Intelligent Management System for the Entire Vascular Access Process Based on a Large Model

Abstract

This invention discloses a large-scale model-based intelligent management system for the entire vascular access process, comprising: a data integration and knowledge graph construction module for building a dynamic knowledge graph of vascular access; a multimodal state prediction and evolution analysis module for identifying key influencing paths and key nodes, and generating state anomaly warning signals; a real-time operation semantic understanding and compliance judgment module for performing semantic understanding and compliance judgment on real-time operation flow data; an adaptive decision optimization and task collaboration module for generating maintenance plan adjustment parameters for vascular access based on state anomaly warning signals with real-time reminder instructions, and updating the nursing task list for the target patient; and a multidimensional visualization interaction and traceability display module for generating an interactive multidimensional visualization view of the entire life cycle of vascular access based on the dynamic knowledge graph, prediction scores, key influencing paths, and maintenance plan adjustment parameters.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, specifically to an intelligent management system for the entire process of vascular access based on a large model. Background Technology

[0002] Vascular access (such as central venous catheters, PICC lines, and arterial fistulas) is a lifeline for critically ill patients, patients requiring long-term intravenous infusion (such as those undergoing chemotherapy or parenteral nutrition), and patients on hemodialysis. Its establishment, maintenance, monitoring, and use are crucial throughout the entire patient care process.

[0003] Currently, the clinical management of vascular access mainly faces the following problems and shortcomings:

[0004] Information regarding vascular access placement, daily assessments, usage records, complication management, and removal is scattered across multiple systems, including electronic medical records, nursing records, and imaging reports. This lack of synchronization between doctors, nurses, and technicians, and the absence of a unified view of the entire process, leads to delayed and biased management decisions.

[0005] Risk assessment for complications such as catheter-related bloodstream infections, thrombosis, and catheter displacement relies primarily on the personal experience of healthcare professionals and regular checkups. It lacks dynamic, quantitative risk warning models based on real-time data, often resulting in reactive intervention only after complications have occurred. This passive monitoring model, lacking dynamic, quantitative risk warning models based on real-time clinical data (such as vital signs, laboratory indicators, and imaging characteristics), often misses the window of opportunity for early intervention.

[0006] The detection of operational deviations usually relies on post-operative quality control, lacking real-time, accurate compliance judgment and immediate reminders. For example, the assessment of catheter puncture sites, the timing of dressing changes, and the operation procedures for flushing and sealing catheters, although there are guidelines, rely entirely on manual verification and recording in busy clinical work, making it difficult to ensure that all operations meet the standards 100%, which poses a medical safety hazard.

[0007] The vast amount of vascular access-related data (procedure records, test results, imaging features, etc.) is scattered across various isolated systems, failing to be effectively integrated and analyzed. This prevents the provision of data-driven decision support for individualized access selection, extubation timing, and quality improvement.

[0008] While some existing electronic medical record systems integrate simple recording modules or provide auxiliary equipment for single steps (such as ultrasound-guided puncture), they lack a comprehensive management system that covers the entire process of vascular access—from establishment and use to maintenance and removal—and enables intelligent early warning, decision support, and standardized quality control. Therefore, solving these problems has become a pressing technical challenge for those skilled in the art. Summary of the Invention

[0009] The purpose of this invention is to provide a large-scale intelligent management system for the entire vascular access process, in order to solve the problems of fragmented vascular access management processes, delayed complication early warning, difficulty in unifying the execution of operating procedures, and lack of data-driven global decision support in the existing technology.

[0010] To address the aforementioned technical problems, in a first aspect, the present invention provides an intelligent management system for the entire vascular access process based on a large model, comprising: The data integration and knowledge graph construction module is used to collect multi-source heterogeneous data related to the vascular access of the target patient in real time, preprocess the collected multi-source heterogeneous data, and construct a dynamic knowledge graph indexed by the unique identifier of the vascular access and covering the entire process from insertion to removal. The multimodal state prediction and evolution analysis module incorporates a large-scale artificial intelligence model based on the Transformer architecture. This model extracts feature vectors from the dynamic knowledge graph, including textual, image, and time-series numerical modalities, and performs cross-modal fusion to generate a multidimensional state representation of the vascular pathway. This multidimensional state representation is then input into the time-series prediction layer to calculate the prediction score for vascular functional disorders occurring within a preset prediction time window. Simultaneously, the module uses a graph neural network to perform evolutionary analysis on entity relationships in the dynamic knowledge graph, identifying key influencing paths and critical nodes that lead to increased prediction scores, and generating early warning signals for abnormal states. The real-time operation semantic understanding and compliance judgment module is used to acquire real-time operation flow data generated by medical staff when performing vascular access operations. The real-time operation flow data is input into the operation semantic understanding model built based on temporal convolutional network, and operation time sequence features are extracted. The operation time sequence features are matched with standardized operation process nodes in the standard library, and the semantic deviation between the current operation and the standard operation process is calculated. If the semantic deviation exceeds the preset deviation tolerance, a real-time reminder instruction containing specific deviation steps and correction suggestions is generated and sent to the medical staff terminal. The adaptive decision optimization and task collaboration module is used to input the abnormal status warning signal and the real-time reminder instruction into the dynamic decision model based on deep reinforcement learning to generate maintenance plan adjustment parameters for vascular access; send the maintenance plan adjustment parameters to the nursing plan system to update the nursing task list of the target patient; and feed back the key influence paths and key nodes to the knowledge graph construction step to optimize the feature weight allocation in the subsequent dynamic knowledge graph construction process. The multidimensional visualization and traceability module is used to generate an interactive multidimensional visualization view of the entire life cycle of vascular access based on dynamic knowledge graphs, prediction scores, key impact paths, and maintenance plans. The interactive multidimensional visualization view allows users to trace the corresponding original data and historical operation records by clicking on any node.

[0011] Furthermore, the data integration and knowledge graph construction module includes: The data acquisition unit is used to acquire multi-source heterogeneous data related to the vascular pathways of the target patient in real time from at least two heterogeneous data sources. The data preprocessing unit is used to perform data cleaning, entity alignment, semantic mapping, and time series calibration on the collected multi-source heterogeneous data. The dynamic knowledge graph construction unit uses the Neo4j graph database as the underlying storage engine to build a dynamic knowledge graph. Nodes in the graph represent entities including target patients, vascular access routes, operational events, and clinical observation indicators, while edges represent the relationships between entities. The dynamic knowledge graph update mechanism employs an incremental calculation strategy. When new data flows in, it checks whether the corresponding unique identifier for a vascular access route already exists in the knowledge graph. If it does, the node and its relationships are updated or added; otherwise, a new vascular access route node is created. Version control is applied to attributes, and the weights of associated edges are updated using a dynamic adjustment formula with a decay factor. The dynamic adjustment formula is as follows: ; in, λ represents the weight of the original edge, Δw represents the contribution of the new event to the association strength, and λ is the decay factor.

