Artificial intelligence-based medical cloud platform resource scheduling method and system

Abstract

The application discloses an artificial intelligence-based medical cloud platform resource scheduling method and system, relates to the field of resource data analysis, and comprises the following steps: collecting regional medical information of a target region; performing standardized granularity splitting on medical entity data to obtain a plurality of entity data blocks; extracting entity access feature vectors of all the entity data blocks, performing data correlation analysis on all the entity data blocks through the entity access feature vectors, and constructing an entity correlation topology graph according to a data correlation analysis result; dividing all the entity data blocks into a plurality of data communities, and extracting community feature vectors of all the data communities respectively; and combining the community feature vectors and an artificial intelligence reasoning engine pre-installed on the medical cloud platform to generate a resource scheduling strategy for the medical cloud platform. The application can effectively improve the resource utilization efficiency of the medical cloud platform.

Description

Technical Field

[0001] This application relates to the field of resource data analysis, and in particular to a resource scheduling method and system for a medical cloud platform based on artificial intelligence. Background Technology

[0002] With the continuous advancement of regional integrated healthcare construction, medical cloud platforms have become the core carrier for aggregating diagnostic and treatment data from medical institutions at all levels within a region and realizing cross-institutional data sharing and business collaboration. At the same time, medical cloud platforms carry massive amounts of complex data, which are frequently accessed and interacted with in different application scenarios. How to efficiently schedule the computing, storage, and network resources of medical cloud platforms to meet the differentiated needs of various medical services for response latency, throughput, and reliability is one of the core challenges currently facing medical cloud technology.

[0003] Currently, most mainstream cloud platform resource scheduling methods are based on static configuration rules or single-dimensional load metrics for resource allocation. For example, elastic scaling is triggered by setting CPU utilization or memory usage thresholds, or data is migrated to different storage tiers based on preset hot and cold data labels. These methods rely on fixed thresholds and category labels preset by human experience, making it difficult to adapt to the continuous and dynamic changes in medical business load. Resource scheduling exhibits significant lag, easily leading to insufficient resource supply during peak business periods, causing data processing delays, data interaction stutters, and other problems, affecting clinical diagnosis and treatment efficiency. They also easily result in long-term idle resources during off-peak business periods, increasing platform operating costs. Summary of the Invention

[0004] This application provides an artificial intelligence-based method and system for scheduling medical cloud platform resources, which is used to improve the efficiency of medical cloud platform resource utilization.

[0005] To achieve the above objectives, the embodiments of this application adopt the following technical solutions: Firstly, an artificial intelligence-based resource scheduling method for a medical cloud platform is provided, the method comprising: Regional medical information of the target area is collected through a medical cloud platform. This regional medical information includes medical entity data and medical interaction data. The medical entity data is split into multiple entity data blocks by standardizing the granularity of the data. Based on the medical interaction data, the entity access feature vectors of all entity data blocks are extracted. The data association analysis of all entity data blocks is completed through the entity access feature vectors, and an entity association topology graph is constructed based on the data association analysis results. Based on the entity association topology diagram, all entity data blocks are divided into multiple data communities, and community feature vectors of all data communities are extracted based on entity access feature vectors. By combining community feature vectors and the AI ​​inference engine pre-installed on the medical cloud platform, a resource scheduling strategy is generated for the medical cloud platform.

[0006] Optionally, entity access feature vectors are extracted from all entity data blocks based on medical interaction data. Data association analysis of all entity data blocks is then performed using these entity access feature vectors. Based on the data association analysis results, an entity association topology graph is constructed, including the following steps: For any entity data block, the entity multidimensional feature vector of the entity data block is extracted based on the medical interaction data, and the entity access feature vector of the entity data block is constructed based on the entity multidimensional feature vector. The entity access feature vector includes access impulse intensity, causal transmission efficiency and correlation stability, and the entity multidimensional feature vector includes access density sequence, data co-occurrence set and causal event chain. The access heat value of the entity data block is calculated by combining the access pulse intensity and causal transmission efficiency; The access time bias is extracted from the access density sequence, and the access time bias and access popularity value are used as the block access features of the entity data block. By combining entity multidimensional feature vectors, entity access feature vectors, and access popularity values, a comprehensive association strength analysis of all entity data blocks is completed, and the data block association degree of all entity data blocks is calculated based on the results of the comprehensive association strength analysis. Calculate the data offset difference between all entity data blocks based on the access timing offset; Treat all entity data blocks as entity nodes, combine data bias difference, data block correlation degree and data co-occurrence set to complete the access correlation strength verification between all entity nodes, and construct the associated directed edges between all entity nodes that pass the access correlation strength verification according to the causal event chain. Based on the access association strength verification results, association edge weights are assigned to all directed edges to obtain the entity association topology graph.

[0007] Optionally, extracting entity multidimensional feature vectors from entity data blocks based on medical interaction data, and constructing entity access feature vectors from entity data blocks based on entity multidimensional feature vectors, includes the following steps: Based on the medical interaction data, the causal chain verification between the entity data block and all other entity data blocks is completed, and a multidimensional feature vector of the entity is constructed based on the causal chain verification results. The smoothed access energy and baseline access activity of the entity data block are calculated cumulatively based on the access density sequence. The access pulse strength of the entity data block is calculated by combining the smooth access energy and the baseline access activity; The causal propagation delay between the entity data block and all entity data blocks in the causal event chain is calculated based on the medical interaction data, and the causal propagation efficiency of the entity data block is calculated based on the causal propagation delay. Calculate the Jaccard similarity among all adjacent data co-occurrence sets within a fixed time slice, and calculate the association stability of entity data blocks based on the Jaccard similarity; The entity access feature vector of the entity data block is constructed by combining access impulse intensity, causal transmission efficiency and correlation stability.

[0008] Optionally, the following steps are taken to perform causal chain verification between the entity data block and all other entity data blocks based on the medical interaction data, and to construct the entity multidimensional feature vector based on the causal chain verification results: Traverse the medical interaction data, count the number of times the entity data block is accessed within a preset fixed time slice, and construct an access density sequence based on all access counts; The time-delay cross-correlation function between the entity data block and all other entity data blocks is calculated based on the access density sequence, and the initial causal set of the entity data block is constructed based on the time-delay cross-correlation function. Based on medical interaction data, the frequency of data co-occurrence between entity data blocks within a fixed time frame and all other entity data blocks is calculated, and a data co-occurrence set of entity data blocks is constructed based on the data co-occurrence frequency. The intersection of the initial causal set and the data co-occurrence set is used to obtain the causal candidate set. Based on medical interaction data, the causal chain verification between the entity data block and all candidate data blocks in the causal candidate set is completed, and the causal event chain of the entity data block is constructed based on the causal chain verification results. By integrating access density sequences, data co-occurrence sets, and causal event chains, a multidimensional feature vector of entity data blocks is constructed.

[0009] Optionally, a comprehensive association strength analysis of all entity data blocks is performed by combining entity multidimensional feature vectors, entity access feature vectors, and access popularity values. The data block association degree of all entity data blocks is then calculated based on the comprehensive association strength analysis results, including the following steps: The data co-occurrence rate among all entity data blocks is calculated based on the data co-occurrence set. The Pearson correlation coefficient between all entity data blocks is calculated based on the access pulse intensity, and the synchronization degree between all entity data blocks is calculated by combining the data co-occurrence rate, Pearson correlation coefficient and access heat value. The causal propagation delay between all entity data blocks was calculated based on medical interaction data. The causal direction between all entity data blocks is determined based on the causal event chain, and the causal transmission value is constructed between all entity data blocks by combining the causal transmission delay, causal direction and access popularity value. The block feature spacing between all entity data blocks is calculated based on the entity access feature vector and the Euclidean distance formula. Taking any entity data block as the center, calculate the local feature density of the entity data block in the feature space based on the block feature spacing; The clustering balance penalty among all entity data blocks is calculated by combining the block feature spacing, local feature density, and access popularity value; The correlation degree between all entity data blocks is calculated by combining the synchronization degree, causal transmission value, and cluster balance penalty.

[0010] Optionally, dividing all entity data blocks into multiple data communities based on the entity association topology graph, and extracting the community feature vectors of all data communities based on the entity access feature vectors, includes the following steps: The Laplace matrix is ​​constructed based on the entity relationship topology graph, and the eigenvalues ​​and eigenvectors of the Laplace matrix are solved. An entity embedding matrix is ​​constructed by combining matrix eigenvalues ​​and eigenvectors and using the eigenvalue gap method. The spatial mapping coordinates of entity data blocks in the clustering feature space are determined based on the row vectors of the entity embedding matrix. Based on all spatial mapping coordinates and using clustering algorithms, all entity data blocks are divided into data communities to obtain multiple data communities; For any data community, the community feature vector is calculated by combining the entity association topology graph and the entity access feature vectors of all community entity data blocks within the data community. The community feature vector includes the community access popularity value, the topology convergence coefficient, and the community access synchronization degree.

[0011] Optionally, calculating the community feature vector of a data community by combining the entity association topology graph and the entity access feature vectors of all community entity data blocks within the data community includes the following steps: The community access heat value of the data community is calculated by combining the access impulse intensity and causal propagation efficiency of all community entity data blocks within the data community. Extract the associated directed edges corresponding to all community entity data blocks from the entity association topology graph, and calculate the directed edge density of the data community based on the results of the associated directed edge extraction. The topological convergence coefficient of the data community is calculated by combining directed edge density, causal transmission efficiency, and correlation stability. The access timing bias of all community entity data blocks is taken as the unit circle vector, and the vector and mean of all unit circle vectors are calculated to obtain the average composite vector. Calculate the vector magnitude of the average composite vector, and calculate the community access synchronization degree of the data community based on the vector magnitude; The community access popularity value, topological convergence coefficient, and community access synchronization degree are integrated into the community feature vector of the data community.

[0012] Optionally, generating resource scheduling strategies for the medical cloud platform by combining community feature vectors and the AI ​​inference engine pre-installed on the medical cloud platform includes the following steps: All community feature vectors are sequentially transmitted to the AI ​​inference engine pre-installed on the medical cloud platform, and the AI ​​inference engine outputs multi-dimensional resource scheduling weights. The AI ​​inference engine is built based on an online learning model, and the multi-dimensional resource scheduling weights include computing power allocation weights, storage level weights, network bandwidth weights, and prefetch depth weights. Multidimensional resource scheduling parameters are calculated by combining community feature vectors and multidimensional resource scheduling weights; Based on the pre-acquired full-platform resource parameters of the medical cloud platform, global resource optimization of multi-dimensional resource scheduling parameters is completed, and a resource scheduling strategy for the medical cloud platform is generated based on the multi-dimensional resource scheduling parameters after global resource optimization. The resource scheduling strategy includes computing power allocation parameters, storage level parameters, network bandwidth parameters, and prefetch depth parameters for all data communities.

[0013] Secondly, this application provides an artificial intelligence-based medical cloud platform resource scheduling system, comprising: The memory is configured to store instructions; and The processor is configured to retrieve the instructions from the memory and, when executing the instructions, to implement the AI-based medical cloud platform resource scheduling method as described in the first aspect.

[0014] Thirdly, this application provides a machine-readable storage medium storing instructions for causing a machine to execute the AI-based medical cloud platform resource scheduling method described in the first aspect.

[0015] The above technical solution collects regional medical information and breaks down medical entity data within this information into entity data blocks. Through lossless splitting at a fixed granularity, all forms of medical data are transformed into the smallest processing units of uniform size, providing a standardized operational platform for subsequent quantitative analysis and unified scheduling across the entire process. This also reduces the resource overhead of data transport, laying the foundation for improving the effective output rate of computing resources. Next, by extracting entity access feature vectors from the entity data blocks, not only are data access behaviors quantified into calculable feature indicators, but deep intrinsic relationships between data are also uncovered through data association analysis, reconstructing the true linkage logic of data in medical operations. Furthermore, by extracting entity access feature vectors, resource supply and access demand can be accurately matched, avoiding resource mismatch and ensuring resources flow to data that truly needs them, reducing unnecessary resource waste and improving the resource utilization efficiency of the medical cloud platform. The construction of the entity association topology graph provides a structured basis for subsequent community division and collaborative scheduling. Next, by extracting community feature vectors from each data community, a grouping basis is provided for differentiated resource strategies. Furthermore, scheduling by community significantly reduces the number of decision-making units and the overhead of the scheduling system itself, freeing up more computing power for medical business computation and indirectly improving the overall effective computing power supply of the platform. Community feature vectors provide a data foundation for precise resource scheduling. Combining community feature vectors with the AI ​​inference engine pre-installed on the medical cloud platform generates resource scheduling strategies, avoiding global resource imbalances caused by local community resource contention. This maximizes the overall utilization efficiency of platform resources while ensuring the resource needs of high-priority communities. In conclusion, this application can effectively improve the resource utilization efficiency of the medical cloud platform and reduce its operating costs.

