A personalized cognitive intervention and narrative nursing method based on a generative large language model

Abstract

This invention discloses a personalized cognitive intervention and narrative care method based on a generative large language model. Addressing the core shortcomings of existing intelligent elderly care systems, such as weak semantic understanding, insufficient integration of professional knowledge, uncontrollable output, high knowledge maintenance costs, and a lack of personalized humanistic care, this invention integrates technologies such as generative large language models, domain-adaptive RAG, knowledge graphs, and vector databases to construct a personalized cognitive intervention and narrative therapy system for the elderly care industry. Through core designs including industry-specific semantic encoding, hybrid precise retrieval, structured prompt word construction, automated knowledge base management, and narrative therapy integration, this invention fundamentally solves the pain points of existing systems from a technical perspective. Simultaneously, it adapts to the professional, refined, and humanistic service needs of the elderly care industry, possessing both technical practicality, industry adaptability, and scenario extensibility.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and elderly care technology, and in particular to a personalized cognitive intervention and narrative care method based on a generative large language model. Background Technology

[0002] Existing intelligent question-and-answer and nursing assistance systems used in elderly care institutions mainly fall into three categories, all providing information response services based on consultation and care needs in elderly care scenarios. Specific solutions are as follows: 1. Keyword-matching FAQ systems: These systems use preset keywords as the core matching basis. After receiving user input, they can only perform simple keyword comparison and result feedback, lacking natural language parsing modules. They cannot handle non-standardized natural language expressions generated by the elderly, family members, or caregivers in actual elderly care scenarios, including vague needs descriptions, implicit care intentions, and colloquial consultation expressions. 2. Traditional knowledge base systems: These systems rely on manual input as the core knowledge acquisition method, requiring professionals to manually input elderly care-related content. After being organized into structured knowledge items and entered into the system, knowledge maintenance still relies on manual operation, resulting in problems such as lagging knowledge updates and low response efficiency. It is difficult to adapt to the rapid iteration of knowledge in the elderly care industry and the localized procedure adjustment needs of different elderly care institutions. General-purpose large language model question-answering systems that have not been enhanced by domains rely on the open-domain dialogue capabilities of general-purpose large language models to respond to user questions. They have not been specifically optimized and integrated with knowledge specific to the elderly care industry. They cannot connect with industry-specific professional knowledge such as elderly care procedures, elderly risk assessment standards, and personalized elderly files. This can easily create content "illusions," and the output suggestions may be inaccurate or non-compliant, or even contain content that violates elderly care safety regulations, creating safety risks.

[0003] The core working principle of the aforementioned existing systems is a linear processing flow. That is, after receiving user input, keyword matching is directly completed through preset rules, or a general large language model is directly called to generate content, and finally the matching result or generated content is directly output. The entire processing lacks a deep understanding of the semantics of user input, dynamic enhancement of dialogue context, and intelligent association mechanism with the knowledge base dedicated to the elderly care industry, and cannot meet the professional and personalized service needs of the elderly care scenario.

[0004] The various systems currently used by elderly care institutions have revealed numerous technical deficiencies in practical application, making them ill-suited to the high-quality and refined nursing service requirements of the elderly care industry. Specific objective shortcomings include: First, weak semantic understanding capabilities. The systems can only recognize standardized textual expressions and preset keywords, lacking the ability to analyze non-standardized expressions. For vague requests from the elderly during actual care, such as "I want water, but not the cold kind," which carries a clear context, the systems cannot accurately identify the core intent, leading to a disconnect between service responses and actual needs. Second, a lack of integration of industry professional knowledge. The systems do not integrate the professional knowledge system of the elderly care field, nor do they connect to core professional content such as standard operating procedures for elderly care, regulations for the management of medications for the elderly, and procedures for the care of special elderly individuals. The output responses lack the authority of the elderly care industry and practical operability in actual care, failing to provide caregivers with professional execution guidelines. Third, information consistency and controllability are poor. In particular, general-purpose large language model question-and-answer systems without domain enhancement are prone to "fabricated" content when lacking industry knowledge constraints and output rule restrictions. The output results are likely to conflict with the local nursing procedures and industry standards of elderly care institutions, which not only undermines the consistency of elderly care information but also creates compliance loopholes and safety risks in care operations and risk management, affecting the safety of elderly care services. Fourth, knowledge update and maintenance costs are high. The knowledge base of the existing system is maintained mainly by humans. Whether it is the initial knowledge entry or subsequent content updates and version adjustments, it must be completed manually by professionals. It is impossible to achieve automatic synchronization of knowledge sources in the elderly care industry and version management of the knowledge base. This not only increases the human and time costs of elderly care institutions but also causes the knowledge base content to be out of sync with the latest industry standards and local adjustments of the institution, further reducing the practical value of the system.