[0012] Furthermore, the multi-source heterogeneous data includes insertion operation records, daily assessment records, usage and maintenance records, complication management records, removal records, laboratory test results, medical imaging images, and physiological monitoring signals.

[0013] Furthermore, the multimodal state prediction and evolution analysis module includes: The data input unit is used to input dynamic knowledge graphs into large-scale artificial intelligence models based on the Transformer architecture; The modality feature extraction unit is used to decompose the data in the dynamic knowledge graph into text modality, image modality, and time-series numerical modality according to modality type; and to extract feature vectors of text modality, image modality, and time-series numerical modality respectively using a multi-head attention mechanism; The feature fusion unit is used to interact feature vectors from different modalities using a cross-attention mechanism to generate a multidimensional state representation of the vascular pathway that integrates information from all modalities. The prediction score generation unit is used to input multi-dimensional state representations into the time-series prediction layer, calculate the probability values ​​of decreased tubing patency, weakened catheter fixation stability, and increased risk of puncture site leakage within a preset prediction time window, and generate corresponding prediction scores. The evolutionary analysis unit is used to update the feature representation of nodes in the dynamic knowledge graph using a graph attention network, identify key influence paths and key nodes that lead to an increase in the prediction score through a gradient-based interpretation method, and generate anomaly warning signals including key influence paths and key nodes.

[0014] Furthermore, the gradient-based interpretation method includes: calculating the gradient of the multidimensional state representation with respect to each node feature, using the norm of the gradient as the node importance score, and marking the nodes with the highest scores as key nodes; then using the shortest path algorithm to search for paths from key nodes to target patient nodes or vascular access nodes in the dynamic knowledge graph to obtain key influence paths.

[0015] Furthermore, the real-time operation semantic understanding and compliance determination module includes: The data acquisition unit is used to acquire real-time operation flow data generated by medical staff when performing vascular access procedures, and at the same time retrieve standardized operation process nodes that are pre-stored in the standard library. The temporal feature extraction unit is used to input the real-time operation flow data and standardized operation process nodes into the pre-trained operation semantic understanding model, and extract the real-time operation temporal features and standardized operation temporal features respectively. The compliance determination unit is used to calculate the optimal alignment path and cumulative distance of real-time operation timing features and standardized operation timing features in the time dimension using a dynamic time warping algorithm, and to use the cumulative distance as the semantic deviation. The semantic deviation is compared with a preset deviation tolerance. If the semantic deviation exceeds the deviation tolerance, the current operation is determined to be non-compliant. At this time, the specific steps that caused the deviation are back located by tracing back the optimal alignment path in the dynamic time warping algorithm. The reminder instruction generation unit is used to generate correction suggestions for each specific deviation step according to the predefined mapping relationship in the specification library; and send the real-time reminder instruction containing the specific deviation step and correction suggestions to the medical terminal.

[0016] Furthermore, real-time operation stream data includes operation timestamps, operator identifiers, operation types, and information on consumables used.

[0017] Furthermore, the dynamic decision-making model based on deep reinforcement learning adopts a deep Q-network architecture; the state space of the dynamic decision-making model consists of the multidimensional state representation of the current vascular access, historical operation compliance statistics, and nursing plan execution status; the action space of the dynamic decision-making model includes the adjustment range of maintenance plan adjustment parameters; the goal of the dynamic decision-making model is to minimize the long-term accumulated risk of tubing functional impairment and maximize operation compliance; the maintenance plan adjustment parameters include the assessment frequency adjustment value, dressing change cycle adjustment value, and flushing fluid dosage adjustment value.

[0018] Furthermore, the interactive multidimensional visualization view includes event distribution on the timeline, trend curves of pathway status, risk evolution tracing paths, and statistical charts of operational norms; The event distribution on the timeline is based on time, with the insertion operation record, daily assessment record, usage and maintenance record, complication handling record, and removal record marked as event nodes on the timeline. The access status trend curve is plotted with the predicted score as the vertical axis and the future prediction time window as the horizontal axis, showing the changes in the predicted score as the tube patency decreases, the catheter fixation stability weakens, and the risk of leakage at the puncture point increases. The risk evolution tracing path visualizes key impact paths in a graph structure, highlighting key nodes and the relationships between nodes, and using color intensity to represent the importance score of nodes; Operational compliance statistics charts, presented in bar or pie chart format, show the frequency of various deviations in historical operations and the overall compliance rate, and compare the expected changes before and after maintenance plan adjustments.

[0019] The beneficial effects of this invention are as follows: By constructing a dynamic knowledge graph indexed by unique vascular access identifiers, multi-source heterogeneous data originally scattered across multiple isolated systems are deeply integrated and standardized, breaking down information barriers and establishing a unified, complete, and real-time updated data foundation for the full lifecycle management of vascular access. Based on the Transformer architecture and graph neural network-based multimodal state prediction model, it can not only proactively calculate the predicted scores of functional impairments such as decreased tubing patency and weakened fixation stability, but also automatically identify and visualize key influencing paths leading to increased scores, achieving a fundamental shift from passively responding to complications to proactively identifying risk evolution paths. By introducing an operation semantic understanding model based on temporal convolutional networks, deep semantic matching is performed between operation trajectory data collected by IoT sensors and standard operating procedures, enabling accurate, real-time judgment and intelligent reminders of clinical operation compliance, effectively ensuring the unified execution of operating standards. Combined with a deep reinforcement learning-based adaptive decision optimization module, it can dynamically generate personalized maintenance plan adjustment parameters, including assessment frequency and dressing change cycle, based on real-time predicted states and operational deviations, and directly coordinate with the nursing task system, realizing data-driven and closed-loop execution of management decisions. Ultimately, through a multi-dimensional visual interactive interface, the event distribution, risk trends, traceability paths, and operational procedures are integrated and presented, providing clinicians with an intuitive and traceable decision support tool. Attached Figure Description

[0020] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, use the same reference numerals to denote the same or similar parts. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0021] Figure 1 This is an architecture diagram of a vascular access end-to-end management system based on a large-scale artificial intelligence model. Detailed Implementation

