Construction method of medical examination recommendation model based on progressive spatial-temporal feature fusion

Abstract

The invention discloses a construction method of a medical examination recommendation model based on progressive spatial-temporal feature fusion, and the method comprises the steps: obtaining the sequence data of a plurality of patients, and constructing a training set after determining a training label; constructing a global collaborative heterogeneous graph, and generating heterogeneous sub-graphs by using a task adaptive denoising strategy based on a diffusion model; processing by using a graph encoder to obtain global collaborative spatial features of the patient and global collaborative spatial features of the project; re-represented patient training sequence data representation is obtained on the basis of the training sequence data representation, and a KANsformer sequence encoder module is used for processing to obtain patient space-time fusion features; processing the patient spatio-temporal fusion features and the patient global collaborative spatial features of the patient by using a cross attention spatio-temporal feature fusion module to obtain patient deep spatio-temporal fusion features; a check prediction vector is obtained; and obtaining a trained medical examination recommendation model based on the examination prediction vectors of all patients and the training label training model, so that the recommendation accuracy can be improved.

Description

Technical Field

[0001] The present invention belongs to the field of medical examination recommendation, and particularly relates to a method for constructing a medical examination recommendation model based on progressive spatio-temporal feature fusion. Background Art

[0002] Medical examination recommendation, as a core method for implementing health risk assessment based on patients' electronic medical records, aims to provide customized examination plans according to the individual characteristics of each patient to ensure that the most appropriate medical examination items can be carried out in the subsequent stage. This process not only depends on the algorithm's accurate prediction ability of health risks, but also needs to integrate multi-dimensional information such as the patient's age, gender, medical history, genetic background, etc., so as to achieve highly personalized recommendations. An efficient medical examination recommendation method can assist doctors in optimizing patients' examination plans by deeply integrating and analyzing these multiple information, ensuring that necessary medical examinations are carried out at the best time. This precise recommendation mechanism not only effectively avoids the risks of over-examination or missed diagnosis, but also significantly improves the treatment effect and health prognosis of patients.

[0003] In clinical practice, comprehensively considering the patient's medical record data is a key step in evaluating their health status. This includes, but is not limited to, considering the patient's past medical history, current symptom manifestations, and various medical examination information received before. These information together constitute the basis for diagnosis, helping doctors understand the evolution process of diseases, identify potential health risks, formulate personalized treatment plans, and predict possible disease progress. In addition, the patient's personal attribute information is also important for their personalized health status. In the actual diagnosis and treatment process, patients of different ages or genders may show the same symptoms, but their disease risks may vary greatly. For example, relevant statistics show that the probability of middle-aged and elderly groups suffering from acute myocardial infarction is much greater than that of teenagers, and the probability of young women suffering from malignant tumors is much greater than that of young men. Therefore, in the medical examination recommendation task, it is necessary to consider heterogeneous data such as the patient's past medical history, symptoms, examinations, as well as their age and gender at the same time.

[0004] At the same time, the temporal relationship between heterogeneous entities during the patient interaction process can effectively depict the dynamic evolution process of the patient's condition. This temporal information not only reveals the change trend of the patient's health status, but also provides an important basis for the assessment of the disease progress. By modeling these temporal relationships, the key change nodes of the patient's condition can be captured more accurately, thus providing more reliable support for medical decision-making.

[0005] In existing medical examination recommendation solutions, etc., the patent "Medical Examination Recommendation Method, Device, Equipment and Medium Based on Artificial Intelligence" (Application No.: CN202111011018.8) discloses a medical examination recommendation method, device, equipment and medium based on artificial intelligence. This method analyzes users' health information, evaluates disease risks, screens and weights medical examination items, scores based on historical physical examination data, and finally recommends appropriate auxiliary medical examinations to help users maintain their health. The patent "Medical Examination Recommendation System and Method Based on Explicit Topic Allocation Technology" (Application No.: CN202111129774.0) discloses a medical examination recommendation system and method based on explicit topic allocation technology. This system and method use historical cases and treatment plans to train the LLDA model. After inputting patient information, personalized medical examination recommendations and treatment plans are generated. This method breaks through the limitations of traditional hospitals relying on medical staff or unified processes, uses data-driven methods to provide decision-making references for doctors, and alleviates the problem of tight medical resources.

