Large language model-based shift change information intelligent completion and error early warning system

Abstract

The invention discloses a shift change information intelligent complementation and error early warning system based on a large language model, which belongs to the technical field of medical information processing and comprises a terminal acquisition module, an intelligent distribution complementation module and an early warning control module. Structure and semantic fingerprints are generated for the data table, a shift exchange asset meta model is constructed through fusion similarity clustering merging, and a semantic enhancement model is generated in combination with medical shift exchange specifications to complete information field completion; the intelligent distribution and completion module configures an optimal path with the minimum comprehensive overhead for the acquisition module by relying on a regional dynamic shift information network formed by double sub-networks; the early warning control module is combined with a large language model semantic analysis and rule engine, monitors an acquisition track and identifies information abnormity to execute error early warning, and an early warning result is fed back to the intelligent distribution complementation module, so that dynamic adjustment of an optimal path is realized, and the integrity, accuracy and acquisition efficiency of medical shift information are improved.

Description

Technical Field

[0001] This invention belongs to the field of medical information processing technology, specifically a handover information intelligent completion and error early warning system based on a large language model. Background Technology

[0002] Medical handover is a crucial step in clinical diagnosis and treatment. The completeness and accuracy of handover information directly affect the continuity of treatment and patient safety. Currently, in medical handover, medical staff mostly enter raw handover information through text, voice, and structured forms. This is prone to problems such as missing fields and non-standard terminology due to human error. Furthermore, handover data is stored in multiple tables in different databases, with differences in table structure and semantic expression, making it difficult to quickly integrate and retrieve information, which significantly reduces the efficiency of information collection and completion.

[0003] Existing methods for processing handover information lack intelligent path planning capabilities. Data collection often uses fixed paths without dynamic optimization based on the characteristics of the handover task and resource status. This results in problems such as high computational latency, unreasonable bandwidth usage, and low data acquisition success rate, making it unsuitable for the handover needs of different departments and different disease levels. At the same time, traditional handover information error verification relies heavily on manual methods, which can only detect explicit errors and cannot identify implicit anomalies such as logical contradictions, inconsistencies with medical knowledge graphs, or deviations from patient historical data trends. Furthermore, the verification process is disconnected from path adjustment, and when deviations in the collection trajectory or information anomalies occur, the collection path cannot be dynamically optimized in a timely manner, which can easily lead to errors in the transmission of medical information and increase medical risks.

[0004] Furthermore, existing systems only perform simple transcription of voice-based handover information, failing to standardize terminology or structure text for medical scenarios, further reducing the usability of unstructured handover information. Therefore, there is an urgent need to build an intelligent handover information processing system to address industry pain points such as inefficient information completion, rigid path planning, and delayed error warnings. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes an intelligent handover information completion and error early warning system based on a large language model. The system comprises three main modules: terminal acquisition, intelligent allocation and completion, and early warning control. The terminal acquisition module obtains raw handover information and data source metadata including database configuration and data tables. It generates structural and semantic fingerprints for the data tables, and constructs a handover asset meta-model through similarity clustering. Combined with medical handover standards, it generates a semantic enhancement model to complete information field completion. The intelligent allocation and completion module relies on a regional dynamic handover information network composed of two subnets to configure the optimal path with the lowest overall cost for the acquisition module. The early warning control module combines large language model semantic analysis and a rule engine to monitor the acquisition trajectory and identify information anomalies to execute error early warnings. The early warning results are fed back to the intelligent allocation and completion module, enabling dynamic adjustment of the optimal path and improving the completeness, accuracy, and acquisition efficiency of medical handover information.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The handover information intelligent completion and error early warning system based on a large language model includes: a terminal acquisition module, an intelligent allocation and completion module, and an early warning control module;

[0008] The terminal acquisition module collects raw shift handover information and shift handover data source metadata; the shift handover data source metadata includes database connection configuration and shift handover data table.

[0009] For each handover data table, generate structural and semantic fingerprints, calculate fusion similarity and cluster and merge under an adaptive threshold to construct a handover asset meta-model, perform semantic parsing on the handover asset meta-model, generate a semantic enhancement model in combination with medical handover specifications, and complete the fields of the original handover information to obtain the completed handover information.

[0010] The intelligent allocation and completion module combines the completed handover information with the regional dynamic handover information network to configure the optimal path for the terminal acquisition module. The regional dynamic handover information network consists of two subnets connected by the historical optimal path. The connection relationship is configured with a hash fingerprint security control key to adjust the acquisition and completion trajectory in case of anomalies.

[0011] The early warning control module, based on the optimal path and the completed handover information, combines semantic analysis of the large language model and the rule engine to execute error warnings. The warning results are fed back to the intelligent allocation and completion module to dynamically adjust the optimal path.

[0012] Specifically, the terminal acquisition module is further configured as follows:

[0013] Obtain text, voice, or pre-filled structured forms entered by medical staff through terminals as raw handover information;

[0014] Obtain handover data source metadata from the pre-configured data source connection pool; the handover data source metadata includes the database connection address, access credentials, data mode, and one or more handover data tables; the handover data tables store patient diagnosis and treatment records, medical orders, test results, and vital signs data.

[0015] Specifically, the generation of structural and semantic fingerprints for each shift handover data table includes:

[0016] For each shift handover data table, perform structure parsing to extract structure information. Concatenate the extracted structure information into a string according to the field order, and use a first preset hash algorithm to calculate the string to generate the structure fingerprint of the shift handover data table. The structure information includes field names, field data types, primary key and foreign key relationships, and inter-table association mappings.

[0017] Semantic parsing is performed on each shift handover data table. The textual meaning of the field names is parsed based on natural language processing technology. Synonyms and hypernyms are obtained by combining medical knowledge graphs. The textual meaning is transformed into a vector representation and a semantic fingerprint is generated through a second preset hash algorithm.

[0018] Specifically, the calculation of fusion similarity and clustering and merging under an adaptive threshold to construct a handover asset meta-model includes:

[0019] For any two handover data tables, a structural similarity score is calculated based on their structural fingerprints, and a semantic similarity score is calculated based on their semantic fingerprints. The structural similarity scores and semantic similarity scores are then weighted and fused using preset weight coefficients to obtain the fused similarity.

[0020] The similarity between each pair of all handover data tables is used to construct a similarity matrix. The silhouette coefficient method is used to perform cluster analysis on the similarity matrix, and the adaptive clustering threshold that optimizes the clustering effect is dynamically determined.

[0021] All shift handover data tables with a similarity greater than the adaptive clustering threshold are grouped into the same cluster, and all shift handover data tables within the same cluster are merged to generate a unified shift handover asset meta-model. The shift handover asset meta-model integrates the common fields, data types, and semantic descriptions of all shift handover data tables within the cluster.

[0022] Specifically, semantic parsing is performed on the aforementioned handover asset meta-model, and a semantically enhanced model is generated by combining it with medical handover standards, including:

[0023] The handover asset meta-model is analyzed by calling a pre-trained large language model in the medical field to identify the core medical entity types and relationships between entities represented by the handover asset meta-model.

[0024] Query the standard handover field definitions, mandatory field rules, and logical dependencies between fields related to the core medical entity type from the pre-built medical handover standard knowledge base;

[0025] Based on the parsing results of the large language model and the query results of the medical handover standard knowledge base, a semantic tag set and a set of field mapping rules are generated for each handover asset meta-model; the semantic tag set, field mapping rules and the handover asset meta-model together constitute a semantic enhancement model.

[0026] Specifically, the step of completing the fields of the original shift handover information to obtain the completed shift handover information includes:

[0027] The semantic enhancement model is used to parse the original handover information to identify the filled fields and their corresponding medical entity identifiers;

[0028] For the identified medical entity identifier, the list of associated fields that need to be completed for the entity is determined according to the field mapping rules in the semantic enhancement model;

[0029] Based on the handover data source metadata obtained by the terminal acquisition module, the system connects to the corresponding database and uses a multi-table join SQL query statement based on the medical entity identifier to retrieve the specific data values ​​corresponding to the list of associated fields from the handover data table.

[0030] The retrieved data values ​​are filled into the corresponding missing fields of the original shift handover information to form the completed shift handover information.

[0031] Specifically, the regional dynamic shift handover information network includes a task logic subnet and a physical resource subnet; the nodes in the task logic subnet represent historical shift handover tasks of different departments or wards, and the nodes in the physical resource subnet represent data source servers, network exchange nodes, and database storage resources; the nodes in the task logic subnet and the nodes in the physical resource subnet are connected by directed edges, each directed edge representing a historical data collection path and associated with a hash fingerprint security control key used to identify the current security status of the path.

[0032] Specifically, configuring the optimal path for the terminal acquisition module includes:

[0033] The completed handover information is analyzed to extract the features of the current handover task, including the department, the severity of the patients involved, the data type required, and the estimated data volume.

[0034] The characteristics of the current shift handover task are matched with the task logic subnet in the regional dynamic shift handover information network to determine the current task node;

[0035] The overall cost is evaluated from the current task node, through the directed edges of each historical data acquisition path, to the target resource node in the physical resource subnet; the overall cost is calculated based on the weighted sum of the historical calculation delay of the path, the network bandwidth utilization rate, and the data acquisition success rate.

[0036] The historical data acquisition path with the lowest overall cost is selected as the optimal path for configuring the terminal acquisition module.

[0037] Specifically, the early warning control module, based on the optimal path and the completed shift handover information, combines semantic analysis from a large language model with a rule engine to execute error warnings, including:

[0038] The actual trajectory of the data collected by the terminal acquisition module along the optimal path is monitored, and the path deviation between the actual trajectory and the preset trajectory of the optimal path is calculated.

[0039] The large language model is invoked to perform semantic analysis on the completed handover information to identify abnormal expressions in the information; the abnormal expressions include logical contradictions, expressions that violate the established medical knowledge graph association relationships, or expressions that differ from the trends of patients' historical data;

[0040] The path deviation and the abnormal expression are input into the rule engine; the rule engine has pre-stored warning rules of different levels. When the input deviation exceeds the first threshold or a specified type of abnormal expression is identified, the corresponding level of warning information is triggered.

[0041] Specifically, the early warning result is fed back to the intelligent allocation and completion module to dynamically adjust the optimal path, including:

[0042] When the warning level triggered by the warning control module exceeds the preset second threshold, the intelligent allocation completion module receives a warning result containing an abnormal path identifier or an abnormal data field; the second threshold is greater than the first threshold;

[0043] The intelligent allocation completion module marks the corresponding hash fingerprint security control key status as invalid based on the abnormal path identifier, and temporarily removes the path from the current set of available paths;

[0044] The intelligent allocation and completion module recalculates the overall cost based on the updated set of optional paths, generates a new optimal path, and sends it to the terminal acquisition module.

