An interactive processing system and method for multi-hospital patient data visualization

Abstract

This invention discloses an interactive processing system and method for visualizing patient data from multiple hospitals, specifically relating to the field of medical information technology. The system includes: a distributed data aggregation module for real-time aggregation of patient clinical data from multiple hospitals according to a unified standard; a medical knowledge graph module for storing a network of medical entity relationships to provide logical association parameters; a dynamic simulation and deduction module that responds to front-end interactive adjustments to the first clinical parameter in the graph, obtains associated parameters based on the graph, and generates a composite prediction task; a real-time prediction calculation module that executes the task using a lightweight model obtained through federated learning and knowledge distillation, outputting predicted values ​​and explanatory information; and a visualization module that drives the interface to integrate and display parameter simulation adjustments, associated parameter prediction trends, and medical relationship paths. This invention achieves a leap from static data display to dynamic intelligent deduction, improving the depth and efficiency of clinical decision-making.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, and more specifically, to an interactive processing system and method for visualizing patient data from multiple hospitals. Background Technology

[0002] In the context of multi-hospital collaborative diagnosis and treatment and clinical research, integrating and utilizing patient data scattered across different medical institutions, and conducting effective visualization analysis and decision support, is a significant challenge in the field of medical informatics. Although existing technologies include federated learning for multi-center data modeling, knowledge graphs for medical relationship representation, and visualization for data display, they still have some shortcomings in practical use. For example, most systems only support static display of historical data and generation of fixed reports, failing to enable interactive adjustment of single clinical parameters by users and lacking the ability to automatically discover multi-level related parameter chains based on medical knowledge graphs. This results in the system being unable to provide forward-looking and dynamic sandbox simulation support for clinical decision-making.

[0003] On the other hand, even if some systems have predictive capabilities, their models often rely on centralized training, making it difficult to integrate multi-center data to achieve accurate predictions under compliant conditions. At the same time, the prediction results and medical logic are often isolated from each other, failing to achieve an integrated and linked display of prediction trends and knowledge graph relationships, resulting in a lack of interpretability and systematicity in the clinical decision-making process. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an interactive processing system and method for visualizing patient data from multiple hospitals, which addresses the problems of data integration delays and lack of dynamic simulation mentioned in the background art through the following solutions.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an interactive processing system for visualizing patient data from multiple hospitals, comprising: Distributed data aggregation module: used to extract and aggregate patient clinical data from multiple hospital clinical information systems according to a unified medical data standard; Medical Knowledge Graph Module: This module stores a graph network of medical entities and the relationships between them. It is used to respond to queries and return associated clinical parameters and relationship paths that have a medical logical connection with the target clinical parameter. Dynamic simulation and deduction module: used to receive interactive adjustment instructions for the first clinical parameter in the patient visualization chart, and obtain related clinical parameters from the medical knowledge graph module according to the adjustment instructions, and generate a composite prediction task containing the first clinical parameter and its related clinical parameters. Real-time prediction calculation module: It encapsulates a lightweight medical prediction model to perform the composite prediction task and output the predicted value and explanation information. Dynamic visualization module: used to receive the predicted value and explanation information, and drive the front-end interface to display the simulated adjustment of the first clinical parameter, the predicted change trend of the associated clinical parameter, and the corresponding medical logic relationship path in an integrated manner.

[0006] Preferably, the distributed data aggregation module extracts and aggregates patient clinical data according to a unified medical data standard, specifically including: By using standardized data interfaces deployed in various hospital information systems, changes in patient data are captured in real time using incremental streaming processing. By using terminology standardization services, heterogeneous clinical descriptions from different hospitals can be mapped to a unified medical terminology standard system.

[0007] Preferably, the medical entities stored in the medical knowledge graph module specifically include: Diseases, clinical symptoms, medications, examination and testing indicators, and treatment procedures; The relationships between entities specifically include pathophysiological relationships, treatment plan correlations, and causal reasoning relationships.

[0008] Preferably, the dynamic simulation and deduction module obtains relevant clinical parameters from the medical knowledge graph module according to the adjustment command, specifically including: The interactive adjustment instructions are analyzed to determine the target clinical parameters and their simulated adjustment values; Starting with the target clinical parameter, the medical knowledge graph is traversed to obtain a set of related clinical parameters that are directly affected by it in medical logic. The target clinical parameter, the simulated adjustment value, and the set of associated clinical parameters are encapsulated into a structured composite prediction task.

[0009] Preferably, the structured composite prediction task has a data format that includes a task identifier, a target parameter simulation scenario, a list of associated parameters, and a corresponding patient data index pointer.

[0010] Preferably, the lightweight medical prediction model in the real-time prediction computing module is transformed from a global prediction model obtained by collaborative training on multiple hospital data sources based on a federated learning framework through knowledge distillation technology.