[0006] The patent "A Multimodal Recommendation Method Based on the Information Bottleneck Principle" (Application No.: CN202411620567.9) discloses a multimodal recommendation method based on the information bottleneck principle. This method includes a contrastive learning information bottleneck module. A loss function based on contrastive learning is designed in the module, aiming to optimize the overall performance of the recommendation system by maximizing the compression efficiency of modal information while reducing redundancy and noise. The patent "Recommendation Model with Adaptive Denoising and Interaction Enhancement Based on Contrastive Learning" (Application No.: CN202411581994.0) discloses a recommendation model with adaptive denoising and interaction enhancement based on contrastive learning. The core of this model lies in the design of an adaptive denoising module. This module significantly reduces the negative impact of noisy data on model training by detecting and eliminating noisy interactions where the predicted value deviates significantly from the true value. In addition, Yangqin Jiang et al. proposed a new recommendation model DiffKG in their published paper DiffKG: Knowledge Graph Diffusion Model for Recommendation. DiffKG uses a diffusion model to eliminate noisy data in the knowledge graph, thus preventing this noise from misleading subsequent recommendation tasks. However, this model implicitly guides the diffusion model at the feature level to achieve adaptive denoising, which not only introduces additional computational burdens but also limits the performance of denoising.

[0007] The patent "A method for classifying and identifying bone marrow cells based on the SCKansformer neural network" (Application No.: CN202410826656.2) discloses a method for classifying and identifying bone marrow cells based on the SCKansformer neural network. This method proposes a neural network called SCKansformer. The SCKansformer neural network combines Kolmogorov-Arnold Networks (KAN) with Transformer to effectively extract discriminative features in bone marrow cell images. In addition, Donghai Hu et al. proposed a new prediction model, CL-Kansformer, in their published paper "CL-Kansformer model for SOC prediction of hydrogen refueling process in fuel cell vehicles". This model improves the prediction accuracy by replacing the traditional multi-layer perceptron (MLP) with KAN in Transformer. However, both SCkansformer and CL-Kansformer only optimize the feed-forward neural network part in Transformer and ignore the core multi-head attention module, thus affecting some performance.

[0008] The patent "A Personalized Sequential Recommendation Method and System Based on Transformer for Multi-Behavior of Users" (Application No.: CN202411211531.5) discloses a personalized sequential recommendation method and system based on Transformer for multi-behavior of users. This method uses Transformer to perform multi-scale temporal modeling on the shopping behaviors of users, aiming to capture the dynamic changes in the multi-behavior interaction sequences of users, so as to extract the behavior patterns at multi-scale time intervals. Subsequently, these dynamic behavior patterns at multi-scale time intervals are integrated into a unified latent representation space, ultimately improving the accuracy of personalized recommendations. The patent "An Adversarial Training and Contrastive Learning Sequential Recommendation Method Based on Bidirectional Transformer" (Application No.: CN202410825189.1) discloses an adversarial training and contrastive learning sequential recommendation method based on bidirectional Transformer. This method uses a bidirectional Transformer model, through adversarial training and contrastive learning methods of policy gradients, to enable it to more accurately capture the relationship between user preferences and sequence context, improving the recommendation effect. The patent "An API Sequence Recommendation Method and Device Based on Retrieval Enhancement and Temperature Loss" (Application No.: CN202410825189.1) discloses an API sequence recommendation method and device based on retrieval enhancement and temperature loss. This method uses a pre-trained Transformer model as a sequence model to generate the output distribution of API sequences. Xue Li et al. proposed a new medication recommendation model Trans-GAHNet in their published paper "Transformer-based medication recommendation with a multiple graph augmentation strategy". This model uses a Transformer network architecture to comprehensively encode sequence electronic health record (EHR) data, thereby generating efficient patient representations.

[0009] However, the existing technologies currently have the following problems:

[0010] 1. The performance of existing medical examination recommendation methods is limited:

[0011] Medical examination recommendation data has both heterogeneity and temporality. However, existing medical examination recommendation methods lack specially designed modules to effectively model these characteristics, resulting in inaccurate assessment of the patient's health status and limited recommendation performance of the methods.