[0045] Compared with the prior art, the beneficial effects of the present invention are:

[0046] 1. This system utilizes structural and semantic fingerprint generation and clustering to construct a handover asset meta-model, and combines it with medical handover standards to generate a semantic enhancement model to achieve intelligent completion of handover information. It can automatically retrieve and fill missing fields from scattered handover data tables, and at the same time, it completes the standardization and structuring of voice-based handover information, effectively solving problems such as missing fields and non-standard terminology that are prone to occur in manual input, greatly improving the completeness and standardization of handover information, and reducing the manual workload of medical staff.

[0047] 2. The regional dynamic handover information network built by the system integrates task logic and physical resource dual subnets. It can combine the characteristics of the current handover task and configure the optimal data acquisition path by calculating the comprehensive cost. It breaks through the limitations of the traditional fixed acquisition path. It can dynamically plan the path according to the needs of department, disease level, data type, etc., reduce computing latency and network bandwidth consumption, improve the data acquisition success rate, make the handover data collection more in line with the actual clinical needs, and improve the efficiency and rationality of information collection.

[0048] 3. The early warning and control module combines semantic analysis from a large language model with a rule engine. It can monitor the actual deviation of data collection trajectory and accurately identify implicit abnormal expressions such as logical contradictions and discrepancies with the medical knowledge graph in the handover information. This enables dual error warnings for both the handover information collection process and content. Compared with the shortcomings of traditional manual verification, which can only find explicit errors, this module improves the comprehensiveness and accuracy of error identification, reduces errors in the transmission of medical information from the source, and lowers medical safety risks.

[0049] 4. The system has constructed a closed-loop linkage mechanism between early warning results and path adjustment. When a high-level early warning is triggered, the failed path can be quickly marked and a new optimal path can be recalculated and generated. At the same time, the hash fingerprint security control key provides security for path adjustment, realizing dynamic adaptive optimization of the collected path. This mechanism enables the system to respond to collection anomalies and information errors in a timely manner, ensuring the continuity and stability of the handover information collection and completion process, and further improving the intelligence and reliability of the entire medical handover information processing flow. Attached Figure Description

[0050] Figure 1 This is an architecture diagram of the intelligent handover information completion and error early warning system based on a large language model according to the present invention;

[0051] Figure 2 This is a flowchart illustrating the construction process of the handover asset meta-model for the intelligent handover information completion and error early warning system based on a large language model, as described in this invention. Detailed Implementation

[0052] Example 1:

[0053] Please see Figure 1The present invention provides an embodiment of an intelligent handover information completion and error early warning system based on a large language model, comprising: a terminal acquisition module, an intelligent allocation completion module, and an early warning control module;

[0054] The terminal acquisition module collects raw shift handover information and shift handover data source metadata; the shift handover data source metadata includes database connection configuration and shift handover data table.

[0055] Furthermore, the specific process of collecting raw handover information and data source metadata through the terminal acquisition module includes: deploying dedicated handover acquisition terminals for medical staff in various departments and wards of the hospital, including computer terminals, tablet terminals, and handheld terminals for voice acquisition. All terminals are connected to the hospital's intranet dedicated medical data transmission channel, and the terminal operation permissions are bound to the medical staff's employee ID number. An identity verification mechanism ensures the security and traceability of data collection. Each acquisition terminal supports three methods for collecting raw handover information: manual text input, real-time voice input, and uploading of pre-filled structured forms. Medical staff can flexibly choose according to the handover work scenario. For example, voice acquisition is given priority during morning oral handover. The system allows users to upload pre-made forms during written night shift handovers. Upon receiving the original handover information, the terminal acquisition module automatically retrieves the corresponding department's handover data source metadata from the pre-configured data source connection pool in the hospital's data platform. This metadata is encrypted and accessible only to authorized terminals. The data source connection pool is categorized by department and ward to ensure accurate matching between the metadata and the collected original handover information. This ultimately completes the synchronous acquisition and temporary storage of the original handover information and the handover data source metadata. The hospital's intranet authentication mechanism uses a public-private key-based asymmetric encryption method, which is existing technology in this field and not a creative solution of this application; therefore, it will not be elaborated upon here.

[0056] Furthermore, this embodiment takes the typical medical scenario of pediatric nursing morning handover as an example. Pediatric nursing handover is characterized by young patients, rapid changes in their conditions, numerous details of diagnosis and treatment information, and frequent use of colloquial expressions. At the nurse station in the pediatric inpatient department, the on-duty nurse uses a handheld voice acquisition terminal to conduct morning handover. This terminal has a built-in noise-canceling microphone, which can clearly capture the handover voice stream in the complex environment of multiple nurses communicating at the nurse station. The terminal has completed nurse identity verification and automatically binds the data source access permissions of the pediatric ward. While collecting voice handover information, it retrieves the pediatric-specific handover data source metadata from the hospital's data platform, including the connection address, access credentials, and data mode of the pediatric patient diagnosis and treatment database, as well as five core handover data tables storing basic information of pediatric patients, execution of medical orders, laboratory examinations, vital signs, and nursing operations, thus preparing the data for subsequent voice transcription and information processing.

[0057] For each handover data table, generate structural and semantic fingerprints, calculate fusion similarity and cluster and merge under an adaptive threshold to construct a handover asset meta-model, perform semantic parsing on the handover asset meta-model, generate a semantic enhancement model in combination with medical handover specifications, and complete the fields of the original handover information to obtain the completed handover information.

[0058] The intelligent allocation and completion module combines the completed handover information with the regional dynamic handover information network to configure the optimal path for the terminal acquisition module. The regional dynamic handover information network consists of two subnets connected by the historical optimal path. The connection relationship is configured with a hash fingerprint security control key to adjust the acquisition and completion trajectory in case of anomalies.

[0059] The early warning control module, based on the optimal path and the completed handover information, combines semantic analysis of the large language model and the rule engine to execute error warnings. The warning results are fed back to the intelligent allocation and completion module to dynamically adjust the optimal path.

[0060] The terminal acquisition module is further configured as follows:

[0061] A1: Obtain the text, voice, or structured forms entered by medical staff through the terminal as the original handover information;

[0062] Furthermore, if the collected original handover information is in voice format, the terminal acquisition module performs a full-process preprocessing operation through its built-in voice processing unit. First, it enables real-time voice stream reception, performing preprocessing such as frame segmentation and windowing on the collected pediatric nursing handover voice stream to reduce environmental noise interference with the voice signal. Then, it calls a medical-specific automatic speech recognition model to transcribe the processed voice stream sentence by sentence, generating initial text. This medical-specific automatic speech recognition model is fine-tuned based on a large language model and trained with a dedicated corpus for pediatric medical scenarios. It includes professional content such as common pediatric disease names, diagnostic and treatment terms, nursing operation vocabulary, and names of children's medications. It can effectively recognize pediatric-specific expressions in handover voice, solving the problem of low accuracy in recognizing medical professional vocabulary by general speech recognition models.

[0063] After generating the initial text, it is input into a large language model for deep processing. The large language model first standardizes and replaces colloquial words in the text. Pediatric nursing handover contains a large number of colloquial expressions. For example, "fever 38.5 degrees" is replaced with "fever, body temperature 38.5℃", "the baby didn't eat much milk today" is replaced with "the child's milk intake was insufficient today", and "the injection failed" is replaced with "the intravenous puncture was unsuccessful", ensuring the professionalism and standardization of the handover information. Subsequently, suspected erroneous medical terms in the text are identified and corrected. For example, the confused terms "cefixime" and "cephalosporin" are corrected. The system differentiates and corrects errors, correcting the incorrect "nebulized inhalation once daily" to "nebulized inhalation twice daily" according to the doctor's orders, thus avoiding deviations in handover information caused by terminology errors. Finally, the large language model reorganizes the processed text according to a pre-set structured template for pediatric nursing. This template, based on pediatric nursing handover standards, includes fixed fields such as patient bed number, name, age, diagnosis, admission time, current condition, doctor's orders execution status, nursing measures, special precautions, and key points of handover. The scattered handover information is filled into the corresponding fields to generate structured original handover information text.

[0064] If the collected original handover information is text or a pre-filled structured form, the terminal acquisition module first processes the text format uniformly, converting texts with different fonts, font sizes, and paragraphing methods into a standard format. It then parses the forms, extracts valid information, marks missing fields, merges duplicate information, and finally reconstructs the text according to the pre-set structured template for pediatric nursing, outputting standardized structured original handover information text with the same format as the processed voice information, ensuring consistency in subsequent completion and warning operations.

[0065] For example, the handover voice recorded by the pediatric nurse via a voice terminal is: "Bed 1, Duoduo, 1 year old, bronchopneumonia, admitted 3 days ago. Yesterday afternoon, her fever reached 38.8 degrees Celsius, which subsided after taking ibuprofen. She had one nebulizer treatment today, and still has a slight cough. Her milk intake is slightly less than yesterday. We will be drawing blood for routine tests later. Please monitor her temperature." After processing by the voice processing unit, the generated structured original handover information text is: Bed No. 1, Name: Duoduo, Age: 1 year old, Diagnosis: Bronchopneumonia, Admission Date: 3 days, Current Condition: Fever yesterday afternoon, temperature 38.8℃, temperature returned to normal after oral ibuprofen, still has a cough today, milk intake is less than yesterday; Doctor's Orders Execution Status: One nebulizer treatment today; Pending Doctor's Orders: Blood Routine Test; Nursing Precautions: Closely monitor temperature changes. This preprocessing process converts the colloquial pediatric handover voice into standardized, structured text, improving the readability and standardization of the handover information.

[0066] A2: Obtain handover data source metadata from the pre-configured data source connection pool; the handover data source metadata includes the database connection address, access credentials, data mode, and one or more handover data tables; the handover data tables store patient diagnosis and treatment records, medical orders, test results, and vital signs data.

