An AI agent-based diabetes intervention scheme generation method

Abstract

The present application relates to a kind of diabetes intervention scheme generation method based on AI intelligent agent, by the operation behavior event generated in the interaction of health management service personnel in question and answer is collected and standardized encapsulation, structured intervention dataset is constructed, and then similar behavior is clustered and analyzed, feature is extracted to generate semantic fingerprint, realize the automatic early warning and hierarchical migration of knowledge applicable boundary.When triggering knowledge failure or high-frequency correction signal, the system automatically initiates knowledge update process, fuses expert consensus, forms double-version coexistence knowledge structure, and matches version by user agency characteristics loading, generates personalized intervention suggestion with national guideline compliance and local practice differentiation.The present application helps to dynamically maintain the applicability of knowledge, improves the intelligent adaptation and continuous improvement capability of diabetes intervention knowledge system in regional practice.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence and clinical knowledge engineering technology, and in particular to a method for generating diabetes intervention plans based on AI agents. Background Technology

[0002] Currently, the dynamic maintenance of diabetes-specific knowledge bases primarily employs two technical approaches: one involves manually reviewing medical literature and authoritative guidelines periodically, with knowledge engineers manually entering and revising knowledge entries, which are then reviewed by experts to ensure their authority; the other utilizes automated web scraping and NLP semantic parsing techniques to extract medical entities and intervention recommendations from the internet or databases, constructing a clinical knowledge graph. The latter has some practical value in scenarios such as initial case screening and rational drug use recommendations, but its content updates depend on the guideline release cycle and the frequency of external data sources.

[0003] The main limitations of existing methods are: manual maintenance is inefficient and susceptible to subjective bias and time lag, failing to keep pace with rapid knowledge evolution and local practice needs; automated data collection methods face higher risks in terms of authority verification and scenario adaptation, easily introducing low-credibility content; existing systems struggle to capture genuine feedback from health management service personnel regarding recommendations during actual interventions, and cannot dynamically reflect the applicability drift of knowledge items in specific regions based on practical data. Key factors such as changes in local medical insurance catalogs and the access of new drugs may cause the "best practices" of authoritative guidelines to become ineffective in frontline clinical practice in the short term, while the knowledge base remains outdated in the long term.

[0004] Current technology lacks a dynamic calibration mechanism for knowledge bases that is driven by real-world intervention behaviors and seamlessly integrated into medical workflows. Existing methods such as manual review, multi-agent discussions, and RAG retrieval enhancement are insufficient to sensitively capture and hierarchically control the microscopic signals of knowledge entries becoming invalid in the local environment in real time. Therefore, there is an urgent need for a solution that centers on the daily intervention decisions of health management service personnel and drives automatic knowledge stratification and dynamic calibration through implicit operational behaviors. This solution would overcome the limitations of traditional methods, such as the lag in authority and the distortion of content, and provide stable, traceable, and dynamically adaptive knowledge support for intelligent and personalized interventions. Summary of the Invention

[0005] This application provides a method for generating diabetes intervention programs based on AI agents, aiming to solve one of the problems or issues of the prior art mentioned in the background section.

[0006] This application provides a method for generating diabetes intervention plans based on AI agents, specifically including: S1: Obtain the action events generated by health management service personnel in response to system-pushed suggestions during diabetes intervention Q&A interactions, and encapsulate them into a raw action event stream; S2: Based on the original operational behavior event flow, associate the unique identifier of the knowledge entry, the original source metadata, and the clinical context parameters to form a structured clinical intervention dataset; S3: Perform cluster analysis on multiple operational behavior events of the same knowledge item in the structured clinical intervention dataset within a preset period and with clinical context parameter similarity greater than a threshold, extract operation type distribution entropy, mean text edit distance and response sensitivity features of clinical context parameters, and generate a semantic fingerprint of clinical intervention behavior. S4: Determine whether the semantic fingerprint of the clinical intervention behavior meets the preset knowledge drift triggering conditions. If so, trigger a knowledge application boundary warning and migrate the corresponding knowledge item from the general knowledge layer to the regional practice adaptation layer to generate a solidified knowledge version. S5: Train a long short-term memory network model using the semantic fingerprint of historical clinical intervention behavior, predict the applicability probability decay curve of the solidified knowledge version in newly emerging clinical situations, and generate a knowledge update requirement package when the predicted decay rate exceeds a preset failure threshold. S6: Obtain the expert's confirmation result for the knowledge update requirement package, update the knowledge graph according to the confirmation result, and form a dual-version coexistence knowledge structure; S7: Obtain user identification information to determine the corresponding organization code, select the corresponding knowledge version from the dual-version coexisting knowledge structure based on the organization code as the reasoning basis, and generate personalized intervention suggestions.

[0007] This application provides a method for generating diabetes intervention plans based on AI agents, which has the following beneficial effects: (1) This application uses the concept of “behavior as evidence” to combine the operational behavior events of service personnel in accepting, fine-tuning, rejecting and marking for review of system-pushed suggestions with clinical context parameters and knowledge sources to construct a “behavior-knowledge-context” three-element event flow, realize the paradigm shift of knowledge base from passive response to proactive perception, maintain consistency with authoritative guidelines and adapt to the needs of grassroots implementation without relying on external crawling or multi-agent debate.

[0008] (2) By performing semantic fingerprint modeling on high-frequency behavioral data, a multi-dimensional "practice adaptation fingerprint" is constructed, which includes operation distribution entropy, mean edit distance, co-occurrence frequency of rejection reasons, and parameter sensitivity gradient. This automatically identifies the knowledge application boundary and triggers a hierarchical transfer mechanism, transferring invalid knowledge from the general layer to the regional adaptation layer and freezing cross-organizational call permissions. Combined with the LSTM model to predict knowledge decay trends, structured knowledge update requirement packages are generated in advance, reducing manual maintenance costs and improving system operation and maintenance efficiency and stability.

[0009] (3) The above mechanism constructs a closed-loop, self-consistent knowledge evolution system, achieving a dynamic balance between national guidelines and local practices. The original authoritative version and the practice consensus version coexist. The front-end engine intelligently loads the corresponding knowledge weight configuration according to the user's institution code, ensuring that the recommendation results conform to medical standards and accurately match the local diagnosis and treatment reality, which is suitable for chronic disease management scenarios with uneven distribution of medical resources. Attached Figure Description

[0010] Figure 1 This is the main flowchart of a method for generating diabetes intervention plans based on AI agents; Figure 2 This is a sub-flowchart of a method for generating diabetes intervention plans based on AI agents; Figure 3 This is another sub-flowchart of a method for generating diabetes intervention plans based on AI agents. Detailed Implementation

[0011] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0012] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0013] like Figure 1 As shown, this application provides a method for generating diabetes intervention plans based on AI agents, specifically including: S1: Obtain the action events generated by health management service personnel in response to system-pushed suggestions during diabetes intervention Q&A interactions, and encapsulate them into a raw action event stream; S2: Based on the original operational behavior event flow, associate the unique identifier of the knowledge entry, the original source metadata, and the clinical context parameters to form a structured clinical intervention dataset; S3: Perform cluster analysis on multiple operational behavior events of the same knowledge item in the structured clinical intervention dataset within a preset period and with clinical context parameter similarity greater than a threshold, extract operation type distribution entropy, mean text edit distance and response sensitivity features of clinical context parameters, and generate a semantic fingerprint of clinical intervention behavior. S4: Determine whether the semantic fingerprint of the clinical intervention behavior meets the preset knowledge drift triggering conditions. If so, trigger a knowledge application boundary warning and migrate the corresponding knowledge item from the general knowledge layer to the regional practice adaptation layer to generate a solidified knowledge version. S5: Train a long short-term memory network model using the semantic fingerprint of historical clinical intervention behavior, predict the applicability probability decay curve of the solidified knowledge version in newly emerging clinical situations, and generate a knowledge update requirement package when the predicted decay rate exceeds a preset failure threshold. S6: Obtain the expert's confirmation result for the knowledge update requirement package, update the knowledge graph according to the confirmation result, and form a dual-version coexistence knowledge structure; S7: Obtain user identification information to determine the corresponding organization code, select the corresponding knowledge version from the dual-version coexisting knowledge structure based on the organization code as the reasoning basis, and generate personalized intervention suggestions.

[0014] It should be noted that this application provides a method for generating diabetes intervention plans based on AI intelligent agents, which is a 3X and (8+N) reversal treatment plan for diabetes. Specifically, 3 refers to three aspects: first, the patient has a strong desire to reverse the condition; second, a method combining traditional Chinese and Western medicine is used to screen for the type and cause of the disease; and third, a precise and personalized intervention plan is developed using a combination of drug and non-drug methods. 8+N refers to eight professionals, including traditional Chinese medicine practitioners, doctors, nutritionists, health managers, exercise therapists, physiotherapists, psychological counselors, and pharmacists, jointly managing a patient. This involves using eight or more methods, including traditional Chinese and Western medicine, nutrition, exercise, rehabilitation therapy, supervision, psychological guidance, dynamic regulation, and new technologies, to manage and repair the patient's eight major systems, including endocrine, nervous, digestive, urinary, musculoskeletal, respiratory, immune, and circulatory systems, and combining these with new technologies, including the Internet, IoT wearable devices, and other information technology monitoring techniques, to improve and achieve the desired blood sugar regulation capacity.

[0015] Step S1: Obtain the action events generated by health management service personnel in response to system-pushed suggestions during the diabetes intervention Q&A interaction, and encapsulate these action events into a raw action event stream. These action events include a direct adoption timestamp, an edit trajectory vector of the adopted suggestion after fine-tuning, a semantic difference backbone of the rejection and rewrite, and a regional quality control review marker. The system-pushed suggestions refer to textual suggestions for diabetes intervention management automatically generated by the system based on the patient's clinical context parameters in the current dialogue round, including fasting blood glucose, 2-hour postprandial blood glucose, glycated hemoglobin, estimated glomerular filtration rate, serum creatinine, urinary microalbumin, systolic blood pressure, diastolic blood pressure, body mass index, disease duration, complication type, and current medication regimen, as well as related knowledge item version information. This textual suggestion content is output after inference calculation by calling clinical guidelines stored in the knowledge graph nodes and combining them with the patient's individualized physiological parameters. It is then presented in structured text form on the health management service personnel's interactive interface for them to directly adopt, fine-tune and adopt, or reject and rewrite.

[0016] Step S1 specifically includes: S1.1: Perform time series alignment processing on the mouse click trajectory, cursor dwell time and confirmation button trigger signal generated by the interaction between health management service personnel and the system interface to extract the direct adoption operation feature vector containing precise millisecond-level timestamps, thereby obtaining direct adoption timestamp data representing unconditional trust behavior.

