Multicenter Clinical Data Collaborative Research System for Complex Endocrine Cases

Abstract

This invention relates to the field of medical data processing technology and discloses a multi-center clinical data collaborative research system for complex endocrine cases. The system integrates heterogeneous case data from multiple centers into a unified time-series dataset through data preprocessing and semantic fusion modules. A collaborative control module uses a sliding window to calculate the gradient of data feature changes and intelligently triggers the collaborative collection phase accordingly. A task scheduling module generates differentiated collection tasks based on gradient information. A dynamic adjustment and transition module monitors task feedback, dynamically adjusts storage resource configuration, and automatically transitions the system from the collection phase to the analysis phase based on the decay of the changing gradient. This system achieves adaptive, closed-loop management of the multi-center data collaboration process, enabling precise resource scheduling based on dynamic data changes and node load, thus improving the efficiency and robustness of large-scale clinical research data collaboration.

Description

Technical Field

[0001] This invention relates to the field of medical data processing technology, specifically a multi-center clinical data collaborative research system for complex endocrine cases. Background Technology

[0002] Currently, multi-center clinical studies of complex endocrine cases require the collection of case data from different medical institutions. This data typically exhibits diverse structures, including structured fields from electronic medical records and a large amount of unstructured text records. Existing multi-center data collaboration platforms generally employ pre-defined data collection plans or rely on researchers' experience to initiate data collection tasks. This model struggles to perceive and respond to the actual dynamic changes in data generated by each center, often resulting in delayed data collection or a mismatch between resource allocation and data activity. Furthermore, platforms often employ static or semi-static resource allocation, where the specifications and deployment locations of data storage and processing resources are determined before the task begins. This lack of flexibility, based on the actual workload of each participating node, can easily lead to resource strain and task backlog at some nodes, while other nodes remain idle.

[0003] In managing multi-center collaborative workflows, existing technologies typically design data collection and data analysis as two relatively independent phases, with transitions between phases relying on fixed schedules or manual decisions. This fixed, open-loop management model lacks a feedback mechanism based on the real-time operational status of the system. When the actual progress of data collection tasks is hindered by uneven node loads, or when the dynamic characteristics of the data itself have changed, the system cannot automatically detect and make adaptive adjustments, resulting in inefficient and slow-responding overall workflows from data collection to analysis. Research processes often experience waiting or interruptions due to unsmooth transitions between phases. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-center clinical data collaborative research system for difficult endocrine cases to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a multi-center clinical data collaborative research system for complex endocrine cases, the system comprising:

[0006] The data preprocessing module acquires raw endocrine case data from different medical institutions and performs structure recognition to separate structured data components from unstructured text components.

[0007] The semantic fusion module performs format alignment on the structured data components and performs entity extraction based on medical ontology on the unstructured text components to form a unified semantic representation. The processed data components are then merged to generate a multi-center case dataset with temporal correlation features.

[0008] The collaborative control module uses a sliding window mechanism to traverse the multi-center case dataset, calculates the gradient of changes in data features within each window, and triggers the collaborative data collection phase when the gradient of changes continuously exceeds a preset active threshold.

[0009] The task scheduling module, during the data collaborative collection phase, determines the configuration specifications and deployment location of the data storage medium based on the value and direction of the changing gradient, and generates differentiated collection task queues for data nodes of different medical institutions.

[0010] The dynamic adjustment and transition module monitors the execution feedback of the differentiated collection task queue, dynamically adjusts the configuration specifications of the data storage medium according to the task backlog of each node, calculates the change gradient based on the adjusted configuration specifications, and transitions the system from the data collaborative collection stage to the data collaborative analysis stage according to the decay of the change gradient.

[0011] Preferably, the structured data components are format aligned, and the unstructured text components are subjected to entity extraction based on medical ontology to form a unified semantic representation, including the following steps:

[0012] Scan the structured data components to identify their inherent data field labels and data value formats;

[0013] Establish a standardized data field label mapping table to map the diverse identified data field labels to unified standardized labels;

[0014] Based on a standardized data field mapping table, the numerical values ​​and units in the structured data components are uniformly converted to generate format-aligned structured data blocks;

[0015] In parallel, the unstructured text components are processed by word segmentation and part-of-speech tagging;

[0016] The segmented and annotated text sequence is matched with a pre-built endocrine medicine ontology knowledge base;

[0017] Locate and extract medical entities and relation triples that match the concepts, attributes, and relations in the endocrine medicine ontology knowledge base from the text sequence;

[0018] Assign standardized semantic identifiers to the extracted medical entities and relation triples;

[0019] All medical entities and relation triples with standardized semantic identifiers are linked according to the logical order in the original text to construct a semantic graph representation of the unstructured text components;

[0020] The structured data blocks aligned to the specified format are associated and stored with the semantic graph representation to complete the construction of a unified semantic representation.

[0021] Preferably, the merged data components generate a multicenter case dataset with temporal correlation characteristics, including the following steps:

[0022] Receive the semantic graph representation of the format-aligned structured data block and the unstructured text components;

[0023] Using the patient's unique identifier and the timestamp of the treatment event as a joint primary key, an index is constructed for the structured data blocks aligned to the specified format;

[0024] Using the unique identifier of the same patient and the timestamp of the diagnosis and treatment event as anchors, the semantic graph representation is fused with the corresponding structured data block;

[0025] During the fusion process, attribute associations are established between the medical entities in the semantic graph representation and the specific detection values ​​in the structured data block;

[0026] Based on the chronological order of the diagnosis and treatment event timestamps, all fused data units of the same patient are sorted to form a longitudinal time-series case data sequence for that patient;

[0027] The longitudinal time-series case data of all patients were collected and grouped according to the medical institution from which the data originated;

[0028] Establish cross-institutional case data association links between case data groups from different medical institutions based on the consistency of disease diagnosis codes and key medical entities;

[0029] Along the cross-institutional case data link, longitudinal time-series case data sequences from different sources that are related are aggregated;

[0030] A cross-center association identifier is added to each aggregated case data unit, and its time span information is recorded, ultimately forming the multi-center case data set with time-series association characteristics.

[0031] Preferably, the multicenter case dataset is traversed using a sliding window mechanism to calculate the gradient of changes in data features within each window, including the following steps:

[0032] Set a time span window and move the window along the time axis of the multicenter case dataset with a fixed step size;

[0033] At each window location, extract all case data units that fall within the time range of the window;

[0034] From the extracted case data units, select core data feature indicators that characterize the disease progression status;

[0035] Calculate the statistical difference between the core data feature indicators in the current window and the corresponding indicators in the previous window;

[0036] The instantaneous change rate of data features at the current window position is obtained by calculating the ratio of the statistical difference to the window time span.

[0037] By continuously moving the window along the complete time axis of the multicenter case dataset, a series of instantaneous rate of change sequences of data features arranged in chronological order are obtained.

[0038] The instantaneous rate of change sequence of the data features is smoothed and filtered to eliminate random fluctuation noise;

[0039] Calculate the average trend of the instantaneous rate of change sequence of the filtered data features over multiple consecutive windows; the average trend is the gradient of change.

[0040] The change gradient is stored together with the metadata of the multicenter case dataset for subsequent judgment of the system's operational stage.

[0041] Preferably, when the change gradient continuously exceeds a preset activity threshold, a data collaborative collection phase is triggered, including the following steps:

[0042] The gradient change was obtained through continuous monitoring and calculation;

[0043] The change gradient is compared with a pre-set activity threshold in real time;

[0044] If the change gradient is higher than the activity threshold for three consecutive monitoring periods, a phase transition determination is triggered.

[0045] Generate system status instructions for entering the data collaborative collection phase;

[0046] Based on the specific value of the change gradient during the trigger phase transition determination, it is divided into high activity level, medium activity level, or low activity level.

[0047] Based on the activity level classification, select the corresponding basic data collection strategy template from the strategy library;

[0048] The basic data collection strategy template defines the basic framework for the collection frequency, data compression ratio, and transmission priority of different types of data.

[0049] Based on the total amount and characteristic distribution of the current multicenter case dataset, the parameters of the selected basic data collection strategy template are fine-tuned;

[0050] Instantiate the fine-tuned strategy parameters into specific, executable data collection strategy instructions;

[0051] The data collection strategy instruction is activated and loaded, enabling the system to enter the working state of the collaborative data collection phase.

[0052] Preferably, during the data collaborative collection phase, the configuration specifications and deployment location of the data storage medium are determined based on the value and direction of the change gradient, including the following steps:

[0053] During the data collaborative collection phase, the latest change gradient and its direction of change are acquired in real time.

