A Medical Examination Resource Supply and Demand Balance Prediction Method Based on Graph Neural Networks

Abstract

This invention relates to the field of distributed deep learning technology, specifically to a method for predicting the supply and demand balance of medical examination resources based on graph neural networks. The method includes: constructing a heterogeneous graph network containing relational information on distributed local medical nodes; aggregating features using a graph convolutional neural network to generate a topological gradient reflecting local resource demand; encrypting and encoding the topological gradient using a neural encryption protocol trained through adversarial co-evolutionary learning to generate encoded gradient data; uploading the data to a server; the server employing a reputation-robust aggregation strategy to remove malicious interference data and then performing multi-gradient fusion to generate unified global parameters; the server distributing the parameters to each local node; and each node, using a differential privacy mechanism, adding noise operators to the global resource prediction model for localized fine-tuning. This improves the robustness and accuracy of the global prediction model.

Description

Technical Field

[0001] This invention relates to the field of distributed deep learning technology, specifically to a method for predicting the supply and demand balance of medical examination resources based on graph neural networks. Background Technology

[0002] With the evolution of geometric deep learning, graph neural networks have become a core tool for processing non-Euclidean spatial data structures. In distributed computing environments, utilizing neural networks for complex spatiotemporal correlation modeling and collaborative parameter updates is a key research direction for achieving large-scale network topology evolution prediction.

[0003] Current neural network models still face significant architectural and learning algorithm bottlenecks when processing large-scale distributed heterogeneous graph data: First, traditional convolutional neural networks struggle to explicitly handle complex graph networks with heterogeneous properties. They are prone to oversmoothing during iterative aggregation, which makes it difficult to accurately extract the characteristics of local nodes and the evolution trend of the global network.

[0004] Secondly, existing federated learning frameworks lack efficient intrinsic encryption mechanisms during neural network parameter synchronization, making model gradients vulnerable to reverse derivation attacks. Furthermore, traditional homomorphic encryption schemes suffer from excessive computational overhead in large-scale neural network applications. Standard neural network aggregation algorithms typically assume that the data distribution of all participating nodes is consistent and trustworthy. When faced with non-independent and identically distributed data or malicious gradient interference, the global resource prediction model parameters are prone to getting trapped in local optima or diverging, lacking effective reputation assessment and weight allocation mechanisms.

[0005] Therefore, this invention aims to solve the technical problems of inaccurate global network parameter training and low robustness in distributed environments, caused by insufficient heterogeneous topology modeling capabilities, high risk of privacy leakage during gradient transmission, and poor stability of distributed aggregation algorithms under heterogeneous data interference.

[0006] To address this, a method for predicting the supply and demand balance of medical examination resources based on graph neural networks is proposed. Summary of the Invention

[0007] The purpose of this invention is to provide a method for predicting the supply and demand balance of medical examination resources based on graph neural networks. By constructing a heterogeneous graph neural network system and a robust federated learning architecture, it achieves privacy computing and collaborative evolution of distributed topological features.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A method for predicting the supply and demand balance of medical examination resources based on graph neural networks includes: Heterogeneous graph networks containing medical institution attributes and patient association information are constructed on multiple distributed local medical nodes. Feature aggregation is performed through graph convolutional neural networks to generate topological gradients that reflect local resource needs. The topological gradients are encrypted and encoded using a neural encryption protocol pre-trained through adversarial co-evolutionary learning to generate encoded gradient data, ensuring that the original medical data features are in an irreversible privacy-preserving state during transmission. Each local medical node uploads the encoded gradient data to the central server; the central server adopts a robust aggregation strategy based on reputation value to quantitatively evaluate the quality and historical contribution of the data uploaded by each node, removes potential malicious interference data, and fuses the gradients from multiple sources to generate unified global resource prediction model parameters. The central server distributes the global resource prediction model parameters to each local medical node; each node, in conjunction with the differential privacy injection mechanism, adds a noise operator to the global resource prediction model for localized fine-tuning.

[0009] Preferably, constructing the heterogeneous graph network includes: Each patient identifier and / or single visit record within the local medical node is defined as a test feature node, and structured features and unstructured clinical descriptions from electronic health records are extracted as the initial attribute vectors of the test feature nodes. The clinical phenotype similarity between different test feature nodes is calculated using the Pearson correlation coefficient. When the clinical phenotype similarity exceeds a preset threshold, an association edge is established between the corresponding two test feature nodes. A medical knowledge graph is introduced, defining medical examination items, drugs, symptoms, and diseases as entity nodes, and semantic association edges between the entity nodes and the test feature nodes are established according to the clinical diagnosis and treatment path to form the heterogeneous graph network.

