National chat dual-channel medicine information semantic management system

Abstract

The invention relates to the technical field of medicine information semantic management, and discloses a national chat dual-channel medicine information semantic management system. The system comprises a state observation unit, a correlation measurement unit, an attribution analysis unit, a feature purification unit and a semantic integration unit. According to the system, semantic features are extracted in a framing mode, a feature profile is synthesized, a multi-scale semantic incidence matrix is constructed to identify key transition points, and a track map is generated. And mapping the transition points to original information elements, and tracking a cross-context evolution path of the original information elements to form an attribution chain. And generating a semantic factor set of causal association based on the chain, and stripping noise similarity factors through feedback iteration to obtain a purified core factor set. And performing time sequence reconstruction and context filling on the core factors according to the evolution sequence, and outputting a standardized semantic record. According to the system, the causal and dynamic process of semantic evolution can be deeply analyzed, and the accuracy and interpretability of information management are improved.

Description

Technical Field

[0001] This invention relates to the field of drug information semantic management technology, specifically a national negotiated dual-channel drug information semantic management system. Background Technology

[0002] Existing technologies for semantic management of drug information typically employ rule-based template matching or statistical machine learning models. These methods segment text, vectorize features, and calculate co-occurrence frequencies or statistical correlations between features to classify and monitor information. Their focus is on identifying keywords and extracting static features from the semantic stream, or determining the category of information through classification models; essentially, they discretize and summarize semantic information in a cross-sectional manner.

[0003] The main shortcomings of existing technologies lie in two aspects. First, the feature extraction and analysis process struggles to distinguish between causal factors and accompanying noise in semantic evolution. Second, conventional methods rely on statistical correlation to construct feature sets, failing to effectively isolate pseudo-features formed by data noise and accidental co-occurrence, leading to semantic understanding deviating from the true driving factors. Third, they lack the ability to characterize the continuous evolution of information entities across contexts. Existing technologies primarily record snapshots of information states or independent events at discrete points in time, failing to establish state relationships of the same information element in different semantic scenarios, and thus unable to reconstruct its complete dynamic evolution path, limiting the realization of deep attribution and accurate source tracing.

[0004] A technical solution is needed that can extract core semantic factors with causal explanatory power from dynamic semantic streams and track the entire evolution process of specific information elements in continuous semantic contexts, so as to achieve in-depth analysis of the essence and path of semantic evolution of drug information. Summary of the Invention

[0005] The purpose of this invention is to provide a semantic management system for dual-channel drug information under national drug negotiations, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a semantic management system for dual-channel drug information under national drug price negotiations, the system comprising: The state observation unit continuously collects the data waveform of the semantic information stream, divides the waveform data into several data frames of fixed duration, extracts and quantizes multi-level semantic features for each data frame, and synthesizes the quantized features of all data frames into a feature profile, which corresponds to the complete evolution process of the drug information stream. The association measurement unit performs multi-scale alignment on the feature profile, constructs semantic association matrices at different time scales, identifies key semantic transition points by calculating the difference between matrices, and generates a trajectory map of information variation. The attribution analysis unit analyzes the trajectory map of information variation, maps the transition points therein to specific information elements in the original drug data stream, and tracks the evolution path of the specific information elements in different semantic contexts to form a structured attribution chain. The feature purification unit generates a set of semantic factors that are causally related to information variation based on the structured attribution chain, and removes semantic factors that are similar to noise patterns from the set of semantic factors through feedback iterative calculation, and outputs the purified core semantic factor set. The semantic integration unit inputs the core semantic factor set into the sequence generation model. The model performs temporal reconstruction and context filling of the core semantic factors based on the evolution order recorded in the structured attribution chain, and outputs standardized drug information semantic records.

[0007] Preferably, the state observation unit includes the following: The system continuously receives raw data from the drug information stream via sensors or interfaces and converts the raw data stream into an analyzable waveform signal. The waveform signal is then uniformly segmented along the time axis to form a series of non-overlapping data frames with fixed durations. For each data frame, multi-level semantic feature extraction operations are performed in parallel. The extracted semantic features include at least word frequency distribution features, syntactic structure features, and contextual association features. The extracted multi-level semantic features are converted into numerical vectors to complete the quantization process; By concatenating and integrating the quantized feature vectors of all data frames in chronological order, a feature profile representing the complete information evolution cycle is constructed.

[0008] Preferably, the correlation metric unit includes the following: The feature profile is received and multiple different observation scales are preset for it, each observation scale corresponding to a specific time window length; At each preset observation scale, the feature profile is divided into continuous segments, the similarity between the feature vectors of adjacent segments is calculated, and all similarity values ​​are arranged to form a semantic association matrix at the observation scale. By comparing multiple semantic association matrices generated at different observation scales, a scale difference matrix is ​​generated by calculating the absolute difference between elements at corresponding positions in the matrices. In the scale difference matrix, peak points whose values ​​exceed a preset threshold are identified, and these peak points are defined as key semantic transition points. Connect all the identified key semantic transition points in chronological order of their occurrence, and label the corresponding scale difference intensity to form a trajectory map of information variation.

[0009] Preferably, the attribution analysis unit includes the following: Based on the timestamps of key semantic transition points recorded in the trajectory map of the information variation, we trace back to the feature profile and the original drug information stream to locate the specific text unit or data field that triggered the key semantic transition point. Starting from the specific text unit, a bidirectional scan is performed along the time axis, forward and backward, to collect all contextual information blocks that contain the specific text unit or are highly semantically related to it. All collected contextual information blocks are arranged chronologically, and the semantic role, attribute modification, and association changes of the specific text unit in different information blocks are analyzed. The semantic role changes, attribute modification changes, and association changes of the specific text units are recorded as a series of event nodes with causal or temporal relationships, and all event nodes are linked together to form a structured attribution chain from the initial state to the transition state.