[0054] If the change direction is positive and the value is at a high activity level, then the high-speed storage media configuration process is initiated, and a high-throughput storage media type and a large-capacity storage space specification are allocated to the high-speed storage media configuration process.

[0055] If the change direction is positive but the value is at a medium or low activity level, then the standard storage media configuration process is initiated, and a storage media type with medium throughput and a storage space specification with appropriate capacity are assigned to the standard storage media configuration process.

[0056] If the change direction is negative, the cache storage medium configuration process is initiated, and a high-response storage medium type and a small-capacity storage space specification are allocated to the cache storage medium configuration process.

[0057] Calculate the total input / output bandwidth requirement of the required storage medium based on the magnitude of the change gradient;

[0058] Based on the calculated total input / output bandwidth requirements and the network topology of each medical institution, a set of data storage media deployment locations that meet the bandwidth latency requirements is determined.

[0059] From the set of data storage media deployment locations, select the node with the lightest network load and the physical location closest to the core data source as the preferred deployment location;

[0060] Generate a data storage media configuration list that includes information on storage media type, storage space specifications, and deployment node location;

[0061] The data storage media configuration list is sent to the infrastructure management layer to complete the allocation and deployment of physical storage resources.

[0062] Preferably, generating differentiated collection task queues for different data nodes in different medical institutions includes the following steps:

[0063] The data storage medium configuration list is analyzed to determine the type and deployment location of the storage medium allocated to each medical institution's data node;

[0064] Based on the historical data output rate and data type preferences of each medical institution's data nodes, the future data generation volume is estimated.

[0065] Based on the throughput specifications of the storage media allocated to each node, calculate the maximum amount of data that each node can receive and process per unit time.

[0066] Based on the estimated data generation volume and the calculated maximum data processing volume, differentiated data collection time windows and upper limits on the amount of data collected in a single instance are set for each data node of a medical institution;

[0067] For nodes containing image data, an image preprocessing subtask is added to the differentiated data collection task, including format standardization and thumbnail generation;

[0068] For nodes containing only text and numerical data, a data encryption and integrity verification subtask is added to the differentiated data collection task;

[0069] All collection instructions configured with specific subtasks, time windows, and data volume limits are grouped according to their associated medical institution data nodes;

[0070] Within each group, data collection instructions are sorted according to their urgency and data priority to form an initial collection task queue for the medical institution's data nodes.

[0071] Coordinate the initial collection task queues of each node to avoid network bandwidth peak conflicts and generate the final, staggered, differentiated collection task queues.

[0072] The differentiated collection task queue is distributed to the corresponding medical institution data nodes to initiate the data collection process.

[0073] Preferably, monitoring the execution feedback of the differentiated collection task queue and dynamically adjusting the configuration specifications of the data storage medium based on the task backlog at each node includes the following steps:

[0074] Receive task execution feedback information from data nodes of various medical institutions. The feedback information includes task completion status, actual amount of data collected, task execution time, and queue backlog length.

[0075] Analyze the task execution feedback information to identify abnormal nodes where the task queue backlog length continues to increase or the task execution time significantly exceeds expectations;

[0076] For the identified abnormal nodes, extract the current configuration specifications and actual performance indicators of their corresponding data storage media;

[0077] By comparing the actual performance indicators with the performance requirements of the differentiated collection task queue for the abnormal nodes, the performance bottleneck can be located.

[0078] If the performance bottleneck is insufficient storage medium throughput, a storage medium upgrade instruction is generated, which instructs the replacement of the storage medium with a higher throughput type at the original deployment location.

[0079] If the performance bottleneck is insufficient storage space, a storage space expansion instruction is generated, which instructs the addition of logical storage units to the existing storage medium.

[0080] If the performance bottleneck is network latency, a deployment location optimization instruction is generated, which instructs to migrate part of the data storage medium to a node with a better network path.

[0081] The generated storage media upgrade instructions, storage space expansion instructions, or deployment location optimization instructions are sent to the infrastructure management layer for execution.

[0082] After the adjustment command is executed, the performance metrics of the abnormal node are retrieved again, and its task queue backlog is monitored until the backlog length returns to normal.

[0083] Preferably, the change gradient is calculated based on the adjusted configuration specifications, and the system transitions from the data collaborative collection phase to the data collaborative analysis phase according to the decay of the change gradient, including the following steps:

[0084] After the data storage medium configuration specifications are dynamically adjusted and the backlog of task queues on each node stabilizes, a new round of gradient calculation is initiated.

[0085] Obtain the multicenter case dataset within the current time window and recalculate the gradient of changes in data features using the same method as the initial calculation.

[0086] The newly calculated gradient of change is compared with the active threshold, and its trend over time is observed.

[0087] If the new gradient continues to decrease over multiple consecutive monitoring periods and eventually stabilizes below the active threshold, then the phase transition condition is triggered.

[0088] Generate system state transition instructions to exit the data collaborative collection phase and enter the data collaborative analysis phase;

[0089] Gradually discontinue the currently active queue of differentiated data collection tasks, but retain the results of completed data collection tasks;

[0090] Initiate the data collaborative analysis preprocessing process, which includes deduplication, normalization, and quality labeling of the collected data.

[0091] Configure dedicated analytical computing resources and algorithm model repositories for the data collaborative analysis phase;

[0092] After the preprocessing process is completed, the system will be switched to the data collaborative analysis stage, and a notification will be issued to indicate that the stage switch is complete.

[0093] Preferably, calculating the statistical difference between the core data feature indicators in the current window and the corresponding indicators in the previous window includes the following steps:

[0094] Extract the specific values ​​of the selected core data feature indicators from all case data units within the current window;

[0095] Extract the specific values ​​of the same core data feature indicators from all case data units within the previous adjacent window;

[0096] Perform descriptive statistical analysis on the set of core data feature indicators within the current window, and calculate their mean, standard deviation, and interquartile range;

[0097] Perform the same descriptive statistical analysis on the set of core data feature indicators in the previous window, and calculate their mean, standard deviation, and interquartile range;

[0098] Calculate the absolute difference between the current window average and the previous window average, and use it as the mean difference component;

[0099] Calculate the relative rate of change between the current window standard deviation and the previous window standard deviation, and use it as the variance component;

[0100] Calculate the ratio of the current window interquartile range to the previous window interquartile range, and use it as the component of distribution morphology difference.

[0101] The mean difference component, the variability difference component, and the distribution pattern difference component are normalized to eliminate the influence of dimensions.

[0102] Assign a first weighting coefficient to the normalized mean difference component, a second weighting coefficient to the normalized variability difference component, and a third weighting coefficient to the normalized distribution shape difference component.

[0103] The three weighted difference components are linearly superimposed to obtain the comprehensive statistical difference.

[0104] The comprehensive statistical difference is logarithmically transformed to compress its numerical range so that it meets the input requirements for subsequent ratio calculations.

[0105] Compared with the prior art, the beneficial effects of the present invention are:

[0106] The collaborative control module continuously calculates the gradient of data feature changes using a sliding window mechanism, and uses this gradient as the core perception indicator of system activity. Subsequent collaborative data collection tasks are only triggered when the gradient consistently exceeds a preset activity threshold. This ensures that system startup and operation are closely dependent on the dynamic characteristics of the data itself, rather than fixed times or instructions. It achieves precise data-driven triggering, avoiding unnecessary resource mobilization and task scheduling during periods of low data change and low value density. Computing, storage, and network communication resources are concentrated on critical windows of truly active data with rich information increments, improving the targeted use of resources and overall collaborative efficiency.

[0107] After the task scheduling module determines the initial resource configuration and generates a task queue based on the change gradient, the dynamic adjustment and transition module dynamically adjusts the configuration specifications of the data storage medium by continuously monitoring the task backlog feedback of each node. The system recalculates the change gradient based on the adjusted new configuration and, according to the decay of this gradient, autonomously decides and executes the transition from the collection phase to the analysis phase. This constitutes a closed-loop control loop including "monitoring-evaluation-adjustment-re-evaluation-transformation." This endows the system with strong adaptive capabilities, enabling it to optimize resource configuration in real time based on the actual resistance to task execution, smooth task load, and automatically advance the workflow based on the achievement of target indicators. It overcomes the rigidity of static configuration and the delay of manual phase switching, allowing the multi-center collaborative system to flexibly cope with complex and unbalanced real-world data environments, ensuring the robustness and smoothness of large-scale data collaboration task execution. Attached Figure Description

[0108] Figure 1 This is a schematic diagram illustrating the working principle of the multi-center clinical data collaborative research system for complex endocrine cases described in this invention.