[0010] Preferably, generating the topological gradient includes: By using the linear transformation layer of the graph convolutional neural network, the initial attribute vectors of the test feature nodes and the attribute features of the entity nodes are mapped to a unified low-dimensional vector space, generating hidden layer representations of the test feature nodes and the entity nodes. An attention mechanism is used to calculate the contribution weights between adjacent nodes in the heterogeneous graph network. Based on the associated edges and semantic associated edges in the heterogeneous graph network, the hidden layer representations of neighboring nodes are iteratively aggregated to update the features of the current node, identifying the characteristics of local nodes and the evolution trend of the global network. The updated node features are input into the prediction loss function, and the partial derivatives of each weight in the graph convolutional neural network are calculated using the backpropagation algorithm to generate the topological gradient.

[0011] Preferably, generating the gradient data includes: The neural encryption protocol, constructed through adversarial training between a sender encryption network and a receiver decryption network, is loaded to achieve nonlinear mapping of gradient features within an asymmetric encryption framework. The encryption network in the neural encryption protocol is used as a mapping operator to map the numerical matrix of the topological gradient to a high-dimensional chaotic space, eliminating the linear correspondence between the topological gradient and the original medical data. Bit-level perturbation and permutation are then applied to the mapped data to generate the encoded gradient data.

[0012] Preferably, the robust aggregation strategy based on reputation value includes: The central server calculates a dynamic reputation score for each local medical node based on its historical prediction accuracy improvement contribution, gradient consistency index, and participation frequency using a reputation assessment algorithm. It compares the distribution differences among the received encoded gradient data, identifying nodes with reputation scores below a preset first safety threshold and gradient distribution deviations exceeding a preset deviation threshold as malicious interference data sources and removing them. The server normalizes the reputation scores of the remaining encoded gradient data nodes, calculating normalized weight coefficients for each node, ensuring a positive correlation between the normalized weight coefficients and the reputation scores. Finally, it multiplies each encoded gradient data point by its corresponding normalized weight coefficient and sums the results linearly to calculate the global gradient increment, thereby updating and generating unified global resource prediction model parameters.

[0013] Preferably, the differential privacy injection mechanism includes: Based on the preset privacy budget and the sensitivity of the global resource prediction model parameters, the scale parameter of the Gaussian noise to be injected is calculated; a random noise vector conforming to the scale parameter is generated using a random number generator, and the random noise vector is superimposed on the gradient update amount in the local fine-tuning process to mask the contributions of the subject feature nodes and local medical nodes; after injecting the random noise vector, iterative optimization is continued using local data, so that the fine-tuned local prediction model can predict the local medical examination resource demand while satisfying differential privacy constraints.

[0014] Preferably, it also includes a dynamic resource balancing and scheduling step: Based on the prediction results output by the local prediction model, combined with the real-time queuing data and maintenance cycle of local medical examination equipment, the expected load rate of each examination department in the future preset time period is calculated; when the expected load rate exceeds the preset second safety threshold, a resource overflow sensing signal is triggered, and the idle capacity of the associated surrounding medical institutions is identified according to the functional flow graph in the heterogeneous graph network; scheduling suggestions are automatically generated based on the resource overflow sensing signal, and the referral guidance instructions for non-emergency patients' examination periods are adjusted to achieve a global balance of medical examination resources in the spatial and temporal dimensions.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention significantly improves the ability to extract medical resource association features in non-Euclidean spaces by constructing a heterogeneous graph network containing subject feature nodes and medical entity nodes. This scheme utilizes the message passing and attention mechanisms of graph convolutional neural networks to accurately identify hidden high-order logical connections in clinical pathways. By multimodal fusion of patient clinical phenotypic similarity and semantic associations with the medical knowledge graph, the model can achieve cross-dimensional data completion by observing the evolutionary trends of topological neighbor nodes when local data is missing.

[0016] 2. This invention constructs a privacy and security barrier covering the entire lifecycle of data transmission and model training through deep coupling of a neural encryption protocol and a differential privacy injection mechanism. Utilizing a neural encryption protocol generated by adversarial co-evolutionary learning, it achieves nonlinear gradient mapping and bit-level perturbation within an asymmetric encryption framework. This significantly reduces the risk of the central server inferring the original sensitive features of the local mechanism based on encoded gradient data, while ensuring model training accuracy. Compared to traditional homomorphic encryption schemes, the neural encryption mechanism employed in this invention solves the bottleneck of computational redundancy in large-scale neural networks in distributed scenarios, while ensuring statistical accuracy and defense against gradient back-inference attacks.