[0012] 2. The performance of existing denoising methods is limited:

[0013] Existing data denoising methods usually rely on contrastive learning to achieve denoising by leveraging complementary information between multiple views. However, this method is more suitable for multimodal data sources and cannot explicitly improve the performance of downstream recommendation tasks. At the same time, some denoising methods reduce the impact of noise on model training by directly discarding interactions where the predicted values ​​deviate significantly from the true values. However, this method has obvious limitations. It is overly dependent on the accuracy of the predicted values, lacks dynamic adjustment capabilities, and easily leads to model overfitting due to the discarding of some data. In addition, some denoising methods achieve adaptive denoising by implicitly guiding the denoising model at the feature level, which not only introduces additional computational burden, but also limits the performance of denoising.

[0014] 3. Existing sequence encoders have limited capabilities in complex sequence modeling:

[0015] Due to the complex heterogeneous relationships between entities in patient sequence data, traditional sequence encoders (such as RNN, Transformer, etc.) have limited modeling capabilities when processing such complex sequences.

[0016] In summary, considering the data characteristics in medical examination recommendation tasks, how to effectively model the heterogeneous and time series information in the data to accurately assess the patient's health status and improve the accuracy of recommendations is a technical problem that needs to be solved urgently in this field. Summary of the invention

[0017] In order to solve the above problems existing in the prior art, the present invention provides a method for constructing a medical examination recommendation model based on progressive spatiotemporal feature fusion and a medical examination recommendation method. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0018] In a first aspect, an embodiment of the present invention provides a method for constructing a medical examination recommendation model based on progressive spatiotemporal feature fusion, and the method for constructing a medical examination recommendation model based on progressive spatiotemporal feature fusion includes:

[0019] Acquire a number of patient sequence data; wherein the patient sequence data represents the diseases, symptoms and examinations of the patients in chronological order and has associated items, including the age and gender of the patients;

[0020] Determine an examination as a training label in each patient sequence data; use the data before the training label as the patient's training sequence data, thereby obtaining a training set consisting of the training sequence data of all patients; wherein the training sequence data of each patient retains associated items;

[0021] Construct a global collaborative heterogeneous graph of the training set using a heterogeneous graph aggregation module, and denoise the global collaborative heterogeneous graph using a task-adaptive denoising strategy based on a diffusion model to generate a heterogeneous subgraph that is beneficial to the downstream examination recommendation task; process the heterogeneous subgraph using a graph encoder to obtain global collaborative spatial features, including patient global collaborative spatial features and item global collaborative spatial features; the items include diseases, symptoms, examinations, age, and gender.

[0022] Based on all item global collaborative spatial features, obtain a re-represented patient training sequence data representation through positional encoding, and process each re-represented patient training sequence data representation using a KANsformer sequence encoder module to obtain corresponding patient spatio-temporal fusion features; wherein, the KANsformer sequence encoder module is obtained by integrating KANs into the Transformer framework.

[0023] For each patient, process the corresponding patient spatio-temporal fusion feature and patient global collaborative spatial feature using a cross-attention spatio-temporal feature fusion module to obtain patient deep spatio-temporal fusion features.

[0024] Based on each patient's deep spatio-temporal fusion feature, obtain an examination prediction vector for the corresponding training sequence data; and perform model training based on the examination prediction vectors of all patients and the corresponding training labels to obtain a trained medical examination recommendation model for recommending the required examinations for patient sequence data.

[0025] In a second aspect, an embodiment of the present invention provides a medical examination recommendation method, and the medical examination recommendation method includes:

[0026] Obtain patient sequence data to be measured.

[0027] Input the patient sequence data to be measured into a pre-constructed medical examination recommendation model to obtain a recommended examination; wherein, the pre-constructed medical examination recommendation model is obtained by the construction method of the medical examination recommendation model based on progressive spatio-temporal feature fusion in the first aspect.

[0028] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0029] 1) More efficient denoising: The present invention fully considers the noise problem in heterogeneous data, designs a task-adaptive denoising strategy based on a diffusion model, and realizes efficient denoising by virtue of the powerful denoising generation ability of the diffusion model and the explicit feedback of the loss function, thus providing a solid foundation for the graph neural network RGAT to extract high-quality spatial features.

[0030] 2) More accurate sequence data modeling: Facing complex sequence data composed of heterogeneous entities, the present invention proposes the KANsformer model. This model ingeniously combines the advantages of Transformer and KAN for modeling in the time dimension. Thanks to the powerful model expression ability of KAN, compared with the traditional Transformer model, KANsformer not only demonstrates more excellent performance in capturing the time features of complex sequence data, but also effectively realizes the implicit preliminary fusion of spatio-temporal features at the project level.

