A CKD clinical pathway recommendation method based on time-series dynamic graph clustering and multi-task joint optimization

Abstract

This invention relates to a CKD clinical pathway recommendation method based on temporal dynamic graph clustering and multi-task joint optimization, belonging to the field of medical big data analysis and smart healthcare technology. The method includes the following steps: acquiring multi-source clinical data; introducing an exponential decay function conforming to the forgetting curve to calculate temporal weights; constructing a clinical diagnosis and treatment interaction graph with time attributes; constructing a two-stream parallel neural network; extracting the static attribute embeddings and dynamic structure embeddings containing temporal dimensions of nodes and performing weighted fusion; subsequently constructing a joint loss function including the main task of recommendation ranking, macro-clustering, and micro-alignment auxiliary tasks; and performing multi-task end-to-end joint optimization of the model; finally, calculating the node embedding scores and selecting Top-K to generate a clinical pathway recommendation list. This invention effectively filters outdated pathological feature noise and significantly improves the accuracy and robustness of the auxiliary diagnosis and treatment system under sparse medical data.

Description

Technical Field

[0001] This invention belongs to the field of medical big data analysis and smart healthcare technology, and relates to a CKD clinical pathway recommendation method based on time-series dynamic graph clustering and multi-task joint optimization. Background Technology

[0002] With the explosive growth of smart healthcare and hospital informatization, electronic medical record systems and chronic disease management platforms have accumulated massive amounts of CKD patient diagnosis and treatment data. How to accurately capture the patterns of patient disease progression and recommend the most appropriate clinical pathways (including examinations, medications, and treatment plans) to doctors in a highly complex and heterogeneous clinical data environment has become a core key to improving treatment outcomes, slowing the progression of kidney disease, and optimizing the allocation of medical resources.

[0003] Traditional clinical pathway recommendation algorithms mainly include methods based on clinical guideline rules and methods based on statistical machine learning. In recent years, with the development of deep learning, graph neural networks have gradually become the mainstream technology in the field of intelligent assisted diagnosis and treatment due to their ability to effectively handle non-Euclidean medical knowledge graphs and capture the complex connectivity between patients and treatment events. Existing graph-based recommendation methods typically model patients and treatment items (such as medications, examinations, and surgeries) as nodes, and historical treatment behaviors as edges, aggregating neighbor features through graph convolution operations to learn the patient's health status embedding.

[0004] However, despite the achievements of existing graph neural network recommendation methods in assisting diagnosis and treatment, they still face the following key technical bottlenecks that urgently need to be addressed in practical CKD chronic disease management applications: Static graph structures fail to capture the timeliness of disease progression: Most existing graph recommendation algorithms, when constructing treatment interaction graphs, typically only statically distinguish between "treated" and "untreated" clinical events, or simply use frequency as weight. This approach ignores the physical time dimension of clinical events, treating a patient's baseline examination a year ago as equivalent to yesterday's acute exacerbation record. According to the progression patterns of CKD and the principle of clinical timeliness, a patient's condition dynamically evolves over time. Long-standing historical treatment records often contain outdated pathological noise. For example, a patient who was initially in CKD stage 1-2 requiring only dietary control may have progressed to CKD stage 3-4 requiring drug intervention or even dialysis preparation. Without this distinction, the recommended clinical pathways lack timeliness and fail to accurately match the patient's current treatment window.

[0005] There is a mismatch between patient subtyping clustering and treatment recommendation goals. To address the sparsity of clinical data (such as rare complications or newly enrolled patients), some studies have attempted to introduce patient clustering tasks as an aid, trying to group patients with similar conditions into the same subtype community to enhance feature representation. However, existing deep attribute graph clustering methods typically only aim to optimize the cluster distribution. While this unsupervised clustering goal can reveal the macroscopic structure of disease subtypes, it is not entirely aligned with the core goal of clinical recommendation systems: ranking optimal treatment options. Optimizing clustering alone often results in models learning features that excel at "disease subtyping" but not at "precise decision-making," failing to directly maximize the score gap between effective and ineffective treatment options, thus limiting the accuracy of clinical pathway recommendations.