[0016] Other features and advantages of the embodiments of this application will be described in detail in the following detailed description section. Attached Figure Description

[0017] Figure 1 A flowchart illustrating a resource scheduling method for a medical cloud platform based on artificial intelligence, provided as an embodiment of this application; Figure 2 A flowchart illustrating an entity association topology graph construction method provided in this application embodiment; Figure 3 This is a flowchart illustrating a community feature vector calculation method provided in an embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for illustration and explanation of the embodiments of this application and are not intended to limit the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0019] It should be noted that if the embodiments of this application involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0020] Furthermore, if the embodiments of this application involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed in this application.

[0021] Figure 1 The illustration shows a flowchart of an artificial intelligence-based medical cloud platform resource scheduling method according to an embodiment of this application. Figure 1 As shown in the figure, this application provides a resource scheduling method for a medical cloud platform based on artificial intelligence, which may include the following steps: S101. Collect regional medical information of the target area through a medical cloud platform, including medical entity data and medical interaction data.

[0022] In this embodiment, a distributed acquisition probe of a medical cloud platform is used to connect to the diagnostic and treatment business systems of all medical institutions within the target area, completing the collection of full-volume regional medical information. Regional medical information is divided into medical entity data and medical interaction data. Medical entity data is the raw medical information generated during the diagnosis and treatment process, encompassing core data directly related to patient diagnosis and treatment, such as medical images, electronic medical records, physiological signal sequences, and laboratory and pathological data. It serves as the service object and data carrier for resource scheduling. Medical interaction data is the full-volume behavioral log data generated during user access, system operation, and computational task invocation. It includes information such as data access timestamps, operation types, origin identifiers, transaction attribution, and operation sequence, serving as the core analytical basis for mining data access patterns and correlations. The collection process adheres to medical data security and compliance requirements, performing desensitization preprocessing on entity data and anonymizing interaction data. Full data access is completed while ensuring data privacy, breaking down data silos between different medical institutions and business systems within the target area. The raw medical data and operational behavior data are decoupled during collection, providing a complete and compliant data foundation for subsequent full-dimensional behavioral analysis and global resource scheduling.

[0023] S102. Standardize the granularity of the medical entity data to obtain multiple entity data blocks.

[0024] In this embodiment, a fixed granularity for data segmentation is first preset. This granularity is determined based on a balance between the fineness of scheduling and the overhead of metadata management. If the granularity is too fine, the total amount of metadata will expand, increasing the system's storage and computational burden; if the granularity is too coarse, fine-grained resource scheduling cannot be achieved. A reference value can be set to 1MB. All medical entity data is losslessly split according to the preset fixed granularity, with each split data unit becoming an entity data block. The splitting process preserves the mapping relationship between each entity data block and the original medical entity data, while maintaining the logical integrity and built-in verification structure within the data block, without disrupting the business continuity and traceability of the medical data, thus meeting the accuracy requirements of medical data. After splitting, a globally unique identifier is assigned to each entity data block for indexing and locating in all subsequent analysis stages.

[0025] S103. Extract the entity access feature vectors of all entity data blocks based on the medical interaction data, complete the data association analysis of all entity data blocks through the entity access feature vectors, and construct the entity association topology graph based on the data association analysis results.

[0026] In this embodiment, medical interaction data is traversed, and basic features are extracted from three dimensions: temporal fluctuation, co-occurrence association, and causal dependency. This constructs a multi-dimensional feature vector for entities, which includes access impulse intensity, causal transmission efficiency, and association stability. Based on this, three types of data association quantitative indicators are further quantified using the multi-dimensional feature vector to construct an entity access feature vector, which includes access impulse intensity, causal transmission efficiency, and association stability. Next, access heat value is calculated by fusing access impulse intensity and causal transmission efficiency, representing the field effect strength of entity data blocks. Access temporal bias is extracted from the access density sequence to represent the rhythmic state of data block access fluctuations. Then, three types of data linkage parameters are calculated and superimposed to obtain a comprehensive association score for a single set of data: a synchronous linkage score, quantifying the degree of business binding between multiple medical data synchronous retrievals and transaction coexistence; a causal transmission score, calculated only for data groups with sequential business dependencies, representing the directional linkage relationship in the diagnosis and treatment process; and a clustering equilibrium constraint value, used to limit excessive aggregation of similar data and prevent the emergence of ultra-large clusters. The correlation between the three types of indicators is obtained by combining them, and the overall business linkage between the two sets of medical data is uniformly quantified.

[0027] The correlation between data blocks is modulated over time, and multiple layers of screening are performed by combining symbiotic hard constraints and correlation strength thresholds to filter out false and weak correlations with no business overlap. For data block pairs that pass the verification, the directional attributes of the correlation edges are determined by combining causal event chains, and the correlation strength after verification is used as the correlation edge weight. Finally, an entity correlation topology graph is assembled with all entity data blocks as nodes and all valid correlation edges as edge sets.

[0028] By constructing an entity-related topology graph, a leap from single-block micro-features to group-related structures was achieved, accurately restoring the data collaboration pattern of medical services and providing structural support that conforms to business logic for subsequent community-level overall scheduling. This solves the problem of traditional medical clouds storing data in a scattered manner according to business systems. Medical records, images, and test data associated with the same diagnosis and treatment are often distributed across different nodes, requiring multiple data transfers across nodes during business calculations. This consumes a large amount of network bandwidth and causes computing resources to wait for data for extended periods, resulting in ineffective idle time. This effectively improves the resource utilization efficiency of the medical cloud platform.

[0029] S104. Divide all entity data blocks into multiple data communities according to the entity association topology diagram, and extract the community feature vectors of all data communities according to the entity access feature vectors.

[0030] In this embodiment, a normalized symmetric Laplacian matrix is ​​constructed based on the adjacency structure of the entity association topology graph. Solving this matrix yields ordered matrix eigenvalues ​​and corresponding eigenvectors. The eigenvalue gap method is used to determine the optimal number of communities. A corresponding number of core feature vectors are selected to construct an entity embedding matrix, mapping each entity data block to spatial coordinates in the clustering feature space. Using all spatial coordinates as input, k-means clustering is performed, ultimately dividing all entity data blocks into multiple data communities with tightly connected internal connections and sparse external connections. Each community corresponds to a set of datasets that are naturally linked in business logic. For each data community, three core indicators are calculated and integrated into a community feature vector, combining the entity access feature vectors of the entity data blocks within the community with the topology graph structure information. The three core indicators are community access popularity, topology convergence coefficient, and community access synchronization.

[0031] Community feature vectors aggregate the access strength, coupling tightness, and concurrency characteristics of the entire community into low-dimensional, standardized quantitative indicators. These can be directly used as input parameters for inference models, supporting the calculation of subsequent multi-dimensional scheduling weights such as computing power, storage, bandwidth, and prefetching. This completes the mapping from data features to scheduling strategies. Furthermore, based on the complete causal links and access synchronization characteristics of the community, unified prefetching and computing power pre-deployment can be performed on the entire set of linked data, resulting in higher accuracy and more complete business coverage compared to single-block prefetching. Completing data loading and resource reservation in advance before tasks arrive reduces business response latency and minimizes fragmented resource allocation, allowing computing power, storage, and bandwidth resources to continuously support business operations and improving continuous resource utilization.

[0032] S105. Combine community feature vectors and the AI ​​inference engine pre-installed on the medical cloud platform to generate resource scheduling strategies for the medical cloud platform.

[0033] In this embodiment, the complete community feature vectors of all data communities are input into the artificial intelligence inference engine, which learns in real time the community load characteristics, access timing characteristics, causal coupling characteristics, and business sensitivity characteristics. Then, based on the community feature vectors, the artificial intelligence inference engine intelligently determines the interval in which the comprehensive scheduling cost value lies, and adaptively and uniformly adjusts the computing power allocation parameters, network bandwidth parameters, storage level parameters, and prefetch depth parameters according to a fixed tier, completing the global adaptive optimization of multi-dimensional resource parameters. The optimized four types of resource scheduling parameters are bound to the corresponding community resource nodes of each data community to generate the final intelligent resource scheduling strategy for the medical cloud platform. Through the above steps, the on-demand and precise allocation of medical cloud platform resources can be achieved, which is a core decision-making link to ensure the stable and efficient operation of medical services. This fundamentally reduces resource mismatch problems, allowing limited platform resources to be tilted towards core medical services, thereby improving the resource utilization efficiency of the medical cloud platform.

[0034] In one embodiment, reference is made to Figure 2 The process involves extracting entity access feature vectors from all entity data blocks based on medical interaction data, performing data association analysis on all entity data blocks using these feature vectors, and constructing an entity association topology graph based on the analysis results. The steps include: S201. For any entity data block, extract the entity multidimensional feature vector of the entity data block based on the medical interaction data, and construct the entity access feature vector of the entity data block based on the entity multidimensional feature vector. The entity access feature vector includes access impulse intensity, causal transmission efficiency and correlation stability, and the entity multidimensional feature vector includes access density sequence, data co-occurrence set and causal event chain. S202. Calculate the access heat value of the entity data block by combining the access pulse intensity and causal transmission efficiency; S203. Extract the access time sequence bias from the access density sequence, and use the access time sequence bias and access popularity value as the block access features of the entity data block. S204. Combine the entity multidimensional feature vector, entity access feature vector and access popularity value to complete the comprehensive association strength analysis of all entity data blocks, and calculate the data block association degree of all entity data blocks based on the comprehensive association strength analysis results. S205. Calculate the data offset difference between all entity data blocks based on the access timing offset; S206. Treat all entity data blocks as entity nodes, combine data bias difference, data block correlation degree and data co-occurrence set to complete the access correlation strength verification between all entity nodes, and construct the associated directed edges between all entity nodes that pass the access correlation strength verification according to the causal event chain. S207. Based on the access association strength verification results, assign association edge weights to all directed edges of association to obtain the entity association topology graph.

[0035] In this embodiment, the medical interaction data is first traversed, and basic features are extracted from three dimensions: temporal fluctuation, co-occurrence association, and causal dependency. These features are then used to construct a multi-dimensional feature vector for entities, which includes access impulse intensity, causal transmission efficiency, and association stability. Specifically, the access density sequence, using a preset fixed time slice as the statistical unit, counts the number of times entity data blocks are accessed segment by segment and arranges them chronologically, representing the fluctuation pattern of data block access popularity over time. The data co-occurrence set is the set of other data blocks that are jointly accessed with the target data block in the same transaction within each time slice, representing the scope of business-related partners of the data block. The causal event chain, through a three-level progressive mechanism of initial screening by temporal cross-correlation, co-occurrence set filtering, and in-transaction read / write verification, identifies the directed causal dependencies between data blocks and splices them into a transmission link, representing the transmission path of the data block in the business flow. Building upon this foundation, three types of data association metrics are further quantified using entity multidimensional feature vectors to construct an entity access feature vector. These metrics include: access impulse intensity, calculated by comparing the short-term smoothed access energy with the long-term baseline access activity using access density sequences, and obtained after logarithmic compression; representing the explosive intensity of data block access behavior; causal transmission efficiency, calculated based on the transmission delay of the causal event chain, and obtained after smoothing and accumulation; representing the speed and reliability of data block triggering downstream data updates; and association stability, calculated smoothly using the Jaccard similarity of adjacent time-slice data co-occurrence sets; representing the robustness of the data block's associated partner set. This three-dimensional feature vector constitutes the core feature space for entity data block data linkage analysis.