[0109] Figure 2 A flowchart for semantic fusion of structured and unstructured data;

[0110] Figure 3 A flowchart for calculating the gradient of data feature changes using a sliding window;

[0111] Figure 4 A chart comparing the nominal throughput and actual available bandwidth of data nodes in multiple medical institutions;

[0112] Figure 5 This is a gradient trend chart of changes in core data features. Detailed Implementation

[0113] 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.

[0114] Please see Figure 1 This invention provides a multi-center clinical data collaborative research system for complex endocrine cases. The system includes: a data preprocessing module, a semantic fusion module, a collaborative control module, a task scheduling module, and a dynamic adjustment and transition module. The data preprocessing module acquires raw, heterogeneous endocrine case data from different medical institutions. This module performs structure identification on the raw data, separating it into structured data components and unstructured text components. The semantic fusion module performs format alignment on the structured data components and entity extraction based on medical ontology on the unstructured text components, forming a unified semantic representation. The processed data components are then merged to ultimately generate a multi-center case dataset with temporal correlation characteristics. The collaborative control module uses a sliding window mechanism to traverse this multi-center case dataset, calculating the gradient of data feature changes within each window. When the gradient continuously exceeds a preset activity threshold, the system triggers the data collaborative collection phase. The task scheduling module is activated during the data collaborative collection phase. This module determines the configuration specifications and deployment location of the data storage medium based on the value and direction of the gradient changes and generates differentiated collection task queues for data nodes in different medical institutions. The dynamic adjustment and transition module monitors the execution feedback of the differentiated collection task queue, dynamically adjusts the configuration specifications of the data storage medium according to the task backlog of each node, recalculates the change gradient based on the adjusted configuration specifications, and transitions the system from the data collaborative collection stage to the data collaborative analysis stage according to the decay of the change gradient.

[0115] In one embodiment of the present invention, see [reference] Figure 2The semantic fusion module constructs a unified semantic representation. First, it scans structured data components, identifies their inherent data field labels and data value formats, and establishes a standardized data field label mapping table. This table maps the diverse identified data field labels to unified standardized labels. Based on this standardized mapping table, it performs a unified dimensional conversion on the numerical values ​​and units in the structured data components, generating format-aligned structured data blocks. In parallel, it performs word segmentation and part-of-speech tagging on unstructured text components. The segmented and tagged text sequence is matched against a pre-built endocrinology medical ontology knowledge base. Medical entities and relation triples that match the concepts, attributes, and relationships in the endocrinology medical ontology knowledge base are located and extracted from the text sequence. Standardized semantic identifiers are assigned to the extracted medical entities and relation triples. All medical entities and relation triples with standardized semantic identifiers are linked according to the logical order in the original text, constructing a semantic graph representation of the unstructured text components. Finally, the format-aligned structured data blocks and the semantic graph representation are associated and stored, completing the construction of the unified semantic representation.

[0116] A multi-center case dataset with temporal correlation characteristics is generated. First, semantic graph representations of format-aligned structured data blocks and unstructured text components are received. Using the patient's unique identifier and treatment event timestamp as the joint primary key, the format-aligned structured data blocks are indexed. Using the same patient unique identifier and treatment event timestamp as anchor points, the semantic graph representations are fused with the corresponding structured data blocks. During the fusion process, attribute associations are established between medical entities in the semantic graph representation and specific detection values ​​in the structured data blocks. Based on the chronological order of treatment event timestamps, all fused data units for the same patient are sorted to form a longitudinal temporal case data sequence for that patient. The longitudinal temporal case data sequences of all patients are collected and grouped according to the medical institution from which the data originates. Between case data groups from different medical institutions, cross-institutional case data association links are established based on the consistency of disease diagnosis codes and key medical entities. Following the cross-institutional case data association links, longitudinal time-series case data sequences from different sources that are related are aggregated. A cross-center association identifier is added to each aggregated case data unit, and its time span information is recorded, ultimately forming a multi-center case dataset with time-series association characteristics. In specific implementation, the data preprocessing module obtains raw endocrinology case data from Hospital A's information system and Hospital B's electronic medical record database. The raw endocrinology case data includes structured data components such as laboratory test results and medical orders, as well as unstructured text components including physician progress notes and imaging diagnostic reports. The data preprocessing module automatically scans and parses the raw endocrinology case data, separating the structured data components and unstructured text components based on data format features and label pattern recognition. The separated structured data components are sent to the format alignment processing pipeline, while the unstructured text components enter the natural language processing channel.

[0117] The semantic fusion module performs format alignment on structured data components. This involves scanning tables from different institutions, identifying the inherent data field labels and data value formats. For example, it identifies the field label "serum GH (ng / mL)" used by Hospital A and the field label "growth hormone concentration (ug / L)" used by Hospital B. A standardized data field label mapping table is established, mapping both "serum GH (ng / mL)" and "growth hormone concentration (ug / L)" to the unified standardized label "growth hormone_concentration". Based on the standardized data field mapping table, the numerical values ​​and units in the structured data components undergo a unified dimensional conversion. Values ​​in Hospital B's report, initially in ug / L, are uniformly converted to ng / mL units using a coefficient of 0.001, generating format-aligned structured data blocks. In parallel, entity extraction based on medical ontology is performed on unstructured text components. This extraction process involves word segmentation and part-of-speech tagging of the unstructured text components. The processed text sequence is then matched against a pre-built endocrinology medical ontology knowledge base, which includes terms such as "pituitary adenoma" and "Cushing's disease". This process involves locating and extracting medical entities and relation triples from a text sequence that match the concepts, attributes, and relationships in an endocrinology medical ontology knowledge base. For example, from a medical record, it extracts the relationship "performed" between the entity "patient" and "MRI examination," and the attribute relationship "hint" between the entity "MRI examination" and "pituitary microadenoma." Standardized semantic identifiers are assigned to the extracted medical entity and relation triples; for example, "pituitary microadenoma" is assigned the identifier "C0019426." All medical entity and relation triples with standardized semantic identifiers are linked according to the logical order in the original text to construct a semantic graph representation of the unstructured text components. The semantic graph representation stores entities and relationships in the form of nodes and edges. Format-aligned structured data blocks are then associated and stored with the semantic graph representation, completing the construction of a unified semantic representation.

[0118] A multi-center case dataset with temporal correlation characteristics is generated. The process involves receiving format-aligned structured data blocks and semantic graph representations of unstructured text components. Using the patient's ID number and treatment event timestamps as joint primary keys, the format-aligned structured data blocks are indexed to ensure rapid data retrieval. Using the same patient ID number and treatment event timestamps as anchors, the semantic graph representations are fused with the corresponding structured data blocks. During fusion, attribute associations are established between medical entities in the semantic graph representation and specific test values ​​in the structured data blocks. For example, the "oral glucose tolerance test" entity in the semantic graph is associated with "0-minute blood glucose level: 5.2 mmol / L" in the structured data block. Based on the chronological order of the treatment event timestamps, all fused data units for the same patient are sorted to form a longitudinal temporal case data sequence for that patient, completely recording all medical events from January 2023 to June 2023. The longitudinal time-series case data of all patients are collected and grouped according to the medical institution from which the data originated. Data from Hospital A forms Group 1, and data from Hospital B forms Group 2. Cross-institutional case data association links are established between case data groups from different medical institutions based on the consistency of disease diagnosis codes and key medical entities. For example, acromegaly cases diagnosed by different hospitals are linked using the International Classification of Diseases (ICD-10:E22.0). Along these cross-institutional case data association links, related longitudinal time-series case data sequences from different sources are aggregated. A cross-center association identifier is added to each aggregated case data unit, and its time span information is recorded, ultimately forming a multi-center case dataset with temporal correlation characteristics.

[0119] In some embodiments, calculating the statistical difference between the core data feature indicators in the current window and the corresponding indicators in the previous window involves the integration of multiple statistical components, which can be calculated using the following formula:

[0120]

[0121] Where: symbol The calculated overall statistical dissimilarity is represented by the symbol. Represents the first weighting coefficient assigned to the mean difference component, symbol The second weighting coefficient assigned to the variance component is represented by the symbol. This represents the third weighting coefficient assigned to the component of distributional morphology difference. (Symbol) This represents the mean difference component after normalization, with the sign... Represents the variance component after normalization, with the symbol... This represents the components of the distribution pattern that differ after normalization. (Function) Used to perform a logarithmic transformation on the results of linear superposition to compress their numerical range.

[0122] In some embodiments, establishing cross-institutional case data association links relies not only on disease diagnosis coding but also on the matching degree of key medical entities in semantic graph representations. The association link assesses the semantic distance between the "pituitary MRI" entity from Hospital A and the "sellar MRI" entity from Hospital B in the medical ontology. When the semantic distance is less than a preset threshold, it is confirmed that the two records point to the same examination item and an association is established. It can be understood that through this deep matching based on ontology semantics, even if there are superficial differences in the medical terminology used by different medical institutions, the system can accurately construct associations between case data, thereby enhancing the integrity and consistency of multi-center datasets.