[0011] Preferably, the dynamic visualization module integrates and displays the simulated adjustments, predicted trends, and medical logical relationship paths in a unified manner, specifically including: Within the same visualization view, clinical indicator trend areas and medical knowledge network areas are displayed side-by-side; In the clinical indicator trend area, the simulated adjustment process of the first clinical parameter and the predicted change trend of the associated clinical parameter are rendered simultaneously in the form of a time-series curve. In the medical knowledge network area, the medical logical relationship path extracted from the medical knowledge graph and connecting the first clinical parameter with each associated clinical parameter is dynamically rendered using an interactive graph structure, and is highlighted in conjunction with the curve in the trend area.

[0012] on the other hand An interactive processing method for visualizing patient data from multiple hospitals, used to implement the aforementioned interactive processing system for visualizing patient data from multiple hospitals, characterized in that it includes: S1. Responding to interactive adjustments to the first clinical parameter on the patient visualization chart, generate a simulation command containing the adjustment value; S2. Based on the simulation instructions, query the pre-built medical knowledge graph to obtain the associated clinical parameters that have a medical logical relationship with the first clinical parameter; S3. Based on the adjusted value, the current patient's clinical data, and the associated clinical parameters, construct a composite prediction query task; S4. Call the pre-trained lightweight medical prediction model to calculate the composite prediction query task and generate prediction results and explanatory information of the first clinical parameter and associated clinical parameter in future time series; S5. The prediction results are integrated with the medical logical relationship paths obtained from the medical knowledge graph, and the visualization interface is updated in real time to present the simulated adjustment effect of the first clinical parameter, the predicted change trend of the associated clinical parameter and the medical logical relationship between them in an associated display manner.

[0013] Preferably, obtaining the relevant clinical parameters in step S2 specifically includes: Obtain a second clinical parameter that has a direct medical logical relationship with the first clinical parameter; Further obtain a third clinical parameter that has a direct medical logical relationship with the second clinical parameter to form a multi-level related parameter chain; In step S4, generating the prediction result specifically involves generating a joint prediction result of the first clinical parameter adjustment on all clinical parameters in the multi-level associated parameter chain.

[0014] The technical effects and advantages of this invention are as follows: 1. This invention constructs an intelligent system architecture with a dynamic simulation and deduction module as the core, which deeply integrates distributed data aggregation, medical knowledge graph and lightweight prediction model. This architecture enables the user's interactive adjustment command for a single clinical parameter to be automatically transformed into a multi-level associated parameter chain joint prediction task based on medical logic. This results in a fundamental technical effect that allows the system to leap from static historical data display to dynamic future scenario deduction. 2. This invention achieves low-latency, high-fidelity aggregation and semantically unified mapping of heterogeneous clinical data from various hospitals by employing an incremental streaming processing mechanism and combining rule-based and BERT model terminology standardization services. This results in providing a high-quality, homogeneous data foundation for upper-level intelligent applications and effectively solving the practical problem of real-time integration and value mining of multi-center medical data. 3. This invention adopts a technical approach of using federated learning to collaboratively train a global model and combining knowledge distillation to obtain a lightweight model. This enables the construction of a prediction model with both high accuracy and millisecond-level response capability without leaving the domain of the original data from each hospital. This achieves a secure and compliant effect of efficiently releasing cross-institutional medical data collaborative intelligence while strictly adhering to data privacy regulations. 4. This invention constructs a complete technical closed loop from data aggregation to intelligent inference by deeply integrating four major technical modules: federated learning, knowledge graph, knowledge distillation, and dynamic visualization. Among them, the knowledge graph provides the medical logical association foundation for dynamic inference, federated learning and knowledge distillation realize a high-precision and lightweight prediction model while ensuring data privacy, and dynamic visualization displays the parameter adjustment, prediction trend and medical relationship in real time, realizing the technical leap from static display to dynamic inference. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the system structure of the present invention; Figure 2 This is a schematic diagram of the method of the present invention; In the diagram: 1. Distributed data aggregation module; 2. Medical knowledge graph module; 3. Dynamic simulation and deduction module; 4. Real-time prediction and calculation module; 5. Dynamic visualization module. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Example 1 Reference Figure 1An interactive processing system for visualizing patient data from multiple hospitals, comprising: Distributed data aggregation module 1: Used to extract and aggregate patient clinical data from the clinical information systems of multiple hospitals according to a unified medical data standard; It should be further explained that the distributed data aggregation module 1 extracts and aggregates patient clinical data according to a unified medical data standard, specifically including: By using standardized data interfaces deployed in various hospital information systems, changes in patient data are captured in real time using incremental streaming processing. By using terminology standardization services, heterogeneous clinical descriptions from different hospitals can be mapped to a unified medical terminology standard system.