[0022] Firstly, such as Figure 1 As shown, this invention discloses an intelligent management system for the entire vascular access process based on a large model, comprising: The data integration and knowledge graph construction module is used to collect multi-source heterogeneous data related to the vascular access of the target patient in real time, preprocess the collected multi-source heterogeneous data, and construct a dynamic knowledge graph indexed by the unique identifier of the vascular access and covering the entire process from insertion to removal. The multimodal state prediction and evolution analysis module incorporates a large-scale artificial intelligence model based on the Transformer architecture. This model extracts feature vectors from the dynamic knowledge graph, including textual, image, and time-series numerical modalities, and performs cross-modal fusion to generate a multidimensional state representation of the vascular pathway. This multidimensional state representation is then input into the time-series prediction layer to calculate a prediction score for vascular functional impairment occurring within a preset prediction time window. Simultaneously, an evolutionary analysis of entity relationships in the dynamic knowledge graph is performed using a graph neural network to identify key influencing paths and critical nodes that lead to an increase in the predicted score, and an abnormal state warning signal is generated. The real-time operation semantic understanding and compliance judgment module is used to acquire real-time operation flow data generated by medical staff when performing vascular access operations. The real-time operation flow data is input into the operation semantic understanding model built based on temporal convolutional network, and operation time sequence features are extracted. The operation time sequence features are matched with standardized operation process nodes in the standard library, and the semantic deviation between the current operation and the standard operation process is calculated. If the semantic deviation exceeds the preset deviation tolerance, a real-time reminder instruction containing specific deviation steps and correction suggestions is generated and sent to the medical staff terminal. The adaptive decision optimization and task collaboration module is used to input the abnormal status warning signal and real-time reminder instruction into the dynamic decision model based on deep reinforcement learning to generate maintenance plan adjustment parameters for vascular access; send the maintenance plan adjustment parameters to the nursing plan system to update the nursing task list of the target patient; and feed back the key impact paths and key nodes to the knowledge graph construction step to optimize the feature weight allocation in the subsequent dynamic knowledge graph construction process. The multidimensional visualization and traceability module is used to generate an interactive multidimensional visualization view of the entire life cycle of vascular access based on dynamic knowledge graphs, prediction scores, key impact paths, and maintenance plans. The interactive multidimensional visualization view allows users to trace the corresponding original data and historical operation records by clicking on any node.

[0023] According to one embodiment of this application, the data integration and knowledge graph construction module includes:

[0024] The data acquisition unit is used to acquire multi-source heterogeneous data related to the vascular pathways of the target patient in real time from at least two heterogeneous data sources.

[0025] The data preprocessing unit is used to perform data cleaning, entity alignment, semantic mapping, and time series calibration on the collected multi-source heterogeneous data.

[0026] The dynamic knowledge graph construction unit uses Neo4j graph database as the underlying storage engine to build a dynamic knowledge graph. Nodes in the graph represent entities (including target patients, vascular access, operation events, clinical observation indicators, etc.), and edges represent the relationships between entities (including insertion relationships, maintenance relationships, complication relationships, temporal relationships, etc.). The dynamic knowledge graph update mechanism adopts an incremental calculation strategy. When new data (such as nursing records) flows in, it is determined whether the unique identifier of the vascular access corresponding to the record already exists in the knowledge graph. If it exists, a new attribute value is added to the corresponding event node, or a new operation event node is created using the CREATE statement and a temporal relationship is established with the original graph. If it does not exist, a new vascular access node is created, and it is associated with the correct patient node based on the demographic characteristics of the target patient.

[0027] When updating attributes, a version control mechanism is used, and each attribute value is appended with a start time validity period. from and end time valid to The end time is empty by default. When the attribute value changes, the end time of the original record will be updated to the current time, and the start time of the new record will be the current time.

[0028] When the association relationship is updated, the temporal association strength is recalculated based on the timestamp of the new data, and the weights of the original association edges are adjusted; the dynamic adjustment formula is:

[0029] ;

[0030] in, The weights of the original edges, The value of the contribution of the new event to the association strength is λ, which is the decay factor. If the weight of the associated edge is lower than the threshold of 0.1 after adjustment, the associated edge is deleted. The entire update process is implemented through the Apache Flink stream processing engine and is completed within seconds, ensuring that the dynamic knowledge graph always reflects the latest state and complete historical trajectory of the vascular pathway.

[0031] According to one embodiment of this application, the multi-source heterogeneous data includes insertion operation records, daily assessment records, usage and maintenance records, complication management records, removal records, laboratory test results, medical imaging images, and physiological monitoring signals.

[0032] Real-time acquisition of multi-source heterogeneous data includes: establishing connections with the electronic medical record system via an interface compliant with the HL7 FHIR standard; connecting with the nursing record system via ODBC or JDBC interfaces; connecting with the image archiving system via the DICOM protocol; and connecting with bedside medical devices via the HL7 v2 or MQTT protocol. The data acquisition process uses Apache Kafka as the message queue technology, supporting high-throughput real-time data stream access. This ensures the real-time capture of insertion operation records, daily assessment records, usage and maintenance records, complication management records, removal records, laboratory test results, medical images, and physiological monitoring signals from the electronic medical record system, nursing records, image archiving system, and bedside medical devices. For each piece of raw data acquired, the system automatically adds a timestamp and data source identifier, forming tamper-proof data traceability information. Data acquisition rate control is achieved through an adaptive flow control algorithm, which dynamically adjusts the production end's sending rate based on the consumer's processing capacity to avoid data backlog or system overload.

[0033] Data cleaning includes: using an outlier detection algorithm based on isolated forests to remove outliers in numerical fields, filling missing values ​​with multiple imputation methods, and using BERT-based natural language processing technology to identify and correct spelling errors in text fields.

[0034] The Isolation Forest algorithm partitions the data by constructing random trees, and the path length for partitioning outliers is typically short. The formula for calculating the outlier score 's' is:

[0035] ;

[0036] Where x is the sample point to be detected, n is the total number of samples, E(h(x)) is the expected path length of sample point x in all random trees, and c(n) is the average path length of the binary search tree for a given number of samples n. The formula for calculating c(n) is:

[0037] ;

[0038] Where γ is Euler's constant, with a value of 0.5772156649. When s(x, n) is greater than the preset threshold of 0.6, the sample point is determined to be an outlier and is removed; missing values ​​are filled using multiple imputation, and the imputation model is constructed based on the historical data distribution of similar patient groups; multiple imputation is implemented using a chain equation, and for the j-th variable with missing values, a regression model is established:

[0039] ;

[0040] in, For the variable to be interpolated, For other complete variables, This is the intercept term (constant term). The regression coefficient, ε, is the random error term, and ε follows a normal distribution with a mean of 0. m complete datasets are generated through iterative sampling, where m is 5. The final interpolated value is the mean of the m simulation results. For text fields in nursing records, BERT-based natural language processing technology is used to identify and correct spelling errors, standardizing the expression of medical terminology. Spelling correction employs the edit distance algorithm, calculated using the following formula:

[0041] ;

[0042] Here, a and b are two strings, and i and j are the current character positions of a and b, respectively. For indicator functions, when Not equal to The value is 1 if the edit distance is less than or equal to 2, otherwise it is 0; when the edit distance is less than or equal to 2, it is automatically replaced with standard words in the medical terminology ontology.