[0017] 3. This invention introduces a robust aggregation strategy based on reputation values, solving the problems of inconsistent data quality and malicious node interference faced by federated learning in heterogeneous medical consortia. By dynamically scoring the predicted contribution and gradient consistency of local medical nodes using reputation scoring, the central server can automatically identify and eliminate abnormal data sources with excessive deviations, ensuring the robustness and convergence stability of the global resource prediction model parameters. Based on this, combined with dynamic resource balancing and scheduling steps, this invention achieves closed-loop optimization from "passive response" to "active perception." By using the predicted load rate to trigger resource overflow signals and achieving precise cross-institutional diversion based on the functional flow graph, patient waiting times are significantly shortened. Attached Figure Description

[0018] Figure 1This is a flowchart of the medical examination resource supply and demand balance prediction method based on graph neural networks proposed in this invention. Figure 2 This is a flowchart of the medical examination resource supply and demand balance prediction method based on graph neural networks proposed in this invention. Figure 3 This is a flowchart of the method for constructing heterogeneous graph networks according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1

[0021] Please see Figures 1 to 3 This invention provides a method for predicting the supply and demand balance of medical examination resources based on graph neural networks. The technical solution is as follows: A method for predicting the supply and demand balance of medical examination resources based on graph neural networks, such as Figure 1 - Figure 2 As shown, it includes: Heterogeneous graph networks containing medical institution attributes and patient association information are constructed on multiple distributed local medical nodes. Feature aggregation is performed through graph convolutional neural networks to generate topological gradients that reflect local resource needs. The topological gradients are encrypted and encoded using a neural encryption protocol pre-trained through adversarial co-evolutionary learning to generate encoded gradient data, ensuring that the original medical data features are in an irreversible privacy-preserving state during transmission. Each local medical node uploads the encoded gradient data to the central server; the central server adopts a robust aggregation strategy based on reputation value to quantitatively evaluate the quality and historical contribution of the data uploaded by each node, removes potential malicious interference data, and fuses the gradients from multiple sources to generate unified global resource prediction model parameters. The central server distributes the global resource prediction model parameters to each local medical node; each node, in conjunction with the differential privacy injection mechanism, adds a noise operator to the global resource prediction model for localized fine-tuning.

[0022] Furthermore, such as Figure 3 As shown, constructing the heterogeneous graph network includes: Each patient identifier and / or single visit record within the local medical node is defined as a test feature node, and structured features and unstructured clinical descriptions from electronic health records are extracted as the initial attribute vectors of the test feature nodes. The clinical phenotype similarity between different test feature nodes is calculated using the Pearson correlation coefficient. When the clinical phenotype similarity exceeds a preset threshold, an association edge is established between the corresponding two test feature nodes. A medical knowledge graph is introduced, defining medical examination items, drugs, symptoms, and diseases as entity nodes, and semantic association edges between the entity nodes and the test feature nodes are established according to the clinical diagnosis and treatment path to form the heterogeneous graph network.

[0023] When constructing the initial attribute vector, a pre-trained medical language model is used to extract semantics from the unstructured clinical description, generating a high-dimensional semantic embedding. A linear mapping layer is then used to project the high-dimensional semantic embedding and the structured features onto a hidden feature space of the same dimension to eliminate dimensional differences between different modalities. When establishing the semantic association edges, based on the entity relationship triples in the medical knowledge graph, a distance vector mapping algorithm is used to calculate the Euclidean distance between the entity node and the subject feature node in the hidden feature space. The connection strength of the semantic association edges is determined based on the mapping relationship between the Euclidean distance and the preset association weights, thereby achieving structured association between cross-modal nodes in the same topological network. The preset threshold is set based on the similarity probability distribution, dynamically filtering weak association noise by balancing network connectivity and computational load. In this embodiment, it is set to 0.75.

[0024] The heterogeneous graph network employs a time-slicing mechanism for multi-temporal topology modeling, dividing the associated edges and semantically associated edges into multiple consecutive temporal topological snapshots according to a preset time span; in this embodiment, the preset time span is 24 hours. When the graph convolutional neural network performs feature aggregation, an adaptive graph learning layer is introduced. Based on the changes in the hidden layer representation of each tested feature node under different time snapshots, the weight coefficients of the associated edges are dynamically adjusted, and a gated recurrent unit is used to perform temporal transfer of topological features between adjacent time snapshots. This enables the heterogeneous graph network to capture the non-stationary cascading effects of medical resource supply and demand changes over time.

[0025] The temporal topology snapshot uses the appointment or actual execution time of the target medical examination as the time axis, dividing the subject feature nodes and their associated edges within the same preset time span into the same temporal snapshot. The adaptive graph learning layer adjusts the weight coefficients of the corresponding associated edges using a preset monotonic function based on the difference in the hidden layer representation of the same subject feature node in adjacent temporal snapshots, so that edges with greater differences in representation receive higher weights during periods of significant load variation.