[0036] The calculation of access popularity value integrates two dimensions: access pulse intensity and causal transmission efficiency. It can reflect both the instantaneous concurrent access intensity of a single piece of medical data and the additional correlation weight brought about by business flow linkage. The higher the transmission efficiency, the greater the increase in data linkage weight. The calculation method is: Access Popularity Value = Access Pulse Intensity × (1 + Causal Transmission Efficiency). A constant of 1 is introduced. When the causal transmission efficiency is 0, the popularity value is still equal to the access pulse intensity, eliminating the problem of zero weight for data without flow.

[0037] The instantaneous temporal code for each entity data block is extracted from the access density sequence, and the temporal code value is uniformly constrained within the interval [0, 2π]. The supporting engineering implementation specifications for the temporal code update rules are as follows: Peak access threshold calculation rules: The 90th percentile of access pulse intensity is statistically analyzed using a 24-hour sliding business window to match the daily outpatient treatment cycle; the streaming access log uses the TDigest approximate quantile algorithm to update the intensity distribution within the window in real time, and expired window data can be discarded every hour to continuously maintain the latest quantile threshold. When the current access pulse intensity of an entity data block exceeds this 90th percentile, it is determined that a batch of concentrated access peaks has occurred, and the current time-series coding value is directly reset to 0. Time-series coding linear increment calculation formula: Set the coding increment corresponding to a fixed statistical step size t. Update the formula: During periods without peak access, an accumulation is performed every fixed statistical step, which can be 1 minute. After accumulation, if... For values ​​greater than or equal to 2π, the code is automatically moduloed by 2π to bring it back to the [0, 2π] range, preventing the value from increasing indefinitely. This code is a pure business time-series labeling parameter, used only to quantify the time misalignment difference of peak retrieval times for different entity datasets. Subsequently, the time-series matching coefficient is calculated using the data bias difference between the two sets of data, serving as the core input indicator for dividing collaborative data communities and pre-allocating pre-fetch bandwidth. All calculations are based on statistical analysis of business access log values ​​and do not involve any physical models or fluctuation simulation logic.

[0038] This design enables the temporal offset encoding to accurately reflect the rhythmic state of data block access fluctuations. When retrieving entity data with overlapping peaks in batches, their temporal encoding values ​​are similar. After extracting the access temporal offset, the access popularity value and the access temporal offset are jointly encapsulated into the block access characteristics of the entity data block. The former describes the field effect strength of the entity data block, and the latter represents the rhythmic state of the entity data block's fluctuations. Together, they constitute the complete data block attribute set for subsequent correlation analysis and correlation verification.

[0039] The comprehensive association strength analysis involves the independent calculation of three types of data association quantification indicators, which are then synthesized to obtain the data block association degree. These three types of data association quantification indicators include synchronization linkage degree, causal transmission value, and clustering balance penalty. The synchronization linkage score characterizes the business binding strength resulting from simultaneous access to multiple data sources; the causal transmission score characterizes the targeted linkage strength resulting from the dependence on the order of diagnosis and treatment; and the clustering balance penalty is a cluster size control parameter used to limit excessive data aggregation and ensure a balanced granularity in data community partitioning.

[0040] The data offset difference between all entity data blocks is calculated based on the access timing offset. First, the absolute value of the difference between two access timing offsets is calculated. Then, the smaller of this difference and 2π minus this difference is taken to ensure that the final data offset difference falls within the range of 0 to π, accurately reflecting the minimum data offset difference between the two oscillation rhythms. The formula for calculating the data offset difference is:

[0041] in, This represents the data offset difference between entity data block i and entity data block j; This represents the access timing offset of entity data block i. This represents the access timing offset of entity data block j. The closer the data offset difference is to 0, the higher the overlap of the peak access times for the two data blocks. When the value approaches π, the peak access times for the two types of services are completely staggered.

[0042] All entity data blocks are treated as entity nodes. The access association strength between all entity nodes is verified by combining data bias difference, data block correlation, and data co-occurrence set. Directed edges are then constructed between all entity nodes that pass the access association strength verification based on the causal event chain. The core function of this step is to filter out entity node pairs with genuine business relationships through a multi-layered progressive mechanism of strength transformation, temporal modulation, hard constraint filtering, and threshold verification. The causal event chain is used to assign a clear directional attribute to the association edges, ensuring that the final association edges have both data linkage basis and business rationality. The specific implementation process consists of four stages: First, the original linkage score is transformed. The positive linkage component in the data block correlation is converted into a positive original association strength. Cluster constraints, i.e., the node pairs with dominant clustering equilibrium penalties, have their original association strength set to 0, indicating they do not have the basis for forming an association. The formula for calculating the original linkage score is:

[0043] in, This represents the initial association strength between entity node i and entity node j; This indicates the degree of correlation between the two data blocks.

[0044] Second, temporal modulation strength correction. A temporal matching coefficient is calculated based on the data bias difference to modulate the original correlation strength, suppressing spurious correlations caused by temporal misalignment. The temporal matching coefficient ranges from 0 to 1. When the peaks of two sets of data overlap, the temporal matching coefficient approaches 1; when the peaks are completely offset, the temporal matching coefficient approaches 0; when the access timing biases are opposite, the coefficient approaches 0, and the correlation strength is significantly weakened. The formula for calculating the temporal matching coefficient is:

[0045] in, Represents the temporal matching coefficient between entity node i and entity node j; This represents the difference in data bias between the two. Multiplying the original association strength by the temporal matching coefficient yields the temporally modulated association strength.

[0046] Third, co-occurrence constraints and strength threshold verification. First, a hard co-occurrence constraint verification is performed: only when two entity nodes have at least one co-occurrence record in the data co-occurrence set, i.e., have been jointly accessed within the same transaction window, is their association strength retained; otherwise, it is directly determined that there is no valid association. This step can eliminate a large number of noisy associations due to accidental similarity in feature vectors but with no business overlap. Then, a soft association strength threshold verification is performed: all non-zero association strengths that pass the co-occurrence verification are collected, and the upper quartile of their distribution is calculated as the association strength threshold. This threshold is determined based on the statistical significance of the association, filtering weak association noise while retaining the core association structure; the reference value is taken as the upper quartile of all non-zero association strengths. Only when the time-series modulated association strength is greater than or equal to this threshold is it determined to have passed the access association strength verification.

[0047] Fourth, constructing directed edges. For all verified node pairs, traverse the causal event chain to confirm their causal relationship: if a causal edge exists from node i to node j, construct a directed edge pointing from i to j, with a corresponding propagation strength attribute; if a reverse causal edge exists, construct a directed edge in the corresponding direction; if there is no causal relationship, construct an edge without a clear causal direction. The correlation strength value verified by accessing the correlation strength is used as the edge weight of the corresponding directed edge. A higher weight indicates a stronger correlation between the two entity nodes. Using all entity data blocks as the entity node set and all weighted, directional edges as the edge set, a formal entity correlation topology graph is formed. The final generated entity correlation topology graph has three core characteristics: first, sparsity, with the number of edges being much smaller than a fully connected graph after multiple layers of filtering, and controllable computational overhead; second, weighting, with edge weights fully integrating correlation strength, peak overlap, and causal propagation characteristics; and third, directional attributes, with edges bearing causal direction labels, supporting the calculation of subsequent features such as the community topology convergence coefficient. This topology diagram can be directly used as input for the subsequent data community segmentation process.

[0048] In one embodiment, extracting entity multidimensional feature vectors of entity data blocks from medical interaction data and constructing entity access feature vectors of entity data blocks based on entity multidimensional feature vectors includes the following steps: Based on the medical interaction data, the causal chain verification between the entity data block and all other entity data blocks is completed, and a multidimensional feature vector of the entity is constructed based on the causal chain verification results. The smoothed access energy and baseline access activity of the entity data block are calculated cumulatively based on the access density sequence. The access pulse strength of the entity data block is calculated by combining the smooth access energy and the baseline access activity; The causal propagation delay between the entity data block and all entity data blocks in the causal event chain is calculated based on the medical interaction data, and the causal propagation efficiency of the entity data block is calculated based on the causal propagation delay. Calculate the Jaccard similarity among all adjacent data co-occurrence sets within a fixed time slice, and calculate the association stability of entity data blocks based on the Jaccard similarity; The entity access feature vector of the entity data block is constructed by combining access impulse intensity, causal transmission efficiency and correlation stability.

[0049] In this embodiment, using any entity data block as the verification core, a three-level progressive causal chain verification is carried out based on medical interaction data: The first level is the initial screening of temporal associations, which calculates the number of time-delayed cross-correspondences based on the access density sequence, and selects candidate data blocks that have a significant lag association with the entity data block in terms of temporal fluctuations, forming an initial causal set; the second level is the co-occurrence feature filtering, which counts the frequency of data co-occurrence between entity data blocks, eliminates false associations that are only similar in temporal waveforms but never co-occur in the same business transaction, and narrows down the candidate range to obtain a causal candidate set; the third level is the transaction operation verification, which matches the operation sequence of "reading the entity data block → writing the candidate data block" for each candidate data block in the causal candidate set, and counts the support, average propagation delay, and propagation purity of the corresponding operation sequence to select effective causal edges, with the direction from the entity data block to the candidate data block. All effective causal edges are spliced ​​together according to the propagation direction to finally form the causal event chain of the entity data block. After the causal event chain is constructed, the three types of features—access density sequence, data co-occurrence set, and causal event chain—are structurally integrated to form the entity multidimensional feature vector of the entity data block.

[0050] First, the access density sequence is extracted from the entity's multi-dimensional feature vector. Using a preset fixed time slice as the unit, the data access energy corresponding to each time slice is calculated. The data access energy is taken as the square of the number of accesses within the corresponding time slice. The squaring operation amplifies the weight of high-concurrency accesses, highlighting the intensity differences of access peaks and more sensitively capturing pulse-like access characteristics. Based on this, exponentially weighted moving average methods with different smoothing coefficients are used to accumulate the data access energy for each time slice, obtaining smoothed access energy and baseline access activity. The smoothed access energy uses a higher update coefficient, which is more sensitive to recent access fluctuations and is used to capture instantaneous access bursts; the baseline access activity uses a lower update coefficient, with a smoother update speed, and is used to reflect the long-term baseline access level of the data block. The update formula for smoothed access energy is:

[0051] in, Let be the smooth access energy corresponding to the t-th fixed time slice; This represents the smooth access energy corresponding to the (t-1)th fixed time slice; The data access energy for the t-th fixed time slice; This is the smoothing coefficient for the first window. Its value is determined based on the capture sensitivity of short-term pulses. The higher the sensitivity requirement, the larger the value should be. The reference value is 0.2.

[0052] The formula for updating the baseline access activity is:

[0053] in, This represents the long-term average activity level corresponding to the t-th fixed time slice, i.e., the baseline access activity level. This represents the long-term average activity corresponding to the (t-1)th fixed time slice; This represents the number of visits within the t-th fixed time slice; This is the smoothing coefficient for the second window, which is smaller than the smoothing coefficient for the first window. Its value is determined based on the stability requirements of the long-term benchmark. The higher the stability requirement, the smaller the value. The reference value is 0.01.

[0054] The access pulse intensity of an entity data block is calculated by combining smoothed access energy and baseline access activity. The core function of this step is to quantify the explosiveness of entity data block access behavior, using a long-term baseline load as a reference to measure the surge rate of instantaneous access, forming a standardized pulse intensity index that represents the severity of simultaneous access to the data block by bursty high-concurrency transactions. Specific steps include: Since smoothed access energy is on the order of the square of the number of accesses, and baseline access activity is on the order of the original number of accesses, their dimensions are not consistent, making direct ratio calculation impossible. Therefore, the baseline access activity is first squared to bring it to the same dimension as the access energy, while a very small regularization constant is introduced to prevent division by zero errors, resulting in the long-term baseline access energy. The ratio of the smoothed access energy of the current time slice to the long-term baseline access energy is calculated to obtain the access pulse multiplier. This multiplier directly reflects the surge rate of the current short-term access energy relative to the long-term normal access level and is the core intermediate index for measuring pulse intensity. Because the original access pulse multiplier has an extremely wide distribution range, increasing by tens or even hundreds of times under extreme peak traffic conditions, such a large numerical range increases the difficulty of subsequent feature calculations and threshold determination. Therefore, the natural logarithm function is used to compress the pulse multiplier, smoothing out the impact of extreme values ​​while preserving the relative magnitude of the values, resulting in a more stable index distribution. The final formula for calculating the access pulse intensity is: ,in, Let t be the access pulse multiplier corresponding to the t-th fixed time slice; a constant 1 is added to the formula to ensure that when the pulse multiplier is 0, the calculated result of the access pulse intensity is exactly 0, which is consistent with the physical meaning of the no-access state.