[0043] The goal of entity alignment is to match and merge the same entity from different data sources. The system uses the unique identifier of the vascular access as its core, combining this with demographic information such as the target patient's name, hospital number, and ID number to construct a unique fingerprint for each entity. For medical images in the image archiving system, image fingerprints are generated using a perceptual hash algorithm. The perceptual hash value is calculated using discrete cosine transform, with the following formula:

[0044] ;

[0045] in, Let be the pixel value of the image at coordinates (x, y), N be the image size, and c(i) and c(j) be the normalization coefficients. When i = 0, ... Otherwise, c(i) = 1; take the low-frequency coefficients to generate a hash sequence and associate it with the image examination records in the electronic medical record system; during entity alignment, a hybrid similarity calculation method based on edit distance and word vectors is adopted, and the formula for calculating the hybrid similarity Sim is:

[0046] ;

[0047] Where s1 and s2 are two text fields to be compared, and |s1| and |s2| are the string lengths, respectively. The edit distance is θ, which is the angle between the word vectors of the two text fields in the embedding space. cos(θ) is calculated by dividing the dot product by the modulus. α is the weight coefficient with a value of 0.4. When the mixed similarity Sim exceeds the preset threshold of 0.85, it is determined to be the same entity, and the entity identifier is unified.

[0048] Semantic mapping is used to convert fields from different data sources that express the same meaning but vary in form into unified standard terms. The system has a pre-built medical terminology ontology, which includes standardized concepts such as operation names, anatomical locations, complication types, and consumable specifications related to vascular access. For operation names in the collected insertion operation records, semantic matching based on a BERT pre-trained model is used to map them to standard operation terms in the ontology. The semantic matching process calculates the similarity between the input text and each standard term in the ontology. The similarity calculation formula is as follows:

[0049] ;

[0050] in, The embedding vector is the output of the last pooling layer of the BERT model from the input text. The embedding vectors of standard terms are used, and the cosine similarity is obtained by dividing the dot product by the modulus. The standard term with the highest similarity is taken as the mapping result, requiring the highest similarity to be no less than the threshold of 0.8. For the test item names in laboratory test results, mapping is performed according to the LOINC coding standard. The mapping adopts a rule-based method, matching keywords in the test item names through regular expressions, such as glucose corresponding to LOINC code 2345-7. For medical images, image features are extracted through a ResNet-152 convolutional neural network. The residual learning unit of ResNet-152 is represented as:

[0051] ;

[0052] Where x is the input feature map, y is the output feature map, and F(x, {Wi}) is the residual mapping to be learned, containing convolutional layers and batch normalization layers. The extracted feature vectors are mapped to labels in a standard image representation library through a fully connected layer, realizing the transformation from unstructured images to structured semantic labels; the matching of feature vectors and standard labels outputs a probability distribution through a softmax function, the formula of which is:

[0053] ;

[0054] Where z is the logits vector output by the fully connected layer, K is the total number of label categories, P(y=j|z) is the probability of being predicted as the j-th class, and the label corresponding to the maximum probability is taken as the mapping result.

[0055] Time series calibration is used to align all events onto a unified timeline to eliminate potential discrepancies in time bases between events recorded from different data sources. The system uses the standard time provided by the Network Time Protocol as a reference and converts the timestamp in each event record to Unix timestamp format. UnixDefined as the total number of seconds from 00:00:00 UTC on January 1, 1970 to the current time. For events with a clear temporal relationship, such as flushing and subsequent infusion, the time interval Δt = t is calculated between the events. post - t pre Verify the logical consistency of the timestamps; if a timestamp deviation is found to exceed the allowable range, i.e. Where δ is the allowable error threshold, with a value of 300 seconds, correction is performed based on the average interval between adjacent events; correction value The calculation formula is:

[0056] ;

[0057] in, For reference to the standard timestamp of the event, The average interval between similar events in historical data is denoted by , where n is the number of historical samples. After time series calibration, each event has a standardized time coordinate, ensuring the accuracy of time series correlations in subsequent analyses.

[0058] Each node in the dynamic knowledge graph contains a corresponding set of attributes. For example, the target patient node contains demographic features such as age, gender, and diagnostic information; the vascular access node contains vascular access attributes such as insertion site, catheter type, and insertion date; the operation event node contains the operator, operation time, and operation details for each operation event; and the clinical observation indicator node contains the specific values ​​of laboratory test results and the waveform characteristics of physiological monitoring signals. The knowledge graph is constructed by creating nodes and edges using Cypher statements. The basic operation for creating a node is CREATE (n:Patient {patient_id: 'P123', age: 65, gender: 'M'}), and the basic operation for creating an edge is CREATE (n)-[:HAS_CATHETER]->(c:Catheter {catheter_id: 'C456'}).

[0059] The relationships in a dynamic knowledge graph include not only static subordinate relationships but also dynamic temporal relationships. For example, the target patient node and the implantation operation node are connected by an implantation relationship edge; the implantation operation node and subsequent daily assessment record nodes are connected by a time sequence edge; and the laboratory test result node and the complication management record node are connected by a causal inference edge. These relationships are automatically discovered during construction using graph algorithms, such as co-occurrence analysis based on time windows. In co-occurrence analysis, the formula for calculating the association strength between two event nodes is:

[0060] ;

[0061] in, For event i With the event The co-occurrence weights are defined as follows: T is the total number of time windows, Δt is the length of the co-occurrence time window (24 hours), and I is an indicator function (1 for events and 0 for events). When the co-occurrence weight exceeds a threshold of 3, an association edge is created between the two event nodes. The PrefixSpan algorithm is used for sequence pattern-based association rule mining. The formula for calculating the support of a sequence pattern is:

[0062] ;

[0063] in, If a candidate sequence pattern has a support greater than 0.1, the pattern is retained, and a corresponding temporal association path is created in the knowledge graph based on the pattern.

[0064] According to one embodiment of this application, the multimodal state prediction and evolution analysis module includes:

[0065] The data input unit is used to input dynamic knowledge graphs into large-scale artificial intelligence models based on the Transformer architecture;

[0066] The modality feature extraction unit decomposes the data in the dynamic knowledge graph into text modality, image modality, and time-series numerical modality according to modality type. It then employs a multi-head attention mechanism to extract feature vectors for each modality. The text modality includes text descriptions in insertion operation records, daily assessment records, usage and maintenance records, complication management records, removal records, and laboratory test results. The image modality includes medical images. The time-series numerical modality includes physiological monitoring signals and numerical time-series data from laboratory test results. For the text modality, the system converts each text record into a text embedding sequence using a word segmenter and word embedding layer, and adds positional encoding to preserve sequence order. For the image modality, the system adjusts each medical image to a uniform size, extracts image feature maps using a convolutional neural network, and then flattens the feature maps into an image feature sequence. For the time-series numerical modality, the system divides the monitoring values ​​at consecutive time points into a time-series numerical sequence according to a fixed time window, and extracts local time-series features using a one-dimensional convolutional layer.