[0031] 3) More accurate recommendation results: The present invention fully considers the heterogeneous and temporal characteristics of medical examination recommendation data, and specifically designs a powerful graph encoder and sequence encoder, and effectively combines the two. Through the progressive spatio-temporal feature fusion strategy, accurate patient health status assessment is achieved, thereby further improving the accuracy of recommendation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a schematic flowchart of a construction method of a medical examination recommendation model based on progressive spatio-temporal feature fusion provided by an embodiment of the present invention;

[0033] Figure 2 It is a schematic diagram of the composition of the main modules adopted in the construction method of the medical examination recommendation model based on progressive spatio-temporal feature fusion provided by an embodiment of the present invention;

[0034] Figure 3 It is a schematic diagram of the principle of the construction method of the medical examination recommendation model based on progressive spatio-temporal feature fusion provided by an embodiment of the present invention;

[0035] Figure 4 It is an example diagram of the dataset division format of an embodiment of the present invention;

[0036] Figure 5 It is a schematic diagram of the processing process of the heterogeneous graph aggregation module of an embodiment of the present invention;

[0037] Figure 6 It is a schematic diagram of the structure of the global collaborative heterogeneous graph in an embodiment of the present invention;

[0038] Figure 7 It is a schematic diagram of the framework of the KANsformer sequence encoder module in an embodiment of the present invention;

[0039] Figure 8 It is an example diagram of the initialization representation of the input sequence data of the KANsformer sequence encoder module in an embodiment of the present invention;

[0040] Figure 9 It is a schematic diagram of the structure of the cross-attention spatio-temporal feature fusion module in an embodiment of the present invention;

[0041] Figure 10 This is a schematic flowchart of a medical examination recommendation method provided by an embodiment of the present invention. Detailed implementation manners

[0042] The following further describes the present invention in detail with reference to specific embodiments, but the implementation manners of the present invention are not limited thereto.

[0043] To solve the performance bottleneck in modeling patient time-series data and heterogeneous data in existing medical examination recommendation methods, the present invention provides a method for constructing a medical examination recommendation model based on progressive spatio-temporal feature fusion and a medical examination recommendation method.

[0044] In a first aspect, an embodiment of the present invention provides a method for constructing a medical examination recommendation model based on progressive spatio-temporal feature fusion, which is mainly completed by using three modules. Please refer to Figure 2 As shown, these three modules are respectively a heterogeneous graph aggregation module, a KANsformer sequence encoder module, and a cross-attention spatio-temporal feature fusion module. Among them, the heterogeneous graph aggregation module is mainly used to extract the global collaborative spatial features of patients and items; the KANsformer sequence encoder module is mainly used to implicitly fuse the spatio-temporal features of patients from the item level while modeling the patient time-series features; the cross-attention spatio-temporal feature fusion module is mainly used to deeply fuse the spatio-temporal features of patients.

[0045] The following is a specific description. Please refer to Figures 1 to 3 It can be understood that the method may include the following steps:

[0046] S1. Obtain a plurality of patient sequence data;

[0047] Among them, the patient sequence data represents the diseases, symptoms, and examinations that a patient has in chronological order and has associated items, including the age and gender of the patient to whom it belongs;

[0048] The patient sequence data can come from data such as historical cases. Please refer to Figure 4 As shown in the schematic diagram of the patient sequence data, the patient sequence data may include different time points t, the diseases, symptoms, and examinations that the patient has at these time points. For example, Figure 4 In the upper sequence, at time point t = 0, the disease of the patient is coronary heart disease; at time point t = 1, the examination is echocardiogram; at time point t = 2, there are symptoms; at time point t = 3, the examination is blood routine,...; at time point t = m, the examination is head CT,...; at the last time point t = T', the examination is head MRI. At the same time, each patient sequence data is associated with the age and gender of the patient as its associated items to represent the patient's personal information, but the associated items do not appear in the patient sequence data.