[0123] In one embodiment of the present invention, see [reference] Figure 3 The collaborative control module utilizes a sliding window mechanism to traverse the multi-center case dataset. A time span window is defined, and the window moves along the time axis of the multi-center case dataset with a fixed step size. At each window position, all case data units falling within the window's time range are extracted. From these extracted data units, core data feature indicators representing disease progression are selected. The statistical difference between the core data feature indicator in the current window and the corresponding indicator in the previous window is calculated. This statistical difference is then rationed to the window's time span to obtain the instantaneous change rate of the data feature at the current window position. The window is continuously moved along the complete time axis of the multi-center case dataset, resulting in a series of chronologically ordered sequences of instantaneous change rates of the data feature. These sequences are then smoothed and filtered to eliminate random fluctuation noise. The average trend of the filtered instantaneous change rate sequence over multiple consecutive windows is calculated; this average trend is the gradient of change. This gradient of change is stored along with the metadata of the multi-center case dataset.

[0124] Calculate the statistical difference between the core data feature indicators in the current window and the corresponding indicators in the previous window. Extract the specific values ​​of the selected core data feature indicators from all case data units in the current window. Extract the specific values ​​of the same core data feature indicators from all case data units in the previous adjacent window. Perform descriptive statistical analysis on the set of core data feature indicator values ​​in the current window, calculating its mean, standard deviation, and interquartile range. Perform the same descriptive statistical analysis on the set of core data feature indicator values ​​in the previous window, calculating its mean, standard deviation, and interquartile range. Calculate the absolute difference between the mean of the current window and the mean of the previous window as the mean difference component. Calculate the relative rate of change between the standard deviation of the current window and the standard deviation of the previous window as the variability difference component. Calculate the ratio of the interquartile range of the current window to the interquartile range of the previous window as the distribution pattern difference component. Normalize the mean difference component, variability difference component, and distribution pattern difference component to eliminate the influence of dimensions. The first weighting coefficient is assigned to the normalized mean difference component, the second weighting coefficient is assigned to the normalized variability difference component, and the third weighting coefficient is assigned to the normalized distribution pattern difference component. The three weighted difference components are then linearly superimposed to obtain the comprehensive statistical difference. The comprehensive statistical difference is then logarithmically transformed to compress its numerical range.

[0125] In its implementation, the collaborative control module utilizes a sliding window mechanism to traverse a multi-center case dataset with temporally correlated characteristics. The sliding window mechanism sets a time span window, such as a 7-day window, and moves the window along the time axis of the multi-center case dataset with a fixed step size, such as moving a 7-day time span window with a 1-day fixed step size. At each window location, all case data units falling within the window's time range are extracted; for example, extracting all patients' laboratory test records and diagnostic events within the window from 00:00 on May 1, 2023 to 24:00 on May 7, 2023. From the extracted case data units, core data feature indicators characterizing disease progression are selected, such as "fasting blood glucose level." The statistical difference between the core data feature indicator in the current window and the corresponding indicator in the previous window is calculated; the statistical difference reflects the degree of change in data distribution. The ratio of the statistical difference to the window's time span is calculated to obtain the instantaneous change rate of the data feature at the current window location; the instantaneous change rate of the data feature characterizes the rate of change of the data feature per unit time. By continuously moving the window along the complete timeline of the multicenter case dataset, a series of transient rate-of-change sequences of data features are obtained, arranged chronologically. For example, from May 1, 2023 to June 30, 2023, one transient rate-of-change for a data feature is calculated each day, forming a sequence of 60 data points. The transient rate-of-change sequence is then smoothed using a moving average method with a window size of 5 to eliminate random fluctuation noise. The average trend of the filtered transient rate-of-change sequence over multiple consecutive windows is calculated; for example, the slope of the linear regression of the most recent 10 transient rate-of-change values ​​is calculated, and the average trend is the gradient of change.

[0126] To calculate the statistical difference between the core data feature indicators in the current window and their corresponding indicators in the previous window, it is necessary to extract the specific values ​​of the selected core data feature indicators from all case data units in the current window. For example, the "fasting blood glucose value" in 100 records in the current window has specific values ​​of [5.1, 5.6, 6.0, 5.3, ...]. Then, extract the specific values ​​of the same core data feature indicators from all case data units in the previous adjacent window. For example, the "fasting blood glucose value" in 95 records in the previous window has specific values ​​of [5.0, 5.5, 5.8, 5.2, ...]. Perform descriptive statistical analysis on the set of core data feature indicator values ​​in the current window, calculating its mean, standard deviation, and interquartile range. Perform the same descriptive statistical analysis on the set of core data feature indicator values ​​in the previous window, calculating its mean, standard deviation, and interquartile range. Calculate the absolute difference between the mean of the current window and the mean of the previous window as the mean difference component. Calculate the relative rate of change between the standard deviation of the current window and the standard deviation of the previous window as the variability difference component. The ratio of the interquartile range of the current window to the interquartile range of the previous window is calculated as the distribution morphology difference component. The mean difference component, variability difference component, and distribution morphology difference component are normalized to eliminate the influence of dimensions. A first weighting coefficient is assigned to the normalized mean difference component, a second weighting coefficient to the normalized variability difference component, and a third weighting coefficient to the normalized distribution morphology difference component. The three weighted difference components are linearly superimposed to obtain the comprehensive statistical difference. A logarithmic transformation is performed on the comprehensive statistical difference to compress its numerical range, making it conform to the input requirements of subsequent ratio calculations.

[0127] In some embodiments, when calculating the gradient of change, the core data feature index can be a multi-dimensional vector, such as simultaneously containing "fasting blood glucose level" and "2-hour postprandial blood glucose level". The calculation process requires calculating the instantaneous rate of change of the data feature for each dimension separately, and then performing weighted summation. Optionally, Gaussian filtering or median filtering methods can be used instead of the moving average method for smoothing filtering to adapt to different data noise characteristics. When calculating the average trend of change, the following formula can be used:

[0128]

[0129] Where: symbol This represents the final calculated gradient of change, with the sign... Indicates the time span of the sliding window. (Symbol) Indicates the number of consecutive windows used to calculate the trend, symbol Indicates the first The center time point of each window, symbol Indicates the first The instantaneous rate of change of data features corresponding to each window after smoothing and filtering. (Symbol) express Time points The average value, sign express Instantaneous change rate of each data feature The average value is calculated as follows. The numerator of the formula calculates the covariance of time and the instantaneous rate of change of data features, while the denominator calculates the variance of time. The overall calculation yields the linear regression slope of the instantaneous rate of change sequence of data features over a specified time interval. The slope value is then divided by the time span. Obtain the final gradient of change .

[0130] In some embodiments, the window's time span and step size can be preset and configured according to disease characteristics. For indicators with large diurnal fluctuations, such as blood glucose, the window time span can be set to 3 days and the step size to 1 day; for long-term monitoring indicators, such as growth hormone, the window time span can be set to 30 days and the step size to 7 days. Optionally, the process of selecting core data feature indicators is based on the list of key disease monitoring items defined in the endocrine medicine knowledge base. The system automatically matches and selects indicators from the list of fields in the multi-center case dataset.

[0131] In one embodiment of the present invention, the collaborative control module triggers the data collaborative collection phase when the gradient change continuously exceeds a preset activity threshold. This module continuously monitors the calculated gradient change and compares it with the preset activity threshold in real time. If the gradient change is higher than the activity threshold for three consecutive monitoring cycles, a phase transition determination is triggered, generating a system status instruction to enter the data collaborative collection phase. Based on the specific value of the gradient change at the time of triggering the phase transition determination, it is classified into high activity level, medium activity level, or low activity level. According to the classified activity level, a corresponding basic data collection strategy template is selected from the strategy library. The basic data collection strategy template defines the basic framework of collection frequency, data compression ratio, and transmission priority for different types of data. Combining the total data volume and characteristic distribution of the current multi-center case dataset, the parameters of the selected basic data collection strategy template are fine-tuned. The fine-tuned strategy parameters are instantiated into specific, executable data collection strategy instructions, which are then activated and loaded, enabling the system to enter the working state of the data collaborative collection phase.