[0067] After completing the serialization representation of each modality, the system inputs the text sequence, image sequence, and time-series numerical sequence into the encoder of the Transformer architecture, respectively. The Transformer encoder contains multiple identical layers, each consisting of a multi-head self-attention sublayer and a feedforward neural network sublayer. The multi-head self-attention mechanism captures contextual information by calculating the correlation between elements within the sequence. Specifically, for the input sequence X, it is transformed into a query matrix Q, a key matrix K, and a value matrix V through three linear transformations. The formula for calculating the attention output is as follows:

[0068] ;

[0069] in, To determine the dimension of the query vector, the softmax function normalizes each row; multi-head attention concatenates the outputs of multiple single-head attention functions and then performs a linear transformation to obtain the final output, as shown in the formula:

[0070] ;

[0071] Among them, each head This is a learnable weight matrix. After multiple layers of encoding, the sequence of each modality is compressed into the corresponding modality feature vector, denoted as follows: , and ;

[0072] The feature fusion unit is used to interact feature vectors from different modalities using a cross-attention mechanism to generate a multidimensional state representation of the vascular pathway that integrates information from all modalities. Specifically, text modal features are used as queries, and image modal features are used as keys and values. Multi-head attention is used to obtain fused features of text and images. Similarly, the fused result is cross-attentioned with temporal modal features again to finally generate a multidimensional state representation that integrates all modal information. The calculation process for cross-attention is the same as that for self-attention, except that the query and key-value pair come from different modalities.

[0073] Predictive scoring generation unit, used to represent multidimensional states The input time-series prediction layer calculates the probability of decreased tubal patency, weakened catheter fixation stability, and increased risk of puncture site leakage occurring within a preset prediction time window, generating corresponding prediction scores. The time-series prediction layer consists of multiple fully connected networks, and the output layer uses a sigmoid activation function to generate a prediction score for each type of tubal functional impairment. The preset prediction time window can be set to the next 24 hours. The system calculates the probability of decreased tubal patency, weakened catheter fixation stability, and increased risk of puncture site leakage occurring within this window, denoted as follows: , and The formula for calculating each score is:

[0074] ;

[0075] in, For the sigmoid function, is the learnable weight vector; g is the learnable bias term.

[0076] The evolutionary analysis unit is used to update the feature representations of nodes in the dynamic knowledge graph using a graph attention network. It identifies key influence paths and key nodes that lead to increased prediction scores through a gradient-based interpretation method, and generates early warning signals for anomalies in the state of these key influence paths and key nodes. Specifically, it treats the dynamic knowledge graph as a graph... , where the set of nodes It includes target patient nodes, vascular access nodes, operation event nodes, clinical observation indicator nodes, etc., and an edge set. It includes various relationships. The system uses a graph attention network to update node features. The update formula is:

[0077] ;

[0078] in, For nodes The set of neighboring nodes, For nodes The original features, where W is a learnable weight matrix. The attention coefficient is calculated using a shared attention mechanism:

[0079] ;

[0080] in, For learnable attention vectors, This represents vector concatenation. After passing through a multi-layer graph attention network, each node obtains a feature representation that incorporates information from its neighbors.

[0081] According to one embodiment of this application, the gradient-based interpretation method includes: calculating a multidimensional state representation. For each node feature gradient The system calculates the importance score of a node by taking the norm of the gradient. The nodes with the highest scores are marked as key nodes. At the same time, the system uses Dijkstra's shortest path algorithm to search for paths from key nodes to target patient nodes or vascular access nodes in the dynamic knowledge graph. The path weights are determined by the node importance score and the edge strength. These paths are the key influence paths.

[0082] According to one embodiment of this application, the real-time operation semantic understanding and compliance determination module includes:

[0083] The data acquisition unit is used to acquire real-time operation flow data generated by medical staff when performing vascular access procedures, and at the same time retrieve standardized operation process nodes that are pre-stored in the standard library.

[0084] The temporal feature extraction unit is used to input real-time operation flow data and standardized operation process nodes into a pre-trained operation semantic understanding model to extract real-time operation temporal features and standardized operation temporal features, respectively. The operation semantic understanding model is built on a temporal convolutional network, which captures long-distance dependencies in the operation sequence through multi-layer causal convolution and dilated convolution. For the input real-time operation flow data, the model first encodes the operation type, consumable information, and operation trajectory data into a vector sequence, converts the operation timestamp into a relative time interval embedding vector, concatenates it with the operation sequence vector, and then inputs it into the temporal convolutional network. The temporal convolutional network gradually extracts high-level operation temporal features through multi-layer convolution operations, and finally outputs a fixed-dimensional vector as the overall semantic representation of the current operation sequence.

[0085] Standardized operating procedure nodes are extracted in a structured manner from clinical guidelines and operating specifications. Each node contains a semantic description of the standard operating steps, the standard operating sequence, the standard operating time range, and the standard operating trajectory pattern. For standardized operating procedure nodes, their semantic representation vectors are also calculated through forward propagation of the operating semantic understanding model and stored in the specification library for subsequent matching.

[0086] The compliance determination unit, after obtaining the semantic representation vector of the current operation sequence, matches it with the semantic representation vector of the standardized operation process nodes. The matching process employs a dynamic time warping algorithm to calculate the optimal alignment path and cumulative distance between the real-time operation sequence features and the standardized operation sequence features in the time dimension. The cumulative distance is used as the semantic deviation. The formula for calculating the cumulative distance is:

[0087] ;

[0088] Where i and j are the time step indices of the current operation sequence and the standardized operation process node, respectively; d(i,j) is the Euclidean distance between the semantic representation vectors corresponding to the two time steps; and D(i,j) is an element in the cumulative distance matrix. After dynamic time warping, the final cumulative distance D(m,n) is the semantic deviation between the current operation and the standard operation process, where m is the length of the current operation sequence and n is the length of the standardized operation process node.

[0089] The semantic deviation is compared with the preset deviation tolerance. If the semantic deviation exceeds the deviation tolerance, the current operation is determined to be non-compliant. At this time, the specific steps that caused the deviation are back located through the optimal alignment path in the dynamic time warping algorithm (e.g., insufficient disinfection time, incorrect flushing sequence, or missed scan). The deviation tolerance is dynamically adjusted according to different operation types (e.g., aseptic operation has a lower tolerance, while record filling has a higher tolerance).

[0090] The reminder instruction generation unit generates correction suggestions for each specific deviation step based on the predefined mapping relationship in the specification library (e.g., please re-disinfect the puncture site, please flush the tube before infusion, or please scan the consumable batch code); and sends the real-time reminder instruction containing the specific deviation step and correction suggestions to the medical staff terminal to remind medical staff to correct the operation in a timely manner.

[0091] According to one embodiment of this application, the real-time operation flow data sources include operation logs automatically recorded by medical terminals and IoT sensors deployed in the operation area, including operation timestamps, operator identification, operation type, and information on consumables used. The IoT sensors include RFID readers, inertial measurement units, and cameras, used to collect operation trajectory data, such as hand movement paths, consumable scanning sequence, and disinfection coverage.

