Hierarchical scheduling diagnosis and treatment method based on heart disease complexity

Abstract

The invention relates to a hierarchical scheduling diagnosis and treatment method based on heart disease complexity. The method comprises the following steps: performing preliminary reasoning on heart fusion features of a case to be diagnosed and treated by using an edge end small model to obtain a heart disease type distribution entropy value; based on a preset dynamic threshold value, obtaining a scheduling candidate queue by using the heart disease type distribution entropy value; obtaining a resource index and a case complexity index corresponding to the scheduling candidate queue on the basis of the diagnosis resource in combination with the heart fusion feature; based on the joint scheduling rule table, obtaining a scheduling result according to the resource index and the case complexity index; the scheduling result comprises edge end small model diagnosis and cloud end large model diagnosis. By adopting the method, an efficient hierarchical diagnosis and treatment mode with local processing as a main part and remote calling as an auxiliary part can be realized, the problems of excessive resource consumption and invalid uploading are avoided, and the diagnosis and treatment efficiency and accuracy are improved.

Description

Technical Field

[0001] The present invention belongs to the field of hierarchical scheduling, and in particular relates to a hierarchical scheduling diagnosis and treatment method based on the complexity of heart diseases. Background Art

[0002] With the development of medical artificial intelligence technology, the integration of edge computing and cloud intelligence has emerged. By deploying small models on the edge and large models on the cloud, accurate diagnosis of heart disease can be achieved.

[0003] In traditional technologies, the diagnostic process is usually fixed with a static path, that is, a small model at the edge is used to collect data, and a large model in the cloud is used for diagnosis to obtain diagnostic results.

[0004] However, although traditional diagnosis and treatment scheduling methods ensure diagnostic accuracy, they have major deficiencies in data transmission, privacy protection, and resource utilization efficiency. They lack adaptive scheduling capabilities based on case complexity and cannot make judgments based on multiple factors such as the diagnostic uncertainty of the current case and the resource status of the equipment. This may lead to problems such as edge model misjudgment, cloud resource congestion, or diagnosis and treatment delays. Summary of the Invention

[0005] Based on this, it is necessary to provide a hierarchical scheduling diagnosis and treatment method based on the complexity of heart disease, which can perform dynamic hierarchical scheduling based on the complexity of the case and resource status, in order to address the above technical problems.

[0006] In a first aspect, the present application provides a hierarchical scheduling diagnosis and treatment method based on the complexity of heart disease, comprising:

[0007] The small edge model is used to perform preliminary inference on the cardiac fusion features of the cases to be diagnosed and treated, and the entropy value of the distribution of cardiac disease types is obtained;

[0008] Based on the preset dynamic threshold, the candidate queue for scheduling is obtained using the entropy value of the heart disease type distribution;

[0009] Based on diagnostic resources and cardiac fusion features, the resource index and case complexity index corresponding to the scheduling candidate queue are obtained;

[0010] Based on the joint scheduling rule table, the scheduling results are obtained according to the resource index and case complexity index; the scheduling results include small model diagnosis at the edge and large model diagnosis in the cloud.

[0011] In one embodiment, a small edge model is used to perform preliminary inference on the cardiac fusion features of the case to be diagnosed and treated, and obtain the entropy value of the cardiac disease type distribution, including:

[0012] Using the edge-end small model to perform preliminary screening based on cardiac fusion features, the probability distribution of cardiac disease types is obtained;

[0013] Calculate the entropy value of heart disease type distribution based on the probability distribution of disease categories;

[0014] The entropy value of the heart disease type distribution is obtained by the following formula:

[0015] H s (x) = -∑P s (y|x)logP s (y|x)

[0016] Among them, H s (x) is the entropy value of the distribution of heart disease types; P s (y|x) is the probability distribution of disease categories.

[0017] In one embodiment, based on a preset dynamic threshold, the candidate scheduling queue is obtained by utilizing the entropy value of the heart disease type distribution, including:

[0018] Comparing the heart disease type distribution entropy value with a preset dynamic threshold to obtain a comparison result;

[0019] If the entropy value of the heart disease type distribution is greater than the preset dynamic threshold, the cases to be diagnosed and treated corresponding to the entropy value of the heart disease type distribution are included in the scheduling candidate queue.