[0049] S2, determining an examination in each patient sequence data as a training label; using the data before the training label as the patient's training sequence data, thereby obtaining a training set consisting of the training sequence data of all patients;

[0050] In this embodiment of the present invention, the data set division follows the leave-one-out strategy in the sequence recommendation task. Figure 4 , for each patient, assume that the length of the patient sequence data of its historical interaction is T′, which contains n examinations. The penultimate examination ( Figure 4 ) as the test label and the penultimate examination ( Figure 4 In the example, “head CT” is used as the training label. All the sequence data of the patient’s sequence data before the training label, i.e. Figure 4 The part from time t=0 to t=ml will constitute the training sequence data of the patient, which will be used as the model input to learn the personalized representation of the patient. In addition, the training sequence data of each patient retains the associated items;

[0051] Through the above processing, the training sequence data of each patient can be obtained, and the associated items of the corresponding patient can be merged, and all the obtained data can form a training set.

[0052] Similarly, the test sequence data of each patient can also be obtained to form a test set, which can be used to subsequently test the trained medical examination recommendation model to achieve model effect evaluation, etc.

[0053] S3, constructing a global collaborative heterogeneous graph of the training set using a heterogeneous graph aggregation module, and denoising the global collaborative heterogeneous graph using a task adaptive denoising strategy based on a diffusion model to generate a heterogeneous subgraph that is beneficial to downstream examination recommendation tasks; processing the heterogeneous subgraph using a graph encoder to obtain global collaborative space features, including patient global collaborative space features and item global collaborative space features; the items include diseases, symptoms, examinations, age and gender;

[0054] Step S3 is completed by the heterogeneous graph aggregation module. The processing process can be found in Figure 5 Understanding can include three steps: building a global collaborative heterogeneous graph, a task-adaptive denoising strategy based on a diffusion model, and a relationship-aware graph attention network, which are explained below.

[0055] (I) Building a global collaborative heterogeneous graph

[0056] Explicitly modeling the dynamic collaboration signals between user sequences in the sequential recommendation scenario is of great significance for modeling user personalized features. By capturing the collaboration signals in the user behavior sequence (such as the behavior patterns of similar users or group preferences), the personalized needs and interest changes of users can be more accurately reflected. In the medical field, from the perspective of collaborative filtering, globally modeling the collaborative features among patients can not only enhance the prediction ability of the recommendation system, but also help identify the potential health risks or demand trends of patients, so as to provide more accurate and personalized recommendation services.

[0057] To this end, in the embodiments of the present invention, the training sequence data S = {s1,..., s U} of all patients and their associated items representing personal information are all connected. Since a large number of patient interactions have common entities, a huge global collaborative heterogeneous graph can be finally constructed, and its structure diagram is as Figure 6 shown. The global collaborative heterogeneous graph contains various heterogeneous entities and complex association relationships among them, involving the interaction information between patients and items, the collaborative information between patients and patients, the potential real association information between items and items, etc.

[0058] Specifically, the global collaborative heterogeneous graph of the training set constructed by using the heterogeneous graph aggregation module is represented as:

[0059]

[0060] Among them, represents the global collaborative heterogeneous graph, which contains 6 types of entity nodes, namely patient nodes, disease nodes, symptom nodes, examination nodes, age nodes and gender nodes. Among them, disease nodes, symptom nodes, examination nodes, age nodes and gender nodes are item nodes. Patient nodes are only connected to item nodes, and item nodes are only connected to patient nodes. Patient nodes are connected to each other through their common nodes; u represents a patient; U represents the set of patients; r represents a relationship edge; R represents the set of relationship edges; i represents an item; represents a disease symptom examination ε xam the overall set of these three item entities; represents the set of ages; represents the set of genders; ∪ represents the union.

[0061] (2) Task Adaptive Denoising Strategy Based on Diffusion Model

[0062] The global collaborative heterogeneous graph contains a large number of different types of entities and complex dependency relationships between entities. In fact, only a partial subset is truly relevant to the downstream inspection recommendation task, while other knowledge relationships will become noise, potentially misleading the model's understanding and assessment of the patient's health status. To address the above problems, the present invention designs a task-adaptive denoising strategy based on a diffusion model.

[0063] First, inspired by the excellent performance of the diffusion model in the denoising generation task, this strategy considers using the diffusion model to reconstruct the original graph to generate graph structure data beneficial to the downstream task. Its processing process can be expressed as:

[0064]

[0065] Among them, represents the global collaborative heterogeneous graph; represents the heterogeneous subgraph generated by the diffusion model; f diffsion (·) represents the reconstruction process of the diffusion model.