[0005] With the deepening of population aging, the demand for intelligent, professional, and personalized nursing assistance systems in elderly care institutions is becoming increasingly urgent. Existing systems, due to the aforementioned shortcomings, can no longer meet the development requirements of the elderly care industry. On the one hand, the service recipients in elderly care settings are seniors, whose language is often ambiguous and colloquial, and different seniors have individualized physical conditions and care needs, placing higher demands on the semantic understanding and personalized adaptation capabilities of the system. On the other hand, elderly care is a highly specialized field, with strict industry norms and operational standards for nursing procedures, medication management, and risk assessment. Furthermore, industry knowledge and local institutional procedures are constantly being updated and iterated, imposing stringent requirements on the system's integration of professional knowledge, efficiency of knowledge updates, and controllability of output content. Therefore, developing an intelligent nursing assistance system that can overcome the shortcomings of existing technologies and adapt to the characteristics and needs of the elderly care industry has become crucial for the intelligent upgrading of elderly care institutions. Summary of the Invention

[0006] This invention provides a personalized cognitive intervention and narrative nursing method based on a generative large language model, comprising: Encode the original question into a semantic vector; Based on the semantic vector, a retrieval is performed to retrieve multiple relevant text fragments from the elderly care knowledge base; Structured enhanced prompts are constructed based on the relevant text fragments; Based on the structured enhanced prompts, structured care suggestions are obtained through a generative large language model.

[0007] In one embodiment of the present invention, the method further includes constructing and automatically updating a knowledge base for elderly care, the steps of which include: Automated import of various data in the elderly care field through multiple channels; Perform standardized preprocessing on all imported data; Segment the preprocessed text; Each processed text segment is input into the BGE-M3 model for encoding, generating a multi-dimensional embedding vector. The generated multi-dimensional embedding vectors are used to construct an elderly care knowledge base according to a standardized field structure. When changes to the content of each data source are detected, the detected changes will be automatically imported into the database according to the above process; And record the version number for each update of the elderly care knowledge base.

[0008] In one embodiment of the present invention, encoding the original question into a semantic vector includes: Input the received original question into the BGE-M3 model; Expand the original question with industry-specific terminology; Fine-tuning was performed using standardized corpora in the field of elderly care. The original question was encoded to include healthcare-related semantics, inferences about the elderly's implicit conditions, and expressions of question intent. Semantic vectors associated with nursing processes.

[0009] In one embodiment of the present invention, the step of retrieving multiple relevant text fragments from the elderly care knowledge base based on the semantic vector includes using vector similarity retrieval + BM25 keyword retrieval.

[0010] In one embodiment of the present invention, the construction of structured enhanced prompt words based on the relevant text fragments includes: Evidence set E' is selected from the relevant text fragments E recalled; Extract the field P' that is directly related to the current problem from the user's personalized profile P; If the current problem involves risky operations, automatically associate the corresponding risk control constraints from the constraint set C; By combining the original question with the user's personalized profile, a set of constraints C' specifically designed for this question is formed; Based on the original question Q, the filtered evidence set E', the extracted personalized information P', and the exclusive constraint set C', structured enhanced prompt words are formed.

[0011] In one embodiment of the present invention, the evidence set is filtered from relevant texts using the following formula: ; in The i-th text segment is given a comprehensive weighted score, with a higher score indicating a stronger match with the question and greater authority. α is the weighting coefficient of vector cosine similarity, used to adjust the proportion of semantic matching in the score; The encoding vector of the user's original question Q and the i-th text fragment The vector cosine similarity between the encoded vectors represents the degree of semantic matching between them; β is a weighting coefficient for the authority of the text source, used to adjust the proportion of authority in the score; For the i-th text segment The source authority coefficient is set according to the authority level of knowledge in the elderly care industry, with priority being SOP; Based on the comprehensive weighted scores, the evidence fragments ranked from highest to lowest are selected, and the top K fragments are used to form the core evidence set E' after screening.

[0012] In one embodiment of the present invention, the step of segmenting the preprocessed text includes: Combine multiple related nursing steps into a single text chunk to maintain the semantic integrity of the nursing process; Each text block is segmented to correspond to the elderly care process to facilitate subsequent retrieval. Add semantic tags specific to the elderly care industry to each text segment.

[0013] In one embodiment of the present invention, the multi-dimensional embedding vector includes: Text vectors are the basic semantic vectors of the text segments themselves; Paragraph tag vectors, generated based on semantic tags for text segments; Process path vector, a vector generated based on the corresponding elderly care process; Personalized vectors are vectors generated based on the individual characteristics of the elderly to adapt to the retrieval needs of personalized care.

[0014] In one embodiment of the present invention, it further includes: With the user's authorization, retrieve the user's personal information from the elderly care knowledge base; Using a large language model, personalized information is dynamically generated based on the retrieved information. Based on the personalized information, we ask and answer questions with users to provide structured care recommendations.