[0132] During the data collaborative collection phase, the task scheduling module determines the configuration specifications and deployment location of the data storage media based on the value and direction of the change gradient. In this phase, the module acquires the latest change gradient and its direction in real time. If the change direction is positive and the value is at a high activity level, the high-speed storage media configuration process is initiated, allocating a high-throughput storage media type and a large-capacity storage space specification. If the change direction is positive but the value is at a medium or low activity level, the standard storage media configuration process is initiated, allocating a medium-throughput storage media type and a moderate-capacity storage space specification. If the change direction is negative, the cache storage media configuration process is initiated, allocating a high-response-speed storage media type and a small-capacity storage space specification. Based on the magnitude of the gradient change, the total input / output bandwidth requirement of the required storage media is calculated. Based on the calculated total input / output bandwidth requirement and the network topology of each medical institution, a set of data storage media deployment locations that meet the bandwidth latency requirements is determined. From the set of data storage media deployment locations, the node with the lightest network load and the physical location closest to the core data source is selected as the preferred deployment location. A data storage media configuration list containing information on storage media type, storage space specifications, and deployment node location is generated and sent to the infrastructure management layer.

[0133] In practical implementation, the collaborative control module continuously monitors the calculated gradient changes. For example, if the hourly updated gradient sequence is [0.03, 0.05, 0.08, 0.09, 0.07, ...], the gradient changes are compared in real time with a pre-set activity threshold of 0.05. If the gradient changes are higher than the activity threshold for three consecutive monitoring periods, for example, if the gradient values ​​for three consecutive periods are 0.08, 0.09, and 0.07, all greater than 0.05, a phase transition determination is triggered. A system status command to enter the data collaborative collection phase is generated, updating the system's main status flag from "monitoring" to "collaborative collection". Based on the specific value of the gradient changes at the time the phase transition determination is triggered, it is classified into high activity level, medium activity level, or low activity level. For example, values ​​greater than 0.08 are classified as high activity level, values ​​between 0.05 and 0.08 as medium activity level, and values ​​less than 0.05 as low activity level. Based on the classified activity levels, a corresponding basic data collection strategy template is selected from the strategy library. For example, the "High-Frequency Low-Latency Collection Template" is selected for the high-activity level. The basic data collection strategy template defines the basic framework for the collection frequency, data compression ratio, and transmission priority of different data types. The "High-Frequency Low-Latency Collection Template" defines image data as collected every 15 minutes, with a compression ratio of 60% and the highest transmission priority. Considering the total data volume and characteristic distribution of the current multi-center case dataset, the parameters of the selected basic data collection strategy template are fine-tuned. For example, based on the characteristic that image data accounts for more than 70% of the current data stream, the image data collection frequency is fine-tuned from the template's 15 minutes to 10 minutes. The fine-tuned strategy parameters are instantiated into specific, executable data collection strategy instructions, with the instruction content being "Node A: Image data, collection interval 10 minutes, compression ratio 60%, priority 5". The data collection strategy instructions are activated and loaded, putting the system into the collaborative data collection phase.

[0134] During the data collaborative collection phase, the task scheduling module determines the configuration specifications and deployment location of the data storage media based on the value and direction of the change gradient. In this phase, the latest change gradient and its direction are acquired in real time; for example, if the latest change gradient is +0.09, and the change direction is positive and the value is at a high activity level, the high-speed storage media configuration process is initiated. This process allocates high-throughput storage media types and large-capacity storage space specifications, such as allocating NVMeSSD as the storage media type, a single-disk throughput greater than 3000MB / s, and a total storage space specification of 10TB. If the change direction is positive but the value is at a medium or low activity level, the standard storage media configuration process is initiated. This process allocates medium-throughput storage media types and moderate-capacity storage space specifications, such as allocating SASHDD as the storage media type, a single-disk throughput of approximately 500MB / s, and a total storage space specification of 4TB. If the change direction is negative, the cache storage media configuration process is initiated, and a high-response-speed storage media type and a small-capacity storage space specification are allocated to the cache storage media configuration process. For example, memory is allocated as the storage media type, with a response time of less than 100 microseconds and a total storage space specification of 64GB.

[0135] Based on the magnitude of the change gradient, the total input / output bandwidth requirement of the required storage media is calculated. For example, based on the current number of active nodes, the estimated data generation rate of each node, and the reservation coefficient, the total input / output bandwidth requirement is calculated to be 360MB / s. Based on the calculated total input / output bandwidth requirement and the network topology of each medical institution, a set of data storage media deployment locations that meet the bandwidth and latency requirements is determined. The network topology shows that the core data center, edge server node A, and edge server node B all meet the bandwidth and latency requirements. From the set of data storage media deployment locations, the node with the lightest network load and the physical location closest to the core data source is selected as the preferred deployment location. For example, the current network load of the core data center is 30%, and the load of edge server node A is 75%, therefore the core data center is selected as the preferred deployment location. A data storage media configuration list containing information on storage media type, storage space specifications, and deployment node location is generated. The list entries are "Type: NVMeSSD; Specification: 10TB; Deployment Location: Core Data Center - Rack 03". The data storage media configuration list is sent to the infrastructure management layer to complete the allocation and deployment of physical storage resources.

[0136] In some embodiments, the threshold for classifying activity levels can be dynamically adjusted based on historical data characteristics. The system can periodically statistically analyze the distribution of the changing gradients, setting the average plus a standard deviation as the lower limit for the high activity level and the average as the lower limit for the medium activity level. Optionally, the basic data collection strategy template library may contain dedicated templates for specific data types, such as a "gene sequencing data collection template," which defines extremely high integrity and encryption requirements but allows for longer collection latency. The total input / output bandwidth requirement of the required storage media can be calculated using the following formula:

[0137]

[0138] Where: symbol This represents the calculated total input / output bandwidth requirement of the storage medium, denoted by the symbol. This represents the system's preset bandwidth reservation factor. (Symbol) This represents the total number of medical institution data nodes that are currently active and require collaborative data collection. This indicates the estimated output rate based on historical data. The data generation rate of each node during the collaborative data collection phase. (Summation term) This represents the sum of the estimated data generation rates of all active nodes.

[0139] In some embodiments, when determining the deployment location of data storage media, in addition to network load and physical distance, the data security level of the nodes is also considered to ensure compliance with preset policies. Only nodes with a security level of "Level 3 Security Protection" or higher are included in the data storage media deployment location set. Optionally, the selection of storage media type can be further refined. For example, in the high-speed storage media configuration process, a specific model of NVMe SSD more suitable for write endurance can be selected based on the data read / write ratio. It can be understood that by using the value and direction of the changing gradient as decision input, the system can dynamically and adaptively configure storage resources that match the data activity level, avoiding performance bottlenecks caused by insufficient resources or waste caused by idle resources.

[0140] In one embodiment of the present invention, the task scheduling module generates differentiated collection task queues for data nodes of different medical institutions, parses the data storage medium configuration list to clarify the storage medium type and deployment location allocated to each medical institution data node, estimates its future data generation volume based on the historical data output rate and data type preferences of each medical institution data node, and calculates the maximum amount of data that each node can receive and process per unit time by combining the throughput specifications of the storage medium allocated to each node. Based on the estimated data generation volume and the calculated maximum processable data volume, differentiated data collection time windows and single collection data volume limits are set for each medical institution data node. For nodes containing image data, an image preprocessing subtask, including format standardization and thumbnail generation, is added to the differentiated data collection task. For nodes containing only text and numerical data, a data encryption and integrity verification subtask is added to the differentiated data collection task. All collection instructions configured with specific subtasks, time windows, and data volume limits are grouped according to their associated medical institution data nodes. Within each group, the collection instructions are sorted according to their urgency and data priority to form an initial collection task queue for the medical institution data nodes. The initial collection task queues of each node are coordinated to avoid network bandwidth peak conflicts, generating a final, staggered, differentiated collection task queue. The differentiated collection task queue is then distributed to the corresponding medical institution data nodes.

[0141] In practical implementation, the task scheduling module generates differentiated collection task queues for data nodes in different medical institutions. The module parses the data storage medium configuration list to clarify the storage medium type and deployment location allocated to each medical institution's data node. For example, the configuration list reveals "Hospital Imaging Center Node Allocation Type: NVMeSSD, Deployment Location: Core Data Center - Rack 03" and "Hospital Laboratory Node Allocation Type: SASHDD, Deployment Location: Edge Node A". Based on the historical data output rate and data type preferences of each medical institution's data node, its future data generation volume is estimated. For instance, based on the logs from the past 7 days, the Hospital Imaging Center node generates approximately 800GB of DICOM image data per day, and the Hospital Laboratory Node generates approximately 50GB of structured test data per day. Based on the storage media throughput specifications allocated to each node, the maximum amount of data that each node can receive and process per unit of time is calculated. For example, if the write throughput of NVMe SSD is 3000MB / s, then the hospital imaging center node can receive a maximum of approximately 3000MB of data per second; if the write throughput of SASHDD is 500MB / s, then the hospital laboratory node can receive a maximum of approximately 500MB of data per second. Based on the estimated data generation volume and the calculated maximum processable data volume, differentiated data collection time windows and single-collection data volume limits are set for each medical institution's data node. For example, a collection window is set every 10 minutes for the hospital imaging center node, with a maximum of 50GB of data collected per window; a collection window is set every 30 minutes for the hospital laboratory node, with a maximum of 2GB of data collected per window.