[0006] Limitations of single-view feature extraction: While existing two-stream network architectures attempt to combine attribute and structural features, their structural flow often relies on manually defined static adjacency matrices or adaptively learned matrices. When faced with the common problems of few samples and chaotic distributions in clinical settings, simply relying on static topological structures can easily lead to oversmoothing, making it difficult to extract stable patient information representations from sparse data.

[0007] In summary, there is an urgent need for a novel recommendation method that can integrate physical time decay mechanisms to construct dynamic maps and simultaneously take into account both macroscopic population clustering structure and microscopic personalized ranking optimization, in order to improve the timeliness and accuracy of CKD clinical pathway recommendation systems. Summary of the Invention

[0008] In view of this, the purpose of this invention is to provide a CKD clinical pathway recommendation method based on temporal dynamic graph clustering and multi-task joint optimization, so as to solve the problem of the time lag in existing graph neural network recommendation technology where static graph structures cannot capture the evolution of patients' conditions, and the technical problem that the feature representation cannot directly serve the accuracy of Top-K clinical pathway recommendation due to the inconsistency between the traditional unsupervised clustering target and the diagnosis and treatment recommendation ranking target.

[0009] To achieve the above objectives, the present invention provides the following technical solution: A CKD clinical pathway recommendation method based on temporal dynamic graph clustering and multi-task joint optimization includes the following steps: S1: Acquire multi-source clinical data in smart healthcare scenarios, calculate the time interval characteristics of the treatment interaction edges based on the physical timestamps in the historical treatment flow, and introduce an exponential decay function that conforms to the forgetting law to map the time interval into the temporal weight of the edge, and construct a clinical treatment interaction graph with time attributes and a temporal weighted adjacency matrix. S2: Construct a two-stream parallel neural network architecture that includes attribute encoding paths and temporal structure encoding paths, extract the static attribute embeddings of nodes and the dynamic structure embeddings that contain temporal dimensions respectively, and then weight and fuse the two to obtain the final node embedding representation. S3: Construct a joint loss function that includes the main task of recommendation ranking, the auxiliary task of macro clustering, and the auxiliary task of micro alignment, and perform end-to-end multi-task joint optimization training on the network model; S4: After the model training is completed, the learned target patient node embedding and candidate treatment item embedding are used to calculate the score, filter historical items and select Top-K to generate a clinical pathway recommendation list.

[0010] Furthermore, in step S1, the multi-source clinical data includes patient baseline information, treatment item attribute information, and patient-treatment item historical interaction log data; the elements in the time-series weighted adjacency matrix The calculation formula is:

[0011] in, This is a preset time decay coefficient used to adjust the decay rate of the impact of historical interaction behavior on the current moment; For the current time and node ,node The time difference between interaction timestamps.

[0012] Furthermore, in step S2, the specific execution process of the dual-stream architecture is as follows: in the attribute encoding path, an autoencoder constructed using a multilayer perceptron extracts the latent low-dimensional static attribute embeddings of the nodes. And reconstruct the loss by reconstructing the computational properties through the decoder. In the temporal structure encoding path, a graph convolutional network is used, based on the temporal weighted adjacency matrix. With the corresponding time-series weighted degree matrix Perform normalized convolution operation, the first The feature propagation formula for a layer is:

[0013] in, To add the weighted matrix after adding self-loops, Learnable weights, thereby generating dynamic structural embeddings. .

[0014] Furthermore, in step S2, the static attribute embedding and the dynamic structure embedding are weighted and fused to obtain the final node embedding representation. The formula is:

[0015] in, This is the introduced fusion coefficient.

[0016] Furthermore, in S3, the main task of constructing the recommendation ranking is specifically as follows: constructing positive and negative sample triples based on the interaction graph, and calculating the Bayesian personalized ranking loss. To maximize the difference in the system's suitability scores for positive and negative samples.

[0017] Furthermore, in step S3, the construction of the macro-clustering auxiliary task specifically involves: constructing a dynamic modularity matrix based on the time-series weighted adjacency matrix. Its element calculation formula is:

[0018] in, For nodes The time-weighted degree, The sum of temporal weights for the entire graph; by minimizing based on macrostructural losses Nodes with recent frequent interaction patterns are forced to cluster into the same community.