[0026] The structured features include at least one patient information related to medical examination resource needs, such as patient age, gender, department visited, key examination item identifiers, and corresponding quantitative results. The unstructured clinical description, after being encoded by a pre-trained medical language model, is output as a fixed-length vector representing the semantics of the symptoms and diagnosis described during the visit. In this embodiment, the structured features are arranged in a preset order to form a first vector, and the fixed-length semantic vector is used as a second vector. The first and second vectors are concatenated and mapped to a hidden feature space of the same dimension through a linear mapping layer to obtain the initial attribute vector of the subject feature node.

[0027] This approach significantly enhances the representation accuracy of neural networks for complex medical scenarios by introducing multimodal alignment and dynamic topology evolution mechanisms. It addresses the issue of dimensional silos between unstructured clinical descriptions and structured data by utilizing pre-trained models and mapping layers, achieving deep fusion of cross-modal features in a unified hidden space. Through time-slicing mechanisms and adaptive graph learning layers, the model overcomes the limitations of static graph structures, enabling real-time capture of dynamic relationships between subject feature nodes over time.

[0028] Furthermore, generating the topological gradient includes: By using the linear transformation layer of the graph convolutional neural network, the initial attribute vectors of the test feature nodes and the attribute features of the entity nodes are mapped to a unified low-dimensional vector space, generating hidden layer representations of the test feature nodes and the entity nodes. An attention mechanism is used to calculate the contribution weights between adjacent nodes in the heterogeneous graph network. Based on the associated edges and semantic associated edges in the heterogeneous graph network, the hidden layer representations of neighboring nodes are iteratively aggregated to update the features of the current node, identifying the characteristics of local nodes and the evolution trend of the global network. The updated node features are input into the prediction loss function, and the partial derivatives of each weight in the graph convolutional neural network are calculated using the backpropagation algorithm to generate the topological gradient.

[0029] When calculating contribution weights using the aforementioned attention mechanism, a heterogeneous attention operator based on metapaths is employed. For both homogeneous node pairs corresponding to the associated edges and heterogeneous node pairs corresponding to the semantic associated edges, the hidden representations of neighboring nodes are transformed to specific semantic subspaces using edge type projection matrices. The cosine similarity score between the current node and its neighboring nodes in the semantic subspace is calculated, and the similarity score is normalized using a flexible maximum activation function to generate the contribution weights. The iterative aggregation process employs a multi-head attention parallel mechanism, which performs aggregation operations in parallel across multiple independent semantic subspaces and concatenates and maps the results to capture the multi-dimensional nonlinear interaction relationships between the tested feature nodes and the entity nodes.

[0030] In the iterative aggregation process, a residual connection structure and a layer normalization operator are introduced after each aggregation operation to linearly superimpose the aggregated features with the hidden layer representation before aggregation, thereby alleviating the gradient vanishing problem in deep networks and maintaining the stability of feature distribution. When calculating the partial derivative, a weight decay regularization term is added to the prediction loss function. By limiting the L2 norm of the partial derivative, overfitting is prevented when the graph convolutional neural network learns the associated edges and semantic associated edges, thereby ensuring that the generated topological gradient can accurately represent the real evolution logic of medical examination resource demand.

[0031] This scheme achieves accurate capture of multi-dimensional nonlinear interactions in heterogeneous medical networks by introducing a heterogeneous attention operator and a multi-head parallel mechanism. Residual connections and layer normalization operators effectively solve the gradient vanishing problem during the training process of deep graph neural networks, ensuring the stability of feature distribution. Weight decay regularization constraints enhance the generalization ability of partial derivative calculations and prevent model overfitting.

[0032] Furthermore, generating the gradient data includes: The neural encryption protocol, constructed through adversarial training between a sender encryption network and a receiver decryption network, is loaded to achieve nonlinear mapping of gradient features within an asymmetric encryption framework. The encryption network in the neural encryption protocol is used as a mapping operator to map the numerical matrix of the topological gradient to a high-dimensional chaotic space, eliminating the linear correspondence between the topological gradient and the original medical data. Bit-level perturbation and permutation are then applied to the mapped data to generate the encoded gradient data.

[0033] During the mapping process in the high-dimensional chaotic space, the encryption network pre-implanted constraint functions based on homomorphism-preserving operators to ensure that the mapped numerical matrix still satisfies the additive homomorphism property in the encrypted state. This allows the central server to directly perform linear aggregation of the encoded gradient data uploaded by multiple local medical nodes. Furthermore, the sending encryption network and the receiving decryption network periodically update the weight parameters generated by adversarial training through a pre-shared synchronization seed. This ensures that the bit-level perturbation and permutation logic is time-varying across different sampling periods, preventing replay attacks while guaranteeing that the central server can accurately reconstruct the aggregated global gradient increment using the corresponding receiving decryption network.