[0067] The generation of structural and semantic fingerprints for each shift handover data table includes:

[0068] B1: Perform structure parsing on each shift handover data table, extract structure information, concatenate the extracted structure information into a string according to the field order, and calculate the string using a first preset hash algorithm to generate the structure fingerprint of the shift handover data table; the structure information includes field names, field data types, primary key and foreign key relationships, and inter-table association mapping; the first preset hash algorithm adopts the MD5 algorithm, wherein the MD5 algorithm is the prior art in this field and is not an inventive solution of this application, and will not be described in detail here;

[0069] Furthermore, the specific steps for B1 include:

[0070] First, a full-dimensional structural analysis is performed on a single shift handover data table to extract its complete structural information. This structural information includes field names, field data types, primary key and foreign key relationships, and inter-table association mappings. The extraction of all structural information strictly adheres to the original storage logic and physical arrangement order of the data table without any arbitrary adjustments. Specifically, when extracting field names, they are extracted sequentially according to the physical storage order of the fields in the data table, forming a field name sequence that perfectly matches the actual field layout. When extracting field data types, each field name in the field name sequence is precisely matched with its defined data type in the data table, ensuring accurate matching between field names and data types. One-to-one correspondence; when extracting primary key and foreign key relationships, first identify the fields defined as primary keys in the data table, clarify the specific composition of the primary key as a single primary key or a composite primary key, then identify the fields defined as foreign keys in the data table, and simultaneously mark the external data tables associated with the foreign key fields and the matching fields in the external data tables, fully presenting the constraint relationship of the primary key and the association relationship of the foreign key; when extracting the inter-table association mapping, comprehensively sort out all the association relationships between the current shift handover data table and other business data tables of the hospital, clarify the name of the target data table to be associated, the names of the fields used for association on both sides, the type of association and the constraint conditions, form complete inter-table association mapping information, and finally output the complete structure information arranged in the original order;

[0071] For example, in the pediatric patient basic information table, the extracted field names include bed number, name, gender, age, admission time, diagnosis result, etc. The field data types include character, numeric, date, etc. The primary key is bed number, which forms a foreign key relationship with the bed number field in the medical order information table and the vital signs data table. The inter-table relationship mapping clarifies the rules for data association between the tables through bed number and patient ID.

[0072] Second, the complete output structure information is standardized and concatenated according to preset rules to form a continuous and uninterrupted structure information string. The concatenation is based on the physical arrangement order of the fields in the data table. First, each field name and its corresponding field data type are concatenated continuously. After all field names and field data types are concatenated, the extracted primary key and foreign key relationships are concatenated. After the primary key and foreign key relationships are concatenated, the extracted inter-table association mapping information is concatenated. The entire concatenation process does not add any delimiters or modifiers. All structure information is integrated through pure text description to ensure that the concatenated structure information string completely restores all extracted structure information without any information additions, deletions, or content deviations. The output is a structure information string.

[0073] Third, the output structure information string is converted into a binary data stream that can be recognized by the MD5 algorithm. The conversion adopts a common character encoding standard to ensure that each character in the structure information string can be accurately converted into the corresponding binary value, forming a continuous and uninterrupted binary data stream.

[0074] Fourth, the MD5 algorithm is used as the first preset hash algorithm to perform hash calculations on the output binary data stream, generating the structural fingerprint of the handover data table. The parameter settings of the MD5 algorithm follow the standard algorithm specifications. The calculation process is as follows: First, the initial vector of the MD5 algorithm is set to four standard 32-bit hexadecimal values, namely 01234567, 89ABCDEF, FEDCBA98, and 76543210. Then, the binary data stream is divided into data groups, with each data group having a length of 512 bits. If the total length of the binary data stream is less than an integer multiple of 512 bits, the binary data stream is padded according to the padding rules of the MD5 algorithm to ensure that the length of the last data group after padding is 512 bits. Next, four rounds of loop processing operations are performed on each 512-bit data group, with each round of operation containing 16 basic logical operations. The four rounds of operations employ four different logic functions: F, G, H, and I. In each basic logic operation, a fixed number of bits of data are selected for the operation, along with an addition constant and a cyclic shift constant. The addition constant is a 32-bit hexadecimal value obtained by taking the square root of the standard sine function. The cyclic shift constant is set to different shift numbers according to the round and order of operations, with shift numbers including 7, 12, 17, 22, 5, 9, 14, 20, 4, 11, 16, 23, 6, 10, 15, and 21. Finally, after completing the four rounds of cyclic operations for all data groups, the result of each round is accumulated and merged with the initial vector to obtain a 128-bit binary hash value. This 128-bit binary hash value is then converted into a 32-bit hexadecimal string, which serves as the structural fingerprint of the handover data table.

[0075] B2: Perform semantic parsing on each shift handover data table, parse the textual meaning of field names based on natural language processing technology, obtain synonyms and hypernyms by combining medical knowledge graph, convert the textual meaning into vector representation, and generate semantic fingerprints through a second preset hash algorithm; the second preset hash algorithm adopts the SimHash algorithm, wherein the SimHash algorithm is the prior art in this field and is not an inventive solution of this application, and will not be described in detail here.

[0076] Furthermore, the specific steps of B2 include:

[0077] First, text meaning parsing is performed on all field names in a single shift handover data table. Natural language processing technology is used to extract the text meaning of the field names. Specifically, the process involves: first, standardizing and preprocessing the field names to remove irrelevant special characters and redundant expressions, retaining only the core text content; then, performing word segmentation on the preprocessed field names to break down the continuous field name text into the smallest semantic units in the medical context; subsequently, part-of-speech tagging and semantic understanding are performed on each segmented semantic unit to clarify the specific meaning of each word in the medical context, ultimately forming a complete text meaning description for each field name.

[0078] For example, the semantics of the "axillary temperature" field is "the body temperature value measured under the armpit of the child", the semantics of the "milk intake" field is "the amount of breast milk / formula milk consumed by the child in a single session or daily", and the semantics of the "nebulization duration" field is "the duration of each nebulization inhalation treatment for the child".

[0079] Second, the output field name text meaning parsing results are matched and associated with the medical knowledge graph to obtain the synonyms and hypernyms corresponding to the text meaning of each field name. Specifically, the following operations are performed: First, based on the core semantics of the field name, the corresponding benchmark term is accurately located in the medical knowledge graph, which contains a standardized terminology system, vocabulary associations, and hierarchical classification structures in the medical field; then, based on the association relationship of the benchmark term in the medical knowledge graph, the synonyms and hypernyms of the term are extracted. At least three synonyms and two hypernyms are extracted for the text meaning of each field name, and finally, complete semantic extension information containing the text meaning of the field name, corresponding synonyms, and corresponding hypernyms is formed.

[0080] Furthermore, the medical knowledge graph is a pediatric-specific graph, containing a vast amount of entities, attributes, and relationships in the pediatric medical field. For example, the synonym for "axillary temperature" is "axillary body temperature," and its superordinate term is "vital signs"; the synonym for "milk intake" is "food intake," and its superordinate term is "nutritional intake index"; the synonym for "nebulized inhalation" is "aerosol inhalation therapy," and its superordinate term is "local respiratory treatment."

[0081] Third, the complete semantic extension information output is converted into a vector representation. Specifically, a pre-trained medical domain word vector model is selected as the basic model for vector conversion. The vector dimension of the medical domain word vector model is set to 300 dimensions, and the training corpus covers various types of medical-related text data, including clinical medical terminology, hospital information system field names, and medical knowledge base text. The obtained field name text meanings, synonyms, and hypernyms are sequentially input into the 300-dimensional word vector model, outputting a 300-dimensional real vector corresponding to each word. Then, all word vectors corresponding to the same field name are weighted and fused. The weight value for the field name text meaning is set to 0.6, the weight value for synonyms is set to 0.2, and the weight value for hypernyms is set to 0.2. The comprehensive vector of the field name is obtained through weighted summation. The comprehensive vectors of all field names together constitute the semantic vector set of the handover data table.

[0082] Fourth, the SimHash algorithm is used as the second preset hash algorithm to perform hash calculations on the output semantic vector set, generating the semantic fingerprint of the handover data table. This includes: firstly, preprocessing each 300-dimensional composite vector in the semantic vector set, using zero as the dividing point, converting dimension values ​​greater than or equal to zero to the number 1, and dimension values ​​less than zero to the number 0, so that each 300-dimensional composite vector is finally converted into a 300-bit binary string; then, initializing the feature vector of the SimHash algorithm as a 300-dimensional zero vector, that is, the initial value of each dimension is zero; then, processing each 300-bit... The binary string is traversed bit by bit. If the binary value at any position is one, the value of the corresponding dimension of the initial feature vector is incremented by one. If the binary value at that position is zero, the value of the corresponding dimension of the initial feature vector is decremented by one. After traversing all binary strings, an updated 300-dimensional feature vector is obtained. Finally, the updated 300-dimensional feature vector is binarized. Using zero as the threshold, the dimension values ​​in the feature vector that are greater than zero are converted to the number one, and the dimension values ​​that are less than or equal to zero are converted to the number zero. This results in a 300-bit binary hash value, which is the semantic fingerprint of the handover data table.

[0083] Example 2:

[0084] Please see Figure 2 In this embodiment, the calculation of fusion similarity and clustering and merging under an adaptive threshold to construct a handover asset meta-model includes:

[0085] C1: For any two handover data tables, calculate the structural similarity score based on their structural fingerprints and the semantic similarity score based on their semantic fingerprints. Use preset weight coefficients to weight and fuse the structural similarity score and the semantic similarity score to obtain the fused similarity.

[0086] Furthermore, based on the structural fingerprints of the two data tables, the Hamming distance is used to calculate their structural similarity. The Hamming distance represents the number of different characters at corresponding positions in two strings of equal length. The smaller the distance, the more similar the structural fingerprints are and the closer the structures of the data tables are. The Hamming distance is normalized and converted into a structural similarity score between 0 and 1. The closer the score is to 1, the higher the structural similarity. The Hamming distance calculation formula is prior art in this field and is not an inventive solution of this application. It will not be elaborated here.

[0087] Furthermore, based on the semantic fingerprints of the two data tables, the semantic similarity between them is calculated using Hamming distance, and the semantic similarity score between 0 and 1 is obtained after normalization.

[0088] In this embodiment, taking into account the characteristics of pediatric handover data, the structural similarity weight is set to 0.4 and the semantic similarity weight is set to 0.6, highlighting the importance of semantic features in the association of pediatric handover data. The weighted value is the fusion similarity between the two data tables.

[0089] C2: Construct a similarity matrix by fusing the pairwise similarities between all the shift handover data tables, and use the silhouette coefficient method to perform cluster analysis on the similarity matrix to dynamically determine the adaptive clustering threshold that optimizes the clustering effect;

[0090] Furthermore, suppose the system acquires a total of M shift handover data tables. Using the data tables as rows and columns, the fusion similarity values ​​between each pair are filled into the corresponding positions in the matrix to construct an M×M similarity matrix. The elements in the diagonal positions of the matrix are the fusion similarity between the data table itself and itself, with a value of 1.

[0091] Furthermore, the silhouette coefficient method clustering analysis process includes: the silhouette coefficient method is a commonly used method for evaluating the quality of clustering. The calculated silhouette coefficient value ranges from [-1, 1]. The closer the value is to 1, the better the clustering effect, meaning that the data tables within the same cluster have high similarity, while the data tables between different clusters have low similarity. The system performs K-means clustering analysis on the similarity matrix, setting the number of clusters K sequentially from 2 to M-1, calculating the overall silhouette coefficient corresponding to each K value, and selecting the K value with the largest silhouette coefficient as the optimal number of clusters. Both the silhouette coefficient method and K-means clustering analysis are existing technologies in this field and are not inventive solutions of this application; therefore, they will not be elaborated upon here.