[0092] According to one embodiment of this application, the adaptive decision optimization and task coordination module specifically performs the following steps during runtime:

[0093] It receives state anomaly warning signals from the multimodal state prediction and evolution analysis module and real-time reminder instructions from the real-time operation semantic understanding and compliance judgment module; the state anomaly warning signals include key impact paths and key nodes identified through graph neural networks, and the real-time reminder instructions include specific deviation steps and correction suggestions;

[0094] The received abnormal status warning signals and real-time reminder commands are used as inputs and passed to a dynamic decision-making model based on deep reinforcement learning. The deep reinforcement learning model adopts a deep Q-network architecture. Its state space consists of a multi-dimensional representation of the current vascular access, historical operation compliance statistics, and nursing plan execution status. The action space includes the adjustment range of maintenance plan parameters such as assessment frequency adjustment value, dressing change cycle adjustment value, and flushing fluid dosage adjustment value. The model learns the optimal decision strategy through interaction with the environment, with the goal of minimizing the long-term accumulated risk of tubing functional impairment while maximizing operation compliance. At the current moment, based on the input abnormal status warning signals and real-time reminder commands, the model outputs a set of maintenance plan adjustment parameters, specifically including assessment frequency adjustment value, dressing change cycle adjustment value, and flushing fluid dosage adjustment value.

[0095] After generating the maintenance plan and adjusting the parameters, the parameters are sent to the nursing plan system through a standard interface. After receiving the parameters, the nursing plan system updates the nursing task list for the target patient (for example, adjusting the catheter puncture site assessment task from once every 8 hours to once every 4 hours, adjusting the dressing change cycle from 7 days to 5 days, or adjusting the flushing fluid dosage from 5 ml to 10 ml), thereby achieving dynamic optimization and closed-loop execution of the nursing plan.

[0096] Simultaneously, the key impact paths and key nodes contained in the abnormal status warning signals are fed back to the data integration and knowledge graph construction module. When constructing the dynamic knowledge graph in the future, the data integration and knowledge graph construction module adjusts the feature weight allocation based on the fed-back key impact paths and key nodes. For example, it increases the weight of entity attributes related to key nodes in semantic mapping and enhances the edge strength of temporal associations related to key impact paths in the knowledge graph. This allows the dynamic knowledge graph to focus more on features related to risk evolution, forming a positive cycle of continuous optimization.

[0097] According to one embodiment of this application, the multi-dimensional visualization interaction and traceability display module specifically performs the following steps during runtime:

[0098] The system acquires a dynamic knowledge graph, predictive scores, key influencing pathways, and maintenance plan adjustment parameters. The dynamic knowledge graph provides the entities and their relationships throughout the entire process of vascular access from insertion to removal. The predictive scores provide future trends of decreased tubing patency, weakened catheter fixation stability, and increased risk of puncture site leakage. The key influencing pathways provide interpretable graph substructures that lead to increased predictive scores. The maintenance plan adjustment parameters provide adjustment values ​​for assessment frequency, dressing change cycle, and flushing fluid dosage.

[0099] Based on the above input data, an interactive multidimensional visualization view of the entire life cycle of vascular pathways is generated through the front-end rendering engine. The interactive multidimensional visualization view includes four core components: event distribution on the timeline, pathway status trend curve, risk evolution tracing path, and operational standard statistical charts.

[0100] The event distribution on the timeline is based on time, with insertion operation records, daily evaluation records, usage and maintenance records, complication handling records, and removal records marked as event nodes on the timeline. Users can zoom in and out of the timeline to view the event distribution at different granularities.

[0101] The access status trend curve plots the predicted score on the vertical axis and the future prediction time window on the horizontal axis, showing the changes in the predicted score as the patency of the tube decreases, the stability of the catheter fixation weakens, and the risk of leakage at the puncture point increases. This helps users intuitively understand the future evolution trend of the access status.

[0102] The risk evolution tracing path visualizes key impact paths in a graph structure, highlighting key nodes and the relationships between them, and using color intensity to represent the importance score of nodes, enabling users to quickly locate the root cause of increased risk.

[0103] Operational compliance statistics charts, presented in the form of bar charts or pie charts, show the frequency of various deviation steps in historical operations and the overall compliance rate, and compare the expected changes before and after the maintenance plan adjustment;

[0104] The interactive multidimensional visualization view also allows users to trace the corresponding original data and historical operation records by clicking on any node displayed in the view; users can click on a daily assessment record node on the timeline, and the system will automatically pop up a window to display the original text content of the record, as well as the corresponding operator's identifier and timestamp; users can click on a clinical observation indicator node in the risk evolution tracing path, and the system will automatically jump to the detailed historical trend chart of that indicator and the corresponding original report of laboratory test results; users can click on a time point on the pathway status trend curve, and the system will automatically locate the medical image in the image archive system near that time point; through this interactive tracing function, users can freely switch between macroscopic views and microscopic data to achieve in-depth insights into the entire life cycle management of vascular pathways.

[0105] The following example, using a specific application scenario, further illustrates the system:

[0106] Example 1 (Dynamic Risk Assessment)

[0107] Scene: Patient Wang, male, 65 years old, ICU patient, 7 days after right internal jugular vein catheterization.

[0108] Data acquisition and processing:

[0109] The data integration and knowledge graph construction module completed the latest batch of data collection at 10:00 AM. The vital signs monitoring system showed a body temperature of 37.8℃ at 8:00 AM. The laboratory information system updated the test results of the blood drawn in the early morning: white blood cell count 12.0 × 10⁻⁶. 9 / L (reference range 3.5-9.5), compared to yesterday's 9.5×10 9 / L significantly elevated; C-reactive protein 50 mg / L (reference value <10), double the 25 mg / L of yesterday. The natural language processing component analyzed the night shift nurse's records: "Patient complains of mild pain at the puncture site, no obvious redness or swelling," and extracted the structured information "Pain: Yes, Redness / Swelling: No." The image recognition component analyzed the puncture site photos uploaded this morning, finding that the percentage of reddened area increased from 2% yesterday to 5%, and the model output "Redness trend: Worsening." Medication records show that vancomycin was started yesterday due to suspected infection.

[0110] The multimodal state prediction and evolutionary analysis module updated all the new data into patient Wang's dynamic knowledge graph and initiated risk prediction analysis. The model's internal attention mechanism focused on several strongly correlated features: elevated body temperature, a sharp increase in white blood cell count, doubling of C-reactive protein, pain at the puncture site, expansion of the redness area, and the use of a new potent antibiotic. These features collectively point to a very high probability of early infection. The module output a prediction score of 78% for catheter-related bloodstream infection within the next 24 hours, marking it as high risk. Simultaneously, evolutionary analysis of the dynamic knowledge graph using a graph neural network identified the top three risk factors: ① a sharp increase in white blood cell and C-reactive protein; ② worsening of local inflammation at the puncture site; ③ initiation of vancomycin due to suspected infection, and generated an abnormal state warning signal containing key influencing pathways.