[0019] Furthermore, in step S3, the micro-alignment auxiliary task is constructed as follows: The node with the highest temporal weighting degree is dynamically selected as the landmark in each community; the similarity distribution between the node and the landmark is calculated using the Student-t distribution; a high-confidence target distribution is constructed; and the KL divergence loss is minimized. Sharpen and align the feature distribution.

[0020] Furthermore, in step S3, the total loss function of multi-task joint optimization... The calculation formula is:

[0021] in, For balance coefficient, The attribute reconstruction loss is used; this function is minimized using the backpropagation algorithm to update the network parameters.

[0022] Furthermore, step S4 specifically includes: calculating the inner product score of the node embedding of the target patient and the candidate treatment item embedding and sorting them in descending order; using a hash table to filter out treatment items that have been performed in the patient's recent history and are not suitable for repetition in the short term; and selecting the top K most suitable items to generate a clinical pathway recommendation list.

[0023] The beneficial effects of this invention are as follows: 1. This invention constructs a time-weighted dynamic graph by introducing an exponential decay mechanism, which can adaptively reduce the weight of long-standing medical history records, effectively filter out outdated pathological feature noise, and make the features aggregated by the graph neural network more reflective of the patient's current disease progression status, thus solving the problem of time lag in traditional static graph recommendation methods in chronic disease management.

[0024] 2. This invention employs a multi-task joint training framework that primarily uses Bayesian ranking and secondarily uses deep patient clustering. It directly optimizes the quality of the clinical pathway recommendation list by utilizing ranking loss, thus solving the problem of inconsistency between the clustering objective (disease classification) and the recommendation ranking objective (i.e., treatment decision) in traditional deep graph clustering methods. At the same time, it uses the clustering task to regularize the feature space, effectively alleviating the overfitting phenomenon under sparse medical data and significantly improving the accuracy and robustness of the auxiliary diagnosis and treatment system.

[0025] 3. This invention utilizes a dual-stream architecture to extract static baseline features and dynamic diagnostic and treatment trajectory features of patients in parallel, and employs an adaptive weighting mechanism for fusion, overcoming the limitation of a single model in simultaneously considering patient demographic information and complex disease course topology. By introducing the Student-t distribution kernel function in the selection and soft assignment process of typical case landmarks, the robustness of the model to abnormal noise in EHR data is enhanced, ensuring the accuracy of patient profiling modeling in complex clinical environments.

[0026] 4. In terms of landmark node selection strategy, this invention adopts a dynamic update mechanism based on modularity contribution, which significantly reduces the computational complexity in the micro-distribution alignment process. By suppressing the weights of long-standing medical records through temporal decay, a sparse dynamic graph structure is naturally formed, reducing the computational load of propagating invalid information and improving the efficiency of the algorithm in training and inference on large-scale medical datasets.

[0027] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0028] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is an overall flowchart of the CKD clinical pathway recommendation method provided in the embodiments of the present invention; Figure 2 This is a schematic diagram of the overall network architecture of the recommendation system provided in an embodiment of the present invention; Figure 3This is a schematic diagram of the dual-stream dynamic feature extraction network structure provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the multi-task joint optimization module provided in an embodiment of the present invention. Detailed Implementation

[0029] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0030] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0031] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0032] Example 1: This invention provides a CKD clinical pathway recommendation method based on temporal dynamic graph clustering and multi-task joint optimization, comprising the following steps: S1. Construct a clinical diagnosis and treatment interaction graph with time attributes.

[0033] Acquire multi-source clinical data in smart healthcare scenarios, including patient baseline information (such as demographic characteristics and medical history), treatment item attribute information (such as drug classification and examination item characteristics), and historical patient-treatment item interaction log data (i.e., electronic medical record (EHR) records). Based on the physical timestamps in the historical treatment log, calculate the time interval characteristics of each treatment interaction edge. By setting the current moment in the clinical decision support system, an exponential decay function conforming to the forgetting curve (or the time-sensitivity of the disease) is introduced to map the time interval to the temporal weights of the edges, thus constructing a temporally weighted adjacency matrix. The elements in the time-weighted adjacency matrix The calculation formula is:

[0034] in, This is a preset time decay coefficient used to adjust the decay rate of the impact of historical interaction behavior on the current moment; For the current time and node ,node The time difference between interactive timestamps. This formula transforms the static topology into a dynamic graph structure that reflects the immediate intensity of a patient's condition.