[0018] Specifically, for the clinical information systems of different hospitals, two types of interfaces were developed: RESTful API and HL7 FHIR. The interface programs were deployed to the servers in each hospital's computer room and included built-in data verification logic. The incremental streaming processing mechanism used Kafka as the message queue and Flink as the streaming processing engine, building a 3-node distributed processing cluster. When data changes in each hospital, the interface program automatically encapsulates a JSON format message with fields including: unique data identifier (data_id), change type (oper_type), change content (content), and data change timestamp (timestamp), and sends it to the Kafka topic "hospital_data_change". The Flink cluster consumes messages from this topic in real time. For newly added data, it is directly included in the aggregated dataset; for modified data, it matches the original data according to the data identifier and updates it; for deleted data, it marks the data status as deleted, preserving data traces for easy tracking.

[0019] The mapping relationship was actually constructed using a combination of manual compilation and model matching: First, three medical coding specialists compiled commonly used terms from five hospitals to form an initial mapping table. Then, the remaining heterogeneous terms were automatically matched using the BERT-base-chinese model. The matching results were reviewed and confirmed by two associate chief physicians. In actual operation, after the streaming engine acquires heterogeneous data, it calls the terminology standardization service interface with a response time of ≤100ms to complete the term replacement. For example, the hypertension blood pressure of a certain hospital is mapped to the systolic blood pressure corresponding to LOINC code 8480-6.

[0020] The standardized patient data is stored in the HBase distributed database, using a row key structure of hospital identifier - patient unique encrypted identifier - data type - timestamp.

[0021] Medical Knowledge Graph Module 2: This module stores a graph network containing medical entities and the relationships between them. It is used to respond to queries and return associated clinical parameters and relationship paths that have a medical logical connection with the target clinical parameter. It should be further explained that the medical entities stored in Medical Knowledge Graph Module 2 specifically include: Diseases, clinical symptoms, medications, examination and testing indicators, and treatment procedures; Relationships between entities specifically include pathophysiological relationships, treatment plan correlations, and causal reasoning relationships.

[0022] Specifically, for entity data collection, data sources include authoritative medical databases, clinical practice guidelines, drug instructions, and hospital clinical data; for unstructured data structuring, a combination of rule parsing and BiLSTM-CRF model is used. The rules are formulated based on the grammatical features of medical texts. The BiLSTM-CRF model is trained on 50,000 annotated medical texts. The annotations were completed by two medical doctors. The model input is a 100-dimensional word vector trained by Word2Vec, and the output is entity type labels. After structuring, disease entities contain 12 attributes and drug entities contain 8 attributes.

[0023] Relation extraction employs a rule engine + machine learning model approach: The rule engine formulates relation extraction rules based on medical logic, such as: Disease A - Clinical Manifestations - Symptom B, Disease A - Treatment Plan - Drug C, Examination Indicator D - Diagnostic Basis - Disease E, etc., to extract relations that conform to conventional medical logic; The machine learning model uses a Transformer-based relation extraction model to extract complex relations that the rule engine cannot cover. The model training data consists of medical text labeled with entity pairs and relation types, with the labeling work completed by medical professionals. The model input is a vector representation of a text fragment containing two entities, and the output is the relation type.

[0024] The knowledge graph is stored using the graph database Neo4j. Medical entities are used as nodes in the graph, and the relationships between entities are used as edges. Node attributes include unique entity identifier, entity name, entity type, and related attribute information. Edge attributes include unique relation identifier, relation type, and relation strength. The relation strength is quantified based on the degree of support from clinical evidence, with a value range of 0-1, and is evaluated by medical experts or obtained through data statistics.

[0025] The latest medical data is collected quarterly, and the entity extraction and relation extraction processes are repeated to incrementally update the knowledge graph. At the same time, an error feedback channel is established, allowing doctors in pilot hospitals to submit graph errors.

[0026] Dynamic simulation and deduction module 3: It is used to receive interactive adjustment instructions for the first clinical parameter in the patient visualization chart, and obtain related clinical parameters from the medical knowledge graph module 2 according to the adjustment instructions, and generate a composite prediction task containing the first clinical parameter and its related clinical parameters. It should be further explained that the dynamic simulation and deduction module 3 obtains relevant clinical parameters from the medical knowledge graph module 2 according to the adjustment instructions, specifically including: Analyze interactive adjustment commands to determine target clinical parameters and their simulated adjustment values; Starting with the target clinical parameter, the medical knowledge graph is traversed to obtain a set of related clinical parameters that are directly affected by it in medical logic. The target clinical parameter, the simulated adjustment value, and the set of associated clinical parameters are encapsulated into a structured composite prediction task.

[0027] The structured composite prediction task has a data format that includes a task identifier, target parameter simulation scenario, a list of associated parameters, and corresponding patient data index pointers.