[0055] Next, the causal propagation delay between the entity data block and all entity data blocks in the causal event chain is calculated based on the medical interaction data, and the causal propagation efficiency of the entity data block is calculated based on the causal propagation delay. From the causal event chain of the entity's multi-dimensional feature vector, all valid causal edges with the entity data block as the upstream node are extracted. For each causal edge, the time difference between each "read upstream data block → write downstream data block" operation pair is extracted from the transaction operation records of the medical interaction data; this is the single causal propagation delay. The single causal propagation delay is converted into an instantaneous propagation score between 0 and 1, with a higher score for shorter delays. When the delay exceeds a reasonable range, the score approaches 0, thus reflecting the impact of propagation speed on propagation efficiency. Then, an exponentially weighted moving average method is used to cumulatively update the instantaneous propagation score, automatically reducing the influence of distant historical events and continuously reflecting the recent causal propagation capability level of the data block. The update formula for causal propagation efficiency is:

[0056] in, This indicates the updated causal transmission efficiency. This represents the causal propagation efficiency before the update, which is the state value maintained at the previous time step. This is the transmission efficiency update coefficient, whose value is based on the weighting of old and new events, with a reference value of 0.1. This represents the instantaneous conduction score.

[0057] At the end of each fixed time slice, the set of co-occurring data blocks corresponding to the current time slice is obtained from the data co-occurrence set of the entity's multidimensional feature vectors, and the cached data co-occurrence set of the previous fixed time slice is read. The Jaccard similarity between the data co-occurrence sets of two adjacent time slices is calculated, which is the ratio of the number of elements in the intersection to the number of elements in the union of the two sets, used to measure the degree of overlap of associated partners within adjacent time slices. An exponentially weighted moving average is used to smoothly accumulate the Jaccard similarity for each time slice to obtain the association stability of the entity data blocks. The update formula for association stability is:

[0058] in, This represents the correlation stability corresponding to the t-th fixed time slice; This represents the correlation stability corresponding to the (t-1)th fixed time slice, which is the state value maintained at the previous time step. This represents the stability smoothing coefficient, and its value is determined based on the response sensitivity to correlated changes. A value of 0.2 is acceptable. Let represent the Jaccard similarity corresponding to the t-th fixed time slice.

[0059] If an entity data block is not accessed within the current fixed time slice and the corresponding data co-occurrence set is empty, then the update is skipped, and the association stability value from the previous time slice is retained. The association stability value ranges from 0 to 1. The closer the value is to 1, the more stable the set of associated partners of the data block, and the stronger the coupling between data; the closer the value is to 0, the more frequently the associated objects change, and the looser the coupling relationship. After the update is completed, the cache is overwritten with the data co-occurrence set of the current time slice to prepare for the next calculation.

[0060] The calculated access impulse strength, causal propagation efficiency, and association stability are structurally combined in a fixed order to form the entity access feature vector of the entity data block. This vector is pushed to the subsequent data association analysis and association calculation module in a streaming manner, and can be directly used as input parameters for block access feature calculation and comprehensive association strength analysis, supporting the construction of entity association topology graph and subsequent resource scheduling decisions.

[0061] In one embodiment, the process of verifying the causal chain between an entity data block and all other entity data blocks based on medical interaction data, and constructing a multidimensional feature vector of the entity based on the causal chain verification results, includes the following steps: Traverse the medical interaction data, count the number of times the entity data block is accessed within a preset fixed time slice, and construct an access density sequence based on all access counts; The time-delay cross-correlation function between the entity data block and all other entity data blocks is calculated based on the access density sequence, and the initial causal set of the entity data block is constructed based on the time-delay cross-correlation function. Based on medical interaction data, the frequency of data co-occurrence between entity data blocks within a fixed time frame and all other entity data blocks is calculated, and a data co-occurrence set of entity data blocks is constructed based on the data co-occurrence frequency. The intersection of the initial causal set and the data co-occurrence set is used to obtain the causal candidate set. Based on medical interaction data, the causal chain verification between the entity data block and all candidate data blocks in the causal candidate set is completed, and the causal event chain of the entity data block is constructed based on the causal chain verification results. By integrating access density sequences, data co-occurrence sets, and causal event chains, a multidimensional feature vector of entity data blocks is constructed.

[0062] In this embodiment, the duration of the preset fixed time slice is determined based on the balance between the minimum response cycle of medical business operations and the platform's statistical calculation overhead. Too coarse a granular time will lose details of temporal fluctuations, while too fine a granular time will amplify random noise and increase computational load. A reference value can be set to 1 minute. The medical interaction data is traversed, and using the preset fixed time slice as the statistical unit, the total number of accesses to the entity data block within each time slice is counted segment by segment. The statistical scope covers all types of access operations, including read, write, and query. The access counts corresponding to all time slices are arranged in chronological order to form a continuous temporal numerical sequence, which is the access density sequence of the entity data block. The access density sequence can intuitively reflect the fluctuation pattern of the access popularity of the entity data block over time and serves as the basic data carrier for subsequent temporal correlation analysis and impulse feature calculation.

[0063] Using the access density sequence of an entity data block as the baseline sequence, and the access density sequence of each other entity data block as the comparison sequence, the time-delay cross-correlation function between the two is calculated to quantify the waveform matching degree between the two access density sequences under different time delays. During the time-delay cross-correlation calculation, the comparison sequence is shifted relative to the baseline sequence. The shift range is set based on the maximum reasonable causal delay of the data access link in medical operations, with a reference value set to ±10 fixed time slices. At the time delay position corresponding to each shift amount, the Pearson correlation coefficient formula is used to calculate the Pearson correlation coefficient between the two sequences. The Pearson correlation coefficient formula is as follows:

[0064] in, Indicates time delay as The Pearson correlation coefficient between two access density sequences. The time lag of the alignment sequence relative to the reference sequence is indicated by a positive value, which means that the alignment sequence lags behind the reference sequence, and a negative value means that the alignment sequence leads the reference sequence. This represents the number of accesses in the baseline sequence, i.e., the access density sequence of the entity data block, at the t-th fixed time slice; This indicates that the alignment sequence, i.e., the access density sequence of the entity data block to be verified, is in the 1st order. The number of visits per fixed time slice; The mean number of visits to the baseline sequence; The average number of visits to the alignment sequence; The fixed total number of time slices for accessing the density sequence.

[0065] After traversing all time delay values, record the maximum correlation coefficient and the optimal time delay corresponding to each alignment sequence. Set a screening threshold for the maximum correlation coefficient, such as 0.65. Screen entity data blocks that simultaneously meet two conditions to form the initial causal set for that entity data block: the first condition is that the maximum correlation coefficient is greater than or equal to the preset screening threshold, indicating that the waveform similarity of the two access density sequences is statistically significant; the second condition is that the optimal time delay is greater than 0, indicating that the access fluctuation of the entity data block to be verified lags behind that entity data block in time sequence, which conforms to the temporal logic of causal transmission. The initial causal set completes the first level of coarse temporal screening, retaining candidate data blocks with potential causal relationships in the temporal pattern.

[0066] Data co-occurrence refers to the phenomenon where two entity data blocks are accessed together within the same fixed time slice and the same business transaction. Data co-occurrence frequency is used to quantify the degree of association between two entity data blocks in business access. By traversing medical interaction data with a statistical period consistent with the access density sequence, the identities of other entity data blocks accessed simultaneously in each access of the entity data block are counted on a time slice basis. The cumulative number of times each entity data block co-occurs with the current entity data block is the data co-occurrence frequency for that entity data block. A filtering threshold for data co-occurrence frequency is set, based on the minimum effective co-occurrence strength of business association, eliminating noise interference from single, accidental co-occurrences. A reference value can be set to a minimum of 3 times in the total statistical time slices. All entity data blocks whose data co-occurrence frequency meets the preset threshold are aggregated to form the data co-occurrence set for that entity data block. The data co-occurrence set filters associated data blocks from the spatial dimension of business co-occurrence, eliminating interference items that are only similar in temporal waveforms but have no actual business association.

[0067] The initial causal set and the data co-occurrence set are intersected, retaining only entity data blocks that appear in both sets simultaneously, forming a causal candidate set. This step further narrows down the candidate data block range through dual constraints of temporal and spatial co-occurrence features: it filters out spurious associations that are only temporally related but never co-occur in business transactions, and also filters out invalid associations that co-occur only occasionally but lack temporal causal logic. This significantly reduces the computational load of subsequent causal verification while maintaining candidate recall, improving overall analysis efficiency and result accuracy. For each candidate data block in the causal candidate set, a refined causal chain verification is performed based on the original medical interaction data, verifying the authenticity of causal dependencies at the level of operation sequences within transactions.

[0068] The verification process first identifies all business transactions that simultaneously contain read operations on the entity data block and write operations on candidate data blocks. Within each transaction, operations are sorted by timestamp and matched against the "read entity data block → write candidate data block" operation sequence pattern. The time difference for each pair of matching operations is then calculated. For each pair of entity data blocks and candidate data blocks, three verification metrics are aggregated: the first is support, which is the total number of times the above read-write operation sequence occurs in all transactions; the second is average propagation latency, which is the arithmetic mean of the time differences of all matching operations; and the third is propagation purity, which is the proportion of the number of times a candidate data block is written after a read of the entity data block to the total number of times a candidate data block is written in all transactions that contain reads of the entity data block.

[0069] The support threshold is determined based on the statistical significance of the causal relationship to avoid misjudgments caused by accidental operations; a reference value of 5 times can be set. The average propagation delay threshold is determined based on the maximum reasonable processing delay of the medical business operation link; a reference value of 10 seconds can be set. The propagation purity threshold is determined based on the inevitability of causal triggering to exclude accidental triggering caused by background noise; a reference value of 0.6 can be set. When a candidate data block meets all three threshold requirements, a directed causal edge is confirmed between the entity data block and the candidate data block, with the direction from the entity data block to the candidate data block, and the corresponding average propagation delay and causal strength parameters are associated and recorded. After verifying all candidate data blocks, all confirmed valid directed causal edges are spliced ​​together. If there are continuous causal edges where the entity data block points to data block A and data block A points to data block B, they are automatically spliced ​​to form a longer dependency link, ultimately constituting the causal event chain of the entity data block. The causal event chain can completely depict the causal propagation path and dependency relationship with the entity data block as the node, providing a direct basis for subsequent causal propagation value calculation and directed edge construction of the topology graph.

[0070] By structurally integrating three types of features—access density sequence, data co-occurrence set, and causal event chain—a multi-dimensional feature vector for the entity data block is formed. Specifically, the access density sequence corresponds to the temporal access fluctuation dimension, the data co-occurrence set corresponds to the spatial co-occurrence association dimension, and the causal event chain corresponds to the causal dependency transmission dimension. These three types of features comprehensively characterize the access behavior and association attributes of the entity data block from different levels. They can directly support the calculation of subsequent entity access feature vector indicators such as access impulse intensity, causal transmission efficiency, and association stability, as well as subsequent steps such as data association analysis and association calculation.

[0071] In one embodiment, the comprehensive association strength analysis of all entity data blocks is completed by combining entity multidimensional feature vectors, entity access feature vectors, and access popularity values. The data block association degree of all entity data blocks is calculated based on the comprehensive association strength analysis results, including the following steps: The data co-occurrence rate among all entity data blocks is calculated based on the data co-occurrence set. The Pearson correlation coefficient between all entity data blocks is calculated based on the access pulse intensity, and the synchronization degree between all entity data blocks is calculated by combining the data co-occurrence rate, Pearson correlation coefficient and access heat value. The causal propagation delay between all entity data blocks was calculated based on medical interaction data. The causal direction between all entity data blocks is determined based on the causal event chain, and the causal transmission value is constructed between all entity data blocks by combining the causal transmission delay, causal direction and access popularity value. The block feature spacing between all entity data blocks is calculated based on the entity access feature vector and the Euclidean distance formula. Taking any entity data block as the center, calculate the local feature density of the entity data block in the feature space based on the block feature spacing; The clustering balance penalty among all entity data blocks is calculated by combining the block feature spacing, local feature density, and access popularity value; The correlation degree between all entity data blocks is calculated by combining the synchronization degree, causal transmission value, and cluster balance penalty.