[0010] Preferably, the feature purification unit includes the following: Receive the structured attribution chain and extract all event nodes recorded as causing changes from the chain; The information change corresponding to each event node is abstracted into an independent semantic factor, and all semantic factors constitute the initial semantic factor set. Establish a noise feature library that runs parallel to the semantic factor set. The noise feature library stores common interference patterns derived from historical data. Each factor in the initial semantic factor set is compared one by one with all patterns in the noise feature library to calculate their morphological similarity; If the similarity between a semantic factor and any noise pattern exceeds a set threshold, a stripping procedure is initiated to subtract the common part of the noise pattern from the feature expression of the semantic factor. After multiple rounds of iterative comparison and stripping operations, the remaining semantic factors that are no longer highly similar to the noise patterns constitute the purified core semantic factor set.

[0011] Preferably, the semantic integration unit includes the following: Configure an attention-based sequence generation model; use the core semantic factor set as the content input of the sequence generation model, and use the evolution order of event nodes recorded in the structured attribution chain as the structural input of the sequence generation model; The sequence generation model first determines the approximate temporal position of the core semantic factors in the output sequence based on the structural input; Based on the content input and location information, the model generates contextual description fragments that conform to the specifications of the pharmaceutical information domain for each core semantic factor; The model combines all generated description fragments according to their defined temporal positions and ensures logical coherence between the fragments, outputting a complete and semantically standardized drug information semantic record.

[0012] Preferably, the system further includes a map updating unit, which includes the following: Continuously monitor newly generated drug information flows and use the state observation unit to generate new feature profiles; The new feature profile is compared with the trajectory map of historical information variation to identify newly emerging semantic transition patterns that are not recorded in the map; the identified new semantic transition patterns are submitted to the attribution analysis unit to generate supplementary structured attribution chains. By using supplementary structured attribution chains, the existing trajectory map of information variation is expanded in terms of nodes and paths, resulting in an updated trajectory map of information variation.

[0013] Preferably, the system further includes a verification closed-loop unit, which includes the following: Obtain the drug information semantic record output by the semantic integration unit and compare it with an authoritative drug information database in the real world; The differences found in the comparison are back-mapped to the structured attribution chain and the core semantic factor set to locate the source node or factor that caused the difference; based on the nature of the difference, a calibration instruction is generated, which is used to adjust the scale difference threshold of the correlation metric unit or the noise stripping intensity of the feature cleanup unit. The calibration command is fed back to the corresponding processing unit of the system, driving the system parameters to perform adaptive optimization.

[0014] Preferably, the system further includes a cache acceleration unit, which includes the following: A feature profile buffer is established between the state observation unit and the associated measurement unit to temporarily store feature profile data generated in the most recent period. When the system needs to process a new drug information flow, it first checks the feature profile cache to see if there is a historical feature profile that is highly similar to the current information flow. If it exists, the trajectory map of pre-calculated information variation corresponding to the historical feature profile and the subsequent processing results are directly called as the initial reference for the current processing, and only the difference part is incrementally calculated.

[0015] Preferably, the system further includes an interface adaptation unit, which includes the following: Define a set of standardized data exchange protocols and semantic description templates; when the system receives external drug information streams, it converts raw data from different sources and in different formats into a unified waveform signal within the system through protocol parsing; When the system outputs semantic records of drug information, it calls the corresponding semantic description template according to the needs of the target application scenario, and converts the internally unified record format into output data that conforms to specific interface specifications.

[0016] Compared with the prior art, the beneficial effects of the present invention are: A semantic factor purification technique based on causal association and feedback iteration generates a set of semantic factors with causal inference relationships to information variation, rather than relying on statistical correlation. By introducing a feedback iteration mechanism, the factors in the set are dynamically compared and similarity calculated with preset or learned noise patterns, and highly similar factors are iteratively removed. This process can filter out pseudo-features formed by accidental associations or data noise, ensuring that the final set of retained semantic factors is pure, essential, and highly correlated with the causal relationship of information variation. This improves the anti-interference ability and interpretive credibility of the semantic feature set, allowing subsequent analysis to be based on core factors that better reflect the true semantic driving force.

[0017] Cross-semantic context information element evolution path tracing technology, after identifying key semantic transition points, maps them back to specific information element entities in the original drug data stream. It continuously tracks the changes in the state, attributes, or relationships of the same entity at different points in time and under different semantic contexts, thereby constructing a structured, dynamic evolution path spanning multiple semantic contexts. This achieves a complete depiction of the continuous changes within the lifecycle of an information entity. This changes the isolated and fragmented nature of information states in traditional analysis, providing a coherent perspective on the information evolution process. It enables semantic analysis to deeply understand the specific stages, connections, and progressive logic of changes, achieving a deeper understanding from result description to process analysis. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the working principle of the dual-channel drug information semantic management system for national drug negotiations as described in this invention. Figure 2 A flowchart illustrating the operation of the state observation unit; Figure 3 A flowchart illustrating the operation of the attribution analysis unit; Figure 4 A bar chart showing the average similarity of feature profiles of different types of drug information flows; Figure 5 A dual-axis composite graph showing the semantic template adaptation effect under different target application scenarios. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 This invention provides a dual-channel drug information semantic management system for national drug negotiations. The system includes: a state observation unit that continuously collects data waveforms of the semantic information stream, divides the waveform data into several data frames of fixed duration, extracts and quantizes multi-level semantic features for each data frame, and synthesizes the quantized features of all data frames into a feature profile, which corresponds to the complete evolution process of the drug information stream. An association measurement unit performs multi-scale alignment on the feature profile, constructs semantic association matrices at different time scales, identifies key semantic transition points by calculating the difference between matrices, and generates a trajectory map of information variation. An attribution analysis unit analyzes the trajectory map of information variation, maps the transition points to specific information elements in the original drug data stream, and tracks the evolution path of specific information elements in different semantic contexts, forming a structured attribution chain. A feature purification unit, based on the structured attribution chain, generates a set of semantic factors causally related to information variation, and through feedback iterative calculation, removes semantic factors similar to noise patterns from the set of semantic factors, outputting a purified core semantic factor set. The semantic integration unit inputs the core semantic factor set into the sequence generation model. Based on the evolution order recorded in the structured attribution chain, the model performs temporal reconstruction and context filling of the core semantic factors and outputs standardized drug information semantic records.