[0035] S2. Dynamic feature extraction based on a two-stream architecture.

[0036] Construct a parallel neural network architecture that includes attribute encoding paths and temporal structure encoding paths.

[0037] In the attribute encoding path, the static attribute embedding of the node is extracted using an autoencoder; In the temporal structure encoding path, graph convolutional networks are used for feature aggregation. Unlike traditional methods, this path is based on the temporal weighted adjacency matrix generated in step S1. With the corresponding time-series weighted degree matrix Perform normalized convolution operation, the first The feature propagation formula for a layer is:

[0038] in, To add the weighted matrix after adding self-loops, These are learnable weights. This formula shows that neighboring nodes with closer interaction times contribute more to feature propagation, thus generating dynamic structural embeddings that contain temporal dimensions.

[0039] Finally, the static attribute embedding and the dynamic structure embedding are weighted and fused to obtain the final node embedding representation.

[0040] S3. Model training based on multi-task joint optimization.

[0041] A joint loss function is constructed that includes the main task of recommendation ranking, the auxiliary task of macro-clustering, and the auxiliary task of micro-alignment, and the network is optimized end-to-end.

[0042] Specifically, it includes: (1) Construct the main task of recommendation ranking: Construct positive and negative sample triples based on the interaction graph and calculate the Bayesian Personalized Ranking (BPR) loss. The aim is to maximize the difference in the system's suitability scores for positive and negative samples; (2) Constructing a macro-clustering auxiliary task: Constructing a dynamic modularity matrix based on a time-weighted adjacency matrix Its element calculation formula is:

[0043] in, For nodes The time-weighted degree, This represents the sum of temporal weights for the entire graph. By minimizing the weights based on... macrostructural losses Nodes with recent frequent interaction patterns are forced to cluster into the same community; (3) Construct a micro-alignment auxiliary task: Dynamically select the node with the highest temporal weighting degree in each community as the landmark, and minimize the KL divergence loss. Sharpen and align the feature distribution; (4) Joint optimization: Constructing the total loss function The function is minimized using the backpropagation algorithm to update the network parameters:

[0044] in, For balance coefficient, The loss is attribute reconstruction.

[0045] S4. Generate a recommendation list.

[0046] After model training is complete, online inference is performed using the learned final node embeddings. The inner product score of the target patient's node embedding and the candidate treatment item embeddings is calculated, and the nodes are sorted in descending order of score. The patient's historical treatment items are filtered, and the Top-K most suitable items for the next step of treatment are selected to generate a clinical pathway recommendation list.

[0047] Example 2: like Figure 1-4 As shown, this embodiment provides a CKD clinical pathway recommendation method based on temporal dynamic graph clustering and multi-task joint optimization, including core steps such as multi-source data preprocessing, temporal weighted graph construction, dual-stream feature extraction, multi-task joint optimization, and recommendation list generation.

[0048] Step S1: Construct a clinical diagnosis and treatment interaction map with time attributes.

[0049] This step aims to extract temporal features from the raw business data and transform the static topology into a dynamic structure that reflects the real-time evolution of the patient's condition. The specific implementation process is as follows: S1.1 Obtain multi-source data. Obtain the following three types of raw clinical data from the hospital's electronic medical record system and chronic disease management platform: Patient information set: defined as ,in This represents the total number of patients. Static attributes of each patient are extracted (e.g., patient ID, age, gender, medical history, complications, etc.).

[0050] Medical service item attribute set: defined as ,in This represents the total number of medical services provided. Extract the attribute characteristics of each medical service (such as service number, drug category, examination type, treatment method, etc.).