[0017] The system acquires the raw input event stream generated by health management service personnel on the system's interactive interface. This event stream includes mouse movement coordinate sequences, click event types, cursor focus switching records, and confirmation button trigger signals. The raw input event stream is then timestamped to normalize the heterogeneous time bases reported by different devices, converting them into a unified system-wide millisecond-level timestamp to eliminate time discrepancies caused by network latency or local clock deviations. Based on the normalized time series, the system extracts the pre-operation window data for the moment the confirmation button is triggered. This window is defined as the continuous time period from the last stable cursor dwell to the confirmation action. The system calculates the dispersion of the mouse trajectory within this window and the proportion of cursor dwell time in key information areas to construct a behavior confidence assessment model. When the mouse trajectory is smooth and the cursor dwell time in the core area of ​​the suggested text exceeds a preset threshold, the confirmation is determined to be a direct adoption behavior after thorough reading. The above processing method transforms the original interaction signal from the previous step into a direct adoption operation feature vector containing precise millisecond-level timestamps, thereby obtaining direct adoption timestamp data that characterizes unconditional trust behavior, providing highly reliable positive sample labels for the subsequent construction of a ternary event flow of behavior-knowledge-clinical context parameters.

[0018] S1.2: Based on the interactive sessions that were not triggered by the direct adoption timestamp data, differential encoding is performed on the keystroke sequence, deletion frequency and text replacement content of the service personnel in the suggested text editing area to calculate the editing distance matrix between the original suggested text and the modified text, thereby obtaining the edit trajectory vector of the fine-tuned adoption that represents the local correction intention.

[0019] Acquire interactive session data that did not trigger a direct adoption timestamp, and use the keystroke sequence, deletion frequency, and text replacement content of the service personnel in the suggested text editing area as raw input signals.

[0020] The keyboard keystroke sequence is reassembled in time, and a character-level operation flow is constructed based on the time interval between key press and release, identifying three basic editing atomic actions: insertion, deletion, and cursor movement.

[0021] Based on the character-level operation flow, a dynamic programming algorithm is used to calculate the minimum editing path between the original suggested text and the modified text, generating an edit distance matrix that includes replacement, insertion, and deletion operation types.

[0022] The Levenstein distance formula is used to quantify the degree of text difference, where D(i,j) represents the minimum editing cost between the first i characters of the source text and the first j characters of the target text: in, The preceding text of the source text (original suggestion text) The characters before the target text (modified text) The minimum edit distance between characters, characters and These represent the first and second parts of the source text, respectively. The first character and the first character representing the target text One character, An empty character. Character Convert to character The cost function. The rules for the cost function are as follows: if a = b (characters are the same or empty), then... If a! = b, then =1. The global minimum edit distance value is obtained by recursively filling the matrix.

[0023] The edit distance matrix is ​​normalized and divided by the original suggested text length to obtain the relative edit density, thus eliminating the influence of text length on the evaluation of the correction magnitude.

[0024] High-frequency replacement phrases and local structural adjustment patterns are extracted from the editing trajectory, and combined with the cursor hovering hotspot areas to construct a multi-dimensional feature vector representing the intention of local correction.

[0025] By using differential coding, the discrete operation signal from the previous step is transformed into a continuous feature space vector, enabling precise quantization of the adopted edit trajectory vector after fine-tuning, and providing fine-grained behavioral basis for subsequent semantic fingerprint modeling.

[0026] S1.3: Using semantic backbone extraction processing in natural language processing, a dependency syntax tree comparison analysis is performed on the new text block generated when the service personnel perform the whole rejection and rewrite operation and the original system suggestion text to identify and remove non-core modifying components, thereby obtaining the semantic difference backbone of rejection and rewrite that represents the fundamental cognitive conflict.

[0027] The new text block generated when service personnel perform a full rejection and rewrite operation on the interactive interface, along with the original system suggestion text, are used as input data objects for semantic difference analysis.

[0028] The original suggested text and the new text block are processed by Chinese word segmentation and part-of-speech tagging respectively. Dependency trees are constructed using Stanford CoreNLP or similar tools to identify the subject-verb-object core structure and the dependency relationships of modifiers in the sentence.

[0029] Based on dependency syntax tree nodes, verb predicates and their associated subjects and objects are extracted to form a semantic backbone set. Non-core modifier nodes such as adjectives and adverbs are stripped to form a denoised core semantic graph.

[0030] The edit distance between the semantic backbone of the original suggested text and the semantic backbone of the new text block is calculated to quantify the degree of structural difference between the two in terms of core medical entities and intervention actions.

[0031] The cosine similarity metric is used to calculate the cosine of the angle between two semantic core vectors in multidimensional space. The formula is as follows: in This is the original proposed backbone vector. This is the new text backbone vector. It is the cosine of the angle between the original suggested backbone vector and the new text backbone vector in multidimensional space.

[0032] When the cosine similarity is below a preset threshold and the edit distance exceeds a set upper limit, a fundamental cognitive conflict is determined, and the semantic node with the greatest difference is extracted as the conflict feature point.

[0033] The conflict feature points and their corresponding clinical context parameters are encoded into structured data to generate a semantic difference backbone representing the rejection and rewriting of fundamental cognitive conflict.

[0034] By using the aforementioned dependency syntax comparison and vector space metric processing methods, unstructured text rewriting behavior is transformed into quantifiable semantic difference backbone data, enabling accurate capture of the deep professional judgment intentions of service personnel and providing high-confidence negative sample evidence for subsequent knowledge credibility verification.

[0035] S1.4: Perform context snapshot processing on the specific quality control mark control operation triggered by the service personnel, and simultaneously lock the patient's physiological parameter status and the associated knowledge item version information in the current dialogue round to generate abnormal behavior identifiers with regional quality control review marks, thereby obtaining regional quality control review mark data that characterizes potential compliance risks.

[0036] Monitor the trigger events of the regional quality control review control in the front-end interactive interface, and capture the raw control signal containing the event type, trigger source ID, and precise millisecond-level timestamp.

[0037] Based on the original control signal, the session state manager is invoked to lock the unique session identifier of the current dialogue round, and a strong association index between the operation behavior event and the specific doctor-patient interaction context is established.

[0038] Read the patient's electronic medical record data snapshot bound to the current session and extract twelve clinical physiological parameters, including fasting blood glucose, 2-hour postprandial blood glucose, glycated hemoglobin, estimated glomerular filtration rate, serum creatinine, urinary microalbumin, systolic blood pressure, diastolic blood pressure, body mass index, disease duration, complication type, and current medication regimen.

[0039] The twelve-dimensional clinical physiological parameters were standardized, cleaned, and imputed for missing values ​​to construct a structured clinical context parameter vector representing the patient's immediate health status.

[0040] The retrieval system pushes suggestions based on knowledge graph nodes, extracts the unique identifier, version number, source guide name, and last update time of the knowledge entry, and forms a knowledge traceability metadata set.

[0041] By spatiotemporally aligning the structured clinical context parameter vector with the knowledge traceability metadata set, a complete context snapshot containing patient status and knowledge version information is generated.

[0042] A hash algorithm is performed on the context snapshot to generate a unique snapshot fingerprint, which is used for subsequent data integrity verification and anti-tampering verification.

[0043] The regional quality control review markers, session identifiers, structured clinical context parameter vectors, knowledge traceability metadata sets, and snapshot fingerprints are encapsulated into structured data objects with abnormal behavior identifiers.

[0044] Through the above processing method, the discrete operation signals of the previous step are transformed into regional quality control review marker data containing rich clinical context and knowledge version information, realizing comprehensive digital retention of potential compliance risk scenarios and providing high-fidelity evidence chain support for the subsequent dynamic calibration of the knowledge base.

[0045] S1.5: Based on the directly adopted timestamp data, the edited trajectory vector of the adopted data after fine-tuning, the semantic difference backbone of the rejected and rewritten data, and the regional quality control review mark data, the standardized encapsulation protocol of multi-source heterogeneous operation signals is executed, and the four types of operation behavior events are arranged in the order of occurrence and a unified event header is added, thereby obtaining the original operation behavior event stream for downstream analysis.

[0046] Step S2: Based on the original operational behavior event flow, associate the unique identifier of the knowledge entry, the original source metadata, and the clinical context parameters to construct a ternary event flow with spatiotemporal attributes of behavior-knowledge-clinical context parameters, forming a structured clinical intervention dataset. The clinical context parameters include glycated hemoglobin trends and renal function staging; specifically: S2.1: Obtain the operation signal sequence in the original operation behavior event stream, and use time window-based sliding alignment processing to perform time-series standardization processing on the direct adoption timestamp, edit trajectory vector and semantic difference backbone to generate a time-series behavior feature vector with a unified time base.

[0047] The discrete operation signal sequence from the original operation behavior event stream is obtained, including the direct adoption timestamp, the edit trajectory vector after fine-tuning adoption, and the semantic difference backbone of rejection and rewriting. The operation signal sequence is subjected to temporal alignment processing based on a sliding time window, with a fixed-length time window to cover a single complete interactive session cycle. Within the time window, the relative time offset of each operation behavior event relative to the reference zero point is calculated, using the system suggestion push time as the reference zero point. Linear interpolation is used to resample the non-uniformly sampled mouse click trajectory and keyboard keystroke sequence, unifying the data sampling frequency to a millisecond-level precision standard. Normalization processing is performed on the resampled edit trajectory vector to eliminate feature scale bias caused by text length differences, generating standardized local correction intent feature components. Dependency tree deep encoding is performed on the semantic difference backbone to extract the topological distance between core predicates and argument structures, mapping it to a numerical index representing the intensity of cognitive conflict. The direct adoption timestamp, standardized edit trajectory vector, and cognitive conflict intensity index are concatenated in chronological order to construct a multi-dimensional temporally sequenced behavior feature vector with a unified time reference. By using the above-mentioned sliding alignment and multi-source signal standardization processing method based on time windows, the heterogeneous original operation behavior event stream obtained in the previous step is transformed into a comparable temporal behavior feature vector, realizing the quantitative representation of different interactive behaviors in a unified spatiotemporal dimension, and providing a high-precision temporal input basis for subsequent associated knowledge metadata.

[0048] S2.2: Based on the time-series behavioral feature vector, the diabetes knowledge base index interface is called to extract the unique identifier of the associated knowledge entry and the original source metadata. Entity linking technology is used to map discrete operation signals to specific knowledge graph nodes to generate behavioral anchoring data with knowledge traceability tags.