[0072] In this embodiment, the total number of time slices in the co-occurrence statistical window is first set. This parameter is determined based on a balance between statistical significance and timeliness, and can be set to a fixed 100 time slices. The data co-occurrence sets of all time slices within the statistical window are traversed, and the total number of times all entity data blocks are jointly accessed in the same time slice and the same transaction is accumulated, obtaining the data co-occurrence frequency of the corresponding data block pair. The data co-occurrence frequency is divided by the total number of time slices in the statistical window to obtain the normalized data co-occurrence rate, which ranges from 0 to 1. A higher value indicates more frequent business co-occurrence between the two entity data blocks. Next, the access pulse intensity time series of the two entity data blocks within the statistical window is extracted, and their Pearson correlation coefficient is calculated to quantify the similarity and synchronicity of their access burst patterns. The value ranges from -1 to 1, with a value closer to 1 indicating higher synchronicity of the burst patterns.

[0073] Next, the synchronization degree between all entity data blocks is calculated by combining the data co-occurrence rate, Pearson correlation coefficient, and access heat value. The access heat value is obtained by fusing access impulse intensity and causal transmission efficiency. The calculation method is that the access heat value equals the sum of access impulse intensity multiplied by 1 and causal transmission efficiency, representing the strength of the entity data block's role in the data linkage field. The magnitude of the synchronization degree is positively correlated with the co-occurrence frequency and access timing matching degree of the two sets of data; the smaller the value, the higher the degree of business binding between the two sets of medical data. Adding a constant 1 to the formula and superimposing it on the correlation coefficient can avoid the score inversion problem caused by negative correlation and ensure that the basic weight of data co-occurrence frequency is stable and effective. The formula for calculating the synchronization degree is as follows:

[0074] in, This indicates the degree of synchronization between entity data block i and entity data block j; , These represent the access popularity values ​​of entity data block i and entity data block j, respectively. This indicates the data co-occurrence rate between two entity data blocks; This represents the Pearson correlation coefficient between the pulse intensity sequences of two entity data blocks.

[0075] For each confirmed directed causal edge in the causal event chain, extract the time difference between all corresponding "read upstream data block → write downstream data block" operation pairs from the transaction operation records of medical interaction data; this is the single causal propagation delay. Calculate the arithmetic mean of all valid single propagation delays for the same directed causal edge to obtain the average causal propagation delay for that edge. For data block pairs without causal association, there is no need to calculate the causal propagation delay; subsequent causal propagation values ​​are directly set to 0. Next, traverse the causal event chain to confirm the upstream and downstream directions of all valid causal edges: if a directed causal edge exists pointing from entity data block i to entity data block j, calculate the corresponding causal propagation value; if no causal edge exists, the causal propagation value is 0. The causal propagation score increases with data access intensity and decreases with business transmission latency. A higher score indicates a closer linkage in the diagnosis and treatment process, i.e., a faster causal link. When a causal edge exists pointing from entity data block i to entity data block j, the formula for calculating its causal propagation value is as follows:

[0076] in, Let be the average causal propagation delay of the directed causal edge between entity data block i and entity data block j. This is a time-sensitive constant, which can be set to 1 second. This is the causal force strength coefficient, used to adjust the overall weight of the causal transmission value relative to the degree of synchronization. It is determined based on the contribution ratio of the causal relationship to the overall binding relationship, and can be set to a value of 0.5. , These represent the access popularity values ​​of entity data block i and entity data block j, respectively.

[0077] A standardized entity data block feature space is constructed using the three dimensions of the entity access feature vector—access impulse intensity, causal propagation efficiency, and correlation stability—as three-dimensional coordinate components. The block feature distance between two entity data blocks is the Euclidean distance of their corresponding feature vectors in three-dimensional space. Since the values ​​of the three dimensions have been normalized to a reasonable range, the calculated distance value can be directly used for subsequent analysis without additional scaling correction. The formula for calculating the block feature distance is as follows:

[0078] in, The block feature spacing between entity data block i and entity data block j; , and These represent the access impulse strength, causal propagation efficiency, and correlation stability of entity data block i, respectively. , and These represent the access impulse strength, causal propagation efficiency, and correlation stability of entity data block i, respectively. Here, "entity data block" refers to the entity data block itself.

[0079] Centered on any entity data block, the local feature density of the entity data block in the feature space is calculated based on the block feature spacing. The core function of this step is to measure the crowding level of the feature space region where the entity data block is located, providing a density basis for clustering balance penalties, preventing all entity data blocks from excessively clustering into a single, indistinguishable giant cluster, and ensuring that the associated topology has a reasonable structural resolution. First, a neighborhood radius parameter is set, which can be set to 1.0. To adapt to streaming computing scenarios with large-scale data blocks and reduce computational overhead, a grid approximation method is used to calculate the density: the three-dimensional feature space is divided into a uniform cubic grid with a side length equal to the neighborhood radius, and an entity data block count is maintained for each grid cell; when the feature vector of an entity data block is updated, causing a change in its grid, the count value of the corresponding grid is atomically increased or decreased. When calculating the local feature density of a data block, it is only necessary to accumulate the total number of entity data blocks in the grid where the entity data block is located and its 26 neighboring grids to obtain an approximate number of neighboring entity data blocks, resulting in low computational complexity. The higher the value of the local feature density, the denser the entity data blocks in the region where the entity data block is located.

[0080] Next, the clustering balance penalty among all entity data blocks is calculated by combining the block feature spacing, local feature density, and access popularity value. The core function of this step is that the clustering balance constraint acts as a cluster control parameter. When the amount of local data is excessive, it automatically reduces the inter-group correlation score to prevent a large amount of data from merging into a single super-large cluster, ensuring a uniform size of each data community. This parameter is positive and automatically decays as the feature spacing between two groups of data increases, only taking effect within a short distance range. In the formula, the smaller value of the local feature density of the two data blocks is taken as the density factor data block. The formula for calculating the clustering balance penalty is:

[0081] in, This represents the clustering balance penalty between entity data block i and entity data block j; This represents the equilibrium control coefficient, a global constant used to adjust the overall weight of cluster constraints relative to the positive linkage score. It serves only as a community equilibrium control item and not as the dominant association weight, with a reference value of 0.1. The constraint range parameter controls the effective range of cluster constraints. When the feature distance is much greater than this value, the constraint effect decays exponentially to a negligible level. It is determined based on the overall scale of the three-dimensional feature space, with a reference value of 0.5. The block feature spacing between entity data block i and entity data block j. , Let represent the local feature densities of entity data block i and entity data block j, respectively. This indicates taking the smaller value of two local feature densities, where the entity data block refers to the entity data block.

[0082] The data block correlation is the algebraic sum of synchronization linkage, causal transmission value, and cluster equilibrium penalty. Among them, the correlation score between two types of business is negative, while the cluster constraint, i.e., the cluster equilibrium penalty, is positive. When the correlation is negative and the larger the absolute value, it means that the comprehensive business binding relationship between the two entity data blocks is stronger, and it is easier to form a stable correlation edge. When the correlation is positive, it means that the cluster size control constraint is dominant, and it is difficult for the two to form a stable correlation structure.

[0083] In one embodiment, reference is made to Figure 3 The process involves dividing all entity data blocks into multiple data communities based on the entity association topology graph, and extracting community feature vectors for each data community based on entity access feature vectors. This includes the following steps: S301. Construct the Laplace matrix based on the entity association topology graph, and solve for the eigenvalues ​​and eigenvectors of the Laplace matrix; S302. Combine matrix eigenvalues ​​and matrix eigenvectors and use the eigenvalue gap method to construct an entity embedding matrix; S303. Determine the spatial mapping coordinates of the entity data block in the clustering feature space based on the row vectors of the entity embedding matrix; S304. Based on all spatial mapping coordinates and using a clustering algorithm, all entity data blocks are divided into data communities to obtain multiple data communities; S305. For any data community, calculate the community feature vector of the data community by combining the entity association topology graph and the entity access feature vector of all community entity data blocks in the data community. The community feature vector includes the community access popularity value, the topology convergence coefficient and the community access synchronization degree.

[0084] In this embodiment, the adjacency matrix and degree matrix are first extracted from the entity association topology graph: the elements of the adjacency matrix correspond to the weights of the associated edges between entity nodes, with a value of 0 if there is no edge connection; the degree matrix is ​​a diagonal matrix, and the elements on the diagonal are the sum of the weights of all associated edges of the corresponding entity node. Based on this, a normalized symmetric Laplace matrix is ​​constructed. This matrix can eliminate the interference of node degree differences on community partitioning, ensuring that data blocks with different access volumes have equal clustering weights, making it more suitable for scenarios with significant differences in medical data access volumes. The formula for calculating the normalized symmetric Laplace matrix is: ,in, This represents the normalized symmetric Laplace matrix, i.e., the Laplace matrix constructed in this scheme. This represents an identity matrix, with dimensions matching the total number of entity nodes. The degree matrix is ​​a diagonal matrix. After constructing the Laplacian matrix, its eigenvalues ​​are decomposed to obtain all matrix eigenvalues ​​sorted in ascending order, and the corresponding eigenvectors for each eigenvalue. Each eigenvector corresponds to a clustering feature dimension of the graph structure; the smaller the eigenvalue, the larger the structural scale of the modality, and the better it reflects the global community partitioning characteristics.

[0085] The core function of constructing an entity embedding matrix by combining matrix eigenvalues ​​and eigenvectors and utilizing the eigenvalue gap method is to automatically determine the optimal number of communities and select the core features that best reflect the community structure from all features to construct a low-dimensional embedding matrix, thus achieving the transformation from a high-dimensional graph structure to a low-dimensional Euclidean space. The implementation process consists of two stages: The first stage determines the optimal number of communities using the eigenvalue gap method. The difference between adjacent matrix eigenvalues ​​is calculated in ascending order, and the position with the largest difference is identified; the corresponding index is the optimal number of communities. A larger eigenvalue gap indicates a more significant structural boundary in the corresponding dimension, thus determining a community number that best matches the natural association structure of the data. To avoid excessively fine scheduling granularity due to an excessive number of communities, an upper limit is preset for the number of communities. This upper limit is determined based on the balance between the classification granularity and scheduling computational overhead of the medical cloud platform business scenario, with a reference value of 10. If the eigenvalue gap is not significant, the value that results in the highest clustering silhouette coefficient within the preset upper limit is selected as the final number of communities. The second stage involves constructing the entity embedding matrix. The eigenvectors corresponding to the k smallest non-zero eigenvalues ​​of the matrix are selected, and the k column vectors are concatenated side-by-side to obtain the initial feature matrix. To eliminate the differences in feature scales between different nodes, L2 normalization is performed on each row of the initial feature matrix, resulting in the entity embedding matrix. Row normalization ensures that the feature vectors of all entity nodes fall on the same scale of the unit hypersphere, guaranteeing the fairness of subsequent clustering. The dimension of the entity embedding matrix is ​​the total number of entity nodes multiplied by the number of communities. Each column corresponds to a core clustering feature dimension, and each row corresponds to a low-dimensional feature representation of an entity node.

[0086] Each entity data block is mapped to a specific coordinate point in a low-dimensional space, transforming the original non-Euclidean graph structure relationships into distance relationships in Euclidean space, providing directly computable coordinate input for subsequent clustering algorithms. The clustering feature space is a low-dimensional feature space constructed using the core clustering feature dimensions as coordinate axes. In this space, the closer the spatial distance between two entity data blocks, the higher their coupling degree and the stronger their business synergy in the association topology; the farther apart they are, the sparser the association. During implementation, each row of the entity embedding matrix corresponds to an entity data block, and the row vector of that row is the spatial mapping coordinate of the corresponding entity data block in the clustering feature space. The dimension of the coordinates is consistent with the number of communities, and the value of each dimension corresponds to the projection value of the data block on the corresponding clustering feature dimension. The k-means clustering algorithm is used to cluster all spatial mapping coordinates, and the number of clusters is equal to the number of communities determined in the previous steps. Since the spectral embedding process has already transformed the community relationships in the graph structure into a spatial clustered distribution, k-means clustering can efficiently and accurately complete the community division. All entity data blocks are categorized into corresponding data communities. Each data community contains a set of entity data blocks that are highly coordinated in terms of access behavior and causal logic, corresponding to datasets that are naturally linked in the business.