[0051] Historical medical record data: Extracting the interaction records between patients and medical procedures, each record is defined as a quadruple. .in, The type indicates the type of clinical event (such as prescribing, performing examinations, or conducting surgery). This indicates the physical time when the medical treatment occurred.

[0052] S1.2 Preprocessing of relevant data. Traversing the interactive pipeline data, to ensure the connectivity of the graph structure and the stability of training, an interaction threshold is set. (This example selects) (Excluding cases with fewer than 100 visits) Occasional patients and extremely rare treatment procedures.

[0053] For discrete non-numerical features (such as "examination classification: complete blood count"), One-Hot Encoding is used to map them into high-dimensional sparse vectors; for numerical clinical laboratory indicators, normalization is used to map them to... Intervals. The processed patient features are concatenated or aligned with the diagnostic and treatment features to construct an attribute feature matrix for all nodes in the graph. ,in For feature dimensions.

[0054] S1.3 Construction of a time-weighted graph based on the timeliness of the patient's condition. Setting the current reference time for the clinical decision support system. For any patient node with diagnostic and treatment interaction. Related to diagnosis and treatment project nodes Calculate their interaction time interval An exponential decay function conforming to the disease progression pattern is introduced to calculate the temporal weights of edges, constructing a temporally weighted adjacency matrix. The first in the matrix Line 1 The formula for calculating column elements is:

[0055] in, This is a time decay coefficient, used to adjust the decay rate of the influence of historical treatment behaviors on the current patient's condition. In this embodiment... The range of values ​​for is selected as follows This formula assigns a weight close to 1 to recent clinical events and a weight close to 0 to more distant historical medical records, effectively filtering outdated pathological noise and making the constructed dynamic graph structure more reflective of the patient's immediate condition. Calculate the time-series weighted matrix Its diagonal elements This is used for normalization in subsequent graph convolutions.

[0056] Step S2: Dynamic feature extraction based on a two-stream architecture.

[0057] This step involves constructing a parallel neural network architecture to extract the static semantic features and dynamic temporal structural features of the nodes.

[0058] S2.1 Static Attribute Feature Encoding. An autoencoder module is constructed using a multilayer perceptron. The attribute feature matrix is ​​then encoded. The input encoder extracts latent low-dimensional attribute embeddings through nonlinear transformation. . No. The formula for layer calculation is:

[0059] in, It is the ReLU activation function. and These are learnable parameters.

[0060] Will Input decoder reconstructs attribute matrix And calculate the attribute reconstruction loss. :

[0061] By minimizing this loss, we ensure that the embedding vector retains the key semantic information of the original data.

[0062] S2.2 Dynamic Structural Feature Encoding. A temporal structural encoding module is constructed using a Graph Convolutional Network (GCN). Unlike existing techniques that use static binary matrices, this step utilizes the temporal weighted matrix generated in step S1 for feature aggregation.

[0063] attribute matrix With time-weighted adjacency matrix Enter GCN. (The following is a list of characters and their meanings, which are not translated as they are not part of the main text.) The feature propagation formula for a layer is:

[0064] in, This is the weighted matrix after adding self-loops. Through this convolution operation, neighboring nodes with closer interaction times contribute more to feature aggregation, thus improving the structural embedding of the output. It can automatically filter out outdated historical behaviors as noise, retaining only the features that best represent the patient's current condition.

[0065] S2.3 Feature Fusion. Introducing a fusion coefficient. (This embodiment selects) ), static attribute features With dynamic structural features Weighted fusion is performed to obtain the final node embedding representation. :

[0066] Step S3: Model training based on multi-task joint optimization.

[0067] This step optimizes the network end-to-end by constructing a joint loss function that includes the main task of recommendation and ranking and the auxiliary task of dual clustering, in order to solve the problem of inconsistency between traditional clustering objectives and recommendation objectives.

[0068] S3.1 Construct the main recommendation ranking task. For the Top-K recommendation scenario, construct a Bayesian personalized ranking task.

[0069] Construct a training triplet set from the interaction graph .in Positive samples (i.e.) ), For negative samples randomly selected through a negative sampling strategy (i.e. Ranking loss calculation: based on node embedding. Calculate the predicted score and minimize the BPR ranking loss. :

[0070] The loss is designed to maximize the difference between the system’s suitability score for positive sample treatments and its score for negative samples.