[0034] To eliminate numerical drift caused by the nonlinear mapping, the neural encryption protocol also includes a precision compensation layer, which performs fixed-point quantization on the topological gradient before the mapping operator is executed, and uses a scaling factor balancing algorithm to maintain the consistency of the numerical dynamic range in the high-dimensional chaotic space. When generating the encoded gradient data, the encryption network adjusts the mapped feature dimension to the standard input dimension of the receiver decryption network through a zero-padding transpose operator, so as to ensure that the gradient direction of the cross-node gradient aggregation result after decryption and restoration maintains the maximum cosine similarity with the average evolution direction of the original topological gradient, thereby maintaining the convergence stability of the global prediction model.

[0035] This scheme achieves direct gradient aggregation in the ciphertext state by combining an adversarial training neural encryption protocol with a homomorphism-preserving operator, improving the computational efficiency and privacy security of distributed training. Utilizing a precision compensation layer and a scaling factor balancing algorithm, it effectively eliminates numerical drift caused by nonlinear mapping, ensuring a high degree of consistency between the decrypted gradient increment and the original topological direction. Combining time-varying bit-level perturbation and dimension alignment operators not only enhances the ability to resist replay attacks but also guarantees the convergence stability and generalization performance of the global prediction model under complex encryption frameworks.

[0036] Furthermore, the robust aggregation strategy based on reputation value includes: The central server calculates a dynamic reputation score for each local medical node based on its historical prediction accuracy improvement contribution, gradient consistency index, and participation frequency using a reputation assessment algorithm. It compares the distribution differences among the received encoded gradient data, identifying nodes with reputation scores below a preset first safety threshold and gradient distribution deviations exceeding a preset deviation threshold as malicious interference data sources and removing them. The server normalizes the reputation scores of the remaining encoded gradient data nodes, calculating normalized weight coefficients for each node, ensuring a positive correlation between the normalized weight coefficients and the reputation scores. Finally, it multiplies each encoded gradient data point by its corresponding normalized weight coefficient and sums the results linearly to calculate the global gradient increment, thereby updating and generating unified global resource prediction model parameters.

[0037] In this embodiment, the preset first safety threshold is 0.6, which aims to exclude low-contribution and high-risk nodes and ensure the reliability of the aggregation parameters.

[0038] When quantifying the gradient consistency index, the central server calculates the cosine similarity in parameter space between the currently received encoded gradient data and the parameters of the global resource prediction model generated in the previous iteration, and constructs a reference baseline distribution by combining the median aggregation operator to identify outlier gradients that deviate significantly from the baseline characteristics. The reputation evaluation algorithm introduces a time decay factor to perform weighted decay processing on historical contributions, making the reputation score more sensitive to the recent gradient quality of nodes. In addition, before performing the linear summation, the central server uses a momentum correction term to smooth the calculated global gradient increment, offsetting the drastic fluctuations in the training direction of the global resource prediction model caused by removing some node data, and ensuring the collaborative robustness of the distributed training system.

[0039] This scheme improves the accuracy of the central server in identifying outlier gradients and Byzantine attacks by introducing cosine similarity quantification and median aggregation operators, solving the problem of malicious data interfering with the global resource prediction model in heterogeneous environments. A reputation algorithm with a time decay factor is used to achieve real-time and agile evaluation of node quality, ensuring a high degree of consistency between weight allocation and recent node performance. A momentum correction term is used to smooth the aggregation process, effectively eliminating model oscillations caused by node removal and ensuring the collaborative robustness and fast convergence of the distributed neural network in noisy, non-independent, and identically distributed environments.

[0040] Furthermore, the differential privacy injection mechanism includes: Based on the preset privacy budget and the sensitivity of the global resource prediction model parameters, the scale parameter of the Gaussian noise to be injected is calculated; a random noise vector conforming to the scale parameter is generated using a random number generator, and the random noise vector is superimposed on the gradient update amount in the local fine-tuning process to mask the contributions of the subject feature nodes and local medical nodes; after injecting the random noise vector, iterative optimization is continued using local data, so that the fine-tuned local prediction model can predict the local medical examination resource demand while satisfying differential privacy constraints.

[0041] Before calculating the scale parameter, the local medical node first performs a gradient pruning operation, using a preset norm threshold to forcibly constrain the gradient update amount, thereby quantifying the influence boundary of the subject feature node on the parameters of the global resource prediction model and determining the upper bound of the sensitivity. In addition, the differential privacy injection mechanism also introduces an adaptive noise adjustment strategy, dynamically adjusting the size of the scale parameter according to the number of iterations during the fine-tuning process. In the early stage of training, a larger scale parameter is used to ensure privacy strength, while in the later stage of training, noise interference is gradually reduced. A moment kernel is used to track and audit the accumulated privacy consumption in real time, thereby maximizing the convergence accuracy of the fine-tuned local prediction model while satisfying strong differential privacy constraints.