[0049] The system receives the time-series behavioral feature vector output from S2.1 and extracts the operation signal timestamp and session ID as retrieval keys. It then calls the diabetes knowledge base indexing interface to generate a unique identifier for the knowledge entry referenced in the suggestion based on the session ID. The system reads the original source information corresponding to this identifier from the knowledge base metadata table, including the guideline version number, issuing organization, and effective date. Entity linking technology is used to perform semantic parsing on the text content in the operation signal, identifying the medical entities involved, such as drug names or physiological indicators. The cosine similarity between the identified medical entities and the node vectors in the knowledge graph is calculated to determine the best-matching knowledge graph node ID. The unique identifier of the knowledge entry, the original source metadata, and the matching knowledge graph node ID are encapsulated into a knowledge traceability tag. This knowledge traceability tag is then merged with the time-series behavioral feature vector at the field level to generate behavioral anchoring data with the knowledge traceability tag. Through this processing, the result of the previous step is transformed into behavioral anchoring data with a clear knowledge attribution, achieving a precise association between behavioral data and the knowledge ontology, providing a structured foundation for the subsequent construction of a ternary event flow.

[0050] S2.3: Obtain patient electronic medical record data in the current interactive session, use a multimodal feature extraction engine to analyze clinical context parameters such as glycated hemoglobin trend curve and renal function staging indicators, and transform unstructured medical record text into a structured clinical context parameter feature matrix through a vector space model.

[0051] The system acquires the raw data stream of the patient's electronic medical record associated with the current interactive session. This data stream contains structured test indicators and unstructured medical record text. A multimodal feature extraction engine is invoked to perform named entity recognition on the unstructured text, locating key clinical terms such as glycated hemoglobin and estimated glomerular filtration rate, along with their corresponding values ​​and timestamps. A clinical context parameter vector space is constructed, defining dimensions including the three most recent glycated hemoglobin values, recent renal function stage, comorbidity markers, and medication history. Linear interpolation is used to process missing time-series data, ensuring the continuity and integrity of parameters across dimensions over time. Z-score standardization is applied to the extracted numerical parameters to eliminate the impact of dimensional differences on subsequent vector distance calculations. A pre-trained medical BERT model maps the unstructured medical record text into a high-dimensional semantic vector, capturing implicit clinical context features. The standardized numerical parameter vector is concatenated and fused with the text semantic vector to generate a structured clinical context parameter feature matrix containing explicit indicators and implicit semantics. Through the above processing method, the behavioral anchoring data of the previous step is transformed into a structured feature matrix with multi-dimensional clinical context representation, realizing the transformation of unstructured medical record information into computable quantitative indicators, and providing accurate context input for constructing a ternary event flow of behavior-knowledge-clinical context parameters.

[0052] S2.4: Based on the behavior anchoring data with knowledge traceability tags and the structured clinical context parameter feature matrix, perform multidimensional tensor splicing operations to construct a three-dimensional data cube containing behavior dimension, knowledge dimension and context dimension, so as to generate a behavior-knowledge-clinical context parameter ternary event flow with spatiotemporal coupling characteristics.

[0053] The system receives behavioral anchoring data with knowledge source tags and a structured clinical context parameter feature matrix as input sources for constructing a three-dimensional data cube. Dimensional parsing is performed on the behavioral anchoring data to extract operation type codes, edit trajectory vectors, and semantic difference backbones, mapping them to behavioral dimension feature tensors. These behavioral dimension feature tensors are normalized to eliminate differences in the signal dimensions of different operations, forming a standardized sequence of behavioral feature vectors. Sparse imputation is performed on the structured clinical context parameter feature matrix; for missing glycated hemoglobin trends or renal function staging indicators, a nearest-neighbor mean interpolation method is used to generate complete clinical parameter vectors. The complete clinical parameter vectors are projected onto a high-dimensional latent space through a linear transformation to generate a contextual dimension embedding representation. A knowledge graph embedding model is invoked to transform the unique identifiers of knowledge entries and original source metadata into fixed-dimensional knowledge dimension vectors. Based on a timestamp alignment mechanism, the behavioral feature vector sequence, contextual dimension embedding representation, and knowledge dimension vectors are precisely matched along the time axis. A multidimensional tensor concatenation operation is performed, concatenating the matched three-dimensional vectors along the feature axis to construct the initial three-dimensional data block. Spatiotemporal location encoding is introduced, encoding the time segment of the interaction and the institution ID of the service personnel into location embedding vectors. These location embedding vectors are then superimposed onto the initial 3D data block to enhance the spatiotemporal coupling characteristics of the data cube. Tensor reshaping is performed on the superimposed data cube, converting it into a structured triplet format with rows and columns as event indices and columns as multidimensional features. Through the above tensor concatenation and spatiotemporal encoding processing, the discrete behavioral, knowledge, and contextual data from the previous step are transformed into a ternary event stream of behavioral-knowledge-clinical context parameters with spatiotemporal coupling characteristics. This achieves deep fusion and structured representation of multi-source heterogeneous data, providing a unified high-dimensional data foundation for subsequent clustering analysis.

[0054] S2.5: Perform integrity verification and temporal sorting on the behavior-knowledge-clinical context parameter ternary event stream, and use a serialization encapsulation protocol to aggregate the scattered triples into a continuous structured clinical intervention dataset, forming a standard input data stream that supports cluster analysis and fingerprint modeling.

[0055] like Figure 2As shown, step S3: Cluster analysis is performed on multiple operational behavior events of the same knowledge item in the structured clinical intervention dataset within a preset period, where the similarity of clinical context parameters is greater than a threshold. Operation type distribution entropy, mean text edit distance, and response sensitivity features of clinical context parameters are extracted to generate a semantic fingerprint of clinical intervention behavior. This semantic fingerprint is used to characterize regional practice adaptability. Specifically, it includes: S3.1: Based on the structured clinical intervention dataset, the dynamic time warping algorithm is used to perform vector space mapping on the clinical context parameters, which include glycated hemoglobin trends and renal function staging, in order to calculate the Euclidean distance between any two operational behavior events and screen out a set of candidate operational behavior events with clinical context parameter similarity greater than a preset threshold, forming a high-confidence subset of similar situational behaviors.

[0056] Extract all operational behavior events of the same knowledge item within a preset period from the structured clinical intervention dataset, and obtain the clinical context parameter vector associated with each event, including the glycated hemoglobin trend slope, renal function stage code, and comorbidity markers.

[0057] The clinical context parameter vector is preprocessed using the Dynamic Time Warping (DTW) algorithm, which performs linear interpolation to align parameters with inconsistent time series dimensions, thereby eliminating temporal misalignment noise caused by differences in detection frequency.

[0058] A local constraint window is constructed to limit the deviation range of the regular path. The optimal matching distance of the clinical context parameter sequence between any two operational events is calculated, and a DTW distance matrix reflecting the similarity of physiological state evolution is generated.

[0059] Based on the DTW distance matrix, the geometric distance in the high-dimensional feature space is calculated using the Euclidean distance formula, which is expressed as: in and Representing two events in the 1st century Normalized values ​​of clinical parameters, denoted as the geometric distance in the high-dimensional feature space.

[0060] By setting a similarity threshold parameter, a set of candidate operational behavior events with both DTW distance and Euclidean distance below the threshold is selected, and outlier samples are removed to eliminate interference from atypical cases.

[0061] The selected candidate operational behavior events are aggregated into a high-confidence subset of similar situational behaviors to ensure that the physiological states of patients within the subset are highly homogeneous.

[0062] By employing a dual verification mechanism of dynamic time warping and Euclidean distance, discrete operational events are transformed into high-confidence subsets with strict clinical context consistency, enabling effective control of contextual variables in subsequent fingerprint extraction and significantly improving the accuracy of regional practice suitability analysis.

[0063] S3.2: For the high-confidence similar situational behavior subset, the Shannon information entropy calculation formula is used to perform probability distribution modeling on the frequency of occurrence of direct adoption, fine-tuned adoption, rejection and rewriting, and marking as requiring regional quality control review, so as to quantify the dispersion of different intervention strategies and generate operation type distribution entropy, forming a first-dimensional feature vector reflecting the consistency of group decision-making.

[0064] The system receives a high-confidence subset of similar contextual behaviors selected in S3.1. This subset contains multiple sets of operation records for the same knowledge item under similar clinical context parameters. The frequency of direct adoption, adoption after fine-tuning, rejection and rewriting, and marking as requiring regional quality control review is statistically analyzed within the subset. The total sample size of the operational behavior events within the subset is calculated, i.e., the sum of the frequencies of the four types of operations, which is used as the basic denominator to construct the probability distribution. Based on the frequencies and the total sample size, the probability P of each operation type in the subset is calculated. i , where i represents the operation type index, with a value ranging from 1 to 4, and the probability value is obtained by dividing the frequency by the total number of samples.

[0065] The operation type distribution entropy is calculated using Shannon's information entropy formula, as follows: in, For the first i The probability of class operations, when When the value is 0, the contribution value of this item is defined as 0 to avoid the logarithm being meaningless. The entropy is the distribution entropy of the operation type. Through the above calculation, the dispersion of decisions made by health management service personnel when faced with specific knowledge suggestions is quantified. A higher entropy value indicates greater disagreement among group opinions, while a lower entropy value indicates stronger consensus. The calculated operation type distribution entropy H is normalized and mapped to the [0,1] interval, forming the first-dimensional feature vector component reflecting the consistency of group decision-making. Using Shannon information entropy modeling, the frequency statistics of behavior from the previous step are transformed into a quantitative indicator representing the fluctuation of the credibility of knowledge items in a specific context. This achieves a mathematical expression of implicit clinical consensus and provides core consistency dimension data for subsequent fingerprint generation.

[0066] S3.3: Based on the edit trajectory vector adopted after fine-tuning in the high-confidence similar situation behavior subset, the number of character editing operations between the original suggested text and the modified text is calculated using the Levenstein distance formula, and the number of character editing operations is normalized to obtain the average text editing distance, forming a second-dimensional feature vector representing the localization correction magnitude.

[0067] Obtain the edit trajectory vectors adopted after fine-tuning from the high-confidence subset of similar situational behaviors, and extract the character sequences of the original suggestion text and the modified text as input objects for the Levenstein distance formula.

[0068] Perform dynamic programming matrix initialization on the original suggested text and the modified text to construct an edit distance matrix of dimension (m+1)×(n+1), where m and n are the character lengths of the two texts, respectively.

[0069] Iterate through each cell of the edit distance matrix and recursively calculate the minimum number of operations required to transform a prefix substring into another prefix substring, based on the unit cost of insertion, deletion, and replacement operations.

[0070] The minimum edit distance value at any position (i,j) in the matrix is ​​calculated using the Levenstein distance formula. For details of the Levenstein distance formula, please refer to S1.2. The final Levinstein distance value is obtained by tracing back to the bottom right element of the matrix. This value represents the substantial correction made by the service personnel to the system's recommendations in a single fine-tuning operation.