[0087] For any data community, a community feature vector is calculated by combining the entity association topology graph and the entity access feature vectors of all community entity data blocks within the data community. This community feature vector includes a community access heat value, a topology convergence coefficient, and a community access synchronization degree. The community access heat value measures the overall access intensity and causal driving capability of the data community, reflecting its overall energy intensity and serving as a core indicator for assessing the community's computing power and storage resource requirements. A higher community access heat value indicates stronger overall access explosiveness and a more significant driving effect on downstream businesses, corresponding to higher resource demand intensity. The topology convergence coefficient quantifies the tightness of causal coupling and structural convergence within the data community, comprehensively reflecting the collaborative tightness of data within the community and serving as a core basis for determining whether to adopt a collaborative storage layout and in-situ computing power scheduling. The community access synchronization degree measures the consistency of temporal offset encoding during peak access periods for all data blocks within the community, reflecting the concurrent alignment characteristics of community access and serving as a core basis for assessing the community's network bandwidth requirements and concurrent computing power reservation scale. The community access popularity value, topological convergence coefficient, and community access synchronization degree are structurally combined in a fixed order to form the community feature vector of the current data community. This vector is a three-dimensional numerical vector that fully characterizes the data community's data linkage behavior from three dimensions: overall energy intensity, internal coupling tightness, and access concurrency synchronization. It can be directly used as input to the subsequent artificial intelligence inference engine to support the calculation of multi-dimensional resource scheduling weights and the generation of resource scheduling strategies.

[0088] In one embodiment, calculating the community feature vector of a data community by combining the entity association topology graph and the entity access feature vectors of all community entity data blocks within the data community includes the following steps: The community access heat value of the data community is calculated by combining the access impulse intensity and causal propagation efficiency of all community entity data blocks within the data community. Extract the associated directed edges corresponding to all community entity data blocks from the entity association topology graph, and calculate the directed edge density of the data community based on the results of the associated directed edge extraction. The topological convergence coefficient of the data community is calculated by combining directed edge density, causal transmission efficiency, and correlation stability. The access timing bias of all community entity data blocks is taken as the unit circle vector, and the vector and mean of all unit circle vectors are calculated to obtain the average composite vector. Calculate the vector magnitude of the average composite vector, and calculate the community access synchronization degree of the data community based on the vector magnitude; The community access popularity value, topological convergence coefficient, and community access synchronization degree are integrated into the community feature vector of the data community.

[0089] In this embodiment, the core function of calculating the community access heat value is to quantify the overall access activity and causal driving capability of the entire data community, reflecting the overall energy level of the community and serving as a core indicator for measuring the intensity of community resource demand. First, the average access pulse intensity and average causal propagation efficiency of all community entity data blocks within the community are calculated. These two metrics respectively reflect the overall access surge and causal propagation capability of the community. Based on these, the community access heat value is calculated. This indicator comprehensively reflects the overall synchronous access activity level of the community; a higher value indicates stronger access intensity and a more significant driving effect on downstream businesses, corresponding to higher demands for computing power, storage, and bandwidth resources. The formula for calculating the community access heat value is:

[0090] in, This represents the community access popularity value of data community c; This represents the arithmetic mean of the access pulse strengths of all community entity data blocks within data community c; This represents the arithmetic mean of the causal propagation efficiency of all community entity data blocks within data community c. The formula uses a squaring operation to amplify the weight of high-access impulses, highlighting the differences in resource demand among high-burst communities; a constant of 1 is added to ensure that communities without causal propagation capabilities still possess basic correlation coefficients.

[0091] The system extracts all directed edges corresponding to community entity data blocks from the entity association topology graph and calculates the directed edge density of the data community based on the extraction results. The core function of this step is to quantify the density of causal connections within the data community, reflecting the coverage of causal links within the community and providing structural dimension parameters for the subsequent calculation of the topology convergence coefficient. During implementation, the entity association topology graph is first traversed, filtering out all associated edges where both ends belong to the current data community. The causal direction marker attached to each edge is then checked: if a causal direction exists from node i to node j, it is counted as one directed edge; if there is a causal relationship in both directions, it is counted as two directed edges; otherwise, it is not counted. The total number of directed edges within the community is then calculated. The directed edge density is the ratio of the actual number of directed edges within the community to the theoretical maximum number of directed edges, where the theoretical maximum is the total number of directed permutations of all different node pairs within the community.

[0092] The average causal propagation efficiency and average association stability of all community entity data blocks within a community are calculated. These two values ​​reflect propagation capability and association stability at the individual level, respectively. Their product is then calculated and multiplied by the directed edge density as a structural penalty factor to obtain the topological convergence coefficient. This design ensures that the topological convergence coefficient only reaches a high value when community members simultaneously possess high propagation capability and high association stability, and these capabilities are significantly applied within the community. This avoids misclassifying causally sparse communities as tightly coupled structures. A higher topological convergence coefficient indicates tighter causal coupling and stronger structural convergence within the community, making it more suitable for collaborative layout and in-situ scheduling strategies. To facilitate subsequent threshold determination and cross-community comparison, the topological convergence coefficients of all communities can be min-max normalized, mapping the values ​​to the range of 0 to 1.

[0093] Then, using a cyclic statistical method, the overall consistency of the access fluctuation phase of all data blocks within the community is calculated, providing an intermediate vector result for calculating the community access synchronization. Since the access timing offset is a periodic angle value ranging from 0 to 2π, the arithmetic mean cannot be directly used to calculate overall consistency. Therefore, a unit circle vector synthesis method is adopted: the access timing offset of each data block is mapped to a unit circle point vector on the unit circle, where the vector's angle equals its phase value, and its magnitude is 1. The mean values ​​of all unit circle point vectors in the horizontal and vertical axes are calculated separately, and then the two mean values ​​are combined into an average composite vector. The formula for calculating the average composite vector is:

[0094]

[0095] in, This represents the mean of the x-coordinate components of the average composite vector; This represents the mean of the ordinate components of the average composite vector; This represents the total number of community entity data blocks within data community c; This represents the access timing bias of the i-th community entity data block. and The resulting two-dimensional vector is the average resultant vector.

[0096] The magnitude of the average composite vector directly reflects the peak temporal overlap of all data within the community: if the peaks of all entity data blocks are completely synchronized, the magnitude of the average composite vector is 1; if the phases are completely random and uniformly distributed, the magnitude approaches 0. The community access synchronization degree is equal to the vector magnitude of the average composite vector, ranging from 0 to 1. The formula for calculating the community access synchronization degree is:

[0097] The community access popularity value, topological convergence coefficient, and community access synchronization degree are structurally combined in a fixed order to form a community feature vector for the corresponding data community. The three dimensions of this vector fully characterize the data linkage behavior of the data community from three aspects: overall energy intensity, internal coupling tightness, and access concurrency synchronization. It can be directly input into the artificial intelligence inference engine to generate a multi-dimensional resource scheduling strategy that matches the characteristics of the community.

[0098] In one embodiment, generating a resource scheduling strategy for the medical cloud platform by combining community feature vectors and the AI ​​inference engine pre-installed on the medical cloud platform includes the following steps: All community feature vectors are sequentially transmitted to the AI ​​inference engine pre-installed on the medical cloud platform, and the AI ​​inference engine outputs multi-dimensional resource scheduling weights. The AI ​​inference engine is built based on an online learning model, and the multi-dimensional resource scheduling weights include computing power allocation weights, storage level weights, network bandwidth weights, and prefetch depth weights. Multidimensional resource scheduling parameters are calculated by combining community feature vectors and multidimensional resource scheduling weights; Based on the pre-acquired full-platform resource parameters of the medical cloud platform, global resource optimization of multi-dimensional resource scheduling parameters is completed, and a resource scheduling strategy for the medical cloud platform is generated based on the multi-dimensional resource scheduling parameters after global resource optimization. The resource scheduling strategy includes computing power allocation parameters, storage level parameters, network bandwidth parameters, and prefetch depth parameters for all data communities.

[0099] In this embodiment, the AI ​​inference engine employs a three-layer feedforward neural network as its basic computing architecture, including an input layer, a hidden layer, and an output layer. First, a baseline sample set is constructed based on historical operational data from the medical cloud: three community features are categorized into quantile levels, and corresponding resource weights are matched with benchmark values, covering typical scenarios such as daily diagnosis and treatment, peak business periods, and archive queries, forming standardized baseline samples. Using the baseline sample set as training data, the network parameters are fitted offline using the least squares method to obtain initial weights and bias parameters. During the fitting process, L2 regularization constraints are added to improve generalization ability, with a reference regularization coefficient of 0.001. The pre-trained parameters and the manually configured rule baseline are stored together as a fallback baseline parameter for policy rollback in abnormal scenarios. Then, based on service quality feedback for each scheduling cycle, the model parameters are updated incrementally in small steps, allowing the scheduling weights to adaptively match the dynamic changes in business load. After each scheduling cycle, the service level agreement achievement status of each community is collected as a feedback signal, generating achievement rate indicators for the four dimensions of computing power, storage, network, and prefetching. The loss function is constructed with the goal of minimizing service quality deviation and avoiding excessive resource allocation. The stochastic gradient descent method is used to incrementally update the network parameters. An upper limit is set for the single-cycle parameter update step size to prevent resource jitter caused by excessive adjustment in a single cycle. The reference value for the upper limit of the step size is 0.1.

[0100] In practical applications, the community feature vectors of all data communities are sequentially transmitted to the AI ​​inference engine pre-installed on the medical cloud platform. The community feature vectors include three normalized features: community access popularity, topological convergence coefficient, and community access synchronization, representing the overall activity intensity, internal causal tightness, and access concurrency synchronization of the community, respectively. The AI ​​inference engine is built on an online learning model and employs a lightweight network structure, ensuring that the latency of a single inference is less than one-tenth of the scheduling cycle length, thus preventing inference delays from affecting the real-time performance of scheduling. The model uses the community feature vectors of the current cycle as input and incorporates the service level agreement (SLA) achievement status of each community in the previous cycle as feedback signals to incrementally update the multi-dimensional resource scheduling weights. During the update process, a step size upper limit constraint is set to ensure that the weight change amplitude of a single adjustment does not exceed a preset threshold, avoiding drastic fluctuations in resource allocation. The step size upper limit is determined based on the stability requirements of resource scheduling to prevent frequent and large adjustments from causing business fluctuations; a reference value can be set to a maximum adjustment amplitude of no more than 0.1 per cycle. To ensure the operational reliability of the scheduling system, the engine also pre-configures a set of rule baseline parameters based on the statistical median. When the service level agreement (SLA) achievement rate corresponding to the scheduling strategy output by the engine for three consecutive cycles falls below the preset threshold, it automatically reverts to the baseline weight parameters. The SLA achievement rate threshold is determined based on the minimum service quality requirements of medical services, with a reference value of 0.85. The engine can also periodically and asynchronously output feature importance ranking reports, indicating the proportion of the impact of the three community features on the weights of each resource, providing interpretability for decision-making. This process does not block the online scheduling process. Finally, the AI ​​inference engine outputs four-dimensional multi-dimensional resource scheduling weights for each data community, with all weight values ​​ranging from 0 to 1. These weights include computing power allocation weights, storage tier weights, network bandwidth weights, and prefetch depth weights. The computing power allocation weight corresponds to the degree of allocation bias in computing resources, the storage tier weight corresponds to the degree of priority bias in storage media, the network bandwidth weight corresponds to the degree of guarantee level bias in network resources, and the prefetch depth weight corresponds to the depth and priority bias in data prefetching.