[0015] This invention also provides a personalized cognitive intervention and narrative care system based on a generative large language model, comprising: The encoding module is configured to encode the original question into a semantic vector; The retrieval module is configured to retrieve multiple relevant text fragments from the elderly care knowledge base based on the semantic vector. The prompt word generation module is configured to construct structured enhanced prompt words based on the relevant text fragments; and The nursing suggestion generation module is configured to obtain structured nursing suggestions based on the structured enhanced prompts and through a generative large language model.

[0016] The present invention has the following beneficial effects: (1) By expanding the vocabulary of the elderly care industry, fine-tuning the professional corpus and enhancing the knowledge graph of the BGE-M3 model, the system has achieved in-depth analysis of users' non-standardized expressions. It can accurately identify the vague descriptions, colloquial expressions and implicit care intentions of the elderly, caregivers and their families. It breaks through the limitation of existing systems that can only match preset keywords, effectively solves the problem of demand identification deviation caused by the characteristics of language expression in the elderly care scenario, and ensures the system's accurate response to various consultation needs in the elderly care scenario.

[0017] (2) A multi-source structured knowledge base exclusive to the elderly care industry has been constructed, integrating professional content such as nursing SOP, drug management standards, special care procedures, and risk assessment standards. The system achieves accurate matching between questions and authoritative knowledge through a hybrid retrieval method of vector similarity + BM25, ensuring that all suggestions output by the system are supported by professional knowledge in the elderly care industry. This completely solves the problem of existing systems lacking professional knowledge integration and providing answers that are not practical, providing nursing staff with a basis for care operations that can be directly implemented and improving the professionalism of elderly care services.

[0018] (3) A complete knowledge base management system has been constructed, which includes automatic import of multi-source heterogeneous data, intelligent preprocessing, structured segmentation, vectorized storage and automatic updates. It supports automatic synchronization of knowledge from multiple channels such as enterprise documents, external professional resources and medical records. It also has functions such as change detection, version backtracking and compliance auditing. It completely eliminates the dependence of the existing system on manual input and maintenance, greatly reduces the manpower and time costs of knowledge management in elderly care institutions, and realizes real-time synchronization of knowledge base content with the latest industry standards and local institutional procedures, ensuring the timeliness of knowledge. Attached Figure Description

[0019] Figure 1 The flowchart of a personalized cognitive intervention and narrative nursing method based on a generative large language model is shown in one embodiment of the present invention. Detailed Implementation

[0020] In the following description, the invention is described with reference to various embodiments. However, those skilled in the art will recognize that the embodiments may be practiced without one or more specific details or with other alternatives and / or additional methods, materials, or components. In other instances, well-known structures, materials, or operations are not shown or described in detail so as not to obscure the inventive points of the invention. Similarly, for illustrative purposes, specific quantities, materials, and configurations are set forth to provide a comprehensive understanding of embodiments of the invention. However, the invention is not limited to these specific details.

[0021] In this invention, the various embodiments are merely intended to illustrate the solutions of the invention and should not be construed as limiting.

[0022] In this specification, references to "an embodiment" or "this embodiment" mean that a particular feature, structure, or characteristic described in connection with that embodiment is included in at least one embodiment of the invention. The phrase "in one embodiment" appearing throughout this specification does not necessarily refer to the same embodiment in all instances.

[0023] Furthermore, the numbering of the steps in the methods of the present invention does not limit the execution order of the method steps. Unless otherwise specified, the method steps may be executed in different orders.

[0024] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0025] Figure 1 The flowchart of a personalized cognitive intervention and narrative nursing method based on a generative large language model is shown in one embodiment of the present invention.

[0026] like Figure 1 As shown, in this embodiment, the personalized cognitive intervention and narrative nursing method based on a generative large language model includes: S1. Original problem multi-source input and industry-specific semantic encoding: The system supports diverse input formats for three core stakeholders in elderly care scenarios. It can directly receive text / voice questions from caregivers, voice statements from the elderly (converted to text after speech recognition), and online / offline natural language consultations from family members. It is compatible with non-standardized expressions that are colloquial, ambiguous, or have implicit intentions, and there are no preset keyword input restrictions.

[0027] The received natural language question is input into a BGE-M3 model customized and optimized for the elderly care industry. A three-layer model optimization method is used to complete the semantic encoding of the question, ultimately generating a high-dimensional semantic vector containing professional characteristics of the elderly care industry. The model optimization method is as follows: Industry vocabulary expansion: Incorporate specialized terms for elderly care into the model, such as blood glucose monitoring, dysphagia, Barthel index, pressure ulcer prevention, etc., to enhance the model's ability to identify and associate industry terms. Industry-specific language fine-tuning: The model is trained using standardized language in the elderly care field, including nursing SOPs, chronic disease management manuals, rehabilitation guidelines, psychological intervention texts, and emergency response plans, so that the model can be adapted to the expression logic and professional needs of the elderly care industry. Knowledge Graph Enhancement: Construct a knowledge graph specifically for the elderly care field, establish the relationship between people, symptoms, processes, risks, and measures, and realize the automatic association between problems and nursing processes and risk points.