[0142] For nodes containing image data, an image preprocessing subtask is added to the differentiated data collection task. This subtask includes format standardization and thumbnail generation, such as converting received DICOM images to JPEG2000 format and generating a 64x64 pixel thumbnail for each image. For nodes containing only text and numerical data, a data encryption and integrity verification subtask is added to the differentiated data collection task. For example, CSV files uploaded by the laboratory are encrypted using the AES-256 algorithm, and their SHA-256 hash value is calculated for subsequent integrity verification. All collection instructions configured with specific subtasks, time windows, and data volume limits are grouped according to their associated medical institution data nodes. Instructions from the hospital imaging center node are grouped into the first group, and instructions from the hospital laboratory node are grouped into the second group. Within each group, the collection instructions are sorted according to their urgency and data priority, forming an initial collection task queue for the medical institution data nodes. For example, encryption tasks are prioritized over transmission tasks, and data collection tasks for high-priority research cases are prioritized over routine cases. Coordinate the initial collection task queues of each node to avoid network bandwidth peak conflicts and generate the final, staggered collection task queues. For example, set the collection window start time of the hospital imaging center node to the hour, and the collection window start time of the hospital laboratory node to 10 minutes after the hour.

[0143] In some embodiments, when calculating the maximum amount of data that each node can receive and process per unit time, the actual available bandwidth of the storage medium and system overhead must be considered. The actual available bandwidth is the nominal throughput minus the system reserved overhead, for example, reserving 10% bandwidth for system logs and monitoring. Optionally, for nodes containing a mixture of multiple data types, differentiated data collection tasks can be further broken down. For example, if a node simultaneously generates images and pathological texts, separate sub-queues and transmission channels can be established for image data and text data respectively. The coordination of task queues can be calculated based on the following formula to determine peak-shaving times:

[0144]

[0145] Where: symbol Indicates the first The start time offset of the data collection window for each medical institution's data node, sign Indicates the baseline start time. Symbol Indicates the minimum collection window time interval set by the system, symbol This represents the total number of data nodes that need to be coordinated for peak shifting. This formula ensures... The collection window of each node has a length of The tasks are evenly distributed over a period of time to distribute the network load evenly. Refer to Table 1 for the specific parameter configuration of the differentiated collection task queue.

[0146] Table 1: Differentiated Collection Task Queue Parameter Configuration Table

[0147]

[0148] In some embodiments, a moving average model can be used to improve the accuracy of the prediction when estimating future data generation. In some embodiments, when using a moving average model to predict future data generation, the system maintains a chronologically ordered sequence of historical data generation for each medical institution's data node. This sequence records the amount of data actually collected by the node in each of the past several consecutive monitoring periods. The system sets a moving window size, for example, five monitoring periods. At the beginning of each new monitoring period, the model takes the historical data generation of the node from the most recent five periods, calculates the arithmetic mean of these data, and uses this average as a prediction of the data generation for the next period. As the system continues to run, each time a new monitoring period begins, the moving window slides forward once, automatically removing the oldest historical data point and adding the most recently generated actual data point, recalculating the average, thereby achieving dynamic updates and predictions of future data generation. Optionally, data priorities can be dynamically marked according to the urgency of the case or the level of the research project. Understandably, by setting differentiated time windows and data volume limits for different nodes, the system achieves fine-grained traffic control over heterogeneous data sources, avoiding the impact of data surges from a single node on the central storage.

[0149] See Figure 4 In the bandwidth utilization analysis during the collaborative data collection phase of the system, quantifying the difference between nominal throughput and actual available bandwidth is the core basis for evaluating the transmission capabilities of data nodes in various medical institutions. Specifically, the nominal throughput (blue bars) of different medical data nodes (such as hospital imaging centers and hospital laboratories) represents the theoretical transmission capacity of the storage medium, while the actual available bandwidth (purple bars) is the actual available value after deducting system reserved overhead. Taking the hospital imaging center as an example, its nominal throughput reaches 3000MB / s, and its actual available bandwidth is 2700.0MB / s, reflecting the system's rule of reserving 10% bandwidth for logs, monitoring, etc.; while the community medical center has a nominal throughput of 300MB / s and an actual available bandwidth of 273.0MB / s, also conforming to this reservation logic. The bandwidth differences between nodes directly support the parameter design of the "differentiated collection task queue" in the task scheduling module (e.g., the high throughput of the hospital imaging center corresponds to a short collection window and a large single-time limit), providing fundamental data support for achieving refined flow control of heterogeneous data sources.

[0150] In one embodiment of the present invention, the dynamic adjustment and transition module monitors the execution feedback of the differentiated collection task queue, receives task execution feedback information from data nodes of various medical institutions, including task completion status, actual collected data volume, task execution time, and queue backlog length. The module analyzes the task execution feedback information to identify abnormal nodes where the task queue backlog length is continuously increasing or the task execution time significantly exceeds expectations. For the identified abnormal nodes, the current configuration specifications and actual performance indicators of their corresponding data storage media are extracted. The actual performance indicators are compared with the performance requirements of the differentiated collection task queue for the abnormal nodes to locate the performance bottleneck. If the performance bottleneck is insufficient storage media throughput, a storage media upgrade instruction is generated, instructing the replacement of the storage media type with a higher throughput at the original deployment location. If the performance bottleneck is insufficient storage space, a storage space expansion instruction is generated, instructing the addition of logical storage units to the original storage media. If the performance bottleneck is network latency, a deployment location optimization instruction is generated, instructing the migration of some data storage media to nodes with better network paths. The generated storage media upgrade instructions, storage space expansion instructions, or deployment location optimization instructions are sent to the infrastructure management layer for execution. After the adjustment instructions are executed, the performance indicators of the abnormal nodes are retrieved again, and the backlog of their task queues is monitored.

[0151] The dynamic adjustment and transition module calculates the change gradient based on the adjusted configuration specifications. After the dynamic adjustment of the data storage medium configuration specifications is completed and the backlog of task queues at each node stabilizes, a new round of change gradient calculation is initiated. This involves acquiring the multi-center case data set within the current time window and recalculating the change gradient of data features using the same method as the initial calculation. The recalculated change gradient is compared with the activity threshold, and its trend over time is observed. If the new change gradient continuously decreases over multiple monitoring periods and eventually stabilizes below the activity threshold, the phase transition condition is triggered. A system state transition instruction is generated to exit the data collaborative collection phase and enter the data collaborative analysis phase. This gradually stops the currently active differentiated collection task queues but retains the results of completed data collection tasks. The data collaborative analysis phase preprocessing process is initiated, including deduplication, normalization, and quality labeling of the collected data. Dedicated analytical computing resources and algorithm model repositories are configured for the data collaborative analysis phase. After the preprocessing process is completed, the system's working state is officially switched to the data collaborative analysis phase.

[0152] In practical implementation, the dynamic adjustment and transition module monitors the execution feedback of the differentiated collection task queue. This module receives task execution feedback information from data nodes across various medical institutions, including task completion status, actual collected data volume, task execution time, and queue backlog length. Analyzing the task execution feedback information identifies abnormal nodes where the task queue backlog length is continuously increasing or task execution time significantly exceeds expectations. For example, the task queue backlog length of the hospital imaging center node increased from 50 to 120 over three consecutive monitoring periods, and the task execution time increased from an average of 5 minutes to 15 minutes. For the identified abnormal nodes, the current configuration specifications and actual performance indicators of their corresponding data storage media are extracted. For example, the hospital imaging center node is currently configured as SASHDD, with a measured write speed of 80MB / s and a storage space utilization rate of 95%. Comparing the actual performance indicators with the performance requirements of the differentiated collection task queue for the abnormal node pinpoints the performance bottleneck. The differentiated collection task queue requires the node to achieve a processing speed of 100MB / s, while the current write speed of 80MB / s does not meet the requirement, identifying the performance bottleneck as insufficient storage media throughput. If the performance bottleneck is insufficient storage media throughput, a storage media upgrade command is generated. This command instructs the replacement of the storage media type with a higher throughput at the existing deployment location. The command content is "Upgrade the storage media of the hospital imaging center node from SASHDD to NVMeSSDPCIe4.0". If the performance bottleneck is insufficient storage space, a storage space expansion command is generated. This command instructs the addition of logical storage units to the existing storage media. The command content is "Add a 2TB logical volume to the existing storage array of the hospital imaging center node". If the performance bottleneck is network latency, a deployment location optimization command is generated. This command instructs the migration of some data storage media to a node with a better network path. The command content is "Migrate 50% of the data cache replicas of the hospital imaging center node to edge node B". The generated storage media upgrade command, storage space expansion command, or deployment location optimization command is issued to the infrastructure management layer for execution. After the adjustment command is executed, the performance indicators of the abnormal node are retrieved again, and its task queue backlog is monitored until the backlog length returns to normal levels.