[0021] In one embodiment of the present invention, see [reference] Figure 2The state observation unit continuously receives raw data from the drug information stream via sensors or interfaces. This raw data stream originates from real-time drug price updates or medical insurance catalog change notifications pushed by the hospital information system. The raw data stream is converted into analyzable waveform signals, which represent the frequency and intensity changes of drug information elements in a time-series format. The waveform signals are uniformly segmented along the time axis to form a series of non-overlapping data frames with fixed durations. The fixed duration of each data frame is set to milliseconds to capture rapidly changing information segments. For each data frame, multi-level semantic feature extraction is performed in parallel. The extracted semantic features include at least word frequency distribution features, syntactic structure features, and contextual association features. Word frequency distribution features are obtained by statistically analyzing drug names and indications. The frequency of occurrence of key terms such as dosage form is obtained. Syntactic structure features are obtained by parsing the node depth and number of branches in the sentence dependency tree. Contextual association features are obtained by calculating the co-occurrence probability of specific words within a sliding window. The extracted multi-level semantic features are converted into numerical vectors to complete the quantization process. Word frequency distribution features are converted into normalized word frequency vectors, syntactic structure features are converted into structure encoding vectors, and contextual association features are converted into association strength vectors. The quantized feature vectors of all data frames are concatenated and integrated in chronological order to construct a feature profile representing the complete information evolution cycle. The feature profile is a two-dimensional matrix, where the rows of the matrix correspond to the data frames arranged in chronological order, and the columns of the matrix correspond to the concatenation dimension of all semantic feature quantized vectors.

[0022] In some embodiments, the association measurement unit receives a feature profile and presets multiple different observation scales for it. Each observation scale corresponds to a time window length. The observation scales include short scale, medium scale, and long scale. The short-scale time window length covers ten data frames, the medium-scale time window length covers one hundred data frames, and the long-scale time window length covers one thousand data frames. Under each preset observation scale, the feature profile is divided into continuous segments. Each segment contains feature vectors from multiple consecutive data frames. The similarity between feature vectors of adjacent segments is calculated using the cosine similarity method. All similarity values ​​are arranged to form a semantic association matrix under the observation scale. The semantic association matrix is ​​a symmetric matrix, and the matrix elements represent the feature vectors between different time segments. The similarity of eigenvectors is compared with multiple semantic association matrices generated at different observation scales. By calculating the absolute difference between corresponding elements in the matrix, a scale difference matrix is ​​generated. Each element in the scale difference matrix reflects the degree of difference in similarity assessment at the same time location at different scales. In the scale difference matrix, peak points with values ​​exceeding a preset threshold are identified. The preset threshold is a fixed value obtained through historical data statistics. Peak points are defined as key semantic transition points. All identified key semantic transition points are connected in chronological order of their occurrence, and their corresponding scale difference intensities are labeled to form a trajectory map of information variation. The trajectory map is presented in the form of a directed graph, where nodes represent key semantic transition points, edges represent the temporal relationship between transition points, and the weights on the edges are labeled with the scale difference intensities.

[0023] It is understandable that the following formula is used to calculate the similarity between the feature vectors of adjacent segments: in: This represents the similarity between the feature vectors of the i-th segment and the j-th segment. This represents the value of the feature vector of the i-th segment in the k-th dimension. Let represent the value of the feature vector of the j-th segment in the k-th dimension, and n represent the total number of dimensions of the feature vector.

[0024] Optionally, data comparison reflects the differences in semantic association matrices at different observation scales. Short-scale semantic association matrices show high-frequency fluctuations, reflecting rapid term updates in the drug information flow. Medium-scale semantic association matrices show abrupt changes in a stable trend, reflecting drug price adjustment events. Long-scale semantic association matrices show slow change patterns, reflecting cyclical updates of medical insurance policies. The scale difference matrix calculates the absolute difference between corresponding elements of the short-scale and long-scale semantic association matrices. Regions where the difference exceeds a threshold correspond to key semantic transition points. For example, at the moment of drug price adjustment, the similarity of the short-scale semantic association matrix drops sharply while the similarity of the long-scale semantic association matrix remains stable, resulting in a peak point in the scale difference matrix.

[0025] In some embodiments, when the feature profile is divided into continuous segments, the segment length is consistent with the time window length of the observation scale. The feature vectors within each segment are aggregated into a segment representative vector through average pooling. The similarity calculation of the representative vectors of adjacent segments adopts the above formula. In the process of constructing the semantic association matrix, the diagonal elements of the matrix are fixed to 1 to indicate that the segment is completely similar to itself, and the off-diagonal elements are filled according to the similarity calculation results. The scale difference matrix is ​​generated by element-wise subtraction. For example, the absolute difference matrix is ​​obtained by subtracting the long-scale semantic association matrix from the short-scale semantic association matrix.

[0026] Optionally, peak point identification employs a sliding window local maximum detection method. A window scan is applied to the scale difference matrix, and a peak point is marked when the value of the center element of the window is greater than the values ​​of all its neighbors and exceeds the global threshold. The timestamps of key semantic transition points are obtained by mapping the row and column indices of the peak points in the scale difference matrix back to the time axis of the original data frame. In the trajectory map of information variation, the node attributes record the peak point timestamps and scale difference intensity values, and the edge attributes record the time intervals between nodes and the rate of change of difference intensity.