[0020] In one embodiment, the resource index and case complexity index corresponding to the scheduling candidate queue are obtained based on diagnostic resources combined with cardiac fusion features, including:

[0021] Based on the dual-dimensional monitoring strategy of resources and cases, the diagnostic resources of the device are obtained; diagnostic resources include memory, computing power, and network bandwidth;

[0022] The resource index is calculated based on the diagnostic resources;

[0023] Based on the feature space of the small edge model, the complexity of the heart disease of the case to be diagnosed is constructed according to the cardiac fusion features, and the case complexity index is obtained.

[0024] In one embodiment, the resource index is obtained by the following formula:

[0025] R = ∈ 1 FLOPs avail +∈2Mem free

[0026] Where R is the resource index; ∈1, ∈2 are resource weight coefficients; FLOPs avail The available computing power of the device; Mem free The free memory of the device;

[0027] The case complexity index is obtained by the following formula:

[0028]

[0029] Among them, C(x) is the case complexity index; Z s (s) is the cardiac fusion feature; is the average value of the training set features.

[0030] In one embodiment, the joint scheduling rule table corresponds to the following method:

[0031] When the resource index is greater than or equal to the first preset threshold, and the case complexity index is less than or equal to the second preset threshold, the scheduling result is the edge small model diagnosis;

[0032] When the resource index is less than the first preset threshold, or the case complexity index is greater than the second preset threshold, the scheduling result is the cloud-based large model diagnosis.

[0033] In one embodiment, the method further comprises:

[0034] Obtain diagnostic results of large cloud-based models;

[0035] Update dynamic thresholds based on diagnostic results;

[0036] The pseudo label is obtained based on the diagnosis results combined with the heart disease type distribution entropy value encoding of the small edge model;

[0037] Based on the constraints of the elastic weight solidification algorithm, the parameters of the small edge model are updated through incremental learning using pseudo labels;

[0038] The dynamic threshold is obtained by the following formula:

[0039]

[0040] Among them, θ new is the updated dynamic threshold; is the attenuation coefficient; θ ori To set the dynamic threshold; D hard A collection of difficult cases; H s (x) is the diagnosis result;

[0041] The constraint loss function corresponding to the elastic weight solidification algorithm is obtained through the following formula:

[0042]

[0043] Among them, γ is the penalty coefficient; F i is the Fisher information matrix, which is used to measure the importance of model parameters; θ i is the current parameter value; θ i,old is the old parameter; i is the index of traversing the small model parameters.

[0044] In a second aspect, the present application further provides a hierarchical scheduling diagnosis and treatment device based on the complexity of heart disease, comprising:

[0045] A preliminary screening module is used to use a small edge model to perform preliminary inference on the cardiac fusion features of the case to be diagnosed and treated, and obtain the entropy value of the distribution of cardiac disease types;

[0046] A candidate scheduling module is used to obtain a scheduling candidate queue based on a preset dynamic threshold and using the entropy value of the heart disease type distribution;

[0047] The scheduling indicator module is used to obtain the resource index and case complexity index corresponding to the scheduling candidate queue based on diagnostic resources and cardiac fusion characteristics;

[0048] The hierarchical scheduling module is used to obtain scheduling results based on the joint scheduling rule table, resource index and case complexity index.

[0049] In a third aspect, the present application also provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the steps of any of the above-mentioned hierarchical scheduling diagnosis and treatment methods based on the complexity of heart disease.

[0050] In a fourth aspect, the present application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the above-mentioned hierarchical scheduling diagnosis and treatment methods based on the complexity of heart disease.