[0101] For each data community, the corresponding community characteristics and resource scheduling weights are coupled and calculated from four dimensions: computing power, network, prefetching, and storage. These together constitute the multi-dimensional resource scheduling parameters for that community. The specific calculation methods are as follows: First, computing power dimension scheduling parameters are calculated. The computing power dimension scheduling parameter is the computing power load index, obtained by multiplying the community access popularity value of the data community by the corresponding computing power allocation weight. This comprehensively reflects the combined effect of the community's own activity intensity and computing power resource quota, and is the core basis for subsequent dynamic scaling strategies. Second, network bandwidth dimension scheduling parameters are calculated. The network bandwidth dimension scheduling parameter is the network affinity index, obtained by multiplying the topology convergence coefficient of the data community by the corresponding network bandwidth weight. This comprehensively reflects the combined effect of the tightness of causal dependencies within the community and the available network resources, and is the core basis for subsequent network topology affinity scheduling strategies. Third, prefetching depth dimension scheduling parameters are calculated. The scheduling parameter for the prefetch depth dimension is the prefetch trigger index, which is obtained by calculating the product of the community access synchronization degree of the data community and the corresponding prefetch depth weight. It comprehensively reflects the combined effect of community access synchronization and prefetch resource quota, and is the core triggering basis for subsequent time-dimensional prefetching and caching strategies. Fourth, the calculation of the storage tier dimension scheduling parameter. The scheduling parameter for the storage tier dimension is the storage comprehensive heat value, which comprehensively reflects the combined effect of multi-dimensional community activity characteristics and storage resource quota, and is the core judgment basis for subsequent hot and cold storage tiering strategies. Its calculation formula is as follows:

[0102] in, This represents the overall storage popularity value of data community c; For data community c, the community access popularity value, Let c be the topological convergence coefficient of the data community. For the storage level weight of data community c, , , The storage heat fusion coefficient is obtained by the artificial intelligence inference engine based on historical storage throughput. The sum of the three is 1, and the initial reference values ​​can be set to 0.4, 0.3, and 0.3 respectively.

[0103] The process of optimizing global resource scheduling parameters based on pre-acquired platform-wide resource parameters of the medical cloud platform, and generating a resource scheduling strategy for the medical cloud platform based on the optimized global resource scheduling parameters, includes the following steps: Based on the pre-acquired network topology information, the platform resources of the medical cloud platform are divided into domains to obtain multiple resource management domains; For any resource management domain, collect the resource load parameters of all resource nodes within the resource management domain, including processor utilization, memory bandwidth utilization, and storage queue load. The resource load parameters are mapped to the local resource load values ​​of resource nodes through a logarithmic compression function; Based on the community access synchronization degree, a set of synchronized communities is selected from all data communities, and several pairs of synchronized communities are matched from the synchronized community set by combining the community access synchronization degree and access time sequence bias. Based on the community feature vector, allocate resource offset extreme values ​​to all synchronous community pairs; For any data community, the cross-community causal edge set of the data community is selected based on the entity association topology graph, and the neighbor community set and corresponding community causal coupling degree of the data community are selected based on the cross-community causal edge set and entity access feature vector. The comprehensive scheduling cost function of the data community for all resource nodes is calculated by combining the local resource load value, the community causal coupling degree, and the resource offset extreme value. Based on the comprehensive scheduling cost function, the community resource nodes and the corresponding comprehensive scheduling cost value of the data community are output. Based on the comprehensive scheduling cost, the multi-dimensional resource scheduling parameters of the data community are globally optimized, and the resource scheduling strategy of the medical cloud platform is generated by combining the community resource nodes and the globally optimized multi-dimensional resource scheduling parameters.

[0104] Specifically, the network topology information includes the connections between physical servers, storage arrays, and network switches within the medical cloud platform, as well as the bandwidth capacity and transmission latency of each link. Based on this network topology information, and following the principles of minimum network hop count and physical proximity, the resource nodes of the medical cloud platform are divided into different resource management domains. Each resource management domain corresponds to a clearly defined physical area. Resource nodes within a domain are typically connected to the same top-of-rack switch or the same group of aggregation switches, ensuring low communication latency and sufficient bandwidth between them. After the division, the system obtains a set of resource management domains. Each resource management domain contains several compute nodes, storage nodes, and network devices, and has a unique resource management domain identifier; these are collectively referred to as resource nodes.

[0105] After the resource management domains are partitioned, periodic load information is collected for each domain. For any given resource management domain, a lightweight acquisition agent deployed within that domain sequentially accesses each resource node within the domain, reads the node's current operating status indicators, and extracts three types of core resource load parameters. The first type is processor utilization, which is the proportion of clock cycles during which all processor cores on the node are in a non-idle state at the moment of sampling, with a value ranging from 0 to 1. The second type is memory bandwidth utilization, which is the ratio of the actual amount of data transmitted by the node's memory bus during the sampling period to the theoretical maximum amount of data transmitted, also normalized to between 0 and 1. The third type is storage queue load rate, which is the ratio of the number of commands currently waiting to be processed in the input / output command queue of the node's storage device to the maximum queue depth, normalized to between 0 and 1. The acquisition agent summarizes these three types of parameters to form the raw load vector of the resource node.

[0106] To overcome the limitation of single-node load metrics exerting excessive influence on overall assessment near extreme values, logarithmic compression is employed to transform the original load parameters into local resource load values. The formula for calculating the local resource load value for any resource node within the resource management domain is as follows:

[0107] in, Let n be the processor utilization of resource node n. Let n be the memory bandwidth utilization of resource node n. Let n be the storage queue load rate of resource node n. All three values ​​have been normalized between 0 and 1. The compression intensity coefficient is typically set to 10. This coefficient is determined based on the following: when a node is at full load (the average of the three utilization rates is close to 1), the local resource load value is approximately 2.4, while when a node is idle, the local resource load value approaches 0. This dynamic range effectively distinguishes different load levels without causing the load value of a fully loaded node to be too high and overwhelm the influence of other nodes. For the overall load assessment of the resource management domain, the average local resource load value of all nodes within the domain is taken as the macro-load of the domain, i.e., the domain average load score. Simultaneously, the location information of the resource node with the highest load score is recorded for subsequent fine-grained scheduling. Finally, each resource management domain reports its domain average load score and a summary of key nodes to the global scheduler, forming a global resource load distribution composed of the load scores of each management domain.

[0108] Traverse all data communities and read the community access synchronization degree from the community feature vector of each data community. Data communities with a community access synchronization degree greater than or equal to a preset synchronization inertia threshold are marked as high-synchronization communities and added to the synchronization community set. The synchronization inertia threshold can be 0.6. All communities within the synchronization community set are paired, and the interference index of each pair is calculated to comprehensively assess the risk of resource peak overlap between the two groups. The calculation formula is as follows:

[0109] in, , These represent the community access synchronization degree of the two highly synchronized communities. The minimum data offset difference between the community access time sequence offsets corresponding to two highly synchronized communities, after periodic folding, falls between 0 and... Within the interval, an interference threshold of 0.4 can be set. Two highly synchronized communities with an interference threshold greater than or equal to the threshold are identified as synchronized community pairs, i.e., community combinations with resource conflict risks, and are included in the subsequent phase shift constraint and scheduling optimization scope.

[0110] Based on the community feature vector, resource offset extreme values ​​are assigned to all synchronization community pairs. Its core function is to set the maximum tolerable phase offset boundary for communities with conflict risks. This ensures that phase shift operations effectively smooth out peaks and valleys while strictly controlling the offset to not exceed the tolerance range of business logic and community timing structure, thus guaranteeing the bottom line of service quality for medical services. The resource offset extreme value is the maximum tolerable phase offset for a single community. It is dynamically calculated based on the community's own causal transmission characteristics and access pattern features, without relying on manually set fixed levels. Each community corresponds to an independent offset extreme value. The calculation formula is as follows:

[0111] in, For the resource offset extreme value of data community c; The average causal propagation delay of the community is extracted from the causal event chain within the community. It is the average propagation delay of all valid causal edges within the community. This item constitutes the upper limit of the phase shift business logic, ensuring that the offset will not exceed the inherent time scale of the business causal trigger, and avoiding disrupting the temporal logic of the diagnosis and treatment business. The community access synchronization degree of data community c is derived from the community feature vector. This is the system time base constant, with a reference value of 2000 milliseconds. The value is determined based on the typical batch processing scheduling window size of throughput-level services in the medical cloud platform. This is a regularization constant, with a reference value of 0.01, used to prevent division by zero errors when the community access synchronization degree is 0.

[0112] For any given data community, traverse all confirmed directed edges in the entity association topology graph. Filter all directed edges where one end belongs to the current data community and the other end belongs to another data community, forming a cross-community causal edge set for the current data community. Specifically, this set includes two types of edges: one type is a forward directed edge where the starting entity data block is located in the current data community and the ending entity data block is located in another data community; the other type is a reverse directed edge where the starting entity data block is located in another data community and the ending entity data block is located in the current data community.

[0113] For any causal edge within the cross-community causal edge set, the causal transmission efficiency value attached to that edge is read. The arithmetic mean of the causal transmission efficiencies of all causal edges in the set is taken to obtain the cross-community average causal transmission efficiency between the two data communities. If the cross-community causal edge set is empty, the cross-community average causal transmission efficiency is zero. Next, the average association stability of each of the two data communities is read from their respective community feature vectors, which is the arithmetic mean of the association stability of all entity data blocks within the data community. Then, the smaller of the two average association stability values ​​is calculated, and its product with the cross-community average causal transmission efficiency is calculated to obtain the community causal coupling degree between the two data communities. If the community causal coupling degree between a certain data community and the current data community is greater than a preset coupling degree threshold, such as 0.7, then the data community is included in the neighbor community set of the current data community.

[0114] The formula takes the smaller value of the two constraints as the final extreme value of the resource offset, forming a double guarantee: neither excessive offset will disrupt the causal logic of the business, nor will excessive offset disrupt the temporal structure of access within the community. When allocating offsets, a priority rule is followed: communities with shorter average causal propagation latency and higher community access synchronization are given priority to maintain their original phase and bear smaller offsets; communities with longer latency and lower synchronization bear more offsets.

[0115] The comprehensive scheduling cost function for each data community across all resource nodes is calculated, and the formula for the comprehensive scheduling cost function is as follows:

[0116] in, This represents the local resource load value of resource node n, and the data item containing it is the load cost item. The physical distance between nodes deployed in neighboring community k within the neighboring community set where resource node n and data community c have a causal coupling relationship is measured by network hop count or round-trip latency, and the data item containing it is the causal affinity cost component. Let c be the set of neighboring communities. The indicator function contains a hardware adaptation penalty component, determined based on the hardware supply capacity of the resource node. Specifically, it is a weighted sum of two indicators: community access popularity and community access synchronization, then mapped to the [0,1] interval using min-max scaling. This is a dimensionless comprehensive load indicator; a higher value indicates greater resource consumption pressure from batch concurrent retrieval and synchronous access by the data community. The weights can all be 0.5. Then, the hardware supply index of the resource node is collected, including hardware acceleration availability: if the resource node is equipped with a graphics processor or field-programmable gate array and its current available computing power share is not exhausted, the value is 1; otherwise, the value is 0. This information comes from a comparison between the platform configuration database and real-time resource monitoring data. Dedicated network slice availability: if the switching device where the resource node is located supports software-defined network slicing and there are still remaining available slice resources, the value is 1; otherwise, the value is 0. This information comes from the real-time slice resource status reported by the network controller. Storage media performance level: the ratio of the random read / write throughput of the resource node's main storage device to the baseline throughput of the all-flash array, normalized to between 0 and 1. This information comes from storage device specifications and real-time performance sampling. Adding these values ​​together and dividing by three yields the hardware supply index. Combining the resource demand index and the hardware supply index, an indicator function is calculated. The formula for the indicator function is: ,in, This is a resource demand index. This is a hardware supply index. The extreme value of the resource offset for data community c; the extreme value of the resource offset for data community c.