[0042] In this embodiment, the preset norm threshold is set to 1.0 to limit the maximum intensity of a single gradient update, thereby stabilizing the noise scale and quantizing the upper bound of sensitivity.

[0043] This scheme achieves an optimal balance between privacy protection and model prediction utility at the mathematical level by introducing gradient pruning and adaptive noise adjustment strategies. Gradient pruning explicitly quantifies the upper bound of sensitivity, eliminating the extreme interference of outliers on model parameters; combined with a moment kernel to audit the privacy budget in real time, it ensures that privacy loss remains within a controlled range during multiple iterations of fine-tuning. This mechanism not only completely masks the contribution of subject feature nodes at the underlying architecture, preventing member inference attacks, but also ensures the rapid convergence and high-accuracy prediction capability of the local prediction model under high-intensity privacy constraints through dynamic noise attenuation in the later stages of training.

[0044] Furthermore, it also includes dynamic resource balancing and scheduling steps: Based on the prediction results output by the local prediction model, combined with the real-time queuing data and maintenance cycle of local medical examination equipment, the expected load rate of each examination department within a preset time period is calculated. When the expected load rate exceeds a preset second safety threshold, a resource overflow sensing signal is triggered, and the idle capacity of associated surrounding medical institutions is identified based on the functional flow graph in the heterogeneous graph network. Scheduling suggestions are automatically generated based on the resource overflow sensing signal, and by adjusting the referral guidance instructions for non-emergency patients' examination periods, a global balance of medical examination resources in both spatial and temporal dimensions is achieved. In this embodiment, the preset second safety threshold is 85%, used to trigger an early warning before departmental capacity saturation and initiate cross-institutional resource collaborative scheduling.

[0045] The functional flow graph is a directed weighted graph dynamically generated by extracting historical referral weights and geographical topological constraints between nodes in the heterogeneous graph network and using the maximum flow minimum cut algorithm. It is used to characterize the collaborative elasticity of examination capabilities among medical institutions. When generating the scheduling suggestions, a multi-objective reinforcement learning optimization operator is used, with the joint optimization objectives of minimizing the variance of the expected load rate, minimizing the patient transfer time cost, and maximizing the utilization rate of examination equipment. The optimal patient time slot allocation parameters and referral node paths are determined by performing a heuristic search on the functional flow graph. The scheduling suggestions are pushed to the task queues of each local medical node in real time through an API interface, realizing the feedback dynamic allocation of medical examination resources.

[0046] The multi-objective reinforcement learning optimization operator uses information such as the expected load rate of each medical institution within a preset time period, the current queue length, and the geographical distance between institutions to constitute the environmental state; it uses the selection of target examination institutions and examination time periods for patients to be scheduled as the action space; it combines the load rate variance, the estimated patient transfer time, and the utilization rate of examination equipment calculated based on the scheduling scheme into a scalar reward value, which is used to evaluate the merits of each scheduling decision, thereby obtaining an approximately optimal scheduling strategy on the functional flow graph through iterative training.

[0047] This solution achieves deep coupling between the predictive model and the physical execution end by constructing a closed-loop scheduling system based on a functional flow graph. It utilizes the maximum flow minimum cut algorithm to quantify the collaborative elasticity between departments and, in conjunction with a multi-objective reinforcement learning optimization operator, achieves global optimality in scheduling decisions while balancing load balancing and transfer costs. Real-time feedback and allocation through API interfaces effectively eliminates resource mismatches between departments, shortening the waiting period for non-emergency patients while improving the overall turnover rate of medical examination equipment and the system's overload resilience.

[0048] This solution mitigates the prediction bias issues caused by insufficient structural correlation modeling and data isolation in traditional prediction models when dealing with medical resource supply and demand by constructing a heterogeneous graph convolutional neural network and a privacy-coordinated architecture. It utilizes the message passing mechanism of graph convolutional neural networks to achieve deep extraction of non-Euclidean topological features in complex medical networks, significantly improving prediction accuracy in non-stationary and fluctuating scenarios. Simultaneously, this invention integrates adversarial neural encryption and a reputation-driven robust aggregation strategy, ensuring the robustness of the distributed training system against malicious attacks without disclosing local raw data. Combined with a differential privacy injection mechanism, it provides robust technical support for precise resource allocation and dynamic supply-demand balance within medical consortia.