[0051] This hierarchical scheduling approach for cardiac disease complexity establishes a clear path from data input to decision output by clearly defining three quantitative metrics: entropy, resource index, and complexity index. This transforms diagnostic scheduling from experience-based to data-driven, improving the controllability of hierarchical scheduling. This approach breaks through the traditional fixed distribution mechanism, enabling dynamic adaptation of diagnostic behavior based on computing conditions, effectively alleviating resource bottlenecks at the edge. Case complexity is determined by quantifying the deviation between case features and semantic centers in the feature space of small models, replacing manual labeling or rule construction and improving the generalizability and deployability of scheduling strategies. Only cases with uncertain diagnoses and complex feature structures trigger diagnosis on the cloud-based large model, avoiding resource overconsumption and ineffective uploads, and achieving an efficient tiered diagnosis and treatment model with local processing as the primary focus and remote invocation as the supplementary support. The decoupling of the scheduling mechanism from the model structure facilitates adaptation to different model types, disease types, or medical equipment requirements, ensuring excellent engineering versatility and cross-scenario deployment capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0053] Figure 1 This is a diagram of the application environment of the hierarchical scheduling diagnosis and treatment method based on the complexity of heart disease according to the present invention;

[0054] Figure 2 Schematic diagram of the process of the hierarchical scheduling diagnosis and treatment method based on the complexity of heart disease according to the present invention;

[0055] Figure 3 Schematic diagram of the step-by-step process of step S201;

[0056] Figure 4 Schematic diagram of the step-by-step process of step S203;

[0057] Figure 5 This is a structural diagram of the hierarchical scheduling diagnosis and treatment device based on the complexity of heart disease of the present invention. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0059] The hierarchical scheduling diagnosis and treatment method based on the complexity of heart disease provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown. Among them, the edge local terminal 101 communicates with the cloud server 102 through the network. The data storage system can store the data that the cloud server 102 needs to process. The data storage system can be integrated on the cloud server 102, or it can be placed on the cloud or other network servers. Among them, the edge local terminal 101 can be, but is not limited to, various personal computers, laptops, smart phones, tablets, Internet of Things devices and portable wearable devices. Portable wearable devices can be smart watches, smart bracelets, etc. The cloud server 102 can be implemented as an independent server or a server cluster consisting of multiple servers.

[0060] In an exemplary embodiment, Figure 2 As shown, a hierarchical scheduling diagnosis and treatment method based on the complexity of heart disease is provided. Figure 1Taking the edge local end as an example, the method includes the following steps 201 to 204. Among them:

[0061] S201. Use the small edge model to perform preliminary inference on the cardiac fusion features of the case to be diagnosed and treated, and obtain the entropy value of the distribution of cardiac disease types.

[0062] Schematically, a small edge model performs preliminary inference on the cardiac fusion features of the case being diagnosed and calculates the entropy of the distribution of cardiac disease types based on this information. The cardiac fusion features refer to a unified feature vector formed by feature alignment and semantic fusion of modal information such as electrocardiogram (ECG), cardiac magnetic resonance imaging (CMR), and unstructured text cases. The small edge model, a lightweight diagnostic network deployed at the medical terminal, is tasked with quickly predicting the fusion features and outputting a disease classification probability distribution vector. Furthermore, the entropy value of this distribution vector is calculated. This entropy value is used to measure the small model's confidence in the current case. Lower entropy values ​​indicate a more concentrated judgment on a particular category and a higher confidence in the decision. Conversely, lower entropy values ​​indicate that the case has a high probability of belonging to multiple categories, with fuzzy boundaries or inter-class interference, requiring further diagnostic confirmation.

[0063] S202: Based on a preset dynamic threshold, a candidate scheduling queue is obtained using the heart disease type distribution entropy value.

[0064] Based on the preset dynamic threshold, the entropy value calculated above is judged, and a scheduling candidate queue is generated accordingly. Schematically, the dynamic threshold is not a static setting, and can be dynamically adjusted according to the historical misjudgment rate, the current diagnostic load, and the complexity of the case to define the boundary area where diagnostic risks may exist. Specifically, when the entropy value of a case exceeds the dynamic threshold, it is judged as a sample with insufficient diagnostic confidence and marked as a candidate for further scheduling judgment. The scheduling candidate queue is part of the current diagnosis and treatment task pool, indicating a collection of cases for which the final diagnostic path cannot be determined at the current stage. The buffering decision between initial screening and re-judgment avoids fluctuations in diagnostic quality caused by misjudgment of small models, while controlling the excessive call of cloud resources.

[0065] S203. Obtain resource index and case complexity index corresponding to the scheduling candidate queue based on diagnostic resources and cardiac fusion features.