[0028] The generated high-dimensional semantic vector has four core semantic dimensions, which comprehensively cover the problem analysis needs of the elderly care scenario: medical care related semantics, inference of the elderly’s implicit condition, expression of problem intent, and possibility of related nursing process.

[0029] S2. Semantic retrieval and precise matching with a knowledge base specifically for elderly care: The core of this step is to intelligently associate the encoded question vector with the elderly care knowledge base, providing authoritative and relevant professional knowledge evidence to support the subsequent construction of prompt words, and solving the problem of the lack of industry professional knowledge integration in existing technologies. The specific operation is as follows: The high-dimensional semantic vector encoded by the BGE-M3 model is then used in a hybrid retrieval method combining vector similarity retrieval and BM25 keyword retrieval. This dual retrieval and matching is performed from a dedicated vector knowledge base for elderly care, retrieving multiple authoritative text fragments highly relevant to the question and forming a set of evidence fragments. ; The recalled text fragments cover professional content such as elderly care procedures, operating procedures, assessment standards, and management principles, providing core knowledge basis for the subsequent construction of structured prompts.

[0030] S3, Contextual Enhancement and Intelligent Construction of Structured Prompts: This step is the core technology of the system. It utilizes a proprietary algorithm to fuse multi-source information—including questions, evidence, personalized information, and compliance constraints—to generate a structured, enhanced prompt. This provides precise guidance and strong constraints on the output of the large language model, mitigating the "illusion" problem inherent in large models. This step has clearly defined input, processing, and output requirements. The processing is divided into four sub-steps, with the evidence selection and ranking step including a proprietary weighted scoring formula. The specific operations are as follows: The core inputs for this step include: User's original natural language question Q; The retrieved collection of relevant evidence fragments ; Personalized user profile structured information P (including elderly patient's medical history, medication, nursing dependence level, etc.); The system has a built-in set of security and compliance constraints, C.

[0031] S3.1. Evidence selection and ranking: A semantic similarity + authority weighted scoring mechanism is used to evaluate each evidence fragment e in the evidence fragment set E. i A comprehensive evaluation is conducted to select the core evidence that best matches the question and is of the highest authority. The specific steps are as follows: For each piece of evidence e i The core formula for calculating the comprehensive weighted score is: ; in The i-th text segment is given a comprehensive weighted score, with a higher score indicating a stronger match with the question and greater authority. α is the weighting coefficient of vector cosine similarity, used to adjust the proportion of semantic matching in the score; The encoding vector of the user's original question Q and the i-th text fragment The vector cosine similarity between the encoded vectors represents the degree of semantic matching between them; β is a weighting coefficient for the authority of the text source, used to adjust the proportion of authority in the score; For the i-th text segment The source authority coefficient is set according to the authority level of knowledge in the elderly care industry, with priority being SOP; Based on the comprehensive weighted scores, the evidence fragments ranked from highest to lowest are selected, and the top K fragments are used to form the core evidence set E' after screening.

[0032] S3.2 Personalized information fusion enables precise association between user's personalized profile and current issues, laying the foundation for generating subsequent personalized nursing recommendations. Specific operations include: From the user's personalized profile information P, the structured fields directly related to the current problem are accurately extracted. The core fields include the elderly's medical history, medication records, diet / lifestyle preferences, care dependency level (ADL / Barthel index), and previous care records, to obtain the extracted personalized information P'. If the current issue involves risky operations such as medication advice, activity guidance, or dietary adjustments, the system will automatically trigger a risk label matching mechanism to associate the corresponding risk prevention and control constraints from the constraint set C, thereby proactively avoiding safety risks in personalized care.

[0033] S3.3 Constraint Parsing and Mapping: This step involves selecting constraint entries from the system's built-in constraint library that are suitable for the current problem and instantiating them. This ensures that the content generated by the subsequent large model conforms to elderly care industry standards and local institutional procedures. Specific operations include: The system's built-in general constraint library C contains three core constraints, covering all dimensions of compliance requirements in elderly care scenarios: safety constraints (such as nursing advice cannot replace professional medical diagnosis), compliance constraints (such as answers must cite authoritative knowledge sources), and operational constraints (such as advice must have clear steps and can be immediately executed by caregivers). The system dynamically selects suitable constraint items from the constraint library C based on the problem type (such as gastrointestinal discomfort care, fall prevention, medication consultation, etc.) and evidence type (such as SOP, professional guidelines, assessment criteria, etc.), and instantiates the constraints by combining the current problem with the user's personalized information, forming a unique constraint set C' for this problem.