[0049] Example 2

[0050] Within a certain city's regional medical consortium, there are three general hospitals and five primary healthcare service institutions, which act as distributed local medical nodes. To achieve precise allocation of imaging examination resources, the system operates through the following specific process: Each local node extracts nearly a year's worth of electronic health record data, defining the medical records of each suspected chest pain patient as a subject feature node. The system processes unstructured complaints (such as "squeezing pain behind the sternum, accompanied by cold sweats") using a medical language model, generating high-dimensional semantic embeddings, and fusing them with structured features such as electrocardiograms and blood pressure. The system calculates the Pearson correlation coefficient between patients, setting a preset threshold of 0.75; for example, when the clinical similarity between patient A and patient B is 0.82, an association edge is established. Simultaneously, a medical knowledge graph containing entities such as "coronary CT," "aspirin," and "myocardial infarction" is introduced to construct semantic association edges. The system uses a 24-hour preset time span and utilizes a time-slicing mechanism to divide the association relationships into continuous snapshots. Through an adaptive graph learning layer, the system senses a surge in the strength of such associations in the morning of the current day, dynamically increasing the weight of the association edges to capture the cascading effect of group chest pain visits caused by a sudden drop in temperature.

[0051] Each hospital utilizes a graph convolutional neural network for aggregation. Through an attention mechanism, the system automatically identifies the "acute myocardial infarction" entity node as having the highest contribution weight to surrounding subject feature nodes. During the iteration process, a multi-head attention parallel mechanism is used to extract interaction relationships from multiple semantic subspaces, and residual connections are used to ensure that features are not lost in deep computation. Finally, each hospital calculates the partial derivatives of the weights using the backpropagation algorithm to generate a topological gradient reflecting local imaging examination (such as CT, angiography) needs.

[0052] To protect patient privacy, hospitals implement neural encryption protocols. The topological gradient matrix is ​​mapped to a high-dimensional chaotic space and bit-level perturbation is applied to ensure that interceptors cannot infer the patient's specific symptoms using the gradient. Due to the homomorphic preservation property of the encryption operator, the encrypted data is uploaded to a central server for direct aggregation. Simultaneously, a precision compensation layer corrects numerical drift caused by nonlinear mapping, ensuring that the decrypted global resource prediction model remains accurate.

[0053] After receiving the gradients from each hospital, the central server calculates their cosine similarity to the parameters from the previous round. If a small clinic uploads a gradient with a significantly deviated distribution and a reputation score below 0.6, the server determines it to be malicious interference or low-quality data and removes it. The remaining gradients are linearly summed based on the reputation-normalized weights to generate unified global resource prediction model parameters, which are then distributed to each hospital.

[0054] After receiving the global parameters, each hospital first performs gradient clipping, setting the preset norm threshold to 1.0 to quantize the upper bound of sensitivity. Subsequently, a Gaussian noise vector conforming to the scaling parameters is injected according to the privacy budget. During the fine-tuning phase, the system utilizes an adaptive noise strategy to gradually reduce noise in the later stages of training, ensuring that the local model achieves the highest predictive accuracy for the hospital's imaging examination needs without revealing individual patient information.

[0055] The system's output prediction showed that the cardiac surgery department of the First General Hospital was expected to reach a load rate of 92% in the next 6 hours, exceeding the safety threshold of 85%. The system immediately triggered a resource overflow sensing signal and, based on the functional flow graph and the maximum flow minimum cut algorithm, identified that the Second Hospital within a two-kilometer radius still had 30% of its CT capacity available. Using a multi-objective reinforcement learning optimization operator, the system automatically generated scheduling suggestions: postponing the examination times of 5 non-urgent follow-up patients and directing 3 newly admitted stable patients to the Second Hospital for examination. These suggestions were pushed to the task queues of each hospital in real time via API, achieving a spatiotemporal balance of medical resources within the region.

[0056] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the supply and demand balance of medical examination resources based on graph neural networks, characterized in that, include: Heterogeneous graph networks containing medical institution attributes and patient association information are constructed on multiple distributed local medical nodes. Feature aggregation is performed through graph convolutional neural networks to generate topological gradients that reflect local resource needs. The topological gradients are encrypted and encoded using a neural encryption protocol pre-trained through adversarial co-evolutionary learning to generate encoded gradient data, ensuring that the original medical data features are in an irreversible privacy-preserving state during transmission. Each local medical node uploads the encoded gradient data to the central server; the central server adopts a robust aggregation strategy based on reputation value to quantitatively evaluate the quality and historical contribution of the data uploaded by each node, removes potential malicious interference data, and fuses the gradients from multiple sources to generate unified global resource prediction model parameters. The central server distributes the global resource prediction model parameters to each local medical node; each node, in conjunction with the differential privacy injection mechanism, adds a noise operator to the global resource prediction model for localized fine-tuning.