[0066] Schematically, the corresponding resource index and complexity index are calculated for the cases in the scheduling candidate queue, combining the cardiac fusion characteristics and diagnostic resource status of the case itself. The resource index is used to reflect the current diagnostic resource status of the edge device. It is mainly obtained by a weighted combination of standardized indicators such as the device's remaining computing power, memory, and bandwidth. The resource index reflects whether the local diagnostic capability is currently available and is an external constraint factor for the selection of the diagnostic path. The complexity index reflects whether the case is a difficult and complicated disease sample in the feature semantic space. Specifically, the complexity value is obtained by calculating the Euclidean distance between the case feature and the average vector of the training set. The larger the distance, the greater the deviation of the case in the semantic space, the weaker the model's recognition ability, and the scheduling tends to be sent to the cloud for processing.

[0067] S204. Based on the joint scheduling rule table, according to the resource index and the case complexity index, the scheduling result is obtained; the scheduling result includes the diagnosis of the small model at the edge and the diagnosis of the large model at the cloud.

[0068] The final diagnostic path is determined based on the combined resource index and complexity index of the joint scheduling rule table. Schematically, the joint scheduling rule table can be a static matrix or a dynamically updated strategy mapping table. Its core logic is to process locally as much as possible when resources are sufficient, and to transfer to the cloud for processing when resources are insufficient or the sample complexity is too high. For example, when the resource index is high and the complexity is low, the diagnosis can be directly performed by the small model on the edge; if both are in a high-risk range, they will be transferred to the large cloud model for processing; if resources are tight but the case complexity is medium, the system can also use the large cloud model for processing to improve diagnostic efficiency.

[0069] In the aforementioned hierarchical scheduling and treatment method based on heart disease complexity, the entropy of the heart disease type distribution is calculated from the probability distribution of heart disease categories output by a small edge model to quantify the diagnostic uncertainty of the current case, thus implementing a technical mechanism to derive the diagnostic confidence level from the model output behavior. As a continuously adjustable quantitative indicator, the entropy value provides a stable and adaptive discriminant signal for subsequent scheduling candidate screening, effectively avoiding the rigid setting of diagnostic confidence boundaries and improving the recognition rate of fuzzy prediction samples. By setting a dynamic threshold mechanism, the diagnostic entropy value is linked to adjustable policy parameters, making the judgment logic adaptable to sample distribution drift, model accuracy fluctuations, and load changes. This mechanism frees the generation process of scheduling candidate queues from static constraints, provides a certain degree of adaptability, and improves the scenario robustness of the diagnosis strategy. By introducing a joint evaluation mechanism of resource index and case complexity index, the scheduling logic no longer relies solely on model output results, but also considers the comprehensive judgment of both the current diagnostic capability and the availability of resources. Based on the conditional judgment strategy of the joint scheduling rule table, a scheduling judgment structure with clear logic and unique path is established, which avoids the difficulties in dividing diagnostic responsibilities and the uncertainty of system behavior caused by traditional fuzzy strategies.

[0070] In one embodiment, Figure 3 As shown in the figure, the small edge model is used to perform preliminary inference on the cardiac fusion features of the cases to be diagnosed and treated, and the entropy value of the distribution of cardiac disease types is obtained, including:

[0071] S301. Use the small edge model to perform preliminary screening based on cardiac fusion features to obtain a probability distribution of cardiac disease types.

[0072] The small edge model is called to perform preliminary screening based on the input heart fusion features. Schematically, after receiving the fusion features, the small edge model uses its internal lightweight neural network structure to perform forward inference and output a probability distribution vector for heart disease types. Each dimension of this vector represents the probability of the current case being classified as a corresponding heart disease type, such as myocardial infarction, atrial fibrillation, or valvular disease.

[0073] S302. Calculate the heart disease type distribution entropy value based on the disease category probability distribution.

[0074] The entropy value of the heart disease type distribution is obtained by the following formula:

[0075] H s (x) = -∑P s (y|x)logP s (y|x)

[0076] Among them, H s (x) is the entropy value of the distribution of heart disease types; P s(y|x) is the probability distribution of disease categories.