[0027] In one embodiment of the present invention, see [reference] Figure 3The attribution analysis unit, based on the timestamps of key semantic transition points recorded in the trajectory map of information variation, traces back to the feature profile and the original drug information stream to locate the specific text unit or data field that triggered the key semantic transition point. The timestamp of the key semantic transition point corresponds to a specific time interval in the original data stream. Through index mapping, it searches for dimensions in the semantic feature vector that show significant mutations within that time interval. Based on the predefined mapping relationship between feature dimensions and original information elements, it determines the specific text unit that triggered the mutation as the "payment standard" field or the specific data field as the "price value". Starting from the specific text unit, it performs a bidirectional scan along the time axis, forward and backward. The forward scan obtains historical data fragments before the key semantic transition point occurs, and the backward scan obtains new data fragments after the key semantic transition point occurs. It collects all contextual information blocks containing the specific text unit or those highly semantically related to it. The contextual information blocks are based on sentences or structured data records as basic units, and semantically... High relevance is determined by co-occurrence frequency exceeding a threshold or belonging to the same semantic cluster. All collected contextual information blocks are arranged chronologically, and the semantic role changes, attribute modification changes, and relational changes of specific text units in different information blocks are analyzed. Semantic role changes are analyzed as the specific text unit changes as a subject, object, or attributive in a sentence. Attribute modification changes are analyzed as the addition, deletion, and replacement of adjectives or adverbs preceding the specific text unit. Relational changes are analyzed as the establishment and dissolution of relations between the specific text unit and other entities. The semantic role changes, attribute modification changes, and relational changes of specific text units are recorded as a series of event nodes with causal or temporal connections. Event nodes include timestamps, change types, and descriptions of change content. All event nodes are linked together to form a structured attribution chain from the initial state to the transition state. The structured attribution chain is stored in the form of a linked list, with each node pointing to the next node and storing a pointer to the evidence of change.

[0028] In some embodiments, the feature purification unit receives a structured attribution chain and extracts all event nodes recorded as generating changes from the chain. It traverses each node of the structured attribution chain, reads the change type and change content description stored in the node, and abstracts the information change corresponding to each event node into an independent semantic factor. The information change abstraction process extracts the essential features of the change. All semantic factors constitute an initial semantic factor set, which is a set containing multiple semantic factor feature vectors. A noise feature library is established in parallel with the semantic factor set. The noise feature library stores common interference patterns obtained from historical data statistics. Common interference patterns include, but are not limited to, "replacement of synonyms", "unified date format", and "unit conversion". Each interference pattern is represented by a feature vector template. Each factor in the initial semantic factor set is compared one by one with all patterns in the noise feature library to calculate their morphological similarity. The morphological similarity calculation measures the closeness between the semantic factor feature vector and the noise pattern feature vector template in multidimensional space. If the similarity between a semantic factor and any noise pattern exceeds a set threshold, a stripping process is initiated. The common part of the noise pattern is subtracted from the feature expression of the semantic factor. The common part is calculated through vector projection. After multiple rounds of iterative comparison and stripping operations, the remaining semantic factors that are no longer highly similar to the noise pattern constitute the purified core semantic factor set. The core semantic factor set is used to represent the real and meaningful semantic changes in the drug information flow.

[0029] It is understandable that the following formula is used when calculating morphological similarity: This represents the morphological similarity between a semantic factor Q and a noise pattern R, where Q represents the feature vector of the semantic factor to be compared, and R represents the feature vector template of a noise pattern in the noise feature library. Let represent the projection vector of vector Q onto the direction of vector R. Let represent the Euclidean norm of a vector. When When the value exceeds the set threshold, it is judged as highly similar.

[0030] Optionally, the data comparison reflects the differences before and after semantic factor purification. The initial semantic factor set includes factors generated by the change of "price unit from 'yuan' to 'yuan / bottle'". After comparing this factor with the "unit conversion" pattern in the noise feature library, the morphological similarity is extremely high. The stripping process is triggered to remove this common part from the factor features. After removal, the factor may be filtered out due to its low information content and thus not enter the core semantic factor set. However, the semantic factor generated by "the restriction on the scope of use changed from 'hospitals above level two' to 'designated medical institutions'" has low similarity to any noise pattern and is thus retained in the core semantic factor set.

[0031] In some embodiments, the semantic integration unit is configured with an attention-based sequence generation model. The sequence generation model employs an encoder-decoder architecture. The encoder processes the content input, and the decoder combines the structural input to generate the sequence. The core semantic factor set is used as the content input of the sequence generation model, transforming each factor in the core semantic factor set into an embedding vector sequence. Simultaneously, the evolution order of event nodes recorded in the structured attribution chain is used as the structural input of the sequence generation model. The structural input is a position-encoded sequence representing the chronological order of events. The sequence generation model first determines the approximate temporal position of the core semantic factors in the output sequence based on the structural input. The temporal position is then processed by the decoder's position feedforward network. The location encoding is mapped to the bias of the factor output. Based on the content input and location information, the model generates a contextual description fragment that conforms to the drug information domain specification for each core semantic factor. The generation process uses an attention mechanism to focus on the historical context related to the current factor. The output description fragment follows the domain expression template of "entity-attribute-value" or "event-condition-result". The model combines all generated description fragments according to a determined temporal position and ensures the logical coherence between fragments. Logical coherence is achieved by checking the consistency of the subject and tense of the fragments before and after through a cross-attention mechanism. The output is a complete and semantically standardized drug information semantic record, which is presented in a structured document or JSON format.

[0032] In one embodiment of the present invention, the map update unit continuously monitors newly generated drug information streams. These newly generated drug information streams originate from dynamic update documents issued by medical insurance agencies or real-time transaction logs of hospital pharmacy management systems. A new feature profile is generated using a state observation unit. This new feature profile follows the same construction method as historical feature profiles, contains semantic feature vectors of the same dimension, and is arranged chronologically. The new feature profile is compared with historical information variation trajectory maps. This comparison process involves inputting the new feature profile into an association metric unit to generate temporary semantic association matrices and scale difference matrices. The temporary matrices are then compared with the matrix data upon which the historical information variation trajectory maps are based, point-by-point, to identify newly emerging semantic transition patterns not recorded in the map. These newly emerging semantic transition patterns manifest as... If the historical trajectory map of information variation lacks peak points corresponding to specific time points or intensity features, or if the semantic feature dimension combinations associated with peak points are unprecedented, the identified new semantic transition patterns are submitted to the attribution analysis unit to generate supplementary structured attribution chains. These supplementary structured attribution chains are independent of existing chains and specifically describe the evolution path of information elements corresponding to the new semantic transition patterns. The supplementary structured attribution chains are used to expand the nodes and paths of the existing trajectory map of information variation. Node expansion involves adding nodes representing new semantic transition patterns to the trajectory map of information variation and labeling their scale difference intensity. Path expansion involves establishing directed connections between new nodes and existing upstream and downstream nodes to form an updated trajectory map of information variation. The updated trajectory map of information variation contains a more comprehensive history of semantic change.