[0153] The dynamic adjustment and transition module calculates the change gradient based on the adjusted configuration specifications. After the dynamic adjustment of the data storage medium configuration specifications is completed and the task queue backlog at each node stabilizes, a new round of change gradient calculation is initiated. For example, calculation is initiated after the storage medium upgrade is completed at the hospital imaging center node and its task queue backlog length stabilizes below 10. The multi-center case data set within the current time window is acquired, and the change gradient of data features is recalculated using the same method as the initial calculation. For example, based on the latest 7-day window data, the recalculated change gradient is 0.02. The recalculated new change gradient is compared with the active threshold, and its trend over time is observed. The active threshold is 0.05. The new change gradient of 0.02 is below the threshold and remains [0.02, 0.01, 0.01, 0.02] for the next four monitoring periods. If the new change gradient continues to decrease over multiple consecutive monitoring periods and eventually stabilizes below the active threshold, the stage transition condition is triggered. A system state transition instruction is generated to exit the data collaborative collection stage and enter the data collaborative analysis stage. The instruction updates the system's main state flag from "collaborative collection" to "collaborative analysis". Gradually discontinue the currently active queue of differentiated data collection tasks, but retain the results of completed data collection tasks. For example, reduce the collection frequency of the hospital imaging center node from once every 10 minutes to once every 24 hours, and stop issuing new tasks. Initiate the preprocessing process for the data collaborative analysis phase, which includes deduplication, normalization, and quality labeling of the collected data. Configure dedicated analytical computing resources and algorithm model repositories for the data collaborative analysis phase, such as allocating a computing cluster with 8 GPU nodes and loading algorithm models for diabetes complication prediction, pituitary tumor classification, etc. After the preprocessing process is completed, officially switch the system to the data collaborative analysis phase and issue a notification that the phase switch is complete.

[0154] In some embodiments, identifying abnormal nodes can be done automatically using quantitative indicators. For example, a queue backlog growth rate exceeding 100% can be defined as "continuous increase," and task execution time exceeding twice the standard deviation of the average execution time can be defined as "significantly exceeding expectations." Optionally, performance bottlenecks can be located by combining more detailed performance counter data, such as storage media IOPS and network packet loss rate, for comprehensive diagnosis. Monitoring task queue backlog and determining whether it has stabilized can be done by calculating the queue health index using the following formula:

[0155]

[0156] Where: symbol This represents the calculated queue health index, symbol [symbol missing]. Indicates the backlog length of the task queue at the end of the current monitoring period, symbol [symbol missing]. This represents the average execution time of tasks within the current monitoring period. Queue Health Index This intuitively reflects the degree of congestion in the queue. When the value remains below the system's set stable threshold, it can be determined that the queue backlog has returned to normal levels.

[0157] In some embodiments, the triggering condition for stage transitions can be set to a more stringent mode, such as requiring the gradient change to be below the active threshold for five consecutive monitoring periods without any single-point rebound during this period. Optionally, the preprocessing workflow in the data collaborative analysis stage can be customized according to the type of analysis model to be run. For example, if the main model being run is an image analysis model, the preprocessing workflow will additionally include image denoising and standardized cropping steps. It can be understood that by monitoring feedback and dynamically adjusting storage configuration, the system can adaptively resolve performance bottlenecks during the data collection stage, ensuring a smooth data aggregation process.

[0158] See Figure 5 In the phase transition process of the multi-center clinical data collaborative research system for complex endocrine cases, the gradient of core data feature changes is a key indicator for determining the transition from the collaborative collection phase to the collaborative analysis phase. Specifically, in the figure, the red curve represents the gradient of core indicator changes, the blue dashed line is the active threshold (0.05), the pink area is the interval where the gradient is higher than the threshold, and the green dashed line is the phase transition judgment point. During the monitoring periods from pre-adjustment 1 to pre-adjustment 4, the gradient of core indicator changes remained consistently above the active threshold interval (up to 0.08), corresponding to the system being in the collaborative data collection phase. After dynamic adjustments such as storage media configuration were completed, the gradient of changes in post-adjustment 1 and subsequent monitoring periods rapidly decayed, eventually stabilizing at a low gradient level of 0.02 or below (consistently below the active threshold). Combined with the phase transition judgment logic, when the gradient of changes is consistently below the active threshold for multiple consecutive monitoring periods, a phase transition is triggered (the green dashed line corresponds to the judgment node), and the system thus transitions from the collaborative collection phase to the collaborative analysis phase.

[0159] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0160] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-center collaborative clinical data research system for complex endocrine cases, characterized in that: The system includes: The data preprocessing module acquires raw endocrine case data from different medical institutions and performs structure recognition to separate structured data components from unstructured text components. The semantic fusion module performs format alignment on the structured data components and performs entity extraction based on medical ontology on the unstructured text components to form a unified semantic representation. The processed data components are then merged to generate a multi-center case dataset with temporal correlation features. The collaborative control module uses a sliding window mechanism to traverse the multi-center case dataset, calculates the gradient of changes in data features within each window, and triggers the collaborative data collection phase when the gradient of changes continuously exceeds a preset active threshold. The task scheduling module, during the data collaborative collection phase, determines the configuration specifications and deployment location of the data storage medium based on the value and direction of the changing gradient, and generates differentiated collection task queues for data nodes of different medical institutions. The dynamic adjustment and transition module monitors the execution feedback of the differentiated collection task queue, dynamically adjusts the configuration specifications of the data storage medium according to the task backlog of each node, calculates the change gradient based on the adjusted configuration specifications, and transitions the system from the data collaborative collection stage to the data collaborative analysis stage according to the decay of the change gradient. The multicenter case dataset is traversed using a sliding window mechanism to calculate the gradient of changes in data features within each window, including the following steps: A time span window is defined, and the window is moved along the time axis of the multi-center case dataset with a fixed step size. At each window position, all case data units falling within the time range of the window are extracted. From the extracted case data units, core data feature indicators characterizing the disease progression status are selected. The statistical difference between the core data feature indicator in the current window and the corresponding indicator in the previous window is calculated. The statistical difference is ratioed to the window time span to obtain the instantaneous change rate of the data feature at the current window position. The window is continuously moved along the complete time axis of the multi-center case dataset to obtain a series of data feature instantaneous change rate sequences arranged in chronological order. The data feature instantaneous change rate sequences are smoothed and filtered to eliminate random fluctuation noise. The average change trend of the filtered data feature instantaneous change rate sequences within multiple consecutive windows is calculated, and the average change trend is the change gradient. The change gradient is stored together with the metadata of the multi-center case dataset for subsequent judgment of the system's working stage. Generate differentiated collection task queues for different data nodes in different medical institutions, including the following steps: The data storage medium configuration list is analyzed to clarify the type and deployment location of the storage medium allocated to each medical institution's data node. Based on the historical data output rate and data type preferences of each medical institution's data node, its future data generation volume is estimated. Combining the throughput specifications of the storage media allocated to each node, the maximum amount of data each node can receive and process per unit time is calculated. Based on the estimated data generation volume and the calculated maximum processable data volume, differentiated data collection time windows and single-collection data volume limits are set for each medical institution's data node. For nodes containing image data, an image preprocessing subtask, including format standardization and compression, is added to the differentiated data collection task. A sketch is generated; for nodes containing only text and numerical data, a data encryption and integrity verification subtask is added to the differentiated data collection task; all collection instructions configured with specific subtasks, time windows, and data volume limits are grouped according to their associated medical institution data nodes; within each group, the collection instructions are sorted according to their urgency and data priority to form an initial collection task queue for the medical institution data nodes; the initial collection task queues of each node are coordinated to avoid network bandwidth peak conflicts, generating a final, time-staggered differentiated collection task queue; the differentiated collection task queue is distributed to the corresponding medical institution data nodes to start the data collection process; Monitoring the execution feedback of the differentiated collection task queue and dynamically adjusting the configuration specifications of the data storage medium based on the task backlog at each node includes the following steps: The system receives task execution feedback information from data nodes of various medical institutions. This feedback includes task completion status, actual data collection volume, task execution time, and queue backlog length. It analyzes this feedback to identify abnormal nodes where the task queue backlog length is continuously increasing or the task execution time significantly exceeds expectations. For each identified abnormal node, it extracts the current configuration specifications and actual performance indicators of its corresponding data storage medium. It compares these actual performance indicators with the performance requirements of the differentiated collection task queue for the abnormal node to pinpoint the performance bottleneck. If the performance bottleneck is insufficient storage medium throughput, it generates a storage medium upgrade instruction, which instructs the user to upgrade the storage medium at a specific time. Replace the existing deployment location with a storage medium type that has a higher throughput. If the performance bottleneck is insufficient storage space, generate a storage space expansion instruction, which instructs the addition of logical storage units to the existing storage medium. If the performance bottleneck is network latency, generate a deployment location optimization instruction, which instructs the migration of some data storage media to nodes with better network paths. Issue the generated storage medium upgrade instruction, storage space expansion instruction, or deployment location optimization instruction to the infrastructure management layer for execution. After the adjustment instruction is executed, reacquire the performance metrics of the abnormal node and continue to monitor its task queue backlog until its backlog length returns to normal. The change gradient is calculated based on the adjusted configuration specifications, and the system transitions from the data collaborative collection phase to the data collaborative analysis phase according to the decay of the change gradient, including the following steps: After the dynamic adjustment of the data storage medium configuration specifications is completed and the backlog of task queues at each node stabilizes, a new round of gradient calculation is initiated. A multi-center case data set within the current time window is acquired, and the gradient of data characteristics is recalculated using the same method as the initial calculation. The recalculated gradient is compared with the active threshold, and its trend over time is observed. If the new gradient continuously decreases over multiple monitoring periods and eventually stabilizes below the active threshold, a phase transition condition is triggered. A system state transition instruction is generated to exit the data collaborative collection phase and enter the data collaborative analysis phase. The currently active differentiated collection task queue is gradually stopped, but the results of completed data collection tasks are retained. The preprocessing process for the data collaborative analysis phase is initiated, including deduplication, normalization, and quality labeling of the collected data. Dedicated analytical computing resources and algorithm model repositories are configured for entering the data collaborative analysis phase. After the preprocessing process is completed, the system's working state is officially switched to the data collaborative analysis phase, and a notification of phase transition completion is issued.

