layer, a direct edge is created for the relationship and the semantic type is labeled. As a result, the graph supports the representation of 25 inter-layer mapping modules, enabling the AI ​​model to flexibly retrieve relevant information along different paths when querying.

5. The method according to claim 1, characterized in that, Step f) The DIKWP-API interface is implemented using a RESTful architecture or an equivalent service method, and its input and output use a common data exchange format. The query interface supports parameters including specifying patient identifiers, node type filtering, relationship path length limits, etc., and can return a set of matching nodes and their associated relationship subgraphs. The interpretation interface takes one or more target nodes as input and outputs a causal chain path from the data layer to the intent layer, which is used to generate natural language interpretations.

6. The method according to claim 1, characterized in that, After receiving new information submitted by an external system, the update interface described in step f) performs the following sub-steps: verify the submitter's permissions and data format → insert the new data node or information node into the corresponding layer of the graph and establish a relationship with existing nodes → if the submitted result is a new decision, update the corresponding wisdom layer node or create a new wisdom node → if the submitted result is a new patient preference, update the intent layer node or version → record all changes and return a confirmation of successful update.

7. The method according to claim 1, characterized in that, Step g) The closed-loop update mechanism further includes: pre-setting several business rules to automatically evaluate the necessity and consistency of the graph update; when a significant change in important patient indicators or an update of intent nodes is detected, triggering the AI ​​model to call the query interface again to obtain the latest graph subgraph for adjusting decisions; at the same time, setting subscription notifications for key nodes, reminding the responsible doctor to review when intent layer nodes change, so as to optimize the treatment plan under human-machine collaboration.

8. The method according to any one of claims 1 to 7, characterized in that, The system also includes the construction of a medical intelligent diagnosis and treatment system that incorporates the multimodal semantic graph and interfaces. This system comprises the following modules: a multimodal data parsing module, an intent recognition module, a DIKWP graph storage and inference module, an interface service module, and a user interaction module. Specifically, the multimodal data parsing module performs text, image, and voice data processing in step b); the intent recognition module extracts patient intent in step d); the graph storage and inference module constructs and queries the graph in step e); the interface service module provides the API services described in step f); and the user interaction module provides a visual interface for doctors and patients to view and provide feedback on the semantic graph information. All modules are connected and work collaboratively via a network or bus to achieve proactive medical intelligent diagnosis and treatment functions based on the DIKWP semantic closed loop.

9. A computing device or server system having program code installed thereon, which, when executed on the device, implements the multimodal health data semantic graph generation and intent recognition interface method of claim 1. The device includes a processor, storage, and a network interface to host the DIKWP semantic graph database and API services, receiving requests from medical AI applications through the network interface and returning semantic query results or confirmation updates to support concurrent medical semantic computation by multiple users.

10. A computer-readable storage medium having a computer program stored thereon, which, when run on a computing device, performs the steps of the method described in any one of claims 1 to 7. The program implements functions for semantic parsing, graph construction, intent recognition, and interface services of multimodal medical data, enabling general-purpose computer hardware to provide semantic graph support and intent alignment services for medical AI.

Images