[0077] Illustratively, when predictions are concentrated in a single disease category, such as in the probability distribution of [0.95, 0.03, 0.02], the entropy is low, indicating stable predictions from the small model at the edge. On the other hand, when predictions are more dispersed, such as in the probability distribution of [0.34, 0.33, 0.33], the entropy is high, indicating that the model has difficulty distinguishing between categories and faces the potential risk of misjudgment. This entropy calculation process can quantify the confidence level of the current case diagnosis stage.

[0078] In one embodiment, based on a preset dynamic threshold, the candidate scheduling queue is obtained by utilizing the entropy value of the heart disease type distribution, including:

[0079] S11. Compare the heart disease type distribution entropy value with a preset dynamic threshold to obtain a comparison result.

[0080] Schematically, the heart disease type distribution entropy value of each case to be diagnosed is compared with the currently set preset dynamic threshold to obtain a judgment result.

[0081] S12. If the heart disease type distribution entropy value is greater than a preset dynamic threshold, the cases to be diagnosed and treated corresponding to the heart disease type distribution entropy value are included in the scheduling candidate queue.

[0082] For example, if the entropy value corresponding to a case is greater than the current dynamic threshold, it is determined that the preliminary diagnosis result of the case is uncertain, that is, it is a potential high-risk case with low confidence and fuzzy boundaries, so it is included in the scheduling candidate queue. This queue serves as an input set for resource evaluation and complexity judgment in subsequent steps, indicating that these cases may require the participation of cloud models for more accurate diagnosis confirmation. Through this candidate screening mechanism based on dynamic thresholds, the first layer of filtering of diagnostic confidence perception and intelligent diversion is realized, effectively avoiding the early spread of misdiagnosis risks, while improving the pertinence and computational efficiency of subsequent resource scheduling strategies.

[0083] In one embodiment, Figure 4 As shown in the figure, based on the diagnostic resources combined with the cardiac fusion features, the resource index and case complexity index corresponding to the scheduling candidate queue are obtained, including:

[0084] S401. Based on the resource-case dual-dimensional monitoring strategy, obtain the diagnostic resources of the device; diagnostic resources include memory, computing power, and network bandwidth.

[0085] Schematically, diagnostic resource status information for edge devices is obtained in real time. This monitoring strategy, integrated into the operational module of the edge computing platform, regularly collects key hardware usage information for the current node, primarily including memory usage, computing power utilization, and current network bandwidth margin. Memory metrics reflect the ability to cache small models during loading and inference, computing power metrics correspond to real-time computing capabilities, and bandwidth metrics relate to the time required to send data to the cloud and transmission reliability.

[0086] S402: Calculate a resource index based on the diagnosis resources.

[0087] Schematically, based on the acquired resource data, the resource index of each edge node in the current scheduling cycle is calculated. The resource index reflects the computing power supported by the current node and is an important reference for determining whether the diagnostic task is suitable for local completion. Specifically, by normalizing and weighting each resource type, a comprehensive evaluation function is formed as follows. This index quantifies whether a node currently has the conditions to independently complete a diagnosis for a candidate case.

[0088] S403. Based on the feature space of the small edge model, the complexity of the heart disease of the case to be diagnosed is constructed according to the heart fusion features to obtain a case complexity index.

[0089] Schematically, based on the feature embedding space of the edge model, the structural semantic complexity of the current candidate case is evaluated to generate its corresponding case complexity index. Specifically, the cardiac fusion features of the current case are input into the small model feature extraction network to obtain its embedding vector in the diagnostic semantic space, and the Euclidean distance is calculated with the global embedding mean vector of the training data set. This distance is used as a numerical expression of the case complexity, indicating the degree of deviation of the current case from the center of the training sample in the semantic space. For example, the larger the distance, the more atypical the case feature is, the higher the probability of being outside the cognitive boundary of the small model, and the more likely it is to lead to misjudgment or fuzzy recognition, so it is sent to the cloud for processing first during scheduling.

[0090] In one embodiment, the resource index is obtained by the following formula:

[0091] R = ∈ 1 FLOPs avail +∈2Mem free

[0092] Where R is the resource index; ∈1, ∈2 are resource weight coefficients; FLOPs avail The available computing power of the device; Mem free The free memory of the device;

[0093] The case complexity index is obtained by the following formula:

[0094]

[0095] Among them, C(x) is the case complexity index; Z s (s) is the cardiac fusion feature; is the average value of the training set features.