[0071] S3.2 Constructing a Clustering Auxiliary Task. To discover macroscopic population structure in the feature space, a clustering task based on dynamic modularity is introduced. The embeddings... Inputting the Softmax layer yields the node's membership level. Probability distribution matrix of potential communities Construct a modularity matrix that reflects the temporal weights. Its elements are calculated as follows:

[0072] in For nodes The time-weighted degree (directly taken from step S1) ), The total temporal weights for the entire graph. Structural loss calculation: Define and minimize the macroscopic structural loss. :

[0073] S3.3 Constructing a micro-alignment auxiliary task. To address the feature divergence problem under sparse data, a landmark-based distribution alignment mechanism is introduced.

[0074] Within each cluster, the node with the highest temporal weighted degree is selected as the "landmark node". The similarity distribution between all nodes and landmarks is calculated using the Student-t distribution. And construct a high-confidence target distribution. .calculate and The KL divergence between them serves as the micro-alignment loss. :

[0075] S3.4 Constructing the total loss function The above losses are weighted and combined:

[0076] in, This is the balance coefficient.

[0077] Backpropagation is performed using the Adam optimizer to synchronously update encoder parameters and cluster centers until the model converges.

[0078] Step S4: Generate a recommendation list.

[0079] After the model training is complete, execute online assisted diagnostic reasoning: input the target patient. The real-time attributes are extracted using a trained network to obtain the final patient embedding. And obtain the embedding matrix of all candidate treatment items. Calculate the inner product score vector between the patient and the candidate treatment items. .

[0080] Sort the scores in descending order and use a hash table to filter out treatments that have been performed in the patient's recent history and are not suitable for repetition in the short term. Select the top-ranked scores. Individual medical procedures (such as) A list of recommended clinical pathways is generated and pushed to the doctor's decision support client.

[0081] Example 3: An electronic device, comprising a memory and a processor; The memory is used to store computer programs; The processor is configured to implement the method described in Embodiment 1 when executing the computer program.

[0082] Example 4: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in Embodiment 1.

[0083] Example 5: A computer program product includes a computer program that, when executed by a processor, implements the method described in Example 1.

[0084] In the above embodiments, the reference to "this embodiment" in the specification indicates that a specific feature, structure, or characteristic described in connection with the embodiment is included in at least some embodiments, but not necessarily all embodiments. Multiple appearances of "this embodiment" do not necessarily refer to the same embodiment.

[0085] In the above embodiments, although the invention has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory structures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed. The embodiments of the invention are intended to cover all such substitutions, modifications, and variations falling within the broad scope of the appended claims.

[0086] As will be understood by those skilled in the art, the computer-readable storage medium described in this embodiment allows for the implementation of all or part of the steps in the above method embodiments by computer program-related hardware. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0087] The electronic terminal provided in this embodiment includes a processor, a memory, a transceiver, and a communication interface. The memory and the communication interface are connected to the processor and the transceiver and complete communication between them. The memory is used to store computer programs, the communication interface is used to perform communication, and the processor and the transceiver are used to run the computer programs, so that the electronic terminal performs the steps of the above method.

[0088] In this embodiment, the memory may include random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device.

[0089] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0090] This invention can be used in a wide range of general-purpose or special-purpose computing system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0091] This invention can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This invention can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A CKD clinical pathway recommendation method based on temporal dynamic graph clustering and multi-task joint optimization, characterized in that: Includes the following steps: S1: Acquire multi-source clinical data in smart healthcare scenarios, calculate the time interval characteristics of the treatment interaction edges based on the physical timestamps in the historical treatment flow, and introduce an exponential decay function that conforms to the forgetting law to map the time interval into the temporal weight of the edge, and construct a clinical treatment interaction graph with time attributes and a temporal weighted adjacency matrix. S2: Construct a two-stream parallel neural network architecture that includes attribute encoding paths and temporal structure encoding paths, extract the static attribute embeddings of nodes and the dynamic structure embeddings that contain temporal dimensions respectively, and then weight and fuse the two to obtain the final node embedding representation. S3: Construct a joint loss function that includes the main task of recommendation ranking, the auxiliary task of macro clustering, and the auxiliary task of micro alignment, and perform end-to-end multi-task joint optimization training on the network model; S4: After the model training is completed, the learned target patient node embedding and candidate treatment item embedding are used to calculate the score, filter historical items and select Top-K to generate a clinical pathway recommendation list.