[0033] In some embodiments, the verification closed-loop unit obtains the drug information semantic record output by the semantic integration unit. The drug information semantic record describes the complete information evolution of the drug in a standardized format and compares it with an authoritative drug information database in the real world. The authoritative drug information database is the National Medical Insurance Drug Catalog Database or the official drug instruction database. The comparison process includes field-level precise matching and semantic-level similarity calculation. Field-level precise matching compares whether fields such as "drug name", "dosage form", and "specification" are completely consistent. Semantic-level similarity calculation compares the consistency of meaning of text fields such as "indication description" and "payment restrictions". The differences found in the comparison are back-mapped to the structured attribution chain and the core semantic factor set. The back-mapping is achieved by tracing the core semantic factors on which the drug information semantic record depends and their source nodes in the structured attribution chain. The source node or factor that causes the difference is located. The source node is an event node in the structured attribution chain where a certain descriptive information changes. The source factor is a feature vector in the core semantic factor set that is determined to cause the final description deviation. Based on the nature of the discrepancies, calibration instructions are generated. The nature of the discrepancies includes numerical bias, incomplete semantic coverage, or incorrect relation description. The calibration instructions are used to adjust the scale difference threshold of the correlation metric unit or the noise stripping intensity of the feature cleanup unit. For example, for discrepancies with incomplete semantic coverage, the calibration instructions instruct to increase the scale difference threshold of the correlation metric unit to capture more subtle transitions. For discrepancies with incorrect relation descriptions, the calibration instructions instruct to enhance the noise stripping intensity of the feature cleanup unit to filter misleading correlations. The calibration instructions are fed back to the corresponding processing units of the system to drive the adaptive optimization of system parameters. The adaptive optimization of system parameters is an incremental learning process that fine-tunes internal thresholds and coefficients without interrupting system operation.

[0034] It is understandable that when identifying newly emerging semantic transition patterns, it is necessary to calculate the difference between the pattern derived from the new feature profile and all existing patterns in the trajectory map of historical information variations. If the difference is below a certain threshold, it is considered a recorded pattern; otherwise, it is a new pattern. The difference is calculated using the following formula: Where: Φ represents the minimum difference between the new semantic transition pattern and existing patterns in the historical graph. It is a matrix representing a pattern feature matrix containing multi-scale semantic association features extracted by the new feature profile within the time window of the new transition point. The pattern feature matrix of the k-th existing semantic transition pattern stored in the trajectory map representing the historical information variation. This represents the Frobenius norm of the matrix. When Φ is greater than a preset difference threshold... When this occurs, the pattern is determined to be a newly emerging semantic transition pattern.

[0035] Optionally, data comparison is reflected in the differences in system output before and after the map update and before and after verification calibration. After monitoring the information flow of the new policy interpretation on the "reimbursement amount for outpatient special diseases" of a certain drug, the new feature profile showed a significant peak at a specific time point. Comparison with the trajectory map of historical information variation revealed that the peak pattern and the existing "adjustment of inpatient reimbursement ratio" pattern had a huge difference in the Frobenius norm Φ value of the feature matrix, exceeding the threshold. Therefore, it was identified as a new semantic transition pattern, which in turn generated a supplementary structured attribution chain describing the evolution of "reimbursement scenarios expanding from inpatient to outpatient special diseases". The trajectory map of information variation was subsequently updated with corresponding nodes and paths. In the verification stage, the "reimbursement amount calculation method" field in the semantic record output by the system differed from the description in the authoritative database. The reverse mapping located a factor about "calculation cardinality" in the core semantic factor set. This factor was over-stripped during the feature purification process because it was similar to the noise pattern of "annual total amount calculation", resulting in missing information. Based on this, a calibration instruction was generated to appropriately reduce the noise stripping intensity of the feature purification unit on the features related to "calculation cardinality". In subsequent processing, the integrity of similar information was improved.

[0036] In some embodiments, when a calibration command is fed back to the correlation metric unit, the correlation metric unit receives the command and linearly adjusts its internal scale difference threshold parameter. The adjustment magnitude of the scale difference threshold parameter is proportional to the deviation severity coefficient carried in the calibration command. After receiving the calibration command, the feature cleanup unit dynamically corrects its internal noise stripping intensity coefficient. The correction of the noise stripping intensity coefficient is based on a feedback controller, whose input is the proportion of historical difference points attributed to improper noise stripping. It can be understood that the adaptive optimization process of system parameters records all calibration commands and their effects, forming a parameter adjustment experience base, which is used to provide initial adjustment suggestions when facing similar difference characteristics.

[0037] Optionally, the comparison between the verification closed-loop unit and the authoritative drug information database adopts a combination of regular batch operations and real-time triggering. Regular batch operations perform a full comparison of records at a fixed time every day, while real-time triggering starts the comparison immediately when the semantic integration unit outputs semantic records of high-risk or new drug categories. The reverse mapping of differences depends on the traceability metadata attached to each descriptive fragment by the semantic integration unit when generating records. The traceability metadata contains the unique identifier of the core semantic factors used and their node numbers in the structured attribution chain, thereby ensuring that the source can be accurately located.