[0096] In one embodiment, the joint scheduling rule table corresponds to the following method:

[0097] When the resource index is ≥ the first preset threshold and the case complexity index is ≤ the second preset threshold, the scheduling result is the edge small model diagnosis.

[0098] In principle, when the resource index of a candidate case is greater than or equal to the first preset threshold and the case complexity index is less than or equal to the second preset threshold, it is determined that the case has both local execution conditions and is not a difficult-to-identify or boundary-fuzzy case. Therefore, the scheduling result is set as the edge small model diagnosis, and the final diagnosis result can be directly generated by the edge small model to ensure diagnostic efficiency.

[0099] When the resource index is less than the first preset threshold, or the case complexity index is greater than the second preset threshold, the scheduling result is the cloud-based large model diagnosis.

[0100] In principle, when the resource index is lower than the first preset threshold, or the case complexity index is higher than the second preset threshold, if any condition is met, the current edge node is deemed unsuitable to continue to undertake the diagnosis task, or the case is deemed not to have the edge processing conditions. Therefore, the scheduling result is set to the cloud-based large model diagnosis, and the fusion features of the current case are uploaded to the cloud, and the large model performs high-precision reasoning.

[0101] In one embodiment, the method further comprises:

[0102] S51. Obtain the diagnosis result of the cloud-based large model diagnosis.

[0103] Schematically, a model self-optimization mechanism based on diagnostic feedback is implemented to update and adjust dynamic thresholds, adaptively enhance the capabilities of small edge models, and incrementally correct historical error samples. This mechanism is activated after the model is called and the diagnostic results are generated. After obtaining the diagnostic results from the large cloud model, this result is cross-analyzed with the set of cases identified as difficult cases in the current scheduling. The current dynamic threshold is adjusted accordingly to improve the adaptability of the diagnostic confidence boundary in future scheduling decisions.

[0104] S52: Update the dynamic threshold according to the diagnosis result.

[0105] The dynamic threshold is obtained by the following formula:

[0106]

[0107] Among them, θ new is the updated dynamic threshold; is the attenuation coefficient; θ ori To set the dynamic threshold; D hard A collection of difficult cases; H s (x) is the diagnosis result;

[0108] For example, In a preferred embodiment, This mechanism ensures that when the model's diagnostic performance changes or the data distribution drifts, the diagnostic boundary can be automatically updated based on the diagnostic result feedback, and the scheduling decision accuracy can be continuously optimized.

[0109] S53. A pseudo label is obtained based on the diagnosis result and the heart disease type distribution entropy value encoding of the edge small model.

[0110] Schematically, the large-scale model's diagnostic results for each difficult case are jointly encoded with the heart disease type distribution entropy output by the marginal small-scale model during initial screening. This generates a pseudo-label for the small-scale model's self-learning. This pseudo-label serves as a soft supervisory signal between the fused prediction and the true diagnosis, guiding the small-scale model to gradually approach the large-scale model's diagnostic behavior. Pseudo-label construction methods can include directly assigning high-confidence labels to the current sample, or introducing weighted fusion strategies and probabilistic matching structures, enabling flexible adjustments based on deployment scenarios.

[0111] S54. Based on the elastic weight solidification algorithm constraints, the parameters of the small model at the edge are updated through incremental learning using pseudo labels.

[0112] The constraint loss function corresponding to the elastic weight solidification algorithm is obtained through the following formula:

[0113]

[0114] Among them, γ is the penalty coefficient; F i is the Fisher information matrix, which is used to measure the importance of model parameters; θ i is the current parameter value; θ i,old is the old parameter; i is the index of traversing the small model parameters.

[0115] Schematically, pseudo-labels are fed into the edge model, and a constrained incremental learning process is performed to complete fine-tuning optimization at the parameter level. Furthermore, to avoid catastrophic forgetting in the small model during incremental updates, the Elastic Weight Consolidation (EWC) algorithm is introduced as a structural regularization term to suppress updates to important parameters in the model. By minimizing the constraint loss term of the constraint loss function, weights that are critical to the performance of existing tasks are prioritized when updating the small model parameters, effectively achieving a balance between adapting to new tasks and maintaining old tasks.