[0111] Results and Applications:

[0112] Upon receiving the early warning signal, the adaptive decision optimization and task collaboration module immediately pushes a strong early warning message to the attending physician and responsible nurse's terminals through the multi-dimensional visualization interaction and traceability display module: "High-risk warning: Patient Wang has a 78% high risk of catheter-related bloodstream infection. Key risk factors: sharply elevated white blood cell / C-reactive protein levels, and worsening inflammation at the puncture site. Recommendations: ① Immediately assess indications for catheter removal or replacement; ② Simultaneously collect peripheral venous blood and catheter tip blood cultures." The physician arrives at the bedside within 15 minutes of receiving the warning to assess the infection and executes the prescribed blood and catheter tip cultures. The culture results, 48 ​​hours later, confirmed an early catheter-related bloodstream infection. Due to timely detection, the infection was effectively controlled after catheter removal and antibiotic adjustment based on drug sensitivity results.

[0113] Example 2 (Operational Standardization Review)

[0114] Scene: Nurse Li is performing a flushing and sealing procedure on a patient Zhang's peripherally inserted central venous catheter.

[0115] Data collection:

[0116] After the procedure, nurse Li selected the "flushing and sealing" template on her personal digital assistant and entered the text: "Flushing went smoothly, blood return was good, injection was smooth, and the dressing is dry and fixed." Following the procedure, she used her personal digital assistant to take a photo of the puncture site and uploaded it. Real-time operation flow data, including operation timestamps, operator identification, operation type, information on consumables used, and operation trajectory data collected via IoT sensors, was fully recorded.

[0117] The real-time operation semantic understanding and compliance judgment module processes the data. The image recognition model preprocesses and infers from the photo, detecting a tiny but identifiable curl at the upper right edge where the dressing adheres to the skin, despite the overall dryness of the dressing. This is categorized as "poor dressing fixation" with 85% confidence. The operation semantic understanding model simultaneously receives text and images. The text understanding module determines that the operation process description conforms to standard operating procedures. However, the multimodal fusion module compares the text description "well-fixed" with the image recognition result "poorly fixed" and determines that there is a contradiction. The final output review result is that there is a defect, and the semantic deviation exceeds the preset deviation tolerance. A real-time reminder instruction containing specific deviation steps and correction suggestions is generated: "Photo review prompt: The edge of the dressing has a slight curl, posing a risk of contamination. Please handle it promptly."

[0118] Results and Applications:

[0119] Because the issue was not urgent, the system did not trigger a strong alert. Instead, it sent a quality control notification to Nurse Li and the head nurse through a multi-dimensional visual interaction and traceability display module. Upon receiving the notification, Nurse Li immediately went to the ward and, under strict aseptic conditions, changed and secured the dressing. This incident was recorded in Nurse Li's personal training file and the department's quality control report for subsequent targeted training and process improvement.

[0120] Example 3 (Full-process prediction and decision support)

[0121] Scenario: Patient Zhao suffers from chronic renal failure and needs to establish a long-term hemodialysis pathway.

[0122] Data input:

[0123] The doctor entered the patient's information into the system: 58 years old, female, with a 10-year history of diabetes, and a vascular surgery consultation record indicating "generally fair vascular conditions." Preoperative ultrasound images were automatically analyzed by the image recognition model in the data integration and knowledge graph construction module, outputting key features: cephalic vein diameter 2.0 mm, arterial blood flow 40 ml / min. After integrating all data, a dynamic knowledge graph of the patient, Ms. Zhao, was constructed.

[0124] The multimodal state prediction and evolution analysis module used "diabetes," "advanced age," and "cephalic vein diameter 2.0 mm" as key search criteria. It performed similarity matching in a desensitized case database containing thousands of arteriovenous fistula surgeries, retrieving the 300 historical cases most similar to Zhao's case, and statistically analyzed the results of different pathway choices for these cases. The module outputs decision support information:

[0125] First choice recommendation: left upper arm autogenous arteriovenous fistula. The predicted maturity rate at 6 months is 73%. The main risk is postoperative thrombosis, which occurs 30% more frequently than in non-diabetic patients. Complication data show that the incidence of fistula thrombosis at 6 months is 18% in similar cases.

[0126] Alternative option: Artificial vascular graft. The projected primary patency rate at 6 months is 85%, with the main risks being infection and seroma. Complication data shows an 8% incidence of infection-related events within 1 year in similar cases.

[0127] The module also generates personalized auxiliary suggestions: strengthen hand function exercises before surgery to promote vasodilation; if there are no contraindications after surgery, consider taking a small dose of aspirin to reduce the risk of thrombosis; monitor arteriovenous fistula blood flow once a week for the first 3 months after surgery.

[0128] Application of results:

[0129] The adaptive decision optimization and task collaboration module sent the above suggestions to the nursing plan system to update the nursing task list. After thorough communication with the patient and family, the doctor chose to create an autogenous arteriovenous fistula in the left upper arm and adopted the preoperative advice to guide the patient in performing hand function exercises for two weeks. The surgery was successfully performed, and an ultrasound examination at the 8th week postoperatively showed that the fistula was well-matured with a blood flow of more than 500 ml per minute, and it was successfully used for hemodialysis.

[0130] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A large model-based vascular access whole-process intelligent management system, characterized in that, include: The data integration and knowledge graph construction module is used to collect multi-source heterogeneous data related to the vascular access of the target patient in real time, preprocess the collected multi-source heterogeneous data, and construct a dynamic knowledge graph indexed by the unique identifier of the vascular access and covering the entire process from insertion to removal. The multimodal state prediction and evolution analysis module incorporates a large-scale artificial intelligence model based on the Transformer architecture. This model extracts feature vectors from the dynamic knowledge graph, including textual, image, and time-series numerical modalities, and performs cross-modal fusion to generate a multidimensional state representation of the vascular pathway. This multidimensional state representation is then input into a time-series prediction layer to calculate a prediction score for vascular functional impairment occurring within a preset prediction time window. Simultaneously, an evolutionary analysis of entity relationships in the dynamic knowledge graph is performed using a graph neural network to identify key influencing paths and critical nodes that lead to an increase in the prediction score, and to generate an abnormal state warning signal. The real-time operation semantic understanding and compliance judgment module is used to acquire real-time operation flow data generated by medical staff when performing vascular access operations. The real-time operation flow data is input into the operation semantic understanding model built based on a temporal convolutional network to extract operation temporal features. The operation temporal features are matched with standardized operation process nodes in the standard library to calculate the semantic deviation between the current operation and the standard operation process. If the semantic deviation exceeds the preset deviation tolerance, a real-time reminder instruction containing specific deviation steps and correction suggestions is generated and sent to the medical staff terminal. The adaptive decision optimization and task collaboration module is used to input the abnormal state warning signal and the real-time reminder instruction into the dynamic decision model based on deep reinforcement learning to generate maintenance plan adjustment parameters for vascular access; send the maintenance plan adjustment parameters to the nursing plan system to update the nursing task list of the target patient; and feed back the key influence path and key nodes to the knowledge graph construction step to optimize the feature weight allocation in the subsequent dynamic knowledge graph construction process. The multidimensional visualization interaction and traceability display module is used to generate an interactive multidimensional visualization view of the entire life cycle of vascular access based on the dynamic knowledge graph, prediction score, key impact path and maintenance plan adjustment parameters. The interactive multidimensional visualization view allows users to trace the corresponding original data and historical operation records by clicking on any node.