[0028] Specifically, the interactive adjustment instructions transmitted by the front-end interface are in JSON format. The data structure includes: operation identifier, user identifier (the ID of the operating doctor), patient identifier (the unique ID of the target patient), first clinical parameter identifier (the ID of the clinical parameter being adjusted), simulated adjustment value (the target value after parameter adjustment), and adjustment timestamp (the time when the operation occurred). When parsing the instructions, the completeness and legality of the instruction format are first verified, such as whether required fields are missing and whether the data type of the adjustment value matches the parameter type. If the verification fails, an error message is returned to the front-end. If the verification passes, key information such as the target clinical parameter identifier, simulated adjustment value, and patient identifier are extracted to prepare for subsequent graph queries.

[0029] Starting with the parsed target clinical parameters, the query interface of the medical knowledge graph module is called. A weighted path selection algorithm based on relation strength and clinical evidence level is used to traverse the knowledge graph and discover associated parameters. The specific steps include: Extract all related parameter nodes and their relationship edges that are directly connected to the target clinical parameter from the medical knowledge graph. Each edge includes the relationship type, relationship strength value S (range 0-1), and clinical evidence level E (quantified as: high=1.0, medium=0.6, low=0.3). For each path that starts from a clinical parameter, reaches the associated parameter (denoted as the second clinical parameter) via a relation edge, calculate its priority score. :

[0030] in, and The preset weighting coefficients are initially assigned based on the principle of prioritizing clinical real-time performance and secondarily on the level of evidence. In this system, which is designed for dynamic decision support, the relationship strength S directly reflects the immediacy and causal force of the influence between parameters, and is therefore assigned a weight of 0.6. The level of evidence E reflects the robustness of medical consensus and is assigned a slightly lower weight of 0.4 as an important supplement. This ratio setting has been preliminarily verified: on a test set containing 200 clinical relationships, three clinical experts scored the path relevance. The path selected using the weight combination of α=0.6 and β=0.4 had the highest expert score consistency. The system allows hospitals to fine-tune α and β according to their specialty characteristics.

[0031] The system sets a configurable screening threshold, with a default value of 0.5. This threshold is set based on the principle of balancing clinical significance, aiming to ensure that the selected pathways have sufficient medical importance and can be calibrated through expert experience or historical data; only [the system] retains [the relevant data]. Paths exceeding a set threshold and their corresponding second clinical parameters form a set of directly related parameters; Starting with each selected second clinical parameter, the process of finding directly related edges and calculating new path scores is repeated, using the same logic as when obtaining the first-level associations. This process retrieves the directly related third clinical parameters, forming a set of indirect association parameters. This process can be iterated, but the system sets a maximum traversal depth to ensure computational efficiency and focus on core associations. All identified related parameters are deduplicated, and cyclic dependencies between parameters are detected. If a loop is found, the edge with the lowest score is broken based on the path score, ensuring that the output chain of related parameters forms a clear directed acyclic graph. This graph visually reflects the medical impact path from the user-adjusted first clinical parameter to related clinical parameters at all levels.

[0032] The composite prediction task is encapsulated, with specific fields implemented as follows: Task identifier: Generated using a combination of hospital identifier, patient identifier, timestamp, and random number; Target parameter simulation scenario: includes the first clinical parameter identifier, simulation adjustment value, adjustment method (such as absolute value adjustment, percentage adjustment), and adjustment effective time (default is from the current time); Related parameter list: contains the identifier, parameter type, and current value of all clinical parameters in the multi-level related parameter chain; the current value is obtained by querying the latest clinical data of the patient from the distributed data aggregation module; Patient data index pointer: Points to the location of the complete clinical data storage for this patient in the distributed database, with the index key consisting of the hospital identifier and the patient's unique identifier; The encapsulated composite prediction task is stored in JSON format and sent to the task queue of the real-time prediction calculation module.

[0033] Real-time prediction calculation module 4: It encapsulates a lightweight medical prediction model to perform complex prediction tasks and output predicted values ​​and explanatory information. It should be further explained that the lightweight medical prediction model in the real-time prediction computing module 4 is transformed from a global prediction model that is collaboratively trained on multiple hospital data sources based on a federated learning framework through knowledge distillation technology.