[0034] S3.4, Structured Template Filling: Integrates multi-source information into a standardized template to generate the final structured enhanced prompt. Specific steps: The system calls a predefined "problem-evidence-constraint" structured template, which contains five fixed core fields: problem description, relevant evidence, personalized information, constraints, and task instructions. The user's original question Q, the filtered core evidence set E', the extracted personalized information P', and the exclusive constraint set C' are precisely filled into the corresponding fields of the template to generate the final structured enhanced prompt.

[0035] S4. Industry-specific fine-tuning of large language models to generate compliant output: The core of this step is to generate professional, personalized, and compliant content based on structured prompt words (Prompt). The output can directly provide practical guidance for elderly care work. The specific operation is as follows: The structured enhanced prompt word Prompt is input into a large language model that has been fine-tuned specifically for the elderly care industry. The model has been trained with elderly care industry corpus and is adapted to the professional logic and expression requirements of elderly care. The model generates content under the constraints and guidance of the Prompt, ensuring that the output content meets four core characteristics: professionalism, controllability, auditability, and does not produce false or fabricated content. It conforms to the local operating procedures of elderly care institutions and integrates personalized risk warnings for users. The model generates content in a structured format, making it easy for nursing staff to execute directly. The core output types include: nutrition and lifestyle recommendations, nursing task work orders, risk assessment and reporting prompts, and structured nursing plans.

[0036] This invention also provides a method for the automated construction and updating of a vector knowledge base specifically for elderly care. This process serves as the knowledge foundation for the entire system, enabling the automated integration, structured processing, vectorized storage, and real-time updating of multi-source heterogeneous knowledge in the elderly care field. It addresses the high costs of knowledge updating and maintenance in existing technologies, providing authoritative and real-time knowledge support for all the aforementioned business processes. The specific operation consists of seven sub-steps: The system integrates heterogeneous data from multiple sources and supports automated import of various data from the elderly care field through multiple channels, eliminating the need for manual data entry and ensuring the comprehensiveness of the knowledge base. Data sources are categorized into three types: Internal company documents: Standard Operating Procedures (SOPs), work instructions, quality management manuals, emergency response plans, and medical and nursing operation standards, etc. External professional resources: guidelines for geriatric medical care, local government standards and regulations for elderly care services, industry professional research reports, etc.; Structured medical and nursing records include: basic patient information, nursing level / dependency, medication records, medical and nursing rounds records, meeting minutes, and rectification documents.

[0037] The standardization preprocessing of original documents addresses the issues of chaotic formats, high repetition rates, complex table structures, and sensitive information in the original data of the elderly care industry. It standardizes and preprocesses all imported data to ensure data usability and security. The processing methods include: layout parsing (unifying the layout structure of documents of different formats), sensitive information desensitization (protecting the privacy of the elderly), and chapter and clause structuring (removing duplicate content and sorting out text logic).

[0038] Text segmentation specifically designed for elderly care scenarios addresses the characteristics of elderly care scenarios, such as numerous processes, rigorous steps, and strong cross-chapter connections. It employs three enhanced segmentation strategies to segment preprocessed text, avoiding the semantic fragmentation and low retrieval accuracy issues caused by conventional segmentation. This ensures the semantic integrity and retrieval adaptability of each text chunk: Meta-chunking: Combines multiple related nursing steps into a single chunk, maintaining the semantic integrity of the nursing process; Latechunking: Makes each chunk precisely correspond to a key nursing action, facilitating accurate retrieval in subsequent searches; Chunk Rich Semantic Tag Enhancement: Add semantic tags specific to the elderly care industry to each chunk to improve search accuracy and reliability.

[0039] Multi-dimensional embedding vector generation involves inputting each processed text chunk into the BGE-M3 model for encoding, generating multi-dimensional high-dimensional embedding vectors that not only contain the semantics of the text itself but also integrate the process and personalized features of the elderly care scenario. This supports multi-dimensional semantic retrieval, and the generated vector types include: text vectors, paragraph tag vectors, process path vectors (for "process semantic retrieval"), and personalized vectors (based on the elderly person's situation).

[0040] The system constructs a standardized vector database, supporting mainstream vector engines (FAISS / Milvus / pgvector / Elasticsearch-vector). It builds a dedicated vector database for elderly care by standardizing the generated multi-dimensional embedded vectors according to a standardized field structure, ensuring the database's manageability and searchability. The core fields of the vector database include: unique chunk identifier (id), embedded vector (vector), raw text fragment (raw_text), hierarchy path (hierarchy_path), risk level (risk_level), associated nursing procedure (sops), automatic update date (update_time), and version number (version_tag).

[0041] The knowledge base is updated automatically. Addressing the frequent updates to SOPs, industry standards, and nursing procedures in elderly care institutions, a fully automated knowledge base update strategy is implemented. No manual maintenance is required, ensuring the timeliness and compliance of the knowledge base content. The update strategy includes: change detection (automatically detecting changes in content from various data sources), automatic database entry (automatically importing changed content into the database according to the aforementioned process), version rollback (recording each update version and supporting historical version queries), and compliance and audit logging (automatically recording update content, time, and source to form a complete audit log).