[0071] The Levenstein distances of all fine-tuning events within the subset are calculated using an arithmetic mean to eliminate noise interference caused by differences in individual operating habits.

[0072] The calculated mean text edit distance is divided by the standard character length of the original suggested text, and then normalization is performed to eliminate the dimensional bias caused by long text.

[0073] Through the above processing method, discrete text editing behavior is transformed into a standardized text editing distance mean index, forming a second-dimensional feature vector that represents the magnitude of localization correction, thereby realizing a quantitative assessment of the degree to which knowledge items are corrected in local practice.

[0074] S3.4: Combining the semantic differences between rejection and rewriting in the high-confidence similar situational behavior subset with clinical context parameters, a multiple linear regression model is applied to fit the contribution weights of parameter changes in each dimension to the occurrence of rejection behavior, so as to calculate the sensitivity gradient coefficient of clinical context parameters and form a third-dimensional feature vector that reveals the degree of influence of specific clinical indicators on the applicability of knowledge items.

[0075] Step S3.4 aims to quantify the impact of specific clinical indicators on the applicability of knowledge items by constructing a multiple linear regression model to calculate the sensitivity gradient coefficients of clinical context parameters. It obtains the semantic difference backbones of rejection and rewriting within a high-confidence subset of similar situational behaviors and the corresponding clinical context parameter matrices. Rejection behavior is defined as a binary dependent variable, where performing a rejection operation is marked as 1, and adoption or fine-tuning is marked as 0. The clinical context parameters are standardized and preprocessed to eliminate the influence of dimensional differences on the estimation of regression coefficients, ensuring that parameters of each dimension are compared on the same order of magnitude. A multiple linear regression model is constructed, with standardized clinical parameters as independent variables and rejection behavior labels as dependent variables. The least squares method is used to solve for the regression coefficient vector, which represents the contribution weight of each clinical indicator to knowledge failure. The objective function of the model is defined as minimizing the sum of squared residuals, specifically expressed as follows: in, For loss function, Let be the label for the rejection behavior of the i-th sample. A feature matrix containing clinical context parameters. The vector of regression coefficients to be solved is the sensitivity gradient coefficient of the clinical context parameter.

[0076] Introducing an L2 regularization term to prevent overfitting, the objective function is modified as follows: in, The regularization coefficient is used to control model complexity. The total number of samples. Iteratively updated using gradient descent. The values ​​are calculated until the loss function converges, yielding the final sensitivity gradient coefficient vector. Components with absolute values ​​greater than a preset threshold of 0.5 are extracted from the coefficient vector as key clinical indicators significantly affecting knowledge applicability, forming a third-dimensional feature vector. Through this processing method, the behavioral and contextual data from the previous step are transformed into quantified clinical context parameter sensitivity gradient coefficients, achieving the expected technical effect of revealing the degree of influence of specific clinical indicators on the applicability of knowledge items.

[0077] S3.5: Integrate the first, second, and third feature vectors, construct a multi-dimensional feature matrix through feature splicing and standardized encoding, and extract feature components in the main variation directions using principal component analysis to finally generate a semantic fingerprint of clinical intervention behavior that uniquely identifies the adaptation status of the knowledge item in a specific regional institution.

[0078] The system receives the first, second, and third eigenvectors and performs vector concatenation to construct an initial multidimensional feature matrix. Z-score standardization is applied to each dimension of the initial multidimensional feature matrix to eliminate the impact of dimensional differences on subsequent principal component analysis, generating a standardized feature matrix. The covariance matrix of the standardized feature matrix is ​​calculated to quantify the linear correlation between different behavioral feature dimensions, forming a symmetric positive definite covariance matrix. Eigenvalue decomposition is performed on the covariance matrix to solve for the eigenvalues ​​and their corresponding eigenvectors, obtaining the principal component axes representing the direction of data variation. Based on a preset cumulative variance contribution rate threshold, the eigenvectors corresponding to the top K largest eigenvalues ​​are selected to construct a dimension-reduced projection matrix. Matrix multiplication is performed between the standardized feature matrix and the dimension-reduced projection matrix to map to a low-dimensional feature space, generating a semantic fingerprint of clinical intervention behavior.

[0079] Principal component analysis is used to reduce the dimensionality of multi-source heterogeneous behavioral features and fuse them, transforming high-dimensional sparse behavioral data into low-dimensional dense semantic fingerprints. This enables the unique identification and efficient representation of the adaptation status of knowledge items in specific regional institutions.

[0080] For example, the 30-dimensional original feature vector, which includes the operation type distribution entropy, the mean text edit distance, and the sensitivity gradient of the clinical context parameter, is Z-score standardized, with a mean of 0 and a standard deviation of 1. A 30x30 covariance matrix is ​​calculated, and eigenvalue decomposition is performed. The eigenvectors corresponding to the top 5 largest eigenvalues ​​are selected, resulting in a cumulative variance contribution rate of 92%. The standardized feature vectors are projected into a 5-dimensional space to generate a 5-dimensional semantic fingerprint of clinical intervention behavior. This fingerprint significantly reduces data storage overhead while preserving key behavioral pattern information, improving the computational efficiency and accuracy of subsequent drift detection.

[0081] like Figure 3 As shown, step S4: Determine whether the semantic fingerprint of the clinical intervention behavior meets the preset knowledge drift triggering conditions. If so, trigger a knowledge application boundary warning and migrate the corresponding knowledge item from the general knowledge layer to the regional practice adaptation layer to generate a fixed knowledge version. The preset knowledge drift triggering conditions include: the fine-tuning adoption rate of the semantic fingerprint of the clinical intervention behavior exceeds a preset high-frequency correction threshold within two consecutive cycles, and the density of occurrence of terms related to the local drug catalog not covering specific entries in the co-occurrence frequency of the rejection reason keywords suddenly increases.

[0082] The general knowledge layer refers to a knowledge graph layer that stores initial knowledge entries built based on authoritative clinical guidelines and general medical knowledge. Knowledge entries in this layer have globally open access permissions and can be invoked by question-answering reasoning engines across institutions and regions. The regional practice adaptation layer refers to a knowledge graph layer that stores fixed versions of knowledge migrated after localization verification or drift determination. Knowledge entries in this layer have clear regional affiliation attribute tags and are preferentially invoked only by question-answering engines in specific administrative regions or medical institution clusters.

[0083] Step S4 specifically includes: S4.1: Obtain the semantic fingerprint of clinical intervention behavior within two consecutive monitoring periods, and perform time-series sliding window calculation on the operation type distribution entropy and the mean text editing distance contained in the semantic fingerprint of clinical intervention behavior to generate a periodic drift feature vector sequence that characterizes the stability fluctuation of knowledge items.

[0084] Extract semantic fingerprint records of all clinical intervention behaviors generated within the two most recent preset monitoring periods (such as T-1 period and T period) from the time-series index of the structured clinical intervention dataset, and construct an original fingerprint time-series matrix containing timestamps, operation type distribution entropy and mean text edit distance.

[0085] For the original fingerprint time series, a time series sliding window of fixed length is set, and the window step size is set to a standard monitoring period unit to ensure that the window coverage can completely capture the evolution trajectory of the behavior pattern of knowledge items across periods.

[0086] Within each sliding window, the entropy sequence of operation type distribution is extracted, and moving average filtering is used to eliminate noise fluctuations caused by single accidental operations. The weighted average of the entropy values ​​within the window is calculated to quantify the consistency and stability of the service personnel group's decision-making within that period.

[0087] The system synchronously extracts the mean sequence of text edit distances within the window, uses an exponentially weighted moving average method to assign higher weights to recent behaviors, calculates a dynamic indicator that reflects the changing trend of localization correction magnitude, and captures subtle semantic correction tendency drift.

[0088] The weighted average of the distribution entropy of the processed operation type is concatenated with the dynamic index of text editing distance to form a two-dimensional periodic drift feature vector. This vector fully characterizes the stability and correction intensity of knowledge items within a specific time window.

[0089] Arrange all the periodic drift feature vectors generated by the sliding windows in chronological order to construct a continuous sequence of periodic drift feature vectors, which will serve as the basic data input for subsequent judgments on whether the knowledge application boundary has drifted significantly.

[0090] By using time-series sliding window calculation and weighted smoothing, the discrete clinical intervention behavior semantic fingerprint generated in the previous step is transformed into a periodic drift feature vector sequence that represents the stability fluctuation of knowledge items. This achieves the expected technical effect of quantitatively tracking changes in the timeliness and regional adaptability of knowledge base content.

[0091] S4.2: Based on the periodic drift feature vector sequence, extract the fine-tuning adoption rate features and the co-occurrence frequency of rejection reason keywords, determine whether they meet the preset knowledge drift triggering conditions, and generate a composite drift judgment result containing quantitative over-limit markers and qualitative conflict labels.

[0092] Receive a sequence of periodically drifting feature vectors and extract numerical indicators representing fine-tuning adoption behavior. Construct a fine-tuning adoption rate calculation model by comparing the number of events adopted after fine-tuning within a preset period with the total number of interaction events. Calculate the fine-tuning adoption rate using the following formula: in, To fine-tune the adoption rate, This represents the number of events adopted after fine-tuning. This represents the total number of interaction events. The calculated fine-tuning adoption rate is compared with a preset high-frequency correction threshold; if it exceeds the threshold, a quantitative over-limit status is marked.

[0093] Parallel processing of rejection reason text data is performed, and keywords are extracted from the semantic differences in the rejection and rewrite operations. A local drug catalog is established that does not cover a specific terminology database, and the co-occurrence frequency of this term in the rejection reasons is counted. The occurrence density of a specific term is calculated and compared with the historical baseline density. If the density suddenly increases beyond a preset fluctuation range, a qualitative conflict is marked.

[0094] By combining quantitative exceedance markers and qualitative conflict states, a composite drift determination result is generated through logical AND operations. This result includes the knowledge item ID, drift type identifier, and confidence score.

[0095] Through the above comparison and statistical processing, the drift feature vector from the previous step is transformed into a composite drift judgment result containing both quantitative and qualitative dimensions, thereby achieving accurate identification of the drift of the applicability boundary of the knowledge base content and providing a solid basis for triggering knowledge freezing and migration in the future.

[0096] S4.3: Using the composite drift determination result as a trigger condition, execute a state locking instruction on the target knowledge entry currently in the general knowledge layer, and switch its access permission configuration from global open mode to restricted frozen mode, so as to generate a solidified knowledge version with read-only attributes and prohibited from cross-organizational access.

[0097] The system receives the composite drift determination result output by S4.2 and parses the unique identifier of the target knowledge entry and the state change instruction flag contained therein. Based on the identifier, the system locates the corresponding knowledge graph node object in the general knowledge layer index library and reads the access control list and version metadata information of the current node.