[0038] In one embodiment of the present invention, the caching acceleration unit establishes a feature profile cache area between the state observation unit and the association measurement unit. The feature profile cache area is implemented using a key-value database or an in-memory hash table and is used to temporarily store feature profile data generated within a recent period. The definition of the recent period is determined by system configuration parameters, for example, it is set to cache all feature profiles processed within the last 24 hours. When the system needs to process a new drug information stream, the caching acceleration unit first searches in the feature profile cache area for historical feature profiles that are highly similar to the current information stream. The search process is completed by calculating the similarity between the newly generated feature profile and each historical feature profile in the cache. If a highly similar historical feature profile exists, the pre-calculated trajectory map of information variation corresponding to the historical feature profile and the subsequent processing results are directly called. These pre-calculated results are stored in the association data field of the feature profile cache area as intermediate data of the association map and the initial output of the attribution analysis unit, serving as the initial reference for the current processing. The system then performs incremental calculations only on the differences, which refer to data segments where the new feature profile is inconsistent with the matched historical feature profile in a specific time interval or feature dimension. The scope of incremental calculations is limited to the data frames corresponding to these differences and the processing units on the subsequent impact link.

[0039] In some embodiments, the process of finding highly similar historical feature profiles is based on the structured feature vectors of the feature profiles. The feature profiles are represented as time-feature matrices, and similarity is determined by calculating normalized similarity metrics between matrices. Similarity determination relies on a preset similarity threshold, which is set based on extensive experience processing pharmaceutical information flow data. The index for fast retrieval is constructed based on the fingerprints of the feature profiles, generated by dimensionality reduction and hashing of the feature profile matrix. Each entry in the feature profile cache not only stores the feature profile data itself but also stores the associated pre-calculated trajectory map of information variation, intermediate states of the structured attribution chain, and intermediate results of the core semantic factor set. These associated data are calculated by subsequent processing units and backfilled into the cache entries when historical feature profiles are generated. Refer to Table 1 for the structure of the cache entries.

[0040] Table 1: Feature Profile Buffer Entry Structure Table

[0041] If highly similar historical feature profiles exist, the pre-calculated trajectory map of information variation corresponding to the historical feature profile and the subsequent processing results are directly invoked as the initial reference for the current processing, and incremental calculation is performed only on the difference parts. After incremental calculation is initiated, the system compares the new feature profile with the matched historical feature profile frame by frame or segment by segment, identifying data frame sequences whose feature vector differences exceed the tolerance. These data frame sequences are marked as difference data blocks. The association measurement unit receives the difference data blocks and the pre-calculated semantic association matrix related to similar historical feature profiles retrieved from the cache. Based on the pre-calculated matrix, the semantic association matrix is ​​recalculated only for the time scale interval covered by the difference data blocks, and the scale difference matrix is ​​updated, thereby generating trajectory map fragments of local information variation for the difference parts. The attribution parsing unit, feature purification unit, and semantic integration unit then perform incremental fusion and update based on the local trajectory map fragments and the relevant context chains pre-stored in the cache, finally synthesizing a new complete processing result.

[0042] It is understandable that the following formula is used when calculating the similarity between a new feature profile and a certain historical feature profile in the cache: Where: Ψ represents the overall similarity score between two feature profiles, and T represents the total number of time frames in the feature profile. This represents the aggregated feature vector of the new feature profile at the t-th time frame. This represents the aggregated feature vector of a historical feature profile in the cache at time frame t. The cosine similarity between the two vectors is calculated using the cosine similarity function Ψ. When the value of Ψ is greater than the preset similarity threshold Ω, it is considered highly similar. Optionally, data comparison reflects the difference in processing flow before and after enabling caching acceleration. For the daily batch processing of highly repetitive drug price catalog synchronization information streams, the feature profile generated by the system on the next day has a very high Ψ similarity value with the feature profile cached on the previous day, exceeding the threshold Ω. The system directly calls the pre-calculated trajectory map of information variation and the associated attribution chain in the cache, and detects that only a few drugs have changed in the "manufacturer" field. Then, it triggers incremental calculation only for a limited number of data frames containing these changed fields, updates the local map and attribution chain, and finally generates a new semantic record. This avoids the complete calculation of the entire process from state observation to semantic integration, significantly reducing the consumption of computing resources and processing latency. When processing a brand-new emergency information stream about monitoring serious adverse drug reactions, if the maximum Ψ similarity value between the generated feature profile and all historical feature profiles in the cache is lower than the threshold Ω, the cache acceleration unit cannot hit any valid entries, and the system will automatically switch to the standard, complete processing flow.

[0043] In some embodiments, the feature profile cache uses a least recently used strategy or a reference counting-based strategy to manage cached entries. When the cache space reaches its limit, entries with the earliest last access time or the lowest reference count are evicted first. The associated data of cached entries (such as associated trajectory maps) is not always a complete pre-calculated result. For partial matches or cases where similarity is at a critical value, the associated data may be used as the initialization state or prior knowledge input for subsequent processing units to accelerate convergence rather than completely replace computation. During incremental computation, the identification accuracy of difference data blocks depends on the setting of the difference tolerance parameter. This parameter is dynamically adjusted according to the information flow type. A lower difference tolerance is used for policy-related information flows to ensure that any text revisions are captured, while a higher difference tolerance can be used for price flow-related information flows to ignore unimportant numerical fluctuations.

[0044] It is understandable that the generation process of feature profile fingerprints is a balance between computational efficiency and comparison accuracy. The fingerprint algorithm needs to ensure that highly similar feature profiles produce the same or similar fingerprint values, while different feature profiles produce significantly different fingerprint values. After processing a new information stream, the caching acceleration unit writes the newly generated feature profile and its final processing result as a new entry or an update to an existing entry into the feature profile cache area, thereby realizing the continuous evolution and updating of cached data.