[0042] The privatization of the knowledge base has been initially constructed. After completing the above steps, a private vector knowledge base integrating professional knowledge of elderly care and personalized information of the elderly will be formed, providing data support for subsequent personalized cognitive intervention and narrative therapy applications.

[0043] In another embodiment of the present invention, the process of the present invention is illustrated with reference to a specific example: This embodiment uses a nursing consultation scenario involving 82-year-old Grandma Li, a hypertensive resident in room 3 of a nursing home, who is experiencing abdominal distension and loss of appetite. It details the entire operation of this system. The core of the system consists of a prompt word generation and enhancement layer and a knowledge base construction and retrieval layer. These two layers work together to generate personalized, professional, and compliant elderly care suggestions, while also implementing narrative therapy and cognitive intervention functions. The following embodiment is only for explaining the invention and is not intended to limit the scope of protection of the invention.

[0044] In this embodiment, nursing staff at the elderly care facility initiate a consultation request through this system to obtain a professional care plan for Grandma Li's abdominal distension and loss of appetite. The specific operation flow of this system is as follows: Step 1: Multiple Sources of Input for the Original Problem: Nurses can submit consultation questions in natural language on the terminal interface of this system: "Grandma Li (82 years old) in room 3 has been complaining of bloating and loss of appetite lately. What should we do?" This input format belongs to the multi-source input type of nurses' questions supported by this system. The system can directly receive this type of non-standard natural language expression with specific scenarios and has no preset keyword input restrictions.

[0045] Step 2: Semantic encoding of the original problem in the elderly care industry: The system will input the received natural language questions into a BGE-M3 model, which is customized and optimized for the elderly care industry, for high-dimensional semantic vector encoding. The model uses a three-layer optimization method to achieve in-depth analysis and industry-specific encoding of the questions: Industry vocabulary expansion: The model calls on the professional terminology library of elderly care and automatically associates exclusive professional terms such as "abdominal distension", "loss of appetite", "care for elderly people with hypertension", "Barthel index" and "prevention of drug side effects" to clarify the medical and nursing attributes and industry scenarios of the problem; Industry-specific language fine-tuning: Based on the SOP for elderly gastrointestinal discomfort care, the care manual for elderly with hypertension, and the guidelines for the prevention and control of drug side effects in elderly people with chronic diseases, the model is adapted to the care logic for gastrointestinal problems of elderly patients with chronic diseases. Knowledge Graph Enhancement: Through the knowledge graph of people-symptoms-processes-risks-measures constructed by this system in the field of elderly care, the system automatically establishes the following relationships: abdominal distension → gastrointestinal discomfort nursing process + dietary adjustment principles + constipation risk assessment + activity safety constraints for elderly people with hypertension.

[0046] The final result is a high-dimensional semantic vector containing four semantic dimensions, fully covering the parsing requirements of this problem: Medical and nursing related semantics: professional care for elderly patients with hypertension and gastrointestinal discomfort; Implicit condition in the elderly person: It is highly likely that constipation is caused by the side effects of antihypertensive medication and insufficient daily activity, which in turn leads to abdominal distension and loss of appetite; The intended message of the problem: To obtain practical and step-by-step suggestions for on-site nursing procedures; Possible related nursing procedures: basic assessment of gastrointestinal discomfort, adjustment of low-gas-producing diet, promotion of intestinal motility, and guidance on safe activities for elderly people with hypertension.

[0047] Step 3: Semantic retrieval and precise matching with the elderly-specific knowledge base: The system uses a hybrid retrieval method combining vector similarity search and BM25 keyword search to perform dual matching from the elderly care-specific vector knowledge base built by the system, recalling authoritative evidence fragments highly relevant to this issue. At the same time, based on the knowledge matching needs of the elderly care industry, the system selects Top-K core evidence through weighted scoring to form the evidence set E' constructed by the prompt words. The core evidence obtained in this search are: "Nursing Procedures for Gastrointestinal Discomfort in the Elderly", "Principles for Dietary Adjustment in Elderly Care Institutions", "Standards for Hydration and Intestinal Management in the Elderly", "Standards for Constipation Risk Assessment in the Elderly", and Grandma Li's personal structured file information.

[0048] Step 4: Contextual Enhancement and Intelligent Construction of Structured Prompts: This step is the core technical link of the system. It takes the user's original question Q, the core evidence set E', Grandma Li's personalized profile P, and the system security and compliance constraint set C as inputs. The structured prompt word construction algorithm designed in this invention completes the entire process and finally generates the enhanced structured prompt word Prompt.