[0098] Construct a state-locked transaction context and initialize the knowledge entry state machine to the "frozen and awaiting migration" intermediate state. Invoke the permission management engine to change the global read permission bit of the target knowledge entry from "allowed" to "prohibited," while retaining the read-only access permission of the local instance to support historical tracing.

[0099] Perform a version snapshot solidification operation, copying the attribute set, associated evidence chain, and semantic fingerprint data of the current knowledge entry to generate a solidified knowledge version with a timestamp. This instance is assigned an immutable attribute identifier to ensure that the content is not tampered with during subsequent manual review.

[0100] Update the knowledge graph edge relationships, disconnect the dynamic link between frozen instances and the general inference engine to prevent them from being incorrectly invoked in cross-institutional question-answering scenarios. The system updates the status field of frozen instances to "restricted frozen mode" and records the status change log for auditing purposes.

[0101] By using the aforementioned state locking and permission isolation methods, the drift determination result from the previous step is transformed into a fixed knowledge version with read-only attributes and prohibited from cross-institutional access. This achieves immediate blocking and protection of knowledge with applicability risks, preventing the spread of erroneous medical advice.

[0102] S4.4: Based on the metadata identifier of the solidified knowledge version, perform a node migration operation to strip the solidified knowledge version from the general knowledge layer and remount it to the regional partition node corresponding to the regional practice adaptation layer, so as to generate a hierarchical storage knowledge entry with regional affiliation attributes.

[0103] The system receives the solidified knowledge version and its associated metadata identifier after processing with the state-locking instruction. The metadata identifier includes a unique knowledge entry ID, the original authoritative version number, and the regional code that triggered the drift determination. The system parses the regional code field in the metadata identifier, calls the pre-stored regional index table, retrieves the target node path of the regional practice adaptation layer that matches the regional code, and determines the destination topology location for the knowledge entry migration.

[0104] Perform graph database transaction operations, read the current attribute set and relation edge information of the corresponding knowledge item in the general knowledge layer, and construct a complete subgraph snapshot containing all clinical intervention rules, drug contraindications and applicable conditions to ensure data consistency and integrity during the migration process.

[0105] Disconnect the direct call connection between this knowledge entry and the global inference engine in the general knowledge layer, and remove its access permission tag in the global question-answering scenario to prevent it from continuing to provide services to users in non-specific regions before it has passed the review of regional experts, thereby blocking the potential spread of low-adaptability knowledge.

[0106] The complete subgraph snapshot is copied and instantiated to the specified regional partition node of the regional practice adaptation layer, a new knowledge node instance is created, and a regional affiliation attribute tag is injected into the instance to clearly indicate that the knowledge version is only applicable to a specific administrative region or medical institution cluster.

[0107] Establish version inheritance relationship edges between the newly generated regional practice adaptation layer nodes and the original general knowledge layer nodes, record the evolution trajectory from the general version to the regional frozen version, preserve the knowledge traceability chain, and support subsequent reverse tracing and analysis of the causes of knowledge drift.

[0108] Update the global index cache of the knowledge graph and refresh the hierarchical weight configuration of the knowledge entry. This will enable the front-end question-answering engine to prioritize loading hierarchically stored knowledge entries with regional affiliation attributes when it receives query requests from users in this region, rather than the general authoritative version.

[0109] Through the node migration and permission reconstruction methods described above, the solidified knowledge version generated in the previous step is transformed into hierarchical storage knowledge entries with a clear geographical scope of application. This enables the knowledge base to evolve from a single static structure to a dynamic hierarchical structure, ensuring the localization adaptability and security of knowledge output.

[0110] S4.5: Generate a knowledge application boundary warning signal for the hierarchical knowledge entries, encapsulate the knowledge application boundary warning signal into a structured notification message containing a description of the failure scenario and recommended alternative strategies, and push it to the knowledge collaborative review workbench of the regional diabetes prevention and control center to initiate the manual review process.

[0111] Step S5: Train a long short-term memory network model using the semantic fingerprint of historical clinical intervention behaviors to predict the decay curve of the applicability probability of the solidified knowledge version in newly emerging clinical situations. When the predicted decay rate exceeds a preset failure threshold, a knowledge update requirement package is generated. This knowledge update requirement package includes a sample set of clinical situations where the knowledge is inapplicable and a cluster of high-frequency alternative intervention expressions. Specifically, it includes: S5.1: Obtain the semantic fingerprint of historical clinical intervention behavior and the corresponding spatiotemporal clinical context parameter matrix. Perform temporal alignment processing on the semantic fingerprint of historical clinical intervention behavior based on the time sliding window mechanism to generate a standardized fingerprint input tensor with uniform time granularity, providing a structured data foundation for subsequent model training.

[0112] The historical clinical intervention behavior semantic fingerprint refers to a sequence of fingerprint vectors formed by collecting, encoding, and semantically modeling the operational behavior events of health management service personnel in response to system-pushed suggestions over multiple monitoring periods according to the methods described in steps S1 to S3, and then arranging them in chronological order. Each fingerprint vector in this sequence corresponds to a complete clinical intervention behavior, recording the service personnel's adoption, modification, or rejection patterns of knowledge items in a specific clinical context, and is used to characterize the stability and drift trend of knowledge items in historical application.

[0113] The semantic fingerprints of historical clinical intervention behaviors and their corresponding spatiotemporal clinical context parameter matrices are obtained as the original input data source for temporal alignment processing. Timestamp normalization is performed on the historical clinical intervention behavior semantic fingerprints to map all discrete operational events to a unified time axis coordinate system, eliminating time reference deviations caused by delays in operations by different service personnel. Based on a preset sliding window mechanism, a fixed-length time window and step size are set, and the normalized fingerprint sequence is segmented to ensure that each window contains complete behavioral cycle features. For each segmented time window, linear interpolation is used to fill in missing values ​​caused by data loss, ensuring the continuity of the sequence in the temporal dimension. Z-Score normalization is performed on the filled sequence to eliminate the interference of differences in feature dimensions on model training, generating a standardized fingerprint input tensor with uniform temporal granularity. Through the above temporal alignment and normalization processing, the historical heterogeneous fingerprint data from the previous step is transformed into a structurally regular standardized fingerprint input tensor, achieving the expected technical effect of providing a high-quality structured data foundation for subsequent long short-term memory network model training.

[0114] S5.2: Construct a long short-term memory network model based on the standardized fingerprint input tensor, and use backpropagation to iteratively optimize and train the weight parameters of the forget gate, input gate and output gate of the long short-term memory network model, so as to output an adaptive feature hidden state vector that can represent the evolution of knowledge items over time.

[0115] Receive the standardized fingerprint input tensor generated by S5.1. This tensor contains the semantic fingerprint sequence of clinical intervention behavior within time step T and the corresponding spatiotemporal clinical context parameter matrix.

[0116] A Long Short-Term Memory (LSTM) network architecture is constructed, defining the input layer dimension to be consistent with the fingerprint feature dimension, and setting the number of hidden layer units to 128 to capture long-term dependencies.

[0117] Initialize the weight matrices and bias vectors of the forget gate, input gate, and output gate, and use the Xavier initialization strategy to ensure gradient propagation stability.

[0118] At each time step t, the forget gate activation value is calculated to determine the proportion of information discarded from the cell state at the previous time step.

[0119] The forget gate output is calculated using the following formula: in, It is the Sigmoid activation function. Here is the forget gate weight matrix. This is the hidden state from the previous moment. Input the fingerprint vector at the current time. For bias terms, The output of the forget gate at time step t.

[0120] Calculate the input gate activation value and the new candidate cell state to determine the new information of the cell state that needs to be stored at the current time.

[0121] The input gate and new cell state are calculated using the following formula: in, The input gate activation value, The current cell state, This represents element-wise multiplication, and tanh is the hyperbolic tangent activation function.

[0122] Calculate the output gate activation value and generate the hidden state vector at the current time step based on the updated cell state.

[0123] The output gate and hidden state are calculated using the following formula: in, The output gate activation value. This refers to the adaptive feature hidden state vector that represents the evolution of knowledge items over time.

[0124] The gradient of the loss function with respect to each level of gating weights is calculated using the backpropagation through-time (BPTT) algorithm.

[0125] The Adam optimizer is used to iteratively update the weight parameters, minimizing the cross-entropy loss between the predicted applicability probability and the actual behavior label.

[0126] Through the above chain derivation, the standardized fingerprint input is transformed into an adaptive feature latent state vector rich in temporal evolution information, thus realizing accurate modeling of the dynamic changes in the life cycle of knowledge items.

[0127] S5.3: Receive the structured parameter description of the newly emerging clinical situation, and input the adaptive feature hidden state vector and the structured parameter description of the newly emerging clinical situation into the trained long short-term memory network model for forward inference calculation to generate an applicability probability decay curve that represents the trend of the effectiveness of the solidified knowledge version in the future time period.

[0128] The system receives the weight parameters and adaptive feature hidden state vectors of the long short-term memory network model trained by S5.2 as the initial conditions for forward inference computation.

[0129] Obtain structured parameter descriptions of emerging clinical scenarios, including recent fluctuations in patients' glycated hemoglobin, indicators of changes in renal function stages, and records of medication adjustments for comorbidities, and encode them into fixed-dimensional scenario input vectors.

[0130] The context input vector is concatenated with the adaptive feature hidden state vector from the previous time step to construct a joint input tensor containing historical dependency information and current context features.

[0131] The joint input tensor is fed into the input gate layer of the Long Short-Term Memory network, and the update coefficient of the input gate is calculated through the sigmoid activation function to control the intensity of the influx of new contextual information into the cell state.

[0132] By using a forgetting gate layer to weight and filter the cell state of the previous time step, historical noise interference that is irrelevant to the current new context is eliminated based on the forgetting coefficient, while retaining key time-series dependent features.

[0133] By combining the output gate layer with the updated cell state, the hidden layer output vector at the current time step is calculated. This vector represents the instantaneous fit score of the knowledge item in a specific new context.

[0134] The model iterates through a preset sequence of future time periods, inputting the predicted scenario parameters for each time point into the model in sequence, and recursively calculates the hidden layer output for each time step to form a continuous time series prediction result.

[0135] The fit scores of continuous time steps are normalized and mapped to a probability range of 0 to 1 to generate an applicability probability decay curve that represents the evolution trend of the effectiveness of solidified knowledge versions over time.

[0136] Through the aforementioned forward reasoning calculation process, discrete parameters of newly emerging clinical situations are transformed into continuous probability decay curve data, enabling quantitative prediction of the future applicability of knowledge items and providing accurate data support for subsequent knowledge update instructions.

[0137] S5.4: Perform slope analysis and threshold comparison processing on the applicable probability decay curve. When the predicted decay rate caused by the downward slope of the curve exceeds the preset failure threshold, trigger the knowledge update generation instruction to lock the target solidified knowledge version that needs immediate intervention and its associated high-risk clinical scenario set.