[0034] Specifically, the global prediction model adopts a horizontal federated learning architecture, with each participating hospital acting as a client and a central server serving as the model aggregation node; the model architecture uses an LSTM model based on time-series data prediction, with the following specific structure: Input layer: Receives patients' time-series clinical data. The input dimension is the number of parameters × time step. The number of parameters includes the first clinical parameter and related parameters. The time step is set to 90 days, that is, inputting the daily clinical parameter data of the past 90 days. Hidden layers: Contains 3 LSTM layers, with 128 units in the first LSTM layer, 64 units in the second layer, and 32 units in the third layer; a Dropout layer is added after each LSTM layer with a dropout rate of 0.2 to prevent overfitting; Fully connected layer: It contains two fully connected layers. The first layer has an output dimension of 16 and uses the ReLU activation function. The second layer has an output dimension of the number of associated parameters and uses the Linear activation function. The output is the predicted value of each associated parameter. Loss function: The mean squared error loss function is used to measure the difference between the predicted value and the true value. The formula is as follows:

[0035] in: : This is the mean squared error loss value, a floating-point number reflecting the average squared deviation between the predicted and actual values. It is calculated using this formula and its value ranges from [value missing]. ; n: The number of samples in a single training batch, i.e. the batch size, is an integer. The default value for this stage is 32, and the value range is ≥1. : This is the model's predicted value for the i-th sample, such as the predicted value of a certain related clinical parameter. It is floating-point data, calculated and output by the model, and its value range varies with the parameter type. : This is the true value of the i-th sample, i.e., the actual recorded value in the patient's clinical data. It can be floating-point or integer data, obtained from the distributed data aggregation module, and the value range varies with the parameter type.

[0036] The specific training process is as follows: Initialization: The central server generates initial model parameters and sends them to each client hospital; Local training: Each client hospital uses local time-series clinical data of patients, which has been standardized and privacy information has been removed, and uses the initial model parameters for local training. The training rounds are set to 10 rounds and the batch size is 32. During the training process, only local data is used and no raw data is transmitted to external systems. Model upload: After each client hospital completes local training, it sends the local model parameters and training data volume to the central server; Model aggregation: The central server aggregates the model parameters of each client using a weighted average method, with the weights being the proportion of each client's training data to the total data volume; after aggregation, the global model parameters are obtained and sent to each client hospital; Repeat the above process of local training, model uploading, and model aggregation until the model's validation loss stabilizes, i.e., the rate of change of loss is less than 0.001 for three consecutive iterations. Stop training to obtain the final global prediction model. The model's validation loss is calculated using the validation set data reserved by each hospital.

[0037] Specifically, the lightweight medical prediction model uses the global prediction model trained by federated learning as the teacher model to construct a lightweight student model. The student model architecture is as follows: Output layer: Consistent with the teacher model, the input dimension is the number of parameters × time step; Hidden layers: Contain two CNN-LSTM hybrid layers. The CNN layer uses 3×3 convolutional kernels with a total of 32 kernels to extract local features from temporal data. The LSTM layer has 64 units to capture temporal dependencies. A BatchNorm layer and a Dropout layer are added after the CNN layer, with a dropout rate of 0.2. Fully connected layer: Contains one fully connected layer with the same output dimension as the teacher model, and uses the Linear activation function.

[0038] The knowledge distillation training process is as follows: The training dataset uses time-series clinical data jointly provided by various hospitals, and is divided into a training set (80%) and a validation set (20%). The loss function uses a combination of distillation loss and hard loss: Hard loss: The mean squared error loss function is adopted, and the formula is consistent with that of the global prediction model training stage. It is used to ensure that the prediction accuracy of the student model is consistent with the real clinical data. The batch size n is preset to 64. Distillation loss: The KL divergence loss function is used to measure the difference in output distribution between the teacher model and the student model, thereby realizing the transfer of knowledge from the teacher model. The formula is as follows:

[0039] in:

[0040]

[0041] in: : This is the KL divergence value, a floating-point number reflecting the difference between the teacher model output distribution P and the student model output distribution Q. It is calculated using a formula and its value ranges from [value missing]. ; m: This is the output dimension of the model, i.e., the number of associated clinical parameters. It is an integer data, determined by the list of associated parameters in the composite prediction task, and varies with the task. For example, m=5 means 5 associated parameters, and the value range is ≥1. : This represents the probability distribution value of the teacher model's output for the j-th associated parameter. It is floating-point data, and the original output of the teacher model is processed by the Softmax function. The converted value ranges from [0, 1]. : This represents the probability distribution value of the student model's output for the j-th associated parameter. It is floating-point data, and the original output of the student model is processed by the Softmax function. The converted value ranges from [0, 1]. : This is the original output of the teacher model for the j-th correlation parameter, the logits value before Softmax processing, which is floating-point data and is output by the teacher model through inference; : This is the original output of the student model for the j-th associated parameter, the logits value before Softmax processing, which is floating-point data and is output by the student model through inference; T: Temperature coefficient, used to adjust the smoothness of the output distribution of the Softmax function. It is an integer, with a default value of 10 and a range of ≥1.

[0042] The total loss function combines distillation loss and hard loss, and the calculation formula is as follows:

[0043] in: : This is the total loss value for knowledge distillation training. It is a floating-point number and is calculated by weighting the KL divergence loss and the MSE loss. : This is the weighting coefficient for distillation loss, used to balance the influence of distillation loss and MSE loss. It is a floating-point number with a preset value of 0.7. Experiments have verified that this value can optimally transfer the teacher's model knowledge and ensure prediction accuracy. The value range is [0, 1].