[0049] Step 5: Generating compliant content for industry-fine-tuned large language models: The structured enhanced prompt word Prompt, which has been constructed as described above, is input into a large language model that has been fine-tuned specifically for the elderly care industry. Under the strong constraints and precise guidance of Prompt, the model generates content, avoiding the "illusion" problem of general large models from the source. This ensures that the output content meets the requirements of being professional, controllable, auditable, and free of false information, and strictly matches the local operating procedures of elderly care institutions, while incorporating Grandma Li's personalized risk warnings.

[0050] In response to Grandma Li's abdominal distension, the model generated structured nursing recommendations according to the constraints. These recommendations were divided into four standardized modules: preliminary assessment recommendations, immediate nursing measures, risk warnings and reporting conditions, and record and follow-up recommendations. Each module contained step-by-step operations that nursing staff could directly execute. At the same time, the model fully considered Grandma Li's history of hypertension, drug side effects, and moderate nursing dependence, and clarified the safety boundaries for activity guidance and dietary adjustments.

[0051] Step 6: Support and automated updates of the elderly care-specific knowledge base: In this embodiment, all authoritative evidence fragments and elderly personal file information retrieved by the system originate from the elderly-specific vector knowledge base constructed by this invention. This knowledge base serves as the system's knowledge foundation and has completed the integration of multi-source heterogeneous data (including nursing SOPs for elderly care institutions, external geriatric medical guidelines, elderly medical records, etc.), standardized document preprocessing, elderly care scenario-specific text segmentation, multi-dimensional embedded vector generation, and standardized vector database construction. If the elderly care institution subsequently updates its SOPs related to gastrointestinal discomfort care, the system will complete the content synchronization and version management of the knowledge base through an automated update mechanism of change detection, automatic database entry, version rollback, and compliance auditing, without requiring manual maintenance and ensuring the timeliness and authority of the knowledge.

[0052] Step 7: Implementation of Personalized Cognitive Intervention and Narrative Therapy: With authorization from Grandma Li and the nursing home, this system, based on information stored in a private knowledge base regarding Grandma Li's personal life, interests, and experiences, combined with the nursing scenario of abdominal distension, dynamically generates highly personalized narrative content and interactive dialogue using a large language model, while also incorporating lightweight cognitive training tasks: Based on Grandma Li's preference for soft foods, we generated nostalgic narrative content that resonated with her life experiences. During the care process, we interacted with Grandma Li to alleviate her irritability caused by physical discomfort and improve her cooperation in the care. Based on Grandma Li's moderate nursing dependence, simple cognitive training tasks such as number counting and object recognition were designed during the process of assisting her with bedside activities and abdominal massage, so as to combine nursing operations with cognitive intervention. The system will simultaneously update Grandma Li's feedback during this care process (such as her acceptance of dietary adjustments, physical reactions after activities, and emotional state during interactions) to her personal structured profile, continuously enriching the personalized information in the private knowledge base, forming a virtuous cycle of the system becoming more "understanding" of the user the more it is used, and providing more precise support for subsequent care and cognitive intervention.

[0053] This embodiment provides personalized, professional, and practical nursing suggestions for elderly hypertensive patients with gastrointestinal discomfort through the full-process operation of the system. It effectively solves the problems of weak semantic understanding, insufficient integration of professional knowledge, uncontrollable output content, and high knowledge maintenance costs in existing elderly care systems. At the same time, through the practical application of narrative therapy and cognitive intervention, it achieves a triple effect of "professional nursing operation + psychological comfort for the elderly + delay of cognitive decline". It not only ensures the professionalism and safety of elderly care services, but also takes into account the personalized psychological needs of the elderly, and adapts to the requirements of the elderly care industry for refined and humanized service development.

[0054] Furthermore, all technical modules of this invention support flexible replacement and adaptation. For example, the embedded model can be replaced with an industry-tuned version of ERNIE or Text2Vec, the vector database can be any vector engine such as FAISS or Milvus, and the large language model supports cloud API or private deployment. All of the above replacement schemes do not depart from the protection scope of this invention.

[0055] In another embodiment of the present invention, a personalized cognitive intervention and narrative care system based on a generative large language model is also provided, comprising: The encoding module is configured to encode the original question into a semantic vector; The retrieval module is configured to retrieve multiple relevant text fragments from the elderly care knowledge base based on the semantic vector. The prompt word generation module is configured to construct structured enhanced prompt words based on the relevant text fragments; and The nursing suggestion generation module is configured to obtain structured nursing suggestions based on the structured enhanced prompts and through a generative large language model.

[0056] Although various embodiments of the invention have been described above, it should be understood that they are presented by way of example only and not as limitations. It will be apparent to those skilled in the art that various combinations, modifications, and alterations can be made without departing from the spirit and scope of the invention. Therefore, the breadth and scope of the invention disclosed herein should not be limited by the exemplary embodiments disclosed above, but should be defined solely by the appended claims and their equivalents.