[0138] The system receives a data sequence of applicability probability decay curves generated by S5.3, which includes time steps and corresponding applicability probability values ​​for knowledge items. A first-order difference operation is performed on the decay curves to calculate the probability change rate between adjacent time steps, obtaining an instantaneous slope vector characterizing the knowledge's failure rate. A sliding window mean filter is used to smooth the instantaneous slope vector, eliminating local fluctuations caused by data noise and generating a stable trend slope sequence. A preset failure threshold is set as a fixed negative constant, and each element in the trend slope sequence is compared point-by-point with this threshold. When the trend slope for three consecutive time steps is detected to be lower than the preset failure threshold, the current knowledge item is determined to have entered a rapid failure phase. The structured parameter description of the newly emerging clinical scenario corresponding to the trigger determination time is extracted and marked as a high-risk clinical scenario. The target knowledge version identifier currently in a frozen state is locked, and an association mapping relationship is established between it and the set of high-risk clinical scenarios. By using slope analysis and threshold comparison, the probability prediction results from the previous step are transformed into clear knowledge update trigger signals and high-risk scenario indexes, enabling precise location and immediate interception of potentially invalid content in the knowledge base, significantly improving the timeliness and relevance of knowledge updates.

[0139] S5.5: Based on the set of high-risk clinical scenarios, retrieve the semantic difference backbone of rejection and rewriting and the edit trajectory vector of adoption after fine-tuning in the original operational behavior event flow. Extract high-frequency alternative intervention expression clusters through cluster analysis and integrate the clinical scenario sample set of knowledge inapplicability to encapsulate and generate a knowledge update requirement package.

[0140] The set of clinical scenarios where knowledge is inapplicable refers to the set of specific clinical scenario parameter combinations that cause the probability of applicability of the solidified knowledge version to decay beyond a preset failure threshold. It is obtained as follows: In step S5.4, when the decreasing slope of the applicability probability decay curve is detected, causing the predicted decay rate to exceed the preset failure threshold, the structured parameter description of the newly emerging clinical scenario corresponding to the triggering judgment time (including the patient's recent glycated hemoglobin volatility, renal function stage change identifiers, and comorbidity medication adjustment records) is locked, and this structured parameter description is used as a sample of a clinical scenario where knowledge is inapplicable; the clinical scenario samples accumulated from multiple triggering judgments are aggregated to form the set of clinical scenarios where knowledge is inapplicable.

[0141] The system receives a set of high-risk clinical scenarios locked by step S5.4 and their corresponding target fixed knowledge version identifiers, which serve as index keys for data retrieval. It then invokes the original operational behavior event stream storage engine, performing a reverse query based on the knowledge entry ID and scenario similarity threshold to extract all associated rejection and rewrite operation records and fine-tuned acceptance operation records. The extracted semantic difference backbones of rejection and rewrite are denoised, removing stop words and non-medical entity modifiers while retaining core diagnostic actions and negation logic words. The TF-IDF algorithm is used to calculate the feature weights of keywords in each semantic difference backbone, constructing a high-dimensional sparse text vector matrix. The K-Means clustering algorithm iteratively divides the text vector matrix, setting the number of clusters K to a preset number of common correction patterns. After convergence, several semantically similar alternative expression clusters are formed. The number of operational behavior events contained in each cluster is counted, and clusters with frequencies exceeding a preset threshold are selected as high-frequency alternative intervention expression clusters. The typical text corresponding to their centroid vectors is extracted as prototypes for recommended alternative strategies. The system processes the adopted edit trajectory vectors after fine-tuning in parallel, calculates the quantiles of the edit distance distribution, identifies high-frequency patterns of local corrections, and maps them to specific parameter adjustment suggestions. It integrates a sample set of clinical scenarios where knowledge is inapplicable, containing specific combinations of patient physiological parameters and medication backgrounds that lead to knowledge applicability decay. The high-frequency alternative intervention expression clusters, local correction patterns, and the sample set of clinical scenarios where knowledge is inapplicable are encapsulated into a standard JSON structure, with added metadata tags such as generation timestamps, confidence scores, and source institution codes. Through this data processing workflow, discrete behavioral logs are transformed into structured knowledge update requirement packages, achieving automated generation from behavioral insights to the basis for knowledge iteration, significantly reducing the information load for expert review.

[0142] Step S6: Obtain the expert confirmation results for the knowledge update request package, and update the knowledge graph based on the confirmation results to form a dual-version coexistence knowledge structure. Specifically, parse the knowledge adoption confirmation instructions and applicable condition semantics of the annotations executed by regional experts on the knowledge update request package, map the confirmation results into dynamic attribute increments of applicable condition nodes in the knowledge graph, and construct a practice consensus version superimposed on the original authoritative version to form a dual-version coexistence knowledge structure. This includes: S6.1: Obtain the knowledge adoption confirmation instruction text and the basis annotation natural language description executed by regional experts on the knowledge collaboration workbench for the knowledge update requirement package. Use entity recognition processing in natural language processing technology to perform intent classification processing on the knowledge adoption confirmation instruction text, and extract key medical term entities from the basis annotation natural language description to generate a standardized set of expert decision semantic vectors.

[0143] The text of the knowledge adoption confirmation instruction and the natural language description with annotations submitted by experts in the receiving area on the knowledge collaboration platform for the knowledge update requirement package are used as the raw input data for semantic parsing.

[0144] The knowledge adoption confirmation instruction text is processed by word segmentation and syntactic dependency analysis to construct an instruction semantic tree to identify the core predicate verb and its modifiers, and to preliminarily determine the category of the expert's operational intention.

[0145] Based on a pre-trained BERT model for the medical field, the instruction text is mapped to a high-dimensional semantic space, and the cosine similarity between it and standard intent vectors such as "accept", "reject", and "correct" in the preset intent template library is calculated.

[0146] The intent label with the highest similarity is selected as the classification result of the current instruction, and corresponding confidence weights are assigned to form a structured intent classification label.

[0147] Named entity recognition (NER) is performed on the labeled natural language descriptions, and the BiLSTM-CRF model is used to extract key medical term entities such as drug names, physiological index thresholds, and complication types.

[0148] The extracted medical terminology entities are standardized and normalized, and mapped to the standard concept unique identifier of the Unified Medical Language System (UMLS) to eliminate synonym ambiguity.

[0149] The intent classification identifier is concatenated with the standardized medical terminology entity set to construct a multimodal feature vector containing the operation direction and medical basis.

[0150] By encoding the multimodal feature vectors through weighted fusion processing, a standardized set of expert decision semantic vectors representing the expert decision-making logic is generated.

[0151] By using the above processing method, unstructured expert review opinions are transformed into a standardized set of expert decision semantic vectors that can be understood by computers, realizing the accurate transformation of expert tacit knowledge into explicit structured data, and providing highly reliable semantic basis for the subsequent dynamic attribute updates of the knowledge graph.

[0152] S6.2: Based on the generated set of expert decision semantic vectors, the pre-built medical logic mapping rule base is called to perform semantic alignment operation, mapping the positive intent features and negative intent features in the set of expert decision semantic vectors to attribute expansion operators or attribute constraint operators in the knowledge graph update strategy, so as to output a structured attribute change instruction stream with directional markers.

[0153] The system receives a set of expert decision semantic vectors generated by S6.1. This set contains intent classification labels for knowledge adoption confirmation instructions and key medical terminology entity features from the annotations. A pre-built medical logic mapping rule base is invoked, which defines a bidirectional mapping relationship table between natural language intents and knowledge graph attribute operators. Positive intent features in the semantic vector set are parsed to identify positive operation keywords such as "include" and "extend application." These positive operation keywords are mapped to attribute expansion operators in the knowledge graph update strategy, indicating the addition of new allowed values ​​or the relaxation of constraints in existing applicable condition nodes. Negative intent features in the semantic vector set are parsed to identify negative operation keywords such as "disable," "limited," and "exclude." These negative operation keywords are mapped to attribute constraint operators in the knowledge graph update strategy, indicating the addition of prohibited values ​​or the tightening of constraints in existing applicable condition nodes. Key medical terminology entities, such as specific drug names, complication stages, or test indicator thresholds, are extracted from the annotation natural language descriptions. Entity linking technology is used to map extracted key medical terminology entities to standard concept node IDs in the knowledge graph. The mapped standard concept node IDs are then structurally assembled with their corresponding attribute operators to generate triple instructions containing the operation type, target node ID, and new attribute value. Directional markers are added to the generated triple instructions to distinguish between incremental and decremental update operations. Through this semantic alignment and mapping process, unstructured expert decision text is transformed into a structured attribute change instruction flow with directional markers, achieving precise conversion of expert experience into machine-executable knowledge and ensuring the atomicity and traceability of knowledge base updates.

[0154] S6.3: Receives a structured attribute change instruction stream with directional markers, locates the applicable condition node under the original authoritative version associated with the solidified knowledge version in the knowledge graph, uses the transaction mechanism of the graph database to apply attribute expansion operators or attribute constraint operators to the existing attribute list of the applicable condition node, and calculates and generates dynamic attribute incremental data blocks representing local diagnosis and treatment adaptability.

[0155] Receive a stream of structured attribute change instructions with directional markers, parse the operator types and associated medical entity identifiers in the instructions, and determine the unique index position of the target knowledge node to be updated in the graph database.

[0156] Call the transaction management interface of the graph database to start a read-write transaction session with atomicity guarantee, lock the original authoritative version node corresponding to the target knowledge entry and its subordinate applicable condition child nodes to prevent data inconsistency caused by concurrent writes.

[0157] Retrieve the existing attribute list of the applicable condition node, identify the attribute key-value pairs that match the operators in the change instruction, and if the operator is an attribute extension, prepare to add the new local diagnosis and treatment parameters to the existing set; if it is an attribute constraint, prepare to modify the value range or logical expression of the original attribute.

[0158] For attribute constraint operations, a set theory-based attribute difference calculation model is constructed, and the dynamic attribute incremental data block is calculated using the following formula: in, For dynamic attribute incremental data blocks, This is a set of new attributes confirmed by experts. For the existing set of attributes in the original authoritative version, This is the original attribute constraint function. This is the corrected attribute constraint function.

[0159] When performing attribute extension operations, the new attribute fields, such as the local medical insurance catalog code and the regional specific complication screening rules contained in the instruction, are serialized and encapsulated according to the predefined schema structure to generate attribute extension fragments to be merged.

[0160] When performing attribute constraint operations, a tuple containing instructions to remove old constraints and add new constraints is generated based on the formula calculation results. This ensures minimal intrusion into the original general guideline logic, retaining only the necessary corrections due to regional differences.