2. The multi-center clinical data collaborative research system for complex endocrine cases according to claim 1, characterized in that, The structured data components are format aligned, and the unstructured text components are subjected to entity extraction based on medical ontology to form a unified semantic representation, including the following steps: Scan the structured data components to identify their inherent data field labels and data value formats; Establish a standardized data field label mapping table to map the diverse identified data field labels to unified standardized labels; Based on a standardized data field mapping table, the numerical values ​​and units in the structured data components are uniformly converted to generate format-aligned structured data blocks; In parallel, the unstructured text components are processed by word segmentation and part-of-speech tagging; The segmented and annotated text sequence is matched with a pre-built endocrine medicine ontology knowledge base; Locate and extract medical entities and relation triples that match the concepts, attributes, and relations in the endocrine medicine ontology knowledge base from the text sequence; Assign standardized semantic identifiers to the extracted medical entities and relation triples; All medical entities and relation triples with standardized semantic identifiers are linked according to the logical order in the original text to construct a semantic graph representation of the unstructured text components; The structured data blocks aligned to the specified format are associated and stored with the semantic graph representation to complete the construction of a unified semantic representation.

3. The multi-center clinical data collaborative research system for complex endocrine cases according to claim 2, characterized in that, The merged data components are used to generate a multicenter case dataset with temporal correlation characteristics, including the following steps: Receive the semantic graph representation of the format-aligned structured data block and the unstructured text components; Using the patient's unique identifier and the timestamp of the treatment event as a joint primary key, an index is constructed for the structured data blocks aligned to the specified format; Using the unique identifier of the same patient and the timestamp of the diagnosis and treatment event as anchors, the semantic graph representation is fused with the corresponding structured data block; During the fusion process, attribute associations are established between the medical entities in the semantic graph representation and the specific detection values ​​in the structured data block; Based on the chronological order of the diagnosis and treatment event timestamps, all fused data units of the same patient are sorted to form a longitudinal time-series case data sequence for that patient; The longitudinal time-series case data of all patients were collected and grouped according to the medical institution from which the data originated; Establish cross-institutional case data association links between case data groups from different medical institutions based on the consistency of disease diagnosis codes and key medical entities; Along the cross-institutional case data link, longitudinal time-series case data sequences from different sources that are related are aggregated; A cross-center association identifier is added to each aggregated case data unit, and its time span information is recorded, ultimately forming the multi-center case data set with time-series association characteristics.

4. The multi-center clinical data collaborative research system for complex endocrine cases according to claim 1, characterized in that, When the changing gradient continuously exceeds a preset activity threshold, a data collaborative collection phase is triggered, including the following steps: The gradient change was obtained through continuous monitoring and calculation; The change gradient is compared with a pre-set activity threshold in real time; If the change gradient is higher than the activity threshold for three consecutive monitoring periods, a phase transition determination is triggered. Generate system status instructions for entering the data collaborative collection phase; Based on the specific value of the change gradient during the trigger phase transition determination, it is divided into high activity level, medium activity level, or low activity level. Based on the activity level classification, select the corresponding basic data collection strategy template from the strategy library; The basic data collection strategy template defines the basic framework for the collection frequency, data compression ratio, and transmission priority of different types of data. Based on the total amount and characteristic distribution of the current multicenter case dataset, the parameters of the selected basic data collection strategy template are fine-tuned; Instantiate the fine-tuned strategy parameters into specific, executable data collection strategy instructions; The data collection strategy instruction is activated and loaded, enabling the system to enter the working state of the collaborative data collection phase.

5. The multi-center clinical data collaborative research system for complex endocrine cases according to claim 1, characterized in that, During the collaborative data collection phase, the configuration specifications and deployment location of the data storage medium are determined based on the value and direction of the change gradient, including the following steps: During the data collaborative collection phase, the latest change gradient and its direction of change are acquired in real time. If the change direction is positive and the value is at a high activity level, then the high-speed storage media configuration process is initiated, and a high-throughput storage media type and a large-capacity storage space specification are allocated to the high-speed storage media configuration process. If the change direction is positive but the value is at a medium or low activity level, then the standard storage media configuration process is initiated, and a storage media type with medium throughput and a storage space specification with appropriate capacity are assigned to the standard storage media configuration process. If the change direction is negative, the cache storage medium configuration process is initiated, and a high-response storage medium type and a small-capacity storage space specification are allocated to the cache storage medium configuration process. Calculate the total input / output bandwidth requirement of the required storage medium based on the magnitude of the change gradient; Based on the calculated total input / output bandwidth requirements and the network topology of each medical institution, a set of data storage media deployment locations that meet the bandwidth latency requirements is determined. From the set of data storage media deployment locations, select the node with the lightest network load and the physical location closest to the core data source as the preferred deployment location; Generate a data storage media configuration list that includes information on storage media type, storage space specifications, and deployment node location; The data storage media configuration list is sent to the infrastructure management layer to complete the allocation and deployment of physical storage resources.

6. The multi-center clinical data collaborative research system for complex endocrine cases according to claim 1, characterized in that, Calculating the statistical difference between the core data feature indicators in the current window and the corresponding indicators in the previous window includes the following steps: Extract the specific values ​​of the selected core data feature indicators from all case data units within the current window; Extract the specific values ​​of the same core data feature indicators from all case data units within the previous adjacent window; Perform descriptive statistical analysis on the set of core data feature indicators within the current window, and calculate their mean, standard deviation, and interquartile range; Perform the same descriptive statistical analysis on the set of core data feature indicators in the previous window, and calculate their mean, standard deviation, and interquartile range; Calculate the absolute difference between the current window average and the previous window average, and use it as the mean difference component; Calculate the relative rate of change between the current window standard deviation and the previous window standard deviation, and use it as the variance component; Calculate the ratio of the current window interquartile range to the previous window interquartile range, and use it as the component of distribution morphology difference. The mean difference component, the variability difference component, and the distribution pattern difference component are normalized to eliminate the influence of dimensions. Assign a first weighting coefficient to the normalized mean difference component, a second weighting coefficient to the normalized variability difference component, and a third weighting coefficient to the normalized distribution shape difference component. The three weighted difference components are linearly superimposed to obtain the comprehensive statistical difference. The comprehensive statistical difference is logarithmically transformed to compress its numerical range so that it meets the input requirements for subsequent ratio calculations.

Images