[0092] Furthermore, the adaptive clustering threshold is determined as follows: based on the optimal number of clusters, the fusion similarity threshold that brings the clustering results to this optimal state is derived in reverse; this is the adaptive clustering threshold. In the pediatric application example, cluster analysis is performed on a 6×6 similarity matrix. When K=2, the overall silhouette coefficient reaches its maximum value of 0.91. The corresponding adaptive clustering threshold is derived in reverse as 0.75. That is, pediatric handover data tables with a fusion similarity greater than 0.75 are grouped into the same cluster, while those with a similarity less than 0.75 are grouped into different clusters.

[0093] It should be noted that the reverse derivation, simply put, is as follows: First, the optimal number of clusters for the six handover data tables is determined to be 2 using the silhouette coefficient. Then, multiple candidate fusion similarity thresholds are set with a fixed step size between 0 and 1. Each candidate threshold is used to divide the six data tables one by one. It is determined whether the fusion similarity of any two data tables is greater than or equal to the candidate threshold. If they are, they are classified into the same cluster; otherwise, they are classified into different clusters. The actual number of clusters formed after each division is counted. Finally, the candidate threshold that makes the number of clusters exactly equal to the optimal number of clusters of 2 is found. This threshold is 0.75, which is the final adaptive clustering threshold. At the same time, it is also necessary to verify that this threshold is a critical value. If it is slightly lower, the number of clusters is still 1; if it is slightly higher, the number of clusters will be greater than 2.

[0094] C3: All shift handover data tables with a similarity greater than the adaptive clustering threshold are grouped into the same cluster, and all shift handover data tables within the same cluster are merged to generate a unified shift handover asset meta-model; the shift handover asset meta-model integrates the common fields, data types and semantic descriptions of all shift handover data tables within the cluster.

[0095] Furthermore, the handover asset meta-model comprehensively integrates the handover data features from multiple dimensions, including pediatric patient diagnosis and treatment, medical orders, laboratory tests, vital signs, and nursing care. It clarifies the structural associations and semantic relationships between various data fields, becoming the core foundation for generating the semantic enhancement model. For example, the handover asset meta-model integrates the common field of "bed number," unifies the numerical data type of the "body temperature" field, and standardizes the semantic description of "nebulized inhalation" as "a common local treatment method for pediatric bronchopneumonia and other respiratory diseases, which converts drugs into aerosols through a nebulizer and is inhaled through the child's mouth and nose, acting on the respiratory mucosa."

[0096] Furthermore, the specific steps of C3 include:

[0097] (1) Based on the adaptive clustering threshold of 0.75, the six shift handover data tables were clustered and divided. The shift handover data tables with a fusion similarity greater than 0.75 were assigned to cluster 1, and the shift handover data tables with a fusion similarity less than 0.75 were assigned to cluster 2, thus achieving a reasonable division of the data tables.

[0098] (2) Perform deep schema merging on all data tables within each cluster, integrate the common fields, data types and semantic descriptions of all data tables within the cluster, while retaining the characteristic fields and exclusive semantic information of each data table to form a unified schema specification; taking cluster 1 as an example, firstly extract the common fields of the three data tables as hospital number, patient name, diagnosis result and treatment measures, and unify the data type of the common fields, such as unifying the data type of the "diagnosis result" field to character type; secondly, unify the semantic description of all fields, and add standardized semantic explanations to each field in combination with the medical knowledge graph to form a unified schema of cluster 1;

[0099] (3) The merging patterns of each cluster are integrated to generate the handover asset meta-model of the system. This model is a unified abstraction and integration of the handover data tables of various business systems in the hospital. It integrates the common fields, data types and semantic descriptions of all data tables in the cluster. At the same time, it establishes the relationship between fields, data constraint rules and semantic mapping relationship, and becomes the core basic model for semantic enhancement model generation and handover information completion.

[0100] Semantic parsing is performed on the aforementioned handover asset meta-model, and a semantically enhanced model is generated by combining it with medical handover standards, including:

[0101] D1: Call a pre-trained large language model in the medical field to parse the handover asset meta-model, and identify the core medical entity types and relationships between entities represented by the handover asset meta-model;

[0102] Furthermore, in this embodiment, the core medical entity types parsed include pediatric patients, disease diagnoses, vital signs, medical orders, nursing procedures, and laboratory test indicators. The relationships between entities include "pediatric patient - suffers from - disease diagnosis", "disease diagnosis - corresponds to - medical order", "pediatric patient - test - vital signs", and "medical order - includes - nursing procedure", which clearly outlines the logical connections between the core medical entities.

[0103] D2: Query the standard handover field definitions, mandatory field rules, and logical dependencies between fields related to the core medical entity type from the pre-built medical handover standard knowledge base;

[0104] Furthermore, the method for constructing the medical handover standard knowledge base includes: extracting standard handover fields, key indicators, mandatory field lists, and logical verification rules from nationally issued core medical systems, hospital internal handover standard documents, and historical high-quality handover records using natural language processing technology; and using knowledge graph construction technology to store and manage the extracted fields, indicators, and rules in the form of entities, attributes, and relationships to form a queryable and reasonable medical handover standard knowledge base.

[0105] Furthermore, the construction of the medical handover standard knowledge base involved multi-dimensional data collection and technical processing. First, standardized information was extracted from national medical standards, hospital-established internal regulations, and high-quality historical handover records from the pediatric ward using natural language processing technology. This extracted standard handover fields, key indicators, mandatory field lists, and logical verification rules applicable to pediatric nursing handovers. Subsequently, knowledge graph construction technology was used to store and manage the extracted fields, indicators, and rules in the form of entities, attributes, and relationships, forming a pediatric nursing-specific medical handover standard knowledge base that can be queried in real time and logically reasoned. This knowledge base can be dynamically iterated and updated according to updates to medical standards and changes in clinical practice, ensuring its timeliness and accuracy.

[0106] In this embodiment, the standard handover field definitions related to the entity relationship "pediatric patient - bronchopneumonia" found from the medical handover standard knowledge base include "temperature monitoring frequency", "respiratory symptom observation", "nebulized inhalation therapy standard", and "principles of anti-infective drug use". The mandatory field rules include that when handing over a child with bronchopneumonia, vital signs such as temperature, respiration, and heart rate must be filled in, the execution status of medical orders and pending medical orders must be noted, and the key points of nursing observation must be clearly stated. The logical dependencies between fields include the dependency between the "nebulized inhalation therapy" field and the "respiratory symptoms" field. If the symptoms do not improve after nebulized inhalation therapy, it must be noted in the "special precautions" field. The "temperature" field and the "antipyretic drug use" field are also dependent. If the temperature exceeds 38.5°C, the name, dosage, and time of use of the antipyretic drug must be noted.

[0107] D3: Based on the parsing results of the large language model and the query results of the medical handover standard knowledge base, generate a semantic tag set and a set of field mapping rules for each handover asset meta-model; the semantic tag set, field mapping rules and the handover asset meta-model together constitute a semantic enhancement model.

[0108] Furthermore, the semantic tag set is a highly condensed set of semantics for each field in the handover asset meta-model. Each field corresponds to one or more semantic tags. For example, the semantic tags for the "body temperature" field are "vital signs, fever monitoring, and core pediatric indicators," the semantic tags for the "nebulized inhalation" field are "respiratory care, bronchopneumonia treatment, and topical medication," and the semantic tags for the "milk intake" field are "nutritional intake, pediatric care, and disease observation indicators." The field mapping rules clarify the correspondence between the fields in the handover asset meta-model and the standard fields in the medical handover specifications, as well as the rules for completing the associations between different fields. For example, the "bed number" field can be mapped to all pediatric data tables in the hospital database, and all relevant patient data can be queried through the bed number. The "diagnosis result - bronchopneumonia" field can be mapped to the standard handover field for bronchopneumonia in the medical handover specifications knowledge base. Based on this mapping relationship, the associated fields that need to be completed can be determined. The semantic tag set, field mapping rules, and pediatric ward handover asset meta-model together constitute a semantic enhancement model exclusive to pediatric nursing. This model combines data structure features, semantic features, and medical standard constraints, providing comprehensive support for intelligent field completion.

[0109] The process of completing the fields in the original shift handover information to obtain the completed shift handover information includes:

[0110] E1: The semantic enhancement model is used to parse the original handover information to identify the filled fields and their corresponding medical entity identifiers;

[0111] E2: For the identified medical entity identifier, determine the list of associated fields that need to be completed for the entity based on the field mapping rules in the semantic enhancement model;

[0112] E3: Based on the handover data source metadata obtained by the terminal acquisition module, connect to the corresponding database, and use a multi-table join SQL query statement based on the medical entity identifier to retrieve the specific data values ​​corresponding to the list of associated fields from the handover data table;

[0113] E4: Fill the missing fields of the original handover information with the retrieved data values ​​to form the completed handover information.

[0114] For example, this embodiment uses the structured original handover information text of Duoduo in pediatric bed 1 as an example to perform a full-process field completion operation. First, the semantic enhancement model is used to perform deep analysis on the structured original handover information text to accurately identify the fields that have been filled in the text and their corresponding medical entity identifiers. The fields that have been filled in include bed number, name, age, diagnosis, admission time, current condition, medical order execution status, medical orders to be executed, and nursing precautions. The corresponding medical entity identifier is "bed number 1-Duoduo-bronchopneumonia". Through this identifier, all relevant medical data of the child can be uniquely identified. Subsequently, for the identified medical entity identifier "Bed No. 1-Duoduo-Bronchopneumonia", based on the field mapping rules in the semantic enhancement model and combined with the handover specifications for pediatric bronchopneumonia, a list of associated fields that need to be completed for this entity was determined. In this embodiment, the determined associated fields include the child's gender, age, body temperature upon admission, respiratory rate, heart rate, historical blood routine test results, name and dosage of nebulized inhaled medication, ibuprofen dosage and time of use, name of the responsible nurse, and nursing level. These fields are all standard mandatory fields for the handover of pediatric bronchopneumonia patients and were not filled in in the original handover information, so they need to be completed through database query. Next, based on the pediatric handover data source metadata acquired by the terminal acquisition module, an encrypted database connection channel is used to connect to the pediatric patient diagnosis and treatment database. A multi-table join SQL query based on the medical entity identifier "Bed No. 1-Duoduo-Bronchopneumonia" is employed to precisely retrieve specific data values ​​corresponding to the associated fields from five core handover data tables: patient basic information table, doctor's order information table, laboratory test results table, vital signs data table, and nursing operation record table. During the retrieval process, the table association mapping rules explicitly defined in the handover asset meta-model are utilized to achieve precise association queries of multi-table data through associated fields such as bed number and patient ID. The retrieved data is matched one-to-one with the child to avoid data confusion. For example, the patient's basic information table shows that the child is female, 1 year old, and requires Level 1 nursing care; the vital signs data table shows that the child's temperature upon admission was 39.2℃, current respiratory rate was 28 breaths / minute, and heart rate was 110 beats / minute; the medical orders table shows that the nebulized medication is budesonide suspension + terbutaline nebulizer solution, at doses of 0.5mg and 2.5mg respectively, twice daily, and ibuprofen was administered at a dose of 5ml at 16:30 yesterday; the laboratory test results table shows that the white blood cell count upon admission was 12.5×10⁻⁶. 9 / L, neutrophil percentage 75%; the responsible nurse was identified as Nurse Zhang from the nursing operation record. Finally, all the retrieved specific data values ​​were precisely filled into the corresponding missing fields of the structured original handover information text, supplementing and improving the information in the original fields to form a complete pediatric nursing handover information. In this embodiment, the complete handover information is as follows: Bed No. 1, Name: Duoduo, Gender: Female, Age: 1 year, Diagnosis: Bronchopneumonia, Admission Time: 3 days, Condition upon Admission: Temperature 39.2℃, Blood routine examination showed white blood cell count 12.5×10^9 / L, neutrophil percentage 75%; Current Condition: Fever yesterday afternoon, temperature 38.8℃, ibuprofen oral administration, dose 5ml, temperature returned to normal after 16:30, still coughing today, respiratory rate 28 breaths / minute, heart rate 110 beats / minute, milk intake decreased compared to yesterday; Nursing Level: Level 1 Nursing, Responsible Nurse: Nurse Zhang; Doctor's Order Execution: Nebulized inhalation treatment once today, including medication: budesonide suspension 0.5mg. + Terbutaline nebulizer solution 2.5mg; To be executed: Complete blood count; Nursing precautions: Closely monitor body temperature changes, observe cough symptoms and respiratory status, and record milk intake. This intelligent completion process fills in all the missing key fields in the original handover information, ensuring the completeness, accuracy, and standardization of the handover information, meeting the professional requirements of pediatric nursing handover.