[0161] The generated attribute extension fragments or constraint correction tuples are encapsulated into standardized dynamic attribute increment data blocks. These data blocks contain version numbers, effective timestamps, source organization codes, and specific attribute change payloads, serving as the core input for subsequent version instantiation.

[0162] Submitting a graph database transaction temporarily stores the dynamic attribute incremental data blocks in the transaction log, awaiting final consistency verification and persistent writing, ensuring the integrity and traceability of data operations.

[0163] By using the graph database transaction mechanism and attribute differential calculation processing method described above, the structured attribute change instruction flow of the previous step is transformed into dynamic attribute incremental data blocks that represent local diagnostic adaptability, thereby realizing non-destructive local modification and refined version management of the authoritative knowledge system.

[0164] S6.4: Based on the generated dynamic attribute incremental data block, instantiate a new practice consensus version node in the knowledge graph storage layer, establish an inheritance relationship edge from the practice consensus version node to the original authoritative version node, and merge the dynamic attribute incremental data block into the attribute set of the practice consensus version node to construct a dual-version coexistence knowledge structure with dual version attributes.

[0165] Receive the dynamic attribute incremental data block generated by S6.3, which contains localized diagnostic constraints and extended attributes confirmed by experts. In the graph database storage layer, call the node instantiation interface to create a new practice consensus version node, assign a globally unique identifier (UUID) to the node, and initialize its version status field to "pending activation".

[0166] Establish a directed inheritance relationship between the practice consensus version node and the original authoritative version node, set the edge type to "derived_from", and record the generation timestamp and the regional expert ID that triggered the version update in the edge attributes to ensure the integrity of the knowledge traceability chain.

[0167] The key-value pair structure in the dynamic attribute incremental data block is parsed, and the attribute fields representing local medical insurance coverage, essential drug list restrictions, and regional epidemiological characteristics are merged one by one into the attribute set of the practice consensus version node. For attributes that conflict with the original authoritative version, an overwrite operation is performed; for newly added attributes, an append operation is performed.

[0168] The confidence weight coefficient of the practical consensus version is calculated using the following formula: in, The number of experts who agree to the change. The total number of experts participating in the review. This serves as a moderating factor for regional healthcare levels. This weighting coefficient is written into the specified attribute field of the practice consensus version node.

[0169] Perform schema validation on the merged attribute set to check whether required fields such as "Applicable Population", "Contraindications", and "Recommendation Level" are complete. If the validation passes, update the node status to "Valid"; otherwise, roll back the transaction and throw an exception log.

[0170] Through the above processing method, the structured attribute change instructions in the previous step are transformed into practice consensus version nodes with complete inheritance relationships and high confidence weights, realizing the coexistence and isolated storage of dual versions of general authoritative knowledge and region-specific knowledge in the knowledge base, providing a structured data foundation for subsequent differentiated reasoning.

[0171] S6.5: Traverse the constructed dual-version coexistence knowledge structure, perform consistency verification on the dynamic attribute incremental data blocks in the practice consensus version node, verify whether the dynamic attribute incremental data blocks conflict with the underlying logical constraints of the original authoritative version node, and if the verification passes, activate the reasoning service interface of the practice consensus version node to output the final practice consensus version knowledge object that can be called by the front-end question answering engine.

[0172] Obtain the practice consensus version node and its associated dynamic attribute incremental data block in the dual-version coexistence knowledge structure, and extract the underlying logical constraint rule set of the original authoritative version node as the input object for consistency verification.

[0173] Formal modeling is performed on the medical entities and logical relationships in the dynamic attribute incremental data block, which is then transformed into a set of predicate logic expressions. A descriptive logic inference engine is used to detect whether the newly added attributes have semantic conflicts with the axiomatic system of the original authoritative version.

[0174] Using graph isomorphism-based subgraph matching, all outgoing and incoming edges of the consensus version nodes in the knowledge graph are traversed to verify whether the newly added applicable condition nodes have destroyed the topological integrity of the original knowledge entries and the connectivity of the causal chain.

[0175] Construct a logical mutual exclusion detection matrix. For hard constraints such as contraindications to diabetes medication, dosage limits, and exclusion relationships with complications, calculate the non-emptiness of the intersection between the dynamic attribute increment and the original constraint rules. If mutual exclusion terms exist, mark them as serious logical conflicts.

[0176] For dynamic attribute increments that do not trigger serious logical conflicts, a soft compatibility score is calculated, and its semantic deviation from the original authoritative version is evaluated using the following formula: in, For semantic deviation, This represents the numerical or encoded value of the i-th dynamic attribute. This is the baseline value for the attribute corresponding to the original authoritative version. This represents the total number of attributes involved in the validation.

[0177] When semantic deviation If the value is below the preset security threshold and there is no hard logical conflict, the consistency check is deemed to have passed, the system automatically activates the inference service interface of the practice consensus version node, and marks the node status as "callable".

[0178] The validated practice consensus version knowledge object is encapsulated into a standardized JSON-LD format data packet, which includes the version hash value, effective timestamp, and regional applicability identifier, and output to the knowledge cache layer of the front-end question answering engine.

[0179] Through the above chain-like derivation process, the dual-version structure constructed in the previous step is transformed into a highly reliable knowledge object that has been logically verified, thereby achieving a closed-loop security control in the knowledge base update process and ensuring that localized corrections do not violate basic medical principles.

[0180] Step S7: Based on the user's institutional information, select the corresponding knowledge version from the dual-version coexisting knowledge structure as the reasoning basis to generate personalized intervention suggestions. Specifically, match the hierarchical weight configuration in the dual-version coexisting knowledge structure according to the user's institutional code, load the corresponding practice consensus version or the original authoritative version as the reasoning basis, and generate the personalized intervention suggestions. The personalized intervention suggestions combine national guideline compliance with local treatment adaptability. Specifically, this includes: S7.1: Obtain the identity information of the user who initiated the question-and-answer request, and parse the user identity information based on the pre-set institution mapping table to extract the unique institution code and the corresponding regional medical attribute label, forming a user clinical context parameter set containing regional characteristics.

[0181] The system receives the user's identity information, including the user's login session ID, affiliated medical institution code, and device fingerprint. It then calls a pre-built institution mapping table query interface, using the medical institution code from the user's identity as the primary key for hash retrieval, matching the corresponding institution-level attribute record. The matched institution-level attribute record is parsed to extract a unique institution code field, which identifies the node affiliation in the regional practice adaptation layer of the knowledge graph. Simultaneously, regional medical attribute tags are extracted from the institution mapping record, including the regional medical insurance policy version, the local essential drug list index, and the clinical pathway execution standard version number. The extracted regional medical attribute tags are standardized and cleaned, removing redundant spaces and non-standard characters to ensure the tag format conforms to the knowledge graph ontology definition specifications. The unique institution code and the standardized regional medical attribute tags are combined to construct a user clinical context parameter set containing regional characteristics. Through the above parsing and extraction processes, the discrete user identity is transformed into a structured regional feature parameter set, providing the data foundation for subsequent accurate knowledge version matching.

[0182] S7.2: Based on the institution code in the user clinical context parameter set containing regional characteristics, retrieve the associated knowledge entry status marker in the dual-version coexisting knowledge structure, calculate the dynamic weight coefficient between the general knowledge layer and the practice consensus layer, and generate a hierarchical weight configuration vector for the current institution.

[0183] The system receives institution codes from a set of user clinical context parameters containing regional characteristics, which are used as index keys to retrieve knowledge structures with coexisting dual versions. It then invokes the knowledge graph distributed indexing engine to locate the associated set of regional practice adaptation layer nodes based on the institution codes, extracting the status markers and version number metadata for all knowledge entries within this set. The extracted status markers are validated, filtering out knowledge entries in a frozen or pending-review state, while retaining active practice consensus version nodes and their corresponding original authoritative version nodes. A hierarchical weight calculation model is constructed, taking into account the confidence score of the practice consensus version, the update timestamp decay factor, and the baseline weight value of the original authoritative version. A weighted harmonic average is used to calculate the dynamic weight coefficients, as shown in the following formula: in, For dynamic weighting coefficients, To establish confidence in the consensus version, The original authoritative version serves as the baseline weight. The calculated dynamic weight coefficients are mapped to the configuration vector space corresponding to the current institution's code, generating a hierarchical weight configuration vector containing the weight values ​​of each knowledge dimension. Through the above processing method, the result of the previous step is transformed into a quantitative weight indicator for a specific institution, achieving the expected technical effect of dynamically balancing the knowledge base content between general compliance and local adaptability.

[0184] S7.3: Using the hierarchical weight configuration vector, the original authoritative version node and the practice consensus version node in the dual-version coexistence knowledge structure are weighted and fused to select high-confidence applicable condition attributes and intervention strategy parameters, and to construct a dynamic reasoning knowledge subgraph that adapts to local diagnosis and treatment realities.

[0185] Receive the hierarchical weight configuration vector generated by S7.2. This vector contains the general knowledge layer weight coefficients and practice consensus layer weight coefficients for the current user's organization. Analyze the original authoritative version node and the practice consensus version node in the dual-version coexistence knowledge structure, extracting the common intervention strategy attribute fields and their respective unique applicable condition constraints. Map the attribute values ​​of the original authoritative version node and the practice consensus version node to high-dimensional feature vectors, respectively. Perform a weighted linear combination operation on the two sets of feature vectors based on the coefficients in the hierarchical weight configuration vector. Calculate the confidence score of the fused intervention strategy using the following formula: in, To integrate confidence scores, For the general knowledge layer weight coefficients, The semantic matching degree of the original authoritative version. To implement the consensus layer weight coefficients, To ensure local adaptability of the consensus version, all candidate intervention paths in the knowledge graph are sorted and filtered based on the calculated fusion confidence score, eliminating low-quality inference branches with confidence scores below a preset threshold. High-confidence applicability attributes and intervention strategy parameters are retained to construct a dynamic inference knowledge subgraph containing drug dosage, contraindication examinations, and lifestyle recommendations. Through weighted fusion processing, the weight configuration from the previous step is transformed into regionally adaptable high-confidence knowledge subgraph data, achieving a construction of inference basis that balances the authority of national guidelines with the realities of local diagnosis and treatment.

[0186] S7.4: Based on the dynamic reasoning knowledge subgraph and combined with real-time collected user clinical sign data, perform multi-path logical reasoning operations to generate a preliminary draft of personalized diabetes intervention recommendations. Then, verify the compliance of the medication recommendations in the draft with the local medical insurance catalog and pharmacy inventory rules to form an intermediate state result of the intervention recommendations after localization correction.