[0045] See Figure 4 This is a bar chart showing the average similarity of feature profiles for different types of drug information flows. The chart displays the average similarity of feature profiles for five types of drug information flows, and uses error bars to represent the fluctuation range of similarity, while also indicating a similarity threshold of 0.8 (red dashed line). Adverse reaction monitoring has the longest error bar, indicating poor stability of its information flow features; price catalog synchronization has the shortest error bar, indicating high feature consistency. This chart provides a basis for resource allocation and strategy optimization for the caching acceleration unit. Priority should be given to expanding the cache for high-similarity information flows such as price catalog synchronization to improve acceleration efficiency; for low-similarity information flows such as adverse reaction monitoring, no cache resources need to be allocated, and the entire process can be directly executed; for information flows with critical similarity such as policy document interpretation, the similarity threshold can be appropriately lowered to expand the cache coverage.

[0046] In one embodiment of the present invention, the interface adaptation unit defines a set of standardized data exchange protocols and semantic description templates. The standardized data exchange protocols specify the data transmission format, field structure, encoding method, and verification rules, while the semantic description templates define the organizational structure, terminology dictionary, and expression specifications of the output information. When the system receives external drug information streams, these streams may originate from HL7 messages from hospital information systems, XML files from medical insurance platforms, or JSON data streams from drug traceability systems. Protocol parsing converts raw data from different sources and in different formats into a unified waveform signal within the system. The protocol parsing process sequentially performs format decoding, structure mapping, and semantic normalization. Format decoding identifies and applies the corresponding parser based on the data source. Structure mapping aligns the fields in heterogeneous data with the system's internal data model. Semantic normalization converts the encoding of terms from different sources into a unified encoding system within the system. The converted waveform signal is a numerical sequence aligned with timestamps and feature values, which can be directly processed by the state observation unit. When the system outputs semantic records of drug information, it calls the corresponding semantic description template according to the needs of the target application scenario, which includes medical insurance settlement system, clinical decision support system or drug regulatory data platform. The corresponding semantic description template predefines data structure, field naming and value range constraints that conform to the interface specification of the target system. It converts the internal unified record format into output data that conforms to the specific interface specification. The conversion process includes field reorganization, value format conversion and semantic encoding mapping. Field reorganization reorganizes the order and nesting relationship of information items according to the target template. Value format conversion converts the internal numerical values ​​or codes into strings or enumeration values ​​required by the target interface. Semantic encoding mapping converts the internal concept encoding of the system into the standard encoding agreed by the target system.

[0047] In some embodiments, the standardized data exchange protocol includes a core protocol and multiple extended protocols. The core protocol defines the common fields and basic structure that all drug information flows must possess, while the extended protocols define supplementary fields and parsing rules for specific data source types. During protocol parsing, format decoding first detects the MIME type or file header identifier of the input data to determine the specific protocol to be used. Structure mapping maps the nodes or key-value pairs of the input document to attributes of the internal data model according to the protocol definition. The mapping relationship is implemented through a configurable mapping table. Semantic normalization relies on a central terminology service, which maintains the mapping relationship between the internal standard codes of core concepts such as drug names, dosage forms, and manufacturers and various external codes. When raw data is converted into an internally unified waveform signal, the waveform signal generation follows a fixed algorithm. Each time point, the mapped and normalized internal data model instance is converted into a fixed-dimensional numerical vector through an embedding layer. This vector serves as the amplitude value of the waveform signal at that time point. The vectors from all time points are arranged sequentially to form the waveform signal.

[0048] It is understandable that the accuracy of structural mapping in the protocol parsing process depends on the completeness and precision of the mapping table. The construction and maintenance of the mapping table involves a learning process that analyzes successful parsing cases from historical data streams. A key step in structural mapping is determining the semantic equivalence between the source data and the internal model, which can be achieved by calculating the semantic similarity between field names and description text. The formula used to evaluate whether two fields are mappable is as follows: Where: Λ(A,B) represents the overall mapping confidence between source data field A and internal model field B. This represents the semantic similarity score between the name strings of field A and the name strings of field B. This represents the semantic similarity score between the description text of field A (if it exists) and the description text of field B. α and β are weighting coefficients, satisfying α + β = 1 and α, β > 0. When Λ(A,B) exceeds a preset mapping threshold, a mapping relationship from field A to field B is established. See also... Figure 5 This is a dual-axis composite graph showing the semantic template adaptation effect under different target application scenarios. This graph provides a basis for template optimization in the interface adaptation unit. Priority should be given to maintaining templates for high-frequency scenarios such as "medical insurance settlement" and "drug supervision" to ensure high accuracy. For the "enterprise reporting" scenario, the flexibility of field reorganization can be increased to improve its adaptation accuracy. The positive correlation between the proportion of field reorganization and accuracy indicates that appropriate field reorganization helps adapt to the information structure of the target scenario. The positive correlation between the proportion of field reorganization and adaptation accuracy verifies the business logic that "appropriate field reorganization can improve template adaptation effect," providing a quantitative basis for the design of templates for subsequent new scenarios.

[0049] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0050] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A semantic management system for dual-channel drug information under national drug price negotiations, characterized in that: include: The state observation unit continuously collects the data waveform of the semantic information stream, divides the waveform data into several data frames of fixed duration, extracts and quantizes multi-level semantic features for each data frame, and synthesizes the quantized features of all data frames into a feature profile, which corresponds to the complete evolution process of the drug information stream. The association measurement unit performs multi-scale alignment on the feature profile, constructs semantic association matrices at different time scales, identifies key semantic transition points by calculating the difference between matrices, and generates a trajectory map of information variation. The attribution analysis unit analyzes the trajectory map of information variation, maps the transition points therein to specific information elements in the original drug data stream, and tracks the evolution path of the specific information elements in different semantic contexts to form a structured attribution chain. The feature purification unit generates a set of semantic factors that are causally related to information variation based on the structured attribution chain, and removes semantic factors that are similar to noise patterns from the set of semantic factors through feedback iterative calculation, and outputs the purified core semantic factor set. The semantic integration unit inputs the core semantic factor set into the sequence generation model. The model performs temporal reconstruction and context filling of the core semantic factors based on the evolution order recorded in the structured attribution chain, and outputs standardized drug information semantic records.