[0115] The regional dynamic shift handover information network includes a task logic subnet and a physical resource subnet;

[0116] F1: The nodes in the task logic subnet represent the historical handover tasks of different departments or wards, and the nodes in the physical resource subnet represent data source servers, network exchange nodes and database storage resources.

[0117] Furthermore, the nodes in the task logic subnet represent historical handover tasks from different departments and wards of the hospital. Each node contains detailed characteristic information of the corresponding historical handover task, such as the department, ward, patient condition type, handover information type, data collection volume, and collection time. In this embodiment, the historical handover task nodes of the pediatric ward include handover task characteristics for different disease types such as common cold, bronchopneumonia, diarrhea, and hand-foot-mouth disease. Each node is dynamically updated according to the execution status of historical handover tasks, recording key indicators such as task execution efficiency and data collection success rate.

[0118] Furthermore, the nodes in the physical resource subnet represent various data source servers, network switching nodes, and database storage resources within the hospital. Data source servers are divided by department, such as pediatric data source servers, internal medicine data source servers, and surgical data source servers. Network switching nodes are network devices such as core switches and access switches in the hospital's intranet. Database storage resources are database clusters that store medical data from various departments, including relational databases and non-relational databases. Each physical resource subnet node records basic information such as its hardware performance, network bandwidth, storage capacity, and data access speed, as well as operational indicators such as response time and failure rate during historical data collection.

[0119] F2: The task logic subnet nodes and physical resource subnet nodes are connected by directed edges. Each directed edge represents a historical data collection path and is associated with a hash fingerprint security control key used to identify the current security status of the path.

[0120] Furthermore, the security control key is generated using a hash algorithm to identify the current security and availability status of the path. If a path experiences network failure, data leakage risk, or abnormal access permissions during data collection, its corresponding hash fingerprint security control key status will be immediately marked as invalid, ensuring that abnormal paths will not be selected again.

[0121] Configuring the optimal path for the terminal acquisition module includes:

[0122] G1: Parse the completed handover information and extract the features of the current handover task. The features include the department, the severity of the patients involved, the data type required, and the estimated data volume.

[0123] Furthermore, the intelligent allocation and completion module combines the completed pediatric nursing handover information with the regional dynamic handover information network to configure the optimal data acquisition path for the terminal acquisition module of the pediatric nursing station. In this embodiment, the completed handover information of Duoduo in pediatric bed 1 is used as an example to perform the optimal path configuration operation. First, the completed handover information is deeply analyzed to extract the core features of the current handover task. These features are the key basis for matching task nodes and evaluating path costs. In this embodiment, the extracted features of the current pediatric handover task include: the department is the pediatric inpatient department, the patient's condition level is ordinary, such as bronchopneumonia, which is a common pediatric disease, the child's current condition is stable, and the required data types include patient basic information, vital signs data, medical order execution data, laboratory test results data, and nursing operation data. The estimated data volume is small batch, such as multi-dimensional data of a single patient, with a data volume of approximately 500KB.

[0124] G2: Match the characteristics of the current shift handover task with the task logic subnet in the regional dynamic shift handover information network to determine the current task node;

[0125] In this embodiment, the current task node is determined to be the historical handover task node of "small batch data collection of single patients with bronchopneumonia" in the pediatric ward through feature matching. This historical handover task node is highly consistent with the current handover task in terms of the department, disease type, data type, data volume, etc., and the execution efficiency and success rate of its historical data collection path are at a high level.

[0126] G3: Evaluate the overall cost of reaching the target resource node in the physical resource subnet from the current task node through the directed edges of each historical data acquisition path; the overall cost is calculated based on the weighted sum of the historical calculation delay value of the path, the network bandwidth utilization rate, and the data acquisition success rate.

[0127] In this embodiment, the target resource nodes are the pediatric data source server, the network switching node corresponding to the pediatric ward, and the pediatric medical database storage resources. The calculation of the comprehensive cost is based on the weighted sum of the historical value of the path calculation delay, the network bandwidth utilization rate, and the data acquisition success rate. In this embodiment, combined with the operating characteristics of the hospital intranet and the data collection requirements of the pediatric handover, the weight of the historical value of the calculation delay is set to 0.3, the weight of the network bandwidth utilization rate is 0.4, and the weight of the data acquisition success rate is 0.3, highlighting the importance of the network bandwidth utilization rate in path selection and avoiding data acquisition lag due to excessive bandwidth utilization.

[0128] Furthermore, the historical latency value is calculated as the average time from initiating a query request to obtaining the data result during historical data collection for that path. The shorter the time, the smaller the historical latency value, and the better the path. The network bandwidth utilization rate is the proportion of the actual bandwidth utilization of that path at the current moment. The lower the proportion, the more abundant the network bandwidth resources and the faster the data transmission speed. The data acquisition success rate is the proportion of the number of times the target data was successfully acquired during historical data collection for that path. The higher the proportion, the stronger the stability and reliability of the path. When calculating the overall cost, the three indicators are first standardized, converting indicators of different dimensions into standardized values ​​between 0 and 1. The historical latency value and network bandwidth utilization rate are negative indicators, with smaller values ​​being better. The data acquisition success rate is a positive indicator, with larger values ​​being better. Then, they are weighted and summed according to preset weight coefficients to obtain the overall cost value for each path. The smaller the overall cost value, the higher the collection efficiency and the stronger the stability of the path.

[0129] In this embodiment, starting from the "Single Patient Small Batch Data Collection for Bronchopneumonia" task node in the pediatric ward, there are three available directed edges for historical data collection paths, each connected to a combination of three target resource nodes in the physical resource subnet. Path 1 has a standardized historical value for computational latency of 0.2, a standardized value for network bandwidth utilization of 0.3, a standardized value for data acquisition success rate of 0.9, and a total cost of 0.2×0.3+0.3×0.4+0.9×0.3=0.45. Path 2 has a standardized historical value for computational latency of 0.3, a standardized value for network bandwidth utilization of 0.5, a standardized value for data acquisition success rate of 0.8, and a total cost of 0.3×0.3+0.5×0.4+0.8×0.3=0.53. Path 3 has a standardized historical value for computational latency of 0.4, a standardized value for network bandwidth utilization of 0.2, a standardized value for data acquisition success rate of 0.7, and a total cost of 0.4×0.3+0.2×0.4+0.7×0.3=0.41.

[0130] Finally, the historical data acquisition path with the lowest overall cost is selected as the optimal path for configuring the terminal acquisition module. In this embodiment, the overall cost value of path 3 is 0.41, which is the minimum. Therefore, path 3 is configured as the optimal data acquisition path for the pediatric nurse station terminal acquisition module. This path is: Pediatric ward "Bronchopneumonia Single Patient Small Batch Data Acquisition" task node - Pediatric ward dedicated network exchange node - Pediatric data source server - Pediatric medical database storage resources. The corresponding hash fingerprint security control key is in a valid state to ensure the efficiency and security of data acquisition.

[0131] G4: Select the historical data acquisition path with the lowest overall overhead as the optimal path for configuring the terminal acquisition module.

[0132] The early warning control module, based on the optimal path and the completed shift handover information, combines large language model semantic analysis and rule engine to execute error warnings, including:

[0133] H1: Monitor the actual trajectory of the data collected by the terminal acquisition module along the optimal path, and calculate the path deviation between the actual trajectory and the preset trajectory of the optimal path;

[0134] Furthermore, the preset trajectory is the fixed collection order and network transmission path specified in the optimal path from the task node to each physical resource node, namely, pediatric ward task node - pediatric ward dedicated network exchange node - pediatric data source server - pediatric medical database storage resources. The actual trajectory is the nodes and transmission path traversed by the terminal acquisition module during the actual data collection process. During the monitoring process, the early warning control module obtains the network access records and data query records of the terminal acquisition module in real time through the hospital's intranet network monitoring system, accurately tracks its actual collection trajectory, and then uses a path matching algorithm to compare the node overlap and path consistency between the actual trajectory and the preset trajectory, and calculates the path deviation. The path deviation ranges from 0 to 1. If the actual trajectory is completely consistent with the preset trajectory, the path deviation is 0; if the actual trajectory is completely different from the preset trajectory, the path deviation is 1. The larger the path deviation, the higher the degree of abnormality of the data collection trajectory.

[0135] In this embodiment, if the terminal acquisition module bypasses the dedicated network exchange node for the pediatric ward and directly accesses the pediatric data source server during the actual acquisition process due to a temporary failure of the network exchange node, its actual trajectory is pediatric ward task node - hospital core network exchange node - pediatric data source server - pediatric medical database storage resources, which differs from the preset trajectory by a number of nodes. After calculation, its path deviation is 0.25. If the network failure is repaired and the terminal acquisition module returns to the preset trajectory, the path deviation will return to 0.