[0187] Acquire the node attributes of the dynamic reasoning knowledge subgraph and real-time user clinical sign data, including fasting blood glucose, glycated hemoglobin, eGFR, and a list of comorbidities. Map the sign data into a constraint matrix in vector space and perform semantic matching with the intervention rule nodes in the knowledge subgraph. Calculate the fit score for each candidate intervention strategy using a weighted scoring method, as shown in the following formula: in, For overall fit, The weight of the i-th dimension clinical indicator, For matching functions, Recommended parameters for the guide. The initial draft intervention recommendations are generated based on the user's actual vital signs. Strategies with a fit higher than a threshold are selected. Drug names and dosages are extracted from the draft, and the reimbursement ratio and restrictions are obtained from the local medical insurance database. The real-time drug inventory status is retrieved from the regional pharmacy inventory system. A compliance verification logic gate is constructed; if a drug is not in the medical insurance catalog or has zero inventory, the recommendation is marked as unavailable. All recommendations are iterated through, and non-compliant items are removed, with the reasons for removal recorded. The remaining compliant recommendations are reordered, prioritizing solutions with high fit and local accessibility. Through multi-path logical reasoning and localized compliance verification, the initial draft is transformed into an intermediate intervention recommendation result corrected for local resource constraints, achieving a balance between the economy and feasibility of the treatment plan.

[0188] S7.5: Perform natural language structured recombination and authoritative source annotation on the intermediate results of the localized intervention recommendations, visualize the differences between the national guidelines and local practice consensus, and generate a final personalized intervention recommendation message that combines compliance with national guidelines and adaptability to local diagnosis and treatment.

[0189] Receive intermediate results of intervention recommendations that have been localized and modified, which include structured diagnostic and treatment parameters and compliance verification markers.

[0190] Dependency syntactic analysis is performed on key medical entities in intermediate results to identify core intervention actions and modifying conditions, and a semantic dependency tree is constructed to distinguish between general guideline logic and local adaptation logic.

[0191] Based on the semantic dependency tree, an authoritative source tracing and annotation engine is invoked to map each intervention node to the original authoritative version ID or practice consensus version ID in the knowledge graph, generating a sequence of semantic units with version source tags.

[0192] Using natural language generation templates, semantic unit sequences are reorganized into coherent natural language text, and differentiated visual markup codes are embedded in the text. The original authoritative basis is rendered in a standard font, while differences in local practice consensus are rendered in the form of highlights or footnotes.

[0193] The final message encapsulation is performed, packaging the recombined natural language text, visual markup code, and underlying knowledge tracing metadata into a standard JSON format output message.

[0194] By using natural language structured recombination and authoritative source annotation, the intermediate results from the previous step are transformed into personalized intervention recommendations that are both compliant with national guidelines and adaptable to local diagnosis and treatment, thus achieving a balance between the scientific validity and practicality of the intervention recommendations.

[0195] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0196] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0197] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for generating diabetes intervention plans based on AI agents, characterized in that, include: S1: Obtain the action events generated by health management service personnel in response to system-pushed suggestions during diabetes intervention Q&A interactions, and encapsulate them into a raw action event stream; S2: Based on the original operational behavior event flow, associate the unique identifier of the knowledge entry, the original source metadata, and the clinical context parameters to form a structured clinical intervention dataset; S3: Perform cluster analysis on multiple operational behavior events of the same knowledge item in the structured clinical intervention dataset within a preset period and with clinical context parameter similarity greater than a threshold, extract operation type distribution entropy, mean text edit distance and response sensitivity features of clinical context parameters, and generate a semantic fingerprint of clinical intervention behavior. S4: Determine whether the semantic fingerprint of the clinical intervention behavior meets the preset knowledge drift triggering conditions. If so, trigger a knowledge application boundary warning and migrate the corresponding knowledge item from the general knowledge layer to the regional practice adaptation layer to generate a solidified knowledge version. S5: Train a long short-term memory network model using the semantic fingerprint of historical clinical intervention behavior, predict the applicability probability decay curve of the solidified knowledge version in newly emerging clinical situations, and generate a knowledge update requirement package when the predicted decay rate exceeds a preset failure threshold. S6: Obtain the expert's confirmation result for the knowledge update requirement package, update the knowledge graph according to the confirmation result, and form a dual-version coexistence knowledge structure; S7: Obtain user identification information to determine the corresponding organization code, select the corresponding knowledge version from the dual-version coexisting knowledge structure based on the organization code as the reasoning basis, and generate personalized intervention suggestions.

2. The method for generating a diabetes intervention plan based on an AI agent according to claim 1, characterized in that, The operational events include direct adoption timestamps, edit trajectory vectors of adoption after fine-tuning, semantic difference backbones of rejection and rewriting, and regional quality control review markers.

3. The method for generating a diabetes intervention plan based on an AI agent according to claim 2, characterized in that, The clinical context parameters include glycated hemoglobin trends and renal function staging.

4. The method for generating a diabetes intervention plan based on an AI agent according to claim 3, characterized in that, Step S3 specifically includes: Based on the structured clinical intervention dataset, the dynamic time warping algorithm is used to perform vector space mapping on the clinical context parameters containing the glycated hemoglobin trend and renal function stage, so as to calculate the Euclidean distance between any two operational behavior events and screen out the candidate operational behavior event set with clinical context parameter similarity greater than a preset threshold, forming a high-confidence similar situational behavior subset. For the high-confidence similar situational behavior subset, the Shannon information entropy calculation formula is used to model the probability distribution of the occurrence frequency of direct adoption, fine-tuned adoption, rejection and rewriting, and marking as requiring regional quality control review, forming the first-dimensional feature vector; Based on the fine-tuned adopted edit trajectory vector in the high-confidence similar situation behavior subset, the number of character editing operations between the original suggested text and the modified text is calculated using the Levenstein distance formula, and the number of character editing operations is normalized to obtain the average text editing distance, forming the second-dimensional feature vector; Combining the semantic differences between the rejection and rewritten parts of the high-confidence similar situational behavior subset with the clinical context parameters, a multiple linear regression model is applied to fit the contribution weights of the changes in each dimension of parameters to the occurrence of rejection behavior, so as to calculate the sensitivity gradient coefficient of the clinical context parameters and form a third-dimensional feature vector. By integrating the first, second, and third feature vectors, a multidimensional feature matrix is ​​constructed through feature concatenation and standardized encoding. Principal component analysis is then used to extract feature components in the main variation directions, ultimately generating the semantic fingerprint of the clinical intervention behavior.

5. The method for generating a diabetes intervention plan based on an AI agent according to claim 1, characterized in that, Step S4 specifically includes: Obtain the semantic fingerprint of clinical intervention behavior within two consecutive monitoring periods, and perform temporal sliding window calculation on the distribution entropy of the operation type and the mean of text editing distance contained in the semantic fingerprint of clinical intervention behavior to generate a periodic drift feature vector sequence; Based on the periodic drift feature vector sequence, the fine-tuning adoption rate features and the co-occurrence frequency of rejection reason keywords are extracted to determine whether they meet the preset knowledge drift triggering conditions and generate a composite drift judgment result. Using the composite drift determination result as a trigger condition, a state locking instruction is executed on the target knowledge item currently in the general knowledge layer, switching its access permission configuration from global open mode to restricted frozen mode, and generating the solidified knowledge version; Based on the metadata identifier of the solidified knowledge version, a node migration operation is performed to strip the solidified knowledge version from the general knowledge layer and remount it to the regional partition node corresponding to the regional practice adaptation layer, thereby generating hierarchical storage knowledge entries. A knowledge application boundary warning signal is generated for the hierarchical storage knowledge entries, and the knowledge application boundary warning signal is encapsulated into a structured notification message.

6. The method for generating a diabetes intervention plan based on an AI agent according to claim 5, characterized in that, The preset knowledge drift triggering conditions include: the adoption rate of the clinical intervention behavior semantic fingerprint exceeds the preset high-frequency correction threshold within two consecutive cycles, and the density of the occurrence of the rejection reason keyword co-occurrence frequency related to the local drug catalog not covering specific entries suddenly increases.

7. The method for generating a diabetes intervention plan based on an AI agent according to claim 1, characterized in that, The knowledge update requirement package includes a sample set of clinical scenarios where knowledge is inapplicable and a cluster of frequently used alternative intervention statements.

8. The method for generating a diabetes intervention plan based on an AI agent according to claim 1, characterized in that, Step S5 specifically includes: Obtain the semantic fingerprint of the historical clinical intervention behavior and the corresponding spatiotemporal clinical context parameter matrix. Perform temporal alignment processing on the semantic fingerprint of the historical clinical intervention behavior based on the time sliding window mechanism to generate a standardized fingerprint input tensor. The long short-term memory network model is constructed based on the standardized fingerprint input tensor. The weight parameters of the forget gate, input gate and output gate of the long short-term memory network model are iteratively optimized and trained using the backpropagation algorithm, and the adaptive feature hidden state vector is output. The system receives a structured parameter description of a newly emerging clinical situation, and inputs the adaptive feature hidden state vector and the structured parameter description of the newly emerging clinical situation into the trained long short-term memory network model for forward inference calculation to generate an applicability probability decay curve. Slope analysis and threshold comparison are performed on the applicable probability decay curve. When the predicted decay rate caused by the downward slope of the curve exceeds the preset failure threshold, a knowledge update generation instruction is triggered to lock the target solidified knowledge version that needs immediate intervention and its associated set of high-risk clinical scenarios. Based on the set of high-risk clinical scenarios, the semantic difference backbone of rejection and rewriting and the edit trajectory vector of adoption after fine-tuning in the original operational behavior event flow are retrieved. The high-frequency alternative intervention expression clusters are extracted by cluster analysis and integrated with the clinical scenario sample set of knowledge inapplicability to encapsulate and generate the knowledge update requirement package.

9. The method for generating a diabetes intervention plan based on an AI agent according to claim 1, characterized in that, The process of obtaining expert confirmation of the knowledge update requirement package and updating the knowledge graph based on the confirmation results to form a dual-version coexisting knowledge structure includes: The knowledge adoption confirmation instructions and applicable condition semantics of the annotations executed by the regional experts on the knowledge update requirement package are analyzed. The confirmation results are mapped to the dynamic attribute increments of the applicable condition nodes in the knowledge graph, and a practice consensus version superimposed on the original authoritative version is constructed to form a dual-version coexistence knowledge structure.

10. The method for generating a diabetes intervention plan based on an AI agent according to claim 1, characterized in that, The step of selecting the corresponding knowledge version from the dual-version coexisting knowledge structure based on user organization information as the basis for reasoning and generating personalized intervention suggestions includes: Based on the user's organization code, the hierarchical weight configuration in the dual-version coexistence knowledge structure is matched, and the corresponding practice consensus version or original authoritative version is loaded as the reasoning basis to generate the personalized intervention suggestion.

Images