2. The intelligent management system for the entire vascular access process based on a large model as described in claim 1, characterized in that, The data integration and knowledge graph construction module includes: The data acquisition unit is used to acquire multi-source heterogeneous data related to the vascular pathways of the target patient in real time from at least two heterogeneous data sources. The data preprocessing unit is used to perform data cleaning, entity alignment, semantic mapping, and time series calibration on the collected multi-source heterogeneous data. A dynamic knowledge graph construction unit is used to construct a dynamic knowledge graph using the Neo4j graph database as the underlying storage engine. Nodes in the graph represent entities including target patients, vascular access routes, operational events, and clinical observation indicators, while edges represent the relationships between entities. The dynamic knowledge graph update mechanism employs an incremental calculation strategy. When new data flows in, it checks whether the corresponding unique identifier for a vascular access route already exists in the knowledge graph. If it does, the node and its relationships are updated or added; otherwise, a new vascular access route node is created. Version control is applied to attributes, and the weights of associated edges are updated using a dynamic adjustment formula with a decay factor. The dynamic adjustment formula is as follows: ； in, λ represents the weight of the original edge, Δw represents the contribution of the new event to the association strength, and λ is the decay factor.

3. The intelligent management system for the entire vascular access process based on a large model as described in claim 1 or 2, characterized in that, The multi-source heterogeneous data includes insertion operation records, daily assessment records, usage and maintenance records, complication management records, removal records, laboratory test results, medical imaging images, and physiological monitoring signals.

4. The intelligent management system for the entire vascular access process based on a large model according to claim 1 or 2, characterized in that, The multimodal state prediction and evolution analysis module includes: The data input unit is used to input the dynamic knowledge graph into a large-scale artificial intelligence model based on the Transformer architecture; The modality feature extraction unit is used to decompose the data in the dynamic knowledge graph into text modality, image modality, and time-series numerical modality according to the modality type; and to extract the feature vectors of text modality, image modality, and time-series numerical modality respectively using a multi-head attention mechanism; The feature fusion unit is used to interact feature vectors from different modalities using a cross-attention mechanism to generate a multidimensional state representation of the vascular pathway that integrates information from all modalities. The prediction score generation unit is used to input the multidimensional state representation into the time-series prediction layer, calculate the probability value of decreased tubing patency, weakened catheter fixation stability, and increased risk of puncture site leakage within a preset prediction time window, and generate the corresponding prediction score. The evolutionary analysis unit is used to update the feature representation of nodes in the dynamic knowledge graph using a graph attention network, identify the key influence paths and key nodes that lead to the increase of the predicted score through a gradient-based interpretation method, and generate anomaly warning signals including key influence paths and key nodes.

5. The intelligent management system for the entire vascular access process based on a large model according to claim 4, characterized in that, The gradient-based interpretation method includes: calculating the gradient of the multidimensional state representation with respect to each node feature, using the norm of the gradient as the node importance score, and marking the nodes with the highest scores as key nodes; then using the shortest path algorithm to search for paths from key nodes to target patient nodes or vascular access nodes in the dynamic knowledge graph to obtain key influence paths.

6. The intelligent management system for the entire vascular access process based on a large model according to claim 1, characterized in that, The real-time operation semantic understanding and compliance determination module includes: The data acquisition unit is used to acquire real-time operation flow data generated by medical staff when performing vascular access procedures, and at the same time retrieve standardized operation process nodes that are pre-stored in the standard library. The temporal feature extraction unit is used to input the real-time operation flow data and standardized operation process nodes into the pre-trained operation semantic understanding model, and extract the real-time operation temporal features and standardized operation temporal features respectively. The compliance determination unit is used to calculate the optimal alignment path and cumulative distance of the real-time operation timing features and standardized operation timing features in the time dimension using a dynamic time warping algorithm, and to use the cumulative distance as the semantic deviation. The semantic deviation is compared with a preset deviation tolerance. If the semantic deviation exceeds the deviation tolerance, the current operation is determined to be non-compliant. At this time, the specific steps that caused the deviation are back located by tracing back the optimal alignment path in the dynamic time warping algorithm. The reminder instruction generation unit is used to generate correction suggestions for each specific deviation step according to the predefined mapping relationship in the specification library; and send the real-time reminder instruction containing the specific deviation step and correction suggestions to the medical terminal.

7. The intelligent management system for the entire vascular access process based on a large model as described in claim 6, characterized in that, The real-time operation stream data includes operation timestamps, operator identifiers, operation types, and information on consumables used.

8. The intelligent management system for the entire vascular access process based on a large model according to claim 1, characterized in that, The dynamic decision-making model based on deep reinforcement learning adopts a deep Q-network architecture; the state space of the dynamic decision-making model consists of the multidimensional state representation of the current vascular access, historical operation compliance statistics, and nursing plan execution status; the action space of the dynamic decision-making model includes the adjustment range of the maintenance plan adjustment parameters; the goal of the dynamic decision-making model is to minimize the long-term accumulated risk of tubing functional impairment and maximize operation compliance; the maintenance plan adjustment parameters include the assessment frequency adjustment value, dressing change cycle adjustment value, and flushing fluid dosage adjustment value.

9. The intelligent management system for the entire vascular access process based on a large model according to claim 1, characterized in that, The interactive multidimensional visualization view includes event distribution on the timeline, trend curves of pathway status, risk evolution tracing paths, and statistical charts of operational norms. The event distribution on the timeline is based on time as the horizontal axis, with insertion operation records, daily evaluation records, usage and maintenance records, complication handling records, and removal records marked as event nodes on the timeline. The pathway status trend curve plots the predicted score as the vertical axis and the future prediction time window as the horizontal axis, showing the predicted score changes as the patency of the tube decreases, the stability of the catheter fixation weakens, and the risk of leakage at the puncture point increases. The risk evolution tracing path visualizes key impact paths in a graph structure, highlighting key nodes and the relationships between nodes, and using color intensity to represent the importance score of nodes. The operational standardization statistics charts, presented in the form of bar charts or pie charts, show the frequency of various deviation steps in historical operations and the overall compliance rate, and compare the expected changes before and after the maintenance plan adjustment.

Images