[0136] H2: Call the large language model to perform semantic analysis on the completed handover information and identify abnormal expressions in the information; the abnormal expressions include logical contradictions, expressions that violate the established medical knowledge graph association relationships, or expressions that differ from the trends of patients' historical data;

[0137] Furthermore, abnormal expressions are the core manifestation of errors in handover information. These mainly include logical contradictions, expressions that violate established medical knowledge graph relationships, or expressions that differ from the trends in patients' historical data. During the semantic analysis process, the large language model combines the pediatric medical knowledge graph and the patient's historical medical data to perform logical verification and semantic analysis on the completed handover information field by field and sentence by sentence to ensure the accuracy of identification.

[0138] Logically contradictory statements refer to inconsistencies and conflicts in the handover information. For example, if the completed handover information contains both "5ml of ibuprofen was administered orally at 16:30 yesterday" and "no antipyretics were used yesterday," it is a clear logical contradiction. Statements that violate established medical knowledge graph relationships refer to handover information that conflicts with entity relationships and medical standards in the pediatric medical knowledge graph. For example, if the handover information for a child with bronchopneumonia contains "metformin was administered orally," it violates the knowledge graph relationship of "bronchopneumonia - corresponding - anti-infective, antitussive, nebulized, and other treatment drugs." Statements that differ from the patient's historical data trend refer to unreasonable deviations between the current data in the handover information and the child's historical treatment data trend. For example, if a child's temperature dropped continuously after admission from 39.2℃ to 38.8℃, but the handover information shows a sudden rise to 39.5℃ without any explanation of the cause of the fever, this is an abnormal statement that differs from the historical data trend.

[0139] In this embodiment, semantic analysis of the completed handover information for the patient, Duoduo, in pediatric bed 1, revealed no logical contradictions or violations of the medical knowledge graph's association relationships. The child's current temperature, respiration, and heart rate data showed a consistent trend with historical data, all indicating improvement. Therefore, no abnormal statements were identified in the information. However, if the completed handover information contained the phrase "nebulized inhalation medication is budesonide suspension 10mg," the large language model would immediately identify this statement as abnormal. This is because, according to the pediatric medical knowledge graph, the maximum single dose of budesonide suspension for a 1-year-old child is 1mg, and a dose of 10mg far exceeds the standard dose, violating the association relationships of the medical knowledge graph.

[0140] H3: Input the path deviation and the abnormal expression into the rule engine; the rule engine has pre-stored warning rules of different levels. When the input deviation exceeds the first threshold or a specified type of abnormal expression is identified, the corresponding level of warning information is triggered.

[0141] Furthermore, the rule engine is a logical judgment module with pre-set warning rules of different levels. Its internal warning rules are formulated in combination with hospital medical data management specifications, nursing handover system and system operation requirements. There are three warning levels: Level 1 warning, Level 2 warning and Level 3 warning, which correspond to minor abnormalities, moderate abnormalities and severe abnormalities, respectively. Different warning levels correspond to different triggering conditions.

[0142] In this embodiment, the first threshold is set to 0.2 as the upper limit of the first-level warning, the lower limit of the second-level warning is set to 0.5, and the lower limit of the third-level warning is set to 0.8. When the deviation of the input path exceeds 0.2 but does not exceed 0.5, or when a slight abnormal expression is identified, such as non-standard field name writing or slight deviation of non-critical data values, a first-level warning is triggered. When the deviation of the input path exceeds 0.5 but does not exceed 0.8, or when a moderate abnormal expression is identified, such as deviation of critical data values ​​or slight logical contradictions that do not affect the judgment of the condition, a second-level warning is triggered. When the deviation of the input path exceeds 0.8, or when a serious abnormal expression is identified, such as major data errors, serious logical contradictions, or violations of core medical standards that may affect the safety of patient diagnosis, treatment, and care, a third-level warning is triggered. After receiving path deviation and anomaly descriptions, the rule engine quickly judges them according to the preset warning rules. If the triggering conditions of the corresponding level are met, the corresponding level of warning information is immediately triggered. The warning information includes the warning level, anomaly type, anomaly location, anomaly details, etc. For example, the level 1 warning information is "Path deviation 0.25, exceeding the first threshold of 0.2, anomaly type: abnormal path acquisition trajectory, anomaly location: network exchange node, anomaly details: the terminal acquisition module bypasses the dedicated network exchange node of the pediatric ward and uses the hospital's core network exchange node for data acquisition".

[0143] The early warning results are fed back to the intelligent allocation and completion module to dynamically adjust the optimal path, including:

[0144] K1: When the warning level triggered by the warning control module exceeds the preset second threshold, the intelligent allocation completion module receives a warning result containing an abnormal path identifier or an abnormal data field; the second threshold is greater than the first threshold;

[0145] Furthermore, if the warning level triggered by the warning control module exceeds the preset second threshold, in this embodiment it is set to a level two warning. When the warning level triggered by the warning control module reaches level two or above, that is, the path deviation exceeds 0.5, or a moderate / severe abnormal expression is identified, the warning result is fed back to the intelligent allocation and completion module, which dynamically adjusts the optimal path. Here, 0.5 is the lower limit for triggering a level two warning, which is greater than the upper limit of 0.2 for a level one warning. If the triggered warning level is a level one warning, only a warning prompt is sent, the optimal path is not adjusted, and only the collection trajectory of the terminal collection module is monitored in real time.

[0146] Furthermore, the warning result includes an abnormal path identifier or an abnormal data field. The abnormal path identifier is the hash fingerprint security control key corresponding to the abnormal collection path, and the abnormal data field is the handover information field that identifies the abnormal expression.

[0147] K2: The intelligent allocation completion module marks the corresponding hash fingerprint security control key status as invalid according to the abnormal path identifier, and temporarily removes the path from the current set of available paths;

[0148] K3: The intelligent allocation and completion module recalculates the overall cost based on the updated set of optional paths, generates a new optimal path, and sends it to the terminal acquisition module.

[0149] In this embodiment, if the path deviation of the terminal acquisition module is 0.6, a level-two warning is triggered. The warning control module will send the warning result containing the abnormal path identifier, such as the hash fingerprint security control key corresponding to path 3, to the intelligent allocation and completion module. After receiving the warning result, the intelligent allocation and completion module first marks the status of the corresponding hash fingerprint security control key as invalid according to the abnormal path identifier, and temporarily removes the path from the current set of optional paths to ensure that the abnormal path will not be used again. Subsequently, based on the updated set of optional paths, i.e., the original path 1 and path 2, the intelligent allocation and completion module recalculates the comprehensive cost of the two paths according to the weighted sum of the historical latency value, network bandwidth utilization rate, and data acquisition success rate. In this embodiment, the comprehensive cost value of path 1 after recalculation is 0.43, and the comprehensive cost value of path 2 is 0.51. The path 1 with the smallest comprehensive cost is selected as the new optimal path, and the new optimal path information is immediately sent to the terminal acquisition module. After receiving the new path, the terminal acquisition module immediately adjusts the data acquisition trajectory and performs data acquisition according to the new optimal path to ensure the stability of the acquisition process.

[0150] If the early warning control module identifies serious anomalies in the completed handover information, such as incorrect nebulized medication dosage, triggering a Level 3 early warning, the warning result containing the abnormal data field, such as the nebulized medication dosage, will be fed back to the intelligent allocation completion module. The intelligent allocation completion module will first mark the hash fingerprint security control key of the current optimal path as invalid, remove it from the set of optional paths, and then reconfigure the optimal path. At the same time, it will feed back the abnormal data field information to the terminal acquisition module, reminding medical staff to check and correct the field to ensure the accuracy of the handover information.

[0151] Example 3:

[0152] In this embodiment, the terminal acquisition module includes an information acquisition unit and a voice processing unit. The information acquisition unit supports medical staff to input raw handover information in various ways. It can obtain pre-configured handover data source metadata from the data source connection pool, encrypt and temporarily store all collected raw information and metadata, and perform preliminary format verification. Information with incorrect format is marked and reminded. The voice processing unit is designed specifically for raw handover information in voice type. It can receive handover voice streams in real time, call a medical-specific automatic speech recognition model to complete speech transcription, generate initial text, and then input the initial text into a large language model to complete the standardized replacement of colloquial words, correction of suspected erroneous medical terms, and reconstruct the text according to a preset department-specific structured template to generate standardized structured raw handover information text.

[0153] For example, taking the pediatric nursing night shift handover scenario as an example, the working process of the terminal data acquisition module is explained. Pediatric night shift handover usually involves information such as changes in the condition of children at night, emergency treatment, and execution of medical orders. It is characterized by suddenness and fragmented information. The night shift nurse enters the handover information through a tablet data acquisition terminal at the pediatric inpatient nursing station. Some emergency treatment information is collected by voice. The information acquisition unit first completes the nurse's identity verification and binds the data source access permissions for the pediatric night shift. Then, it receives the nurse's voice collection of "Bed 5, Xiaoyu, 2 years old, acute diarrhea, vomited at 2 am, vomited 3 times in total, vomited gastric contents, given oral rehydration salts, the vomiting has been relieved, but diarrhea still exists, 5-6 times a day, watery stool, body temperature monitored at night 37.8℃". At the same time, it obtains the handover data source metadata of the pediatric night shift from the data source connection pool, including the pediatric emergency database connection configuration and four handover data tables: patient basic information, emergency treatment, vital signs, and nursing operations.

[0154] After receiving the speech stream, the speech processing unit first performs noise reduction processing, then calls an automatic speech recognition model specifically for pediatric medical fields to transcribe it, generating the initial text: "Bed 5, Xiaoyu, 2 years old, acute diarrhea, vomited at 2 am, vomited 3 times, all food, drank rehydration salts, now vomiting less, still has diarrhea 5-6 times a day, watery stools, temperature 37.8 degrees Celsius at night"; then the initial text is input into a large language model for standardization processing, replacing colloquial words with professional terms, such as "vomited at 2 am" to "vomited at 02:00 am", "vomited 3 times, all food" to "vomited 3 times in total, vomit consisted of stomach contents", "drank rehydration salts" to "administered oral rehydration salt solution III", "still has diarrhea, 5-6 times a day, watery stools, temperature 37.8 degrees Celsius at night"; then the initial text is input into a large language model for standardization processing, replacing colloquial words with professional terms, such as "vomited at 2 am" to "vomited at 02:00 am", "vomited 3 times, all food", "vomited 3 times in total, vomit consisted of stomach contents", "administered rehydration salts", "administered oral rehydration salt solution III", "still has diarrhea, 5-6 times a day, watery stools, temperature 37.8 degrees Celsius at night". The phrase "5-6 times, watery stool" was replaced with "Still has diarrhea, 5-6 times a day, stool is watery," and "Nighttime temperature 37.8 degrees Celsius" was replaced with "Nighttime axillary temperature 37.8 degrees Celsius." After checking for medical terminology errors, the text was reconstructed according to the pediatric nursing handover structured template to generate standardized structured original handover information text: Bed No. 5, Name: Xiaoyu, Age: 2 years old, Diagnosis: Acute diarrhea, Current condition: Vomiting occurred at 02:00 AM, vomiting a total of 3 times, vomitus was gastric contents, vomiting symptoms were relieved after oral rehydration solution III was administered, diarrhea persists, 5-6 times a day, stool is watery; Vital signs: Nighttime axillary temperature 37.8 degrees Celsius; Nursing measures: Oral rehydration solution III was administered, vomiting and diarrhea were closely observed. The information collection unit encrypted and stored this structured text along with the handover data source metadata, completing the entire workflow of the terminal collection module.