Claims

1. A personalized cognitive intervention and narrative nursing method based on a generative large language model, characterized in that, include: Encode the original question into a semantic vector; Based on the semantic vector, a retrieval is performed to retrieve multiple relevant text fragments from the elderly care knowledge base; Structured enhanced prompts are constructed based on the relevant text fragments; Based on the structured enhanced prompts, structured care suggestions are obtained through a generative large language model.

2. The personalized cognitive intervention and narrative nursing method based on a generative large language model according to claim 1, characterized in that, It also includes building and automating the updating of a senior care knowledge base, the steps of which include: Automated import of various data in the elderly care field through multiple channels; Perform standardized preprocessing on all imported data; Segment the preprocessed text; Each processed text segment is input into the BGE-M3 model for encoding, generating a multi-dimensional embedding vector. The generated multi-dimensional embedding vectors are used to construct an elderly care knowledge base according to a standardized field structure. When changes to the content of each data source are detected, the detected changes will be automatically imported into the database according to the above process; And record the version number for each update of the elderly care knowledge base.

3. The personalized cognitive intervention and narrative nursing method based on a generative large language model according to claim 1, characterized in that, Encoding the original question into a semantic vector includes: Input the received original question into the BGE-M3 model; Expand the original question with industry-specific terminology; Fine-tuning was performed using standardized corpora in the field of elderly care. The original question was encoded to include healthcare-related semantics, inferences about the elderly's implicit conditions, and expressions of question intent. Semantic vectors associated with nursing processes.

4. The personalized cognitive intervention and narrative nursing method based on a generative large language model according to claim 1, characterized in that, The process of retrieving multiple relevant text fragments from the elderly care knowledge base based on the semantic vector includes using vector similarity retrieval + BM25 keyword retrieval.

5. The personalized cognitive intervention and narrative nursing method based on a generative large language model according to claim 1, characterized in that, The construction of structured enhanced prompt words based on the relevant text fragments includes: Evidence set E' is selected from the relevant text fragments E recalled; Extract the field P' that is directly related to the current problem from the user's personalized profile P; If the current problem involves risky operations, automatically associate the corresponding risk control constraints from the constraint set C; By combining the original question with the user's personalized profile, a set of constraints C' specifically designed for this question is formed; Based on the original question Q, the filtered evidence set E', the extracted personalized information P', and the exclusive constraint set C', structured enhanced prompt words are formed.

6. The personalized cognitive intervention and narrative nursing method based on a generative large language model according to claim 5, characterized in that, The following formula is used to filter the relevant text to obtain a set of evidence: ； in The i-th text segment is given a comprehensive weighted score, with a higher score indicating a stronger match with the question and greater authority. α is the weighting coefficient of vector cosine similarity, used to adjust the proportion of semantic matching in the score; The encoding vector of the user's original question Q and the i-th text fragment The vector cosine similarity between the encoded vectors represents the degree of semantic matching between them; β is a weighting coefficient for the authority of the text source, used to adjust the proportion of authority in the score; For the i-th text segment The source authority coefficient is set according to the authority level of knowledge in the elderly care industry, with priority being SOP; Based on the comprehensive weighted scores, the evidence fragments ranked from highest to lowest are selected, and the top K fragments are used to form the core evidence set E' after screening.

7. The personalized cognitive intervention and narrative nursing method based on a generative large language model according to claim 2, characterized in that, The step of segmenting the preprocessed text includes: Combine multiple related nursing steps into a single text chunk to maintain the semantic integrity of the nursing process; Each text block is segmented to correspond to the elderly care process to facilitate subsequent retrieval. Add semantic tags specific to the elderly care industry to each text segment.

8. The personalized cognitive intervention and narrative nursing method based on a generative large language model according to claim 2, characterized in that, The multi-dimensional embedding vector includes: Text vectors are the basic semantic vectors of the text segments themselves; Paragraph tag vectors, generated based on semantic tags for text segments; Process path vector, a vector generated based on the corresponding elderly care process; Personalized vectors are vectors generated based on the individual characteristics of the elderly to adapt to the retrieval needs of personalized care.

9. The personalized cognitive intervention and narrative nursing method based on a generative large language model according to claim 1, characterized in that, Also includes: With the user's authorization, retrieve the user's personal information from the elderly care knowledge base; Using a large language model, personalized information is dynamically generated based on the retrieved information. Based on the personalized information, we ask and answer questions with users to provide structured care recommendations.

10. A personalized cognitive intervention and narrative nursing system based on a generative large language model, characterized in that, include: The encoding module is configured to encode the original question into a semantic vector; The retrieval module is configured to retrieve multiple relevant text fragments from the elderly care knowledge base based on the semantic vector. The prompt word generation module is configured to construct structured enhanced prompt words based on the relevant text fragments; as well as The nursing suggestion generation module is configured to obtain structured nursing suggestions based on the structured enhanced prompts and through a generative large language model.

Images