[0116] This mechanism converts diagnostic results from the judgment endpoint into a model feedback signal, constructing a self-closed loop path from prediction behavior to parameter update, enabling the small edge model to have the ability to continuously evolve, continuously improving its adaptability to ambiguous cases and new feature combinations, and effectively expanding the timeliness of the system's use and the sustainability of its deployment in real medical environments.

[0117] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0118] Based on the same inventive concept, embodiments of the present application further provide a device for hierarchical scheduling of diagnosis and treatment based on the complexity of heart disease, which is used to implement the aforementioned method for hierarchical scheduling of diagnosis and treatment based on the complexity of heart disease. The solution provided by this device is similar to the solution described in the aforementioned method. Therefore, the specific limitations of one or more embodiments of the device for hierarchical scheduling of diagnosis and treatment based on the complexity of heart disease provided below can be found in the above-mentioned limitations of the hierarchical scheduling of diagnosis and treatment based on the complexity of heart disease, and will not be repeated here.

[0119] In an exemplary embodiment, Figure 5 As shown, a hierarchical scheduling diagnosis and treatment device based on the complexity of heart disease is provided, comprising:

[0120] A preliminary screening module is used to use a small edge model to perform preliminary inference on the cardiac fusion features of the case to be diagnosed and treated, and obtain the entropy value of the distribution of cardiac disease types;

[0121] A candidate scheduling module is used to obtain a scheduling candidate queue based on a preset dynamic threshold and using the entropy value of the heart disease type distribution;

[0122] The scheduling indicator module is used to obtain the resource index and case complexity index corresponding to the scheduling candidate queue based on diagnostic resources and cardiac fusion characteristics;

[0123] The hierarchical scheduling module is used to obtain scheduling results based on the joint scheduling rule table, resource index and case complexity index.

[0124] In one embodiment, the preliminary screening module is further configured to perform preliminary screening based on cardiac fusion features using a small edge model to obtain a probability distribution of cardiac disease types;

[0125] The preliminary screening module is also used to calculate the distribution entropy value of the heart disease type based on the probability distribution of the disease category.

[0126] In one embodiment, it further includes:

[0127] The comparison module is used to compare the heart disease type distribution entropy value with a preset dynamic threshold to obtain a comparison result.

[0128] In one embodiment, it further includes:

[0129] The resource monitoring module is used to obtain the diagnostic resources of the equipment based on the resource-case dual-dimensional monitoring strategy;

[0130] Resource judgment module, used to calculate resource index based on diagnosis resources;

[0131] The difficult and complicated disease module is used to construct the complexity of heart diseases of cases to be diagnosed and treated based on the feature space of the small edge model and the cardiac fusion features, and obtain the case complexity index.

[0132] In one embodiment, it further includes:

[0133] Feedback module, used to obtain the diagnosis results of the cloud-based large model and update the dynamic threshold according to the diagnosis results;

[0134] The optimized training module is used to obtain pseudo labels based on the diagnosis results combined with the entropy value encoding of the heart disease type distribution of the edge model, and to update the parameters of the edge model through incremental learning using the pseudo labels based on the elastic weight solidification algorithm constraints.

[0135] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0136] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0137] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the partial description of the method embodiments. The device embodiments described above are merely illustrative, wherein the components described as separate parts may or may not be physically separated, and the parts displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the disclosed solution. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0138] The above-described embodiments merely represent several implementation methods of the embodiments of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that a person skilled in the art may make various modifications and improvements without departing from the concept of the embodiments of the present application, and these modifications and improvements fall within the scope of protection of the embodiments of the present application.

Claims

1. A hierarchical scheduling diagnosis and treatment method based on the complexity of heart disease, characterized by: The method comprises: The small edge model is used to perform preliminary inference on the cardiac fusion features of the cases to be diagnosed and treated, and the entropy value of the distribution of cardiac disease types is obtained; Based on a preset dynamic threshold, utilizing the heart disease type distribution entropy value to obtain a scheduling candidate queue; Obtaining a resource index and a case complexity index corresponding to the scheduling candidate queue based on diagnostic resources and the cardiac fusion feature; Based on the joint scheduling rule table, a scheduling result is obtained according to the resource index and the case complexity index; the scheduling result includes edge small model diagnosis and cloud large model diagnosis.