2. The method for predicting the supply and demand balance of medical examination resources based on graph neural networks according to claim 1, characterized in that, Constructing the heterogeneous graph network includes: Each patient identifier and / or single visit record within the local medical node is defined as a test feature node, and structured features and unstructured clinical descriptions from electronic health records are extracted as the initial attribute vectors of the test feature nodes. The clinical phenotype similarity between different test feature nodes is calculated using the Pearson correlation coefficient. When the clinical phenotype similarity exceeds a preset threshold, an association edge is established between the corresponding two test feature nodes. A medical knowledge graph is introduced, defining medical examination items, drugs, symptoms, and diseases as entity nodes, and semantic association edges between the entity nodes and the test feature nodes are established according to the clinical diagnosis and treatment path to form the heterogeneous graph network.

3. The method for predicting the supply and demand balance of medical examination resources based on graph neural networks according to claim 2, characterized in that, Generating topological gradients includes: By using the linear transformation layer of the graph convolutional neural network, the initial attribute vectors of the test feature nodes and the attribute features of the entity nodes are mapped to a unified low-dimensional vector space, generating hidden layer representations of the test feature nodes and the entity nodes. An attention mechanism is used to calculate the contribution weights between adjacent nodes in the heterogeneous graph network. Based on the associated edges and semantic associated edges in the heterogeneous graph network, the hidden layer representations of neighboring nodes are iteratively aggregated to update the features of the current node, identifying the characteristics of local nodes and the evolution trend of the global network. The updated node features are input into the prediction loss function, and the partial derivatives of each weight in the graph convolutional neural network are calculated using the backpropagation algorithm to generate the topological gradient.

4. The method for predicting the supply and demand balance of medical examination resources based on graph neural networks according to claim 1, characterized in that, Generating the gradient data includes: The neural encryption protocol, constructed through adversarial training between a sender encryption network and a receiver decryption network, is loaded to achieve nonlinear mapping of gradient features within an asymmetric encryption framework. The encryption network in the neural encryption protocol is used as a mapping operator to map the numerical matrix of the topological gradient to a high-dimensional chaotic space, eliminating the linear correspondence between the topological gradient and the original medical data. Bit-level perturbation and permutation are then applied to the mapped data to generate the encoded gradient data.

5. The method for predicting the supply and demand balance of medical examination resources based on graph neural networks according to claim 1, characterized in that, The reputation-based robust aggregation strategy includes: The central server calculates a dynamic reputation score for each local medical node based on its historical prediction accuracy improvement contribution, gradient consistency index, and participation frequency using a reputation assessment algorithm. It compares the distribution differences among the received encoded gradient data, identifying nodes with reputation scores below a preset first safety threshold and gradient distribution deviations exceeding a preset deviation threshold as malicious interference data sources and removing them. The server normalizes the reputation scores of the remaining encoded gradient data nodes, calculating normalized weight coefficients for each node, ensuring a positive correlation between the normalized weight coefficients and the reputation scores. Finally, it multiplies each encoded gradient data point by its corresponding normalized weight coefficient and sums the results linearly to calculate the global gradient increment, thereby updating and generating unified global resource prediction model parameters.

6. The method for predicting the supply and demand balance of medical examination resources based on graph neural networks according to claim 1, characterized in that, The differential privacy injection mechanism includes: Based on the preset privacy budget and the sensitivity of the global resource prediction model parameters, the scale parameter of the Gaussian noise to be injected is calculated; a random noise vector conforming to the scale parameter is generated using a random number generator, and the random noise vector is superimposed on the gradient update amount in the local fine-tuning process to mask the contributions of the subject feature nodes and local medical nodes; after injecting the random noise vector, iterative optimization is continued using local data, so that the fine-tuned local prediction model can predict the local medical examination resource demand while satisfying differential privacy constraints.

7. The method for predicting the supply and demand balance of medical examination resources based on graph neural networks according to claim 1, characterized in that, It also includes dynamic resource balancing and scheduling steps: Based on the prediction results output by the local prediction model, combined with the real-time queuing data and maintenance cycle of local medical examination equipment, the expected load rate of each examination department in the future preset time period is calculated; when the expected load rate exceeds the preset second safety threshold, a resource overflow sensing signal is triggered, and the idle capacity of the associated surrounding medical institutions is identified according to the functional flow graph in the heterogeneous graph network; scheduling suggestions are automatically generated based on the resource overflow sensing signal, and the referral guidance instructions for non-emergency patients' examination periods are adjusted to achieve a global balance of medical examination resources in the spatial and temporal dimensions.

Images