2. The CKD clinical pathway recommendation method based on temporal dynamic graph clustering and multi-task joint optimization according to claim 1, characterized in that: In step S1, the multi-source clinical data includes patient baseline information, treatment item attribute information, and patient-treatment item historical interaction log data; the elements in the time-series weighted adjacency matrix The calculation formula is: in, This is a preset time decay coefficient used to adjust the decay rate of the impact of historical interaction behavior on the current moment; For the current time and node ,node The time difference between interaction timestamps.

3. The CKD clinical pathway recommendation method based on temporal dynamic graph clustering and multi-task joint optimization according to claim 1, characterized in that: In step S2, the specific execution process of the dual-stream architecture is as follows: in the attribute encoding path, an autoencoder constructed using a multilayer perceptron extracts the latent low-dimensional static attribute embeddings of the nodes. And reconstruct the loss by reconstructing the computational properties through the decoder. In the temporal structure encoding path, a graph convolutional network is used, based on the temporal weighted adjacency matrix. With the corresponding time-series weighted degree matrix Perform normalized convolution operation, the first The feature propagation formula for a layer is: in, To add the weighted matrix after adding self-loops, Learnable weights, thereby generating dynamic structural embeddings. .

4. The CKD clinical pathway recommendation method based on temporal dynamic graph clustering and multi-task joint optimization according to claim 1, characterized in that: In step S2, the static attribute embedding and the dynamic structure embedding are weighted and fused to obtain the final node embedding representation. The formula is: in, This is the introduced fusion coefficient.

5. The CKD clinical pathway recommendation method based on temporal dynamic graph clustering and multi-task joint optimization according to claim 1, characterized in that: In S3, the main task of constructing the recommendation ranking is specifically as follows: constructing positive and negative sample triples based on the interaction graph, and calculating the Bayesian personalized ranking loss. To maximize the difference in the system's suitability scores for positive and negative samples.

6. The CKD clinical pathway recommendation method based on temporal dynamic graph clustering and multi-task joint optimization according to claim 1, characterized in that: In step S3, the construction of the macro-clustering auxiliary task specifically involves: constructing a dynamic modularity matrix based on the temporally weighted adjacency matrix. Its element calculation formula is: in, For nodes The time-weighted degree, The sum of temporal weights for the entire graph; by minimizing based on macrostructural losses Nodes with recent frequent interaction patterns are forced to cluster into the same community.

7. The CKD clinical pathway recommendation method based on temporal dynamic graph clustering and multi-task joint optimization according to claim 1, characterized in that: In step S3, the micro-alignment auxiliary task is constructed as follows: The node with the highest temporal weighting degree is dynamically selected as the landmark in each community; the similarity distribution between the node and the landmark is calculated using the Student-t distribution; a high-confidence target distribution is constructed; and the KL divergence loss is minimized. Sharpen and align the feature distribution.

8. The CKD clinical pathway recommendation method based on temporal dynamic graph clustering and multi-task joint optimization according to claim 1, characterized in that: In step S3, the total loss function of multi-task joint optimization The calculation formula is: in, For balance coefficient, The attribute reconstruction loss is used; this function is minimized using the backpropagation algorithm to update the network parameters.

9. The CKD clinical pathway recommendation method based on temporal dynamic graph clustering and multi-task joint optimization according to claim 1, characterized in that: Step S4 specifically includes: calculating the inner product score of the node embedding of the target patient and the candidate treatment item embedding and sorting them in descending order; using a hash table to filter out treatment items that have been performed in the patient's recent history and are not suitable for repetition in the short term; and selecting the top K most suitable items to generate a clinical pathway recommendation list.

Images