2. The method according to claim 1, characterized in that The small edge model is used to perform preliminary inference on the cardiac fusion features of the case to be diagnosed and treated, and obtain the entropy value of the cardiac disease type distribution, including: Using the edge-end small model to perform preliminary screening based on the heart fusion features to obtain a probability distribution of heart disease types; Calculating the heart disease type distribution entropy value according to the disease category probability distribution; The entropy value of the heart disease type distribution is obtained by the following formula: H s (x)=-∑P s (y|x)logP s (y|x) Among them, H s (x) is the entropy value of the heart disease type distribution; P s (y|x) is the probability distribution of the disease category.

3. The method according to claim 1, characterized in that The method of obtaining a scheduling candidate queue based on a preset dynamic threshold and utilizing the heart disease type distribution entropy value includes: Comparing the heart disease type distribution entropy value with the preset dynamic threshold to obtain a comparison result; If the heart disease type distribution entropy value is greater than the preset dynamic threshold, the case to be diagnosed and treated corresponding to the heart disease type distribution entropy value is included in the scheduling candidate queue.

4. The method according to claim 1, wherein The resource index and case complexity index corresponding to the scheduling candidate queue are obtained based on the diagnostic resources and the cardiac fusion feature, including: Based on the dual-dimensional monitoring strategy of resources and cases, the diagnostic resources of the device are obtained; the diagnostic resources include memory, computing power and network bandwidth; Calculating the resource index based on the diagnostic resources; Based on the feature space of the edge-end small model and according to the heart fusion feature, the complexity of the heart disease of the case to be diagnosed is constructed to obtain the case complexity index.

5. The method according to claim 4, characterized in that: The resource index is obtained by the following formula: R=∈1FLOPs avail +∈2Mem free Where R is the resource index; ∈1, ∈2 are resource weight coefficients; FLOPs avail The available computing power of the device; Mem free The free memory of the device; The case complexity index is obtained by the following formula: Wherein, C(x) is the case complexity index; Z s (s) is the cardiac fusion feature; is the average value of the training set features.

6. The method according to claim 1, characterized in that The joint scheduling rule table corresponds to the following methods: When the resource index is greater than or equal to a first preset threshold, and the case complexity index is less than or equal to a second preset threshold, the scheduling result is the edge small model diagnosis; When the resource index is less than the first preset threshold, or the case complexity index is greater than the second preset threshold, the scheduling result is the cloud-based large model diagnosis.

7. The method according to any one of claims 1 to 6, characterized in that The method further comprises: Obtaining the diagnosis result of the cloud-based large model diagnosis; updating the dynamic threshold according to the diagnosis result; Obtain a pseudo label according to the diagnosis result combined with the heart disease type distribution entropy value encoding of the edge small model; Based on the elastic weight solidification algorithm constraints, the parameters of the edge model are updated through incremental learning using the pseudo labels; The dynamic threshold is obtained by the following formula: Among them, θ new is the updated dynamic threshold; is the attenuation coefficient; θ ori To set the dynamic threshold; D hard A collection of difficult cases; H s (x) is the diagnosis result; The constraint loss function corresponding to the elastic weight solidification algorithm is obtained by the following formula: Among them, γ is the penalty coefficient; F i is the Fisher information matrix, which is used to measure the importance of model parameters; θ i is the current parameter value; θ i,old is the old parameter; i is the index of traversing the small model parameters.

8. A hierarchical scheduling diagnosis and treatment device based on the complexity of heart disease, characterized by: The device comprises: A preliminary screening module is used to use a small edge model to perform preliminary inference on the cardiac fusion features of the case to be diagnosed and treated, and obtain the entropy value of the distribution of cardiac disease types; A candidate scheduling module, configured to obtain a scheduling candidate queue based on a preset dynamic threshold and using the heart disease type distribution entropy value; A scheduling index module, configured to obtain a resource index and a case complexity index corresponding to the scheduling candidate queue based on diagnostic resources and the cardiac fusion feature; The hierarchical scheduling module is used to obtain a scheduling result based on the joint scheduling rule table, the resource index and the case complexity index.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images