[0044] The training optimizer uses the Adam optimizer with a learning rate of 0.001, 15 training epochs, and a batch size of 64. During training, the loss of the validation set is monitored in real time. When the loss of the validation set does not decrease for 5 consecutive epochs, the training is stopped using an early stopping strategy. After training, the student model is compressed using a quantization compression method to convert the model parameters from 32-bit floating-point numbers to 16-bit floating-point numbers, resulting in the final lightweight medical prediction model, which is then encapsulated in the real-time prediction calculation module.

[0045] The execution of the composite prediction task is as follows: The real-time prediction calculation module retrieves the composite prediction task from the task queue, parses the patient data index pointer in the task, and queries the corresponding time-series clinical data for the patient, including daily data for the past 90 days, from the distributed data aggregation module; data preprocessing is then performed on the retrieved time-series clinical data and the simulated adjustment values ​​in the task, specifically including: Time alignment: Ensure that the time nodes of all clinical parameters are consistent. If there are missing time nodes, use linear interpolation to supplement them. Data normalization: The Min-Max normalization method is used to map the values ​​of all parameters to the interval [0, 1], eliminating the impact of differences in the range of different parameter values ​​on model prediction. The formula is as follows:

[0046] in: : Represents the normalized parameter value, with a range of [0, 1]; x: represents the raw values ​​of clinical parameters, such as the blood pressure and blood sugar values ​​of a patient at a single time point. The raw time-series data of the corresponding clinical parameters of the patient are obtained by querying the distributed data aggregation module. The range of values ​​varies with the parameter type. For example, the range of blood pressure is 60-230 mmHg, and the range of blood sugar is 2.8-20 mmol / L. : This is the historical minimum value of the clinical parameter, which is obtained by extracting all historical time-series data of the clinical parameter of the patient from a distributed database and performing a minimum value operation. Its value is not greater than all the original values ​​x of the parameter. : This is the historical maximum value of the clinical parameter. It is also a floating-point or integer data type. It is obtained by extracting all historical time-series data of the clinical parameter of the patient from a distributed database and performing a maximum value calculation. Its value is not less than all the original values ​​x of the parameter.

[0047] The simulated adjusted value of the first clinical parameter replaces the last time node value of its original time series data to form the adjusted time series data; the adjusted time series data is then concatenated with the time series data of the associated parameters according to the parameter dimension to obtain the model input data.

[0048] The preprocessed model input data is fed into the lightweight medical prediction model for real-time inference calculation. The inference process is accelerated by CPU or GPU, depending on the deployment environment. The lightweight model supports real-time CPU inference and outputs daily predicted values ​​of various related clinical parameters for the next 30 days.

[0049] Feature importance analysis was used to calculate the contribution of the primary clinical parameter and each associated parameter to the prediction results, clarifying which parameters are the key factors affecting the prediction results. At the same time, in conjunction with the relationship paths in the medical knowledge graph, the medical logic of how the adjustment of the primary clinical parameter produces predictive changes in each associated parameter was explained. For example, the adjustment of the primary clinical parameter A affects parameter B through pathophysiological relationships, which in turn affects parameter C, leading to an increase in the predicted value of parameter C.

[0050] Finally, the time series prediction results and explanatory information are encapsulated in JSON format and sent to the dynamic visualization module.

[0051] Dynamic visualization module 5: It is used to receive predicted values ​​and interpretation information, and drive the front-end interface to display the simulated adjustment of the first clinical parameter, the predicted change trend of related clinical parameters, and the corresponding medical logical relationship path in an integrated manner.

[0052] It should be further explained that the dynamic visualization module 5 integrates and displays the simulated adjustments, predicted trends, and medical logical relationship paths in a unified manner, specifically including: Within the same visualization view, clinical indicator trend areas and medical knowledge network areas are displayed side-by-side; In the clinical indicator trend area, the simulated adjustment process of the first clinical parameter and the predicted change trend of the associated clinical parameters are rendered simultaneously in the form of time-series curves. In the medical knowledge network area, the medical logical relationship paths extracted from the medical knowledge graph and connecting the first clinical parameter with each related clinical parameter are dynamically rendered using an interactive graph structure, and are highlighted in conjunction with the curves in the trend area.

[0053] Specifically, the visualization interface architecture uses Vue.js as the front-end development framework, ECharts as the time-series curve rendering library, and D3.js as the medical knowledge network rendering library to build the visualization interface. The interface layout adopts a left-right split structure, with the left side displaying the clinical indicator trend area and the right side displaying the medical knowledge network area. The two columns occupy a 1:1 ratio on the screen and support responsive layout. Communication between the front-end and back-end modules uses the WebSocket protocol to ensure real-time linkage between prediction results and the visualization interface. At the same time, Redis is used to cache recent visualization data to improve the interface data loading speed.