[0155] Furthermore, the intelligent allocation completion module includes a fingerprint generation unit, a model building unit, an information completion unit, and a path configuration unit. The fingerprint generation unit performs structural and semantic analysis on each handover data table in the handover data source metadata, generating structural and semantic fingerprints using the MD5 and SimHash algorithms respectively, providing a foundation for constructing the handover asset meta-model. The model building unit calculates the fusion similarity between data tables, determines the adaptive clustering threshold using the silhouette coefficient method, and completes cluster merging to construct the handover asset meta-model. Subsequently, it combines this with the medical handover standard knowledge base to generate a semantically enhanced model. The semantic enhancement model comprises three parts: a semantic tag set, field mapping rules, and a handover asset meta-model. The information completion unit parses the structured original handover information based on the semantic enhancement model, determines the list of missing related fields, retrieves the corresponding data values ​​from the hospital database, completes the filling, and outputs the completed handover information. The path configuration unit combines the completed handover information with the regional dynamic handover information network, extracts the current handover task features, matches task logic subnet nodes, evaluates the comprehensive cost of each historical data collection path, selects the path with the lowest comprehensive cost as the optimal path, and sends it to the terminal collection module.

[0156] For example, using the structured original handover information of Xiaoyu in pediatric bed 5 as an example, the working process of the intelligent allocation and completion module is explained. The fingerprint generation unit first performs structural and semantic fingerprint generation on the four handover data tables in the pediatric night shift handover data source metadata: the patient basic information table, the emergency treatment table, the vital signs data table, and the nursing operation record table. For each table, structural information such as field names, data types, and primary key and foreign key relationships are extracted, concatenated into a string, and then the structural fingerprint is generated by the MD5 algorithm. Semantic parsing is performed on each field name, and synonyms and hypernyms are obtained by combining the pediatric medical knowledge graph. After being converted into vectors, the semantic fingerprint is generated by the SimHash algorithm.

[0157] The model building unit calculates the fusion similarity between the four data tables, setting a structural similarity weight of 0.4 and a semantic similarity weight of 0.6. The calculated fusion similarity of all four data tables is higher than 0.7. The silhouette coefficient method is used to determine the adaptive clustering threshold as 0.7, and the four data tables are grouped into the same cluster. After pattern merging, a meta-model of pediatric emergency department handover assets is constructed, integrating the common fields, data types, and semantic descriptions of the four tables. Subsequently, a pre-trained large language model in the medical field is called to parse the meta-model and identify the core medical entity classes. The model is designed for pediatric emergency patients, acute diarrhea, vomiting, vital signs, and fluid resuscitation. It retrieves standard handover fields, mandatory field rules, and logical dependencies between fields from the medical handover standard knowledge base. This generates a set of semantic tags and field mapping rules for the handover asset meta-model. For example, the semantic tags for "acute diarrhea" are "pediatric emergency, gastrointestinal disease, dehydration risk", and the field mapping rule is "acute diarrhea - association - stool routine examination, electrolyte examination, fluid resuscitation dosage". Finally, a semantic enhancement model specifically for pediatric emergency patients is generated.

[0158] The information completion unit uses this semantic enhancement model to parse the structured original handover information of Xiaoyu in bed 5, identifying the medical entity identifier corresponding to the filled fields as "Bed No. 5-Xiaoyu-Acute Diarrhea". Based on the field mapping rules, it determines the list of related fields that need to be completed, including the child's gender, age, admission time, stool routine examination results, oral rehydration salts (ORS) III solution dosage, nursing level, responsible nurse, and electrolyte test orders to be executed. Subsequently, based on the handover data source metadata, it connects to the pediatric emergency database and retrieves the corresponding data values ​​from the four handover data tables through a multi-table join SQL query: gender male, age 2 years, admission time 1 day ago, stool routine examination upon admission showed leukocytes + / HP, red blood cells negative, ORS III solution dosage 50ml / time, 3 times daily, nursing level level 2, responsible nurse is Nurse Li, and orders to be executed. The doctor's order was for an electrolyte test. These data values ​​were then filled into the corresponding missing fields of the original structured handover information to generate the complete handover information: Bed No. 5, Name: Xiaoyu, Gender: Male, Age: 2 years, Diagnosis: Acute diarrhea, Admission Time: 1 day prior, Admission Examination: Stool routine showed leukocytes + / HP, red blood cells negative; Current condition: Vomiting occurred at 02:00 AM, a total of 3 times, vomitus consisted of gastric contents. After oral rehydration solution III was administered, the vomiting symptoms subsided, but diarrhea persisted, 5-6 times daily, stool consistency was watery; Vital signs: Nighttime axillary temperature 37.8℃; Nursing level: Level II nursing, Responsible nurse: Nurse Li; Nursing measures: Administer oral rehydration solution III, closely observe vomiting and diarrhea, record stool frequency and consistency; Outgoing orders: Electrolyte test; Nursing precautions: Closely monitor the child's mental state and urine output, be alert for dehydration.

[0159] The path configuration unit, combining the completed handover information, extracts the characteristics of the current handover task: the department is the Pediatric Emergency Inpatient Department, the patient's condition level is ordinary, the required data types are basic information of emergency patients, treatment data, vital sign data, and nursing operation data, and the estimated data volume is small batch; this characteristic is matched with the task logic subnet of the regional dynamic handover information network to determine that the current task node is the historical handover task node of "acute diarrhea single patient small batch data collection" in the Pediatric Emergency Department; then, the comprehensive cost of the four historical data collection paths starting from this node is evaluated. The calculation is based on historical latency, network bandwidth utilization, and data acquisition success rate, with weights of 0.3, 0.4, and 0.3, respectively. The calculated comprehensive cost of path 4 is the lowest at 0.38. This path is: Pediatric Emergency Department "Single Patient Small Batch Data Collection for Acute Diarrhea" task node - Pediatric Emergency Department dedicated network exchange node - Pediatric Emergency Department data source server - Pediatric Emergency Department database storage resources. The corresponding hash fingerprint security control key is valid. The path configuration unit selects this path as the optimal path and distributes it to the tablet acquisition terminal at the pediatric nurse station to complete the path configuration.

[0160] Furthermore, the early warning control module includes a trajectory monitoring unit, a semantic analysis unit, a rule engine unit, and a path adjustment unit. The trajectory monitoring unit monitors the actual trajectory of the data collected by the terminal acquisition module along the optimal path in real time, calculates the path deviation between the actual trajectory and the preset trajectory through a path matching algorithm, and provides real-time feedback on the path acquisition status. The semantic analysis unit calls a large language model to perform deep semantic analysis on the completed handover information, and combines the department's medical knowledge graph and the patient's historical treatment data to identify abnormal expressions such as logical contradictions, violations of medical knowledge graph associations, and differences from historical data trends. The rule engine unit pre-stores error early warning rules of different levels, takes the path deviation and abnormal expressions as input, judges based on trigger conditions, and outputs early warning information of the corresponding level. If the path adjustment unit receives early warning information exceeding a preset threshold, it feeds back the early warning result containing abnormal path identifiers or abnormal data fields to the intelligent allocation and completion module, which performs dynamic adjustment of the optimal path, and simultaneously sends the early warning information to the terminal acquisition module to remind medical staff to check and correct.

[0161] For example, taking the optimal path and completed handover information of Xiaoyu in pediatric bed 5 as an example, the working process of the early warning control module is explained. The trajectory monitoring unit first obtains the preset trajectory of the optimal path issued by the path configuration unit: Pediatric Emergency Department "Single Patient Small Batch Data Collection for Acute Diarrhea" task node - Pediatric Emergency Department dedicated network exchange node - Pediatric Emergency Department data source server - Pediatric Emergency Department database storage resources. Then, through the hospital's intranet monitoring system, it tracks the actual data collection trajectory of the tablet acquisition terminal along this path in real time. During the actual collection process, the terminal acquisition module briefly switches to the hospital's backup pediatric data source server due to temporary maintenance of the pediatric emergency department data source server. The actual trajectory is: Pediatric Emergency Department "Single Patient Small Batch Data Collection for Acute Diarrhea" task node - Pediatric Emergency Department dedicated network exchange node - Hospital backup pediatric data source server - Pediatric Emergency Department database storage resources. By comparing the node overlap and path consistency, the trajectory monitoring unit calculates that the path deviation between the actual trajectory and the preset trajectory is 0.3.

[0162] The semantic analysis unit invoked a large language model to perform deep semantic analysis on the completed handover information. First, it combined the pediatric medical knowledge graph to verify whether the medical terminology and treatment measures in the information conformed to the standards. It confirmed that the measures taken for children with acute diarrhea, such as oral rehydration salts III solution, monitoring urine output and mental status, were in accordance with the treatment guidelines and there were no expressions that violated the association relationships of the knowledge graph. Then, it combined the child's historical medical data to compare the current condition with the historical data trend. When the child was admitted to the hospital, he had diarrhea 8-10 times a day, accompanied by vomiting. Currently, the diarrhea has decreased to 5-6 times a day, the vomiting has subsided, and the body temperature is 37.8℃, showing an improving trend. There were no expressions that did not conform to the historical data trend. Finally, the information was logically verified to confirm that there were no inconsistencies or logical conflicts between the fields. Therefore, no abnormal expressions were identified.

[0163] The rule engine unit takes the path deviation of 0.3 calculated by the trajectory monitoring unit and the no-abnormality statement result from the semantic analysis unit as input, and judges according to the preset warning rules. In this embodiment, the first threshold is set to 0.2. The path deviation of 0.3 exceeds the first threshold but does not exceed 0.5, and there is no abnormality statement. Therefore, a level 1 warning is triggered. The generated warning information is as follows: Warning level: Level 1, Abnormality type: Path acquisition trajectory abnormality, Abnormality location: Pediatric emergency data source server, Abnormality details: The pediatric emergency data source server is undergoing temporary maintenance. The terminal acquisition module has switched to the hospital's backup pediatric data source server for data acquisition. The path deviation is 0.3, and there is no handover information abnormality statement.