[0066] To make the generated heterogeneous subgraph have a strong correlation with the downstream recommendation task to achieve task-driven denoising, in the embodiments of the present invention, using the task-adaptive denoising strategy based on the diffusion model to denoise the global collaborative heterogeneous graph and generate a heterogeneous subgraph beneficial to the downstream inspection recommendation task includes:

[0067] By introducing a task-adaptive function during the training process of the diffusion model to guide the diffusion model to generate a heterogeneous subgraph beneficial to the downstream inspection recommendation task, which can be expressed by the formula:

[0068]

[0069]

[0070] Among them, L rec and L diffsion respectively represent the loss functions of the recommendation task and the diffusion model; L d ′ iffsion is the loss function of the diffusion model after being processed by the task-adaptive denoising strategy; represents the difference between the current training Epoch and the last Epoch of the global collaborative heterogeneous graph; represents the average change of L rec in the first δ epochs, where δ > 1 and ε < 1.

[0071] This learning strategy can dynamically adjust the training of the diffusion model according to the impact of the currently generated heterogeneous subgraph on the downstream inspection recommendation task. If It is explained that in the current Epoch, the existing heterogeneous subgraphs are beneficial to the training of downstream recommendation tasks, resulting in a faster decline rate of the loss function of downstream tasks. Therefore, ε < 1 is used to suppress the training of the diffusion model to retain the graph structure of the existing heterogeneous subgraphs. Conversely, the training of the diffusion model is resumed to guide it to reconstruct heterogeneous subgraphs adapted to downstream recommendation tasks.

[0072] (III) Relation-Aware Graph Attention Network

[0073] To make full use of the complex heterogeneous relationships in the heterogeneous subgraphs, the present invention can use a relation-aware graph attention network (RGAT) as a graph encoder. This encoder can effectively capture various relationships inherent in the global collaborative heterogeneous graph connection structure.

[0074] Specifically, the heterogeneous subgraphs are processed by the graph encoder to obtain global collaborative spatial features, including:

[0075] (1) Using the relation-aware graph attention network RGAT as the graph encoder and utilizing its message aggregation mechanism for knowledge aggregation to obtain entity representations of each layer; wherein, the message aggregation mechanism is expressed by the formula:

[0076]

[0077] wherein, the heterogeneous subgraphs are represented by ; in the knowledge aggregation process, N m represents all adjacent entity sets in the heterogeneous subgraphs where entity m is connected by relation r m,n ; the embeddings of entity m and n are respectively represented by α(m, r m,n , n) represents the attention correlation of entities and relations estimated in the knowledge aggregation process, which is used to capture the unique semantics of the relationship between entity m and n; the superscript Τ represents the transpose operation; W represents a learnable weight matrix; LeakyReLU(·) represents the LeakyReLU activation function; Drop(·) represents the Dropout function; Norm(·) represents the normalization process; exp(·) represents the natural exponential function; || represents the concatenation operation (i.e., Concat); represents the entity representation obtained by entity m through fusion at the l g layer of RGAT;

[0078] (2) Fusing the entity representations of each layer as the global collaborative spatial features; the global collaborative spatial features are represented as:

[0079]

[0080] Among them, represents the global collaborative spatial feature of any entity m in the heterogeneous subgraph, including the patient global collaborative spatial feature, denoted as and the project global collaborative spatial feature, denoted as d is the embedding dimension, is the set of patient global collaborative spatial features, is the set of project global collaborative spatial features; L g represents the number of layers of the RGAT.

[0081] Specifically, each patient has a corresponding patient global collaborative spatial feature For the 5 projects mentioned above, namely the specific project entities included in disease, symptom, examination, age, and gender, such as coronary heart disease and acute myocardial infarction, each has its corresponding project global collaborative spatial feature

[0082] It should be noted that in other embodiments, the graph encoder can be considered to be replaced by a Graph Neural Network (GNN), a Graph Convolutional Network (GCN), a Graph Attention Network (GAT), etc. to save computational costs.