2. The national negotiated dual-channel drug information semantic management system according to claim 1, characterized in that, The state observation unit includes the following: The system continuously receives raw data from the drug information stream via sensors or interfaces and converts the raw data stream into analyzable waveform signals. The waveform signal is uniformly divided along the time axis to form a series of non-overlapping data frames with fixed durations. For each data frame, multi-level semantic feature extraction operations are performed in parallel. The extracted semantic features include at least word frequency distribution features, syntactic structure features, and contextual association features. The extracted multi-level semantic features are converted into numerical vectors to complete the quantization process; By concatenating and integrating the quantized feature vectors of all data frames in chronological order, a feature profile representing the complete information evolution cycle is constructed.

3. The national negotiated dual-channel drug information semantic management system according to claim 2, characterized in that, The correlation metric unit includes the following: The feature profile is received and multiple different observation scales are preset for it, each observation scale corresponding to a specific time window length; At each preset observation scale, the feature profile is divided into continuous segments, the similarity between the feature vectors of adjacent segments is calculated, and all similarity values ​​are arranged to form a semantic association matrix at the observation scale. By comparing multiple semantic association matrices generated at different observation scales, a scale difference matrix is ​​generated by calculating the absolute difference between elements at corresponding positions in the matrices. In the scale difference matrix, peak points whose values ​​exceed a preset threshold are identified, and these peak points are defined as key semantic transition points. Connect all the identified key semantic transition points in chronological order of their occurrence, and label the corresponding scale difference intensity to form a trajectory map of information variation.

4. The national negotiated dual-channel drug information semantic management system according to claim 3, characterized in that, The attribution analysis unit includes the following: Based on the timestamps of key semantic transition points recorded in the trajectory map of the information variation, we trace back to the feature profile and the original drug information stream to locate the specific text unit or data field that triggered the key semantic transition point. Starting from the specific text unit, a bidirectional scan is performed along the time axis, forward and backward, to collect all contextual information blocks that contain the specific text unit or are highly semantically related to it. All collected contextual information blocks are arranged chronologically, and the semantic role, attribute modification, and association changes of the specific text unit in different information blocks are analyzed. The semantic role changes, attribute modification changes, and association changes of the specific text units are recorded as a series of event nodes with causal or temporal relationships, and all event nodes are linked together to form a structured attribution chain from the initial state to the transition state.

5. The national negotiated dual-channel drug information semantic management system according to claim 4, characterized in that, The feature purification unit includes the following: Receive the structured attribution chain and extract all event nodes recorded as causing changes from the chain; The information change corresponding to each event node is abstracted into an independent semantic factor, and all semantic factors constitute the initial semantic factor set. Establish a noise feature library that runs parallel to the semantic factor set. The noise feature library stores common interference patterns derived from historical data. Each factor in the initial semantic factor set is compared one by one with all patterns in the noise feature library to calculate their morphological similarity; If the similarity between a semantic factor and any noise pattern exceeds a set threshold, a stripping procedure is initiated to subtract the common part of the noise pattern from the feature expression of the semantic factor. After multiple rounds of iterative comparison and stripping operations, the remaining semantic factors that are no longer highly similar to the noise patterns constitute the purified core semantic factor set.

6. The national negotiated dual-channel drug information semantic management system according to claim 5, characterized in that, The semantic integration unit includes the following: Configure an attention-based sequence generation model; use the core semantic factor set as the content input of the sequence generation model, and use the evolution order of event nodes recorded in the structured attribution chain as the structural input of the sequence generation model; The sequence generation model first determines the approximate temporal position of the core semantic factors in the output sequence based on the structural input; Based on the content input and location information, the model generates contextual description fragments that conform to the specifications of the pharmaceutical information domain for each core semantic factor; The model combines all generated description fragments according to their defined temporal positions and ensures logical coherence between the fragments, outputting a complete and semantically standardized drug information semantic record.

7. The national negotiated dual-channel drug information semantic management system according to claim 6, characterized in that, The system also includes a map update unit, which comprises the following: Continuously monitor newly generated drug information flows and use the state observation unit to generate new feature profiles; The new feature profile is compared with the trajectory map of historical information variation to identify newly emerging semantic transition patterns that are not recorded in the map; the identified new semantic transition patterns are submitted to the attribution analysis unit to generate supplementary structured attribution chains. By using supplementary structured attribution chains, the existing trajectory map of information variation is expanded in terms of nodes and paths, resulting in an updated trajectory map of information variation.

8. The national negotiated dual-channel drug information semantic management system according to claim 7, characterized in that, The system also includes a verification closed-loop unit, which comprises the following: Obtain the drug information semantic record output by the semantic integration unit and compare it with an authoritative drug information database in the real world; The differences found in the comparison are mapped back to the structured attribution chain and the core semantic factor set to locate the source node or factor that caused the difference. Based on the nature of the difference, a calibration instruction is generated, which is used to adjust the scale difference threshold of the associated metric unit or the noise stripping intensity of the feature cleanup unit. The calibration command is fed back to the corresponding processing unit of the system, driving the system parameters to perform adaptive optimization.

9. The national negotiated dual-channel drug information semantic management system according to claim 8, characterized in that, The system also includes a cache acceleration unit, which comprises the following: A feature profile buffer is established between the state observation unit and the associated measurement unit to temporarily store feature profile data generated in the most recent period. When the system needs to process a new drug information flow, it first checks the feature profile cache to see if there is a historical feature profile that is highly similar to the current information flow. If it exists, the trajectory map of pre-calculated information variation corresponding to the historical feature profile and the subsequent processing results are directly called as the initial reference for the current processing, and only the difference part is incrementally calculated.

10. The national negotiated dual-channel drug information semantic management system according to claim 9, characterized in that, The system also includes an interface adaptation unit, which comprises the following: Define a set of standardized data exchange protocols and semantic description templates; when the system receives external drug information streams, it converts raw data from different sources and in different formats into a unified waveform signal within the system through protocol parsing; When the system outputs semantic records of drug information, it calls the corresponding semantic description template according to the needs of the target application scenario, and converts the internally unified record format into output data that conforms to specific interface specifications.

Images