[0164] After receiving the Level 1 warning information, the path adjustment unit only sends the warning information to the tablet data acquisition terminal at the pediatric nursing station to remind nurses to pay attention to the data acquisition status, since the warning level does not exceed the preset second threshold. The warning result is not fed back to the intelligent allocation and completion module, and no optimal path adjustment is performed. The trajectory monitoring unit continuously monitors the acquisition trajectory of the terminal acquisition module. When the pediatric emergency data source server is maintained, the terminal acquisition module returns to the preset trajectory, the path deviation returns to 0, and the trajectory monitoring unit sends the path return to normal to the terminal acquisition module, and the Level 1 warning is lifted.

[0165] In another exemplary scenario, if the completed handover information contains the phrase "Children with acute diarrhea are given ceftriaxone sodium intravenously, 2g, once daily," the semantic analysis unit will immediately identify this statement as a serious aberration. This is because, according to the pediatric medical knowledge graph, antibiotics are not necessary for acute diarrhea without evidence of bacterial infection, and the maximum single dose of ceftriaxone sodium for a 2-year-old child is 1g. This statement violates the medical knowledge graph's association relationships and contains a drug dosage error, potentially leading to serious medical risks. Upon inputting this serious aberration, the rule engine unit triggers a level-three warning, and the path adjustment unit will include aberration data fields, such as the antibiotic name. The warning results regarding the dosage are fed back to the intelligent allocation completion module, and the level 3 warning information is sent to the terminal acquisition module to remind medical staff to immediately verify and correct it. After receiving the warning results, the intelligent allocation completion module marks the hash fingerprint security control key of the current optimal path as invalid and removes it from the set of available paths. Then, it re-evaluates the overall cost of the remaining paths, configures a new optimal path and sends it to the terminal acquisition module. At the same time, based on the abnormal data field information, it re-retrieves the infection indicator test results of the child from the hospital database. After confirming that there is no evidence of bacterial infection, it reminds medical staff to delete the erroneous medical order information to ensure the accuracy of handover information and medical safety.

[0166] This invention discloses a method and system for intelligent handover information completion and error early warning based on a large language model. Through a series of core steps detailed in the embodiments, including multimodal raw information preprocessing, construction of a handover asset meta-model and semantic enhancement model, intelligent field completion, optimal path configuration, and multi-dimensional error early warning, it effectively solves industry pain points in medical handover processes such as incomplete information, non-standard expressions, incorrect terminology, unreasonable data collection paths, and untimely error detection. Especially for pediatric nursing handovers, which involve young patients, rapidly changing conditions, frequent colloquial expressions, and high professional requirements, the system achieves standardized, structured, and intelligent completion of handover information through pediatric-specific model training, corpus optimization, and standardization adaptation. Simultaneously, the system ensures the efficiency and security of data collection paths through a regional dynamic handover information network and hash fingerprint security control keys. Combined with large language model semantic analysis and a rule engine, it achieves comprehensive, multi-level early warning of handover information errors and abnormal collection paths, as well as dynamic adjustment of the optimal path.

[0167] This invention deeply integrates large language models with medical handover operations, constructing a full-process, intelligent, and highly secure handover information processing system. This system improves the completeness, accuracy, and standardization of medical handover information, enhances the efficiency and quality of handover work, reduces medical risks caused by errors in handover information, and provides reliable information support for medical staff's diagnosis and nursing work. It has significant practical value for improving the hospital's medical quality management level and ensuring patient medical safety.

[0168] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of the present invention without departing from the spirit and scope of the present invention. All of these variations are within the protection scope of the present invention.

Claims

1. A handover information intelligent completion and error early warning system based on a large language model, characterized in that, include: Terminal acquisition module, intelligent allocation and completion module, early warning and control module; The terminal acquisition module collects raw shift handover information and shift data source metadata. The handover data source metadata includes database connection configuration and handover data table; For each handover data table, generate structural and semantic fingerprints, calculate fusion similarity and cluster and merge under an adaptive threshold to construct a handover asset meta-model, perform semantic parsing on the handover asset meta-model, generate a semantic enhancement model in combination with medical handover specifications, and complete the fields of the original handover information to obtain the completed handover information. The intelligent allocation and completion module combines the completed handover information with the regional dynamic handover information network to configure the optimal path for the terminal acquisition module. The regional dynamic handover information network consists of two subnets connected by the historical optimal path. The connection relationship is configured with a hash fingerprint security control key to adjust the acquisition and completion trajectory in case of anomalies. The early warning control module, based on the optimal path and the completed handover information, combines semantic analysis of the large language model and the rule engine to execute error warnings. The warning results are fed back to the intelligent allocation and completion module to dynamically adjust the optimal path.

2. The intelligent handover information completion and error early warning system based on a large language model as described in claim 1, characterized in that, The terminal acquisition module is further configured as follows: Obtain text, voice, or pre-filled structured forms entered by medical staff through terminals as raw handover information; Obtain handover data source metadata from the pre-configured data source connection pool; the handover data source metadata includes the database connection address, access credentials, data mode, and one or more handover data tables; the handover data tables store patient diagnosis and treatment records, medical orders, test results, and vital signs data.

3. The intelligent handover information completion and error early warning system based on a large language model as described in claim 2, characterized in that, The generation of structural and semantic fingerprints for each shift handover data table includes: For each shift handover data table, perform structure parsing to extract structure information. Concatenate the extracted structure information into a string according to the field order, and use a first preset hash algorithm to calculate the string to generate the structure fingerprint of the shift handover data table. The structure information includes field names, field data types, primary key and foreign key relationships, and inter-table association mappings. Semantic parsing is performed on each shift handover data table. The textual meaning of the field names is parsed based on natural language processing technology. Synonyms and hypernyms are obtained by combining medical knowledge graphs. The textual meaning is transformed into a vector representation and a semantic fingerprint is generated through a second preset hash algorithm.

4. The intelligent handover information completion and error early warning system based on a large language model as described in claim 3, characterized in that, The calculation of fusion similarity and clustering and merging under an adaptive threshold to construct a handover asset meta-model includes: For any two handover data tables, calculate the structural similarity score based on their structural fingerprints and the semantic similarity score based on their semantic fingerprints. Use preset weight coefficients to weight and fuse the structural similarity score and the semantic similarity score to obtain the fused similarity. The similarity between each pair of all handover data tables is used to construct a similarity matrix. The silhouette coefficient method is used to perform cluster analysis on the similarity matrix, and the adaptive clustering threshold that optimizes the clustering effect is dynamically determined. All shift handover data tables with a similarity greater than the adaptive clustering threshold are grouped into the same cluster, and all shift handover data tables within the same cluster are merged to generate a unified shift handover asset meta-model. The shift handover asset meta-model integrates the common fields, data types, and semantic descriptions of all shift handover data tables within the cluster.

5. The intelligent handover information completion and error early warning system based on a large language model as described in claim 4, characterized in that, Semantic parsing is performed on the aforementioned handover asset meta-model, and a semantically enhanced model is generated by combining it with medical handover standards, including: The handover asset meta-model is analyzed by calling a pre-trained large language model in the medical field to identify the core medical entity types and relationships between entities represented by the handover asset meta-model. Query the standard handover field definitions, mandatory field rules, and logical dependencies between fields related to the core medical entity type from the pre-built medical handover standard knowledge base; Based on the parsing results of the large language model and the query results of the medical handover standard knowledge base, a semantic tag set and a set of field mapping rules are generated for each handover asset meta-model; the semantic tag set, field mapping rules and the handover asset meta-model together constitute a semantic enhancement model.

6. The intelligent handover information completion and error early warning system based on a large language model as described in claim 5, characterized in that, The process of completing the fields in the original shift handover information to obtain the completed shift handover information includes: The semantic enhancement model is used to parse the original handover information to identify the filled fields and their corresponding medical entity identifiers; For the identified medical entity identifier, the list of associated fields that need to be completed for the entity is determined according to the field mapping rules in the semantic enhancement model; Based on the handover data source metadata obtained by the terminal acquisition module, the system connects to the corresponding database and uses a multi-table join SQL query statement based on the medical entity identifier to retrieve the specific data values ​​corresponding to the list of associated fields from the handover data table. The retrieved data values ​​are filled into the corresponding missing fields of the original shift handover information to form the completed shift handover information.

7. The intelligent handover information completion and error early warning system based on a large language model as described in claim 6, characterized in that, The regional dynamic shift handover information network includes a task logic subnet and a physical resource subnet. Nodes in the task logic subnet represent historical shift handover tasks of different departments or wards, while nodes in the physical resource subnet represent data source servers, network exchange nodes, and database storage resources. The nodes in the task logic subnet and the nodes in the physical resource subnet are connected by directed edges. Each directed edge represents a historical data collection path and is associated with a hash fingerprint security control key used to identify the current security status of that path.

8. The intelligent handover information completion and error early warning system based on a large language model as described in claim 7, characterized in that, Configuring the optimal path for the terminal acquisition module includes: The completed handover information is analyzed to extract the features of the current handover task, including the department, the severity of the patients involved, the data type required, and the estimated data volume. The characteristics of the current shift handover task are matched with the task logic subnet in the regional dynamic shift handover information network to determine the current task node; The overall cost is evaluated from the current task node, through the directed edges of each historical data acquisition path, to the target resource node in the physical resource subnet; the overall cost is calculated based on the weighted sum of the historical calculation delay of the path, the network bandwidth utilization rate, and the data acquisition success rate. The historical data acquisition path with the lowest overall cost is selected as the optimal path for configuring the terminal acquisition module.

9. The intelligent handover information completion and error early warning system based on a large language model as described in claim 8, characterized in that, The early warning control module, based on the optimal path and the completed shift handover information, combines large language model semantic analysis and rule engine to execute error warnings, including: The actual trajectory of the data collected by the terminal acquisition module along the optimal path is monitored, and the path deviation between the actual trajectory and the preset trajectory of the optimal path is calculated. The large language model is invoked to perform semantic analysis on the completed handover information to identify abnormal expressions in the information; the abnormal expressions include expressions that are logically contradictory, violate the established medical knowledge graph association relationship, or differ from the trend of the patient's historical data; The path deviation and the abnormal expression are input into the rule engine; the rule engine has pre-stored warning rules of different levels. When the input deviation exceeds the first threshold or a specified type of abnormal expression is identified, the corresponding level of warning information is triggered.

10. The intelligent handover information completion and error early warning system based on a large language model as described in claim 9, characterized in that, The early warning results are fed back to the intelligent allocation and completion module to dynamically adjust the optimal path, including: When the warning level triggered by the warning control module exceeds the preset second threshold, the intelligent allocation completion module receives a warning result containing an abnormal path identifier or an abnormal data field; the second threshold is greater than the first threshold; The intelligent allocation completion module marks the corresponding hash fingerprint security control key status as invalid based on the abnormal path identifier, and temporarily removes the path from the current set of available paths; The intelligent allocation and completion module recalculates the overall cost based on the updated set of optional paths, generates a new optimal path, and sends it to the terminal acquisition module.

Images