[0054] Clinical Indicator Trend Zone: Receives time-series prediction results from the real-time prediction calculation module, parses them to obtain the simulated adjustment process data of the first clinical parameter (the adjusted value at the last time point and historical data) and the predicted trend data of each related parameter (daily predicted values ​​for the next 30 days); uses ECharts' line chart component to render the time-series curve. The horizontal axis represents the timeline, including 90 days of historical data and 30 days of future data. The vertical axis represents the parameter values, and users can switch between displaying normalized or original values. The curve for the first clinical parameter is drawn with a solid red line, and the nodes of numerical change before and after adjustment are marked; the curves for each related parameter are drawn with dashed lines of different colors, and the corresponding parameter name is marked next to each curve; Curve rendering supports interactive operations: when the mouse hovers over a time node on the curve, the specific time, parameter name, and corresponding value of that node are displayed; the timeline can be zoomed in to view details of a certain time period and zoomed out to view the overall trend; a curve can be hidden / shown, which can be controlled by the parameter name checkbox.

[0055] Medical Knowledge Network Area: Receives medical logical relationship path information sent by the real-time prediction calculation module, and queries the corresponding entity nodes and relationship edge data from the medical knowledge graph module; uses D3.js's force-directed graph algorithm to render the interactive graph structure. Node design: The first clinical parameter node is a red circular node with a size of 20px; the associated parameter node is a blue circular node with a size of 15px; the parameter name is labeled on the node, and detailed information about the parameter is displayed when the mouse hovers over it; Edge design: Different colors are used to draw different types of relationships: green for pathophysiological relationships, orange for treatment plan relationships, and purple for causal reasoning relationships; the thickness of the edges is adjusted according to the strength of the relationship, the stronger the relationship, the thicker the edge; the relationship type name is labeled on the edges; Layout optimization: The force-directed graph algorithm is used to automatically adjust the position of nodes to avoid node overlap; users can manually drag and drop nodes to adjust their positions, and the adjusted positions are automatically saved.

[0056] Achieve synchronized highlighting between the clinical indicator trend area and the medical knowledge network area: When a user clicks on a parameter curve in the clinical indicator trend area, the corresponding parameter node and its connected edges in the medical knowledge network area are automatically highlighted, the node color is darkened, and the edge brightness is increased; at the same time, the color of the curve in the clinical indicator trend area is also darkened, making it easier for the user to identify it. When a user clicks on a parameter node in the medical knowledge network area, the corresponding parameter curve in the clinical indicator trend area is automatically highlighted, and the time axis interval corresponding to the curve flashes as a prompt; at the same time, all relationship paths related to that node in the medical knowledge network area are also highlighted.

[0057] Example 2 Reference Figure 2 An interactive processing method for visualizing patient data from multiple hospitals, used to implement the aforementioned interactive processing system for visualizing patient data from multiple hospitals, characterized in that it includes: S1. Responding to interactive adjustments to the first clinical parameter on the patient visualization chart, generate a simulation command containing the adjustment value; S2. Based on the simulation instructions, query the pre-built medical knowledge graph to obtain the associated clinical parameters that have a medical logical relationship with the first clinical parameter; S3. Based on the adjusted value, the current patient's clinical data, and the associated clinical parameters, construct a composite prediction query task; S4. Call the pre-trained lightweight medical prediction model to calculate the composite prediction query task and generate prediction results and explanatory information of the first clinical parameter and associated clinical parameter in future time series; S5. The prediction results are integrated with the medical logical relationship paths obtained from the medical knowledge graph, and the visualization interface is updated in real time to present the simulated adjustment effect of the first clinical parameter, the predicted change trend of the associated clinical parameter and the medical logical relationship between them in an associated display manner.

[0058] It should be further explained that the relevant clinical parameters obtained in S2 specifically include: Obtain a second clinical parameter that has a direct medical logical relationship with the first clinical parameter; Further obtain a third clinical parameter that has a direct medical logical relationship with the second clinical parameter to form a multi-level related parameter chain; In S4, the prediction results are generated, specifically by generating a joint prediction result of all clinical parameters in the multi-level associated parameter chain based on the adjustment of the first clinical parameter.