[0117] For each data community, a cost matrix is ​​formed by calculating the comprehensive scheduling cost function point-by-point across all its resource nodes. Each row in the cost matrix corresponds to a data community, and each column corresponds to a resource node. The value of each matrix element is the comprehensive scheduling cost of deploying the corresponding data community to that candidate resource node. For each data community, the resource node with the lowest comprehensive scheduling cost is selected as its community resource node. After assigning a resource node to a data community, the remaining available resources of the corresponding resource node are deducted synchronously, covering four core resources: computing power, memory bandwidth, storage throughput, and network bandwidth. If the remaining available resources of a resource node are insufficient to meet the resource needs of the current data community, the comprehensive scheduling cost of that resource node is temporarily set to infinity, thus removing it from the competition for node selection in this round. When multiple data communities compete for the same resource node and the node's resources are insufficient to meet the demand simultaneously, they are sorted according to the urgency of their resource needs from high to low, with priority given to data communities with higher urgency. The urgency of resource needs is calculated by multiplying the community access popularity value by the community access synchronization value. Data communities with lower urgency reselect the next best resource node. This sorting rule is consistent with the definition of the comprehensive urgency value in the global resource arbitration process, ensuring the deployment priority of high-priority services.

[0118] Next, the multi-dimensional resource scheduling parameters of the data community are globally optimized based on the minimum comprehensive scheduling cost. For example, if the comprehensive scheduling cost is less than or equal to the minimum value of the preset interval, such as 0.5, it indicates that the community's resource nodes have low load, close causal neighbor deployments, and good hardware matching, resulting in excellent overall deployment conditions. In this case, the multi-dimensional resource scheduling parameters can be uniformly reduced by a certain margin, such as 10%, to compress resource redundancy while ensuring normal business operation. If the comprehensive scheduling cost is within the preset interval, such as (0.5, 1.0), it indicates that the node deployment conditions are at a normal level, and the multi-dimensional resource scheduling parameters remain unchanged. If the comprehensive scheduling cost is greater than or equal to the maximum value of the preset interval, it indicates that the node load is high, the causal neighbor distance is far, or the hardware matching is insufficient, resulting in poor overall deployment conditions. It is necessary to improve resource protection to offset performance losses, and the multi-dimensional resource scheduling parameters need to be increased by a certain margin, such as 20%, to strengthen resource protection. After completing the parameter adjustment, the community resource nodes are combined with the adjusted four types of resource scheduling parameters to form the resource scheduling strategy for the data community.

[0119] This application also provides an artificial intelligence-based medical cloud platform resource scheduling system, including: The memory is configured to store instructions; and The processor is configured to retrieve the instructions from the memory and, when executing the instructions, to implement any of the above-described AI-based medical cloud platform resource scheduling methods.

[0120] This application also provides a machine-readable storage medium storing instructions for causing a machine to execute any of the above-described AI-based medical cloud platform resource scheduling methods.

[0121] The processor can be a central processing unit (CPU). Of course, depending on the actual use, it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc., and this application does not limit it.

[0122] The memory can be an internal storage unit of a computer device, such as a hard disk or RAM, or an external storage device, such as a plug-in hard disk, smart memory card (SMC), secure digital card (SD), or flash memory card (FC) provided on the computer device. Furthermore, the memory can be a combination of internal storage units and external storage devices of a computer device. The memory is used to store computer programs and other programs and data required by the computer device. The memory can also be used to temporarily store data that has been output or will be output. This application does not limit this.

[0123] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0124] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0125] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0126] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0127] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0128] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, like read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0129] Computer-readable media include both permanent and non-permanent, removable and non-removable media that can store information by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0130] It should also be noted that 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. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0131] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A resource scheduling method for a medical cloud platform based on artificial intelligence, characterized in that, The method includes the following steps: Regional medical information of the target area is collected through a medical cloud platform. This regional medical information includes medical entity data and medical interaction data. The medical entity data is split into multiple entity data blocks by standardizing the granularity of the data. Based on the medical interaction data, the entity access feature vectors of all entity data blocks are extracted. The data association analysis of all entity data blocks is completed through the entity access feature vectors, and an entity association topology graph is constructed based on the data association analysis results. Based on the entity association topology diagram, all entity data blocks are divided into multiple data communities, and community feature vectors of all data communities are extracted based on entity access feature vectors. By combining community feature vectors and the AI ​​inference engine pre-installed on the medical cloud platform, a resource scheduling strategy is generated for the medical cloud platform.

2. The method according to claim 1, characterized in that, The process of extracting entity access feature vectors from all entity data blocks based on medical interaction data, performing data association analysis on all entity data blocks using these entity access feature vectors, and constructing an entity association topology graph based on the data association analysis results includes the following steps: For any entity data block, the entity multidimensional feature vector of the entity data block is extracted based on the medical interaction data, and the entity access feature vector of the entity data block is constructed based on the entity multidimensional feature vector. The entity access feature vector includes access impulse intensity, causal transmission efficiency and correlation stability, and the entity multidimensional feature vector includes access density sequence, data co-occurrence set and causal event chain. The access heat value of the entity data block is calculated by combining the access pulse intensity and causal transmission efficiency; The access time bias is extracted from the access density sequence, and the access time bias and access popularity value are used as the block access features of the entity data block. By combining entity multidimensional feature vectors, entity access feature vectors, and access popularity values, a comprehensive association strength analysis of all entity data blocks is completed, and the data block association degree of all entity data blocks is calculated based on the results of the comprehensive association strength analysis. Calculate the data offset difference between all entity data blocks based on the access timing offset; Treat all entity data blocks as entity nodes, combine data bias difference, data block correlation degree and data co-occurrence set to complete the access correlation strength verification between all entity nodes, and construct the associated directed edges between all entity nodes that pass the access correlation strength verification according to the causal event chain. Based on the access association strength verification results, association edge weights are assigned to all directed edges to obtain the entity association topology graph.

3. The method according to claim 2, characterized in that, The process of extracting entity multidimensional feature vectors from entity data blocks based on medical interaction data, and constructing entity access feature vectors from entity data blocks based on entity multidimensional feature vectors, includes the following steps: Based on the medical interaction data, the causal chain verification between the entity data block and all other entity data blocks is completed, and the entity multidimensional feature vector is constructed based on the causal chain verification results. The smoothed access energy and baseline access activity of the entity data block are calculated cumulatively based on the access density sequence. The access pulse intensity of the entity data block is calculated by combining the smooth access energy and the baseline access activity; The causal propagation delay between the entity data block and all entity data blocks in the causal event chain is calculated based on the medical interaction data, and the causal propagation efficiency of the entity data block is calculated based on the causal propagation delay. Calculate the Jaccard similarity among all adjacent data co-occurrence sets within a fixed time slice, and calculate the association stability of entity data blocks based on the Jaccard similarity; The entity access feature vector of the entity data block is constructed by combining access impulse intensity, causal transmission efficiency and correlation stability.

4. The method according to claim 3, characterized in that, The step of verifying the causal chain between an entity data block and all other entity data blocks based on medical interaction data, and constructing a multidimensional feature vector of the entity based on the causal chain verification results, includes the following steps: Traverse the medical interaction data, count the number of times the entity data block is accessed within a preset fixed time slice, and construct an access density sequence based on all access counts; The time-delay cross-correlation function between the entity data block and all other entity data blocks is calculated based on the access density sequence, and the initial causal set of the entity data block is constructed based on the time-delay cross-correlation function. Based on medical interaction data, the frequency of data co-occurrence between entity data blocks within a fixed time frame and all other entity data blocks is calculated, and a data co-occurrence set of entity data blocks is constructed based on the data co-occurrence frequency. The intersection of the initial causal set and the data co-occurrence set is used to obtain the causal candidate set. Based on medical interaction data, the causal chain verification between the entity data block and all candidate data blocks in the causal candidate set is completed, and the causal event chain of the entity data block is constructed based on the causal chain verification results. By integrating access density sequences, data co-occurrence sets, and causal event chains, a multidimensional feature vector of entity data blocks is constructed.

5. The method according to claim 2, characterized in that, The process of combining entity multidimensional feature vectors, entity access feature vectors, and access popularity values ​​to complete the comprehensive association strength analysis of all entity data blocks, and calculating the data block association degree of all entity data blocks based on the comprehensive association strength analysis results, includes the following steps: The data co-occurrence rate among all entity data blocks is calculated based on the data co-occurrence set. The Pearson correlation coefficient between all entity data blocks is calculated based on the access pulse intensity, and the synchronization degree between all entity data blocks is calculated by combining the data co-occurrence rate, Pearson correlation coefficient and access heat value. The causal propagation delay between all entity data blocks was calculated based on medical interaction data. The causal direction between all entity data blocks is determined based on the causal event chain, and the causal transmission value is constructed between all entity data blocks by combining the causal transmission delay, causal direction and access popularity value. The block feature spacing between all entity data blocks is calculated based on the entity access feature vector and the Euclidean distance formula. Taking any entity data block as the center, calculate the local feature density of the entity data block in the feature space based on the block feature spacing; The clustering balance penalty among all entity data blocks is calculated by combining the block feature spacing, local feature density, and access popularity value; The correlation degree between all entity data blocks is calculated by combining the synchronization degree, causal transmission value, and cluster balance penalty.

6. The method according to claim 2, characterized in that, The step of dividing all entity data blocks into multiple data communities based on the entity association topology graph and extracting the community feature vectors of all data communities based on the entity access feature vectors includes the following steps: The Laplace matrix is ​​constructed based on the entity relationship topology graph, and the eigenvalues ​​and eigenvectors of the Laplace matrix are solved. An entity embedding matrix is ​​constructed by combining matrix eigenvalues ​​and eigenvectors and using the eigenvalue gap method. The spatial mapping coordinates of entity data blocks in the clustering feature space are determined based on the row vectors of the entity embedding matrix. Based on all spatial mapping coordinates and using clustering algorithms, all entity data blocks are divided into data communities to obtain multiple data communities; For any data community, the community feature vector is calculated by combining the entity association topology graph and the entity access feature vectors of all community entity data blocks within the data community. The community feature vector includes the community access popularity value, the topology convergence coefficient, and the community access synchronization degree.

7. The method according to claim 6, characterized in that, The process of calculating the community feature vector of a data community by combining the entity association topology graph and the entity access feature vectors of all community entity data blocks within the data community includes the following steps: The community access heat value of the data community is calculated by combining the access impulse intensity and causal propagation efficiency of all community entity data blocks within the data community. Extract the associated directed edges corresponding to all community entity data blocks from the entity association topology graph, and calculate the directed edge density of the data community based on the results of the associated directed edge extraction. The topological convergence coefficient of the data community is calculated by combining directed edge density, causal transmission efficiency, and correlation stability. The access timing bias of all community entity data blocks is taken as the unit point vector, and the vector and mean of all unit point vectors are calculated to obtain the average composite vector. Calculate the vector magnitude of the average composite vector, and calculate the community access synchronization degree of the data community based on the vector magnitude; The community access popularity value, topological convergence coefficient, and community access synchronization degree are integrated into the community feature vector of the data community.

8. The method according to claim 1, characterized in that, The process of generating a resource scheduling strategy for the medical cloud platform by combining community feature vectors and the AI ​​inference engine pre-installed on the medical cloud platform includes the following steps: All community feature vectors are sequentially transmitted to the AI ​​inference engine pre-installed on the medical cloud platform, and the AI ​​inference engine outputs multi-dimensional resource scheduling weights. The AI ​​inference engine is built based on an online learning model, and the multi-dimensional resource scheduling weights include computing power allocation weights, storage level weights, network bandwidth weights, and prefetch depth weights. Multidimensional resource scheduling parameters are calculated by combining community feature vectors and multidimensional resource scheduling weights; Based on the pre-acquired full-platform resource parameters of the medical cloud platform, global resource optimization of multi-dimensional resource scheduling parameters is completed, and a resource scheduling strategy for the medical cloud platform is generated based on the multi-dimensional resource scheduling parameters after global resource optimization. The resource scheduling strategy includes computing power allocation parameters, storage level parameters, network bandwidth parameters, and prefetch depth parameters for all data communities.

9. A resource scheduling system for a medical cloud platform based on artificial intelligence, characterized in that, include: The memory is configured to store instructions; as well as A processor is configured to retrieve the instructions from the memory and, when executing the instructions, to implement the AI-based medical cloud platform resource scheduling method according to any one of claims 1 to 8.

10. A machine-readable storage medium, characterized in that, The machine-readable storage medium stores instructions for causing the machine to execute the AI-based medical cloud platform resource scheduling method according to any one of claims 1 to 8.

Images