[0059] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An interactive processing system for multi-hospital patient data visualization, characterized by, include: Distributed data aggregation module (1): used to extract and aggregate patient clinical data from the clinical information systems of multiple hospitals according to a unified medical data standard; Medical knowledge graph module (2): A graph network storing medical entities and relationships between entities, used to respond to queries and return associated clinical parameters and relationship paths that have medical logical connections with the target clinical parameters; Dynamic simulation and deduction module (3): used to receive interactive adjustment instructions for the first clinical parameter in the patient visualization chart, and obtain related clinical parameters from the medical knowledge graph module (2) according to the adjustment instructions, and generate a composite prediction task containing the first clinical parameter and its related clinical parameters; Real-time prediction calculation module (4): It contains a lightweight medical prediction model, which is used to perform the composite prediction task and output the prediction value and explanation information. Dynamic visualization module (5): used to receive the predicted value and explanation information, and drive the front-end interface to display the simulated adjustment of the first clinical parameter, the predicted change trend of the associated clinical parameter and the corresponding medical logic relationship path in an integrated manner.

2. The interactive processing system for visualizing multi-hospital patient data according to claim 1, characterized in that: The distributed data aggregation module (1) extracts and aggregates patient clinical data according to a unified medical data standard, specifically including: By using standardized data interfaces deployed in various hospital information systems, changes in patient data are captured in real time using incremental streaming processing. By using terminology standardization services, heterogeneous clinical descriptions from different hospitals can be mapped to a unified medical terminology standard system.

3. The interactive processing system for visualizing multi-hospital patient data according to claim 1, characterized in that: The medical entities stored in the medical knowledge graph module (2) specifically include: Diseases, clinical symptoms, medications, examination and testing indicators, and treatment procedures; The relationships between entities specifically include pathophysiological relationships, treatment plan correlations, and causal reasoning relationships.

4. The interactive processing system for visualizing multi-hospital patient data according to claim 1, characterized in that: The dynamic simulation and deduction module (3) obtains relevant clinical parameters from the medical knowledge graph module (2) according to the adjustment instructions, specifically including: The interactive adjustment instructions are analyzed to determine the target clinical parameters and their simulated adjustment values; Starting with the target clinical parameter, the medical knowledge graph is traversed to obtain a set of related clinical parameters that are directly affected by it in medical logic. The target clinical parameter, the simulated adjustment value, and the set of associated clinical parameters are encapsulated into a structured composite prediction task.

5. The interactive processing system for visualizing multi-hospital patient data according to claim 4, characterized in that: The structured composite prediction task includes a data format that includes a task identifier, a target parameter simulation scenario, a list of associated parameters, and a corresponding patient data index pointer.

6. The interactive processing system for visualizing multi-hospital patient data according to claim 1, characterized in that: The lightweight medical prediction model in the real-time prediction calculation module (4) is transformed from a global prediction model that is collaboratively trained on multiple hospital data sources based on a federated learning framework through knowledge distillation technology.

7. The interactive processing system for visualizing multi-hospital patient data according to claim 1, characterized in that: The dynamic visualization module (5) integrates and displays the simulation adjustment, prediction of change trends, and medical logical relationship paths, specifically including: Within the same visualization view, clinical indicator trend areas and medical knowledge network areas are displayed side-by-side; In the clinical indicator trend area, the simulated adjustment process of the first clinical parameter and the predicted change trend of the associated clinical parameter are rendered simultaneously in the form of a time-series curve. In the medical knowledge network area, the medical logical relationship path extracted from the medical knowledge graph and connecting the first clinical parameter with each associated clinical parameter is dynamically rendered using an interactive graph structure, and is highlighted in conjunction with the curve in the trend area.

8. An interactive processing method for visualizing patient data from multiple hospitals, used to implement the interactive processing system for visualizing patient data from multiple hospitals, characterized in that, include: S1. Responding to interactive adjustments to the first clinical parameter on the patient visualization chart, generate a simulation command containing the adjustment value; S2. Based on the simulation instructions, query the pre-built medical knowledge graph to obtain the associated clinical parameters that have a medical logical relationship with the first clinical parameter; S3. Based on the adjusted value, the current patient's clinical data, and the associated clinical parameters, construct a composite prediction query task; S4. Call the pre-trained lightweight medical prediction model to calculate the composite prediction query task and generate prediction results and explanatory information of the first clinical parameter and associated clinical parameter in future time series; S5. The prediction results are integrated with the medical logical relationship paths obtained from the medical knowledge graph, and the visualization interface is updated in real time to present the simulated adjustment effect of the first clinical parameter, the predicted change trend of the associated clinical parameter and the medical logical relationship between them in an associated display manner.

9. The interactive processing method for visualizing multi-hospital patient data according to claim 8, characterized in that: The acquisition of relevant clinical parameters in S2 specifically includes: Obtain a second clinical parameter that has a direct medical logical relationship with the first clinical parameter; Further obtain a third clinical parameter that has a direct medical logical relationship with the second clinical parameter to form a multi-level related parameter chain; In step S4, generating the prediction result specifically involves generating a joint prediction result of the first clinical parameter adjustment on all clinical parameters in the multi-level associated parameter chain.

Images