[0083] S4. Based on all project global collaborative spatial features, obtain the re-represented patient training sequence data representation through positional encoding, and use the KANsformer sequence encoder module to process each re-represented patient training sequence data representation to obtain the corresponding patient spatio-temporal fusion feature; wherein, the KANsformer sequence encoder module is obtained by integrating KANs into the Transformer framework;

[0084] Referring to Figure 7 , on the basis of the global collaborative spatial feature, it is also necessary to use a sequence encoder to model the patient's training sequence data to extract the patient's temporal features. Transformers have been proven to perform excellently in modeling temporal data. Recently, Kolmogorov-Arnold Networks (KANs) enable neural networks to learn complex function mappings more efficiently, demonstrating their great potential in learning complex function mappings. Therefore, the present invention innovatively integrates KANs into the Transformer framework and names it KANsformer. Its detailed architecture is as Figure 7As shown, it can be seen that KANsformer replaces the feed-forward network in the Transformer framework with KANs and correspondingly modifies the multi-head attention to multi-head attention based on KANs. Experiments show that compared with traditional Transformer models, KANsformer demonstrates more excellent performance in learning the temporal and heterogeneous dependence features of complex sequence data.

[0085] In the embodiments of the present invention, before using the KANsformer sequence encoder module, it is necessary to first obtain the re-represented patient training sequence data representation.

[0086] Specifically, obtaining the re-represented patient training sequence data representation through positional encoding based on all-item global collaborative spatial features includes:

[0087] (1) Taking the set of item global collaborative spatial features as the initial representation of each patient training sequence entity;

[0088] (2) For the training sequence data of each patient, determining the position of the item data in the training sequence data of this patient in and obtaining the re-represented patient training sequence data representation of this patient by adding its position embedding representation p i to .

[0089] First, taking the set of item global collaborative spatial features learned by the graph encoder as the initial representation of all patient training sequence entities, as shown by the upper rectangular box in Figure 8 , which contains all the subdivided item global collaborative spatial features.

[0090] Then, for the training sequence data of each patient (as shown by the left rectangular box in Figure 8 ), by adding the position embedding representation p i to the global collaborative spatial representation of the item entity in the training sequence data, that is, using the position embedding representation p i to select the in that matches the patient training sequence data. Therefore, the re-represented patient training sequence data representation can be expressed as where represents the representation of the t-th item entity in the current patient training sequence data, and T is the current sequence length.

[0091] Among them, the KANsformer sequence encoder module adopts the core component of the Transformer architecture, the multi-head self-attention mechanism, and its formula is expressed as follows:

[0092]

[0093] Among them, represents the output of the l s -th layer of the multi-head attention module based on KANs in KANsformer; L s and N respectively represent the total number of multi-head self-attention blocks and the number of heads; A n represents the attention value of the n-th head; φ(·) is a single-layer Efficient KAN function; is the n-th Efficient KAN function corresponding to the query, key, and value in the attention mechanism, which is used for feature mapping transformation; in order to further learn the temporal features of the patient, corresponding to the feed-forward neural network in Transformer, the KANsformer sequence encoder module of the embodiment of the present invention uses a two-layer Efficient KAN function, where φ1 represents the first-layer Efficient KAN function and φ2 represents the second-layer Efficient KAN function; GELU represents the GELU activation function; represents the spatio-temporal fusion feature of the patient obtained by KANsformer; represents the output of the last layer of KANsformer, L s represents the last layer of KANsformer, and T in u T represents the length of the training sequence data of the patient.

[0094] Since the training sequence entity data of the patient uses the global collaborative feature as the initialization representation. Therefore, when KANsformer encodes the training sequence data of the patient, it can implicitly fuse the spatio-temporal dynamic information of the patient through feature interaction at the item level. Finally, referring to the common operation of the sequence recommendation model, the last feature vector of the model output sequence is used as the preliminarily integrated spatio-temporal fusion feature for subsequent calculation.

[0095] S5. For each patient, the corresponding patient spatio-temporal fusion feature and the patient global collaborative spatial feature are processed by the cross-attention spatio-temporal feature fusion module to obtain the patient deep spatio-temporal fusion feature;

[0096] The patient spatio-temporal fusion features extracted by the KANsformer sequence encoder module only implicitly represent the patient's global collaborative spatial features at the project level. However, in the patient node representation learned by the relation-aware graph attention network, the global collaborative spatial features among patients are explicitly modeled. Therefore, in order to make full use of the above two types of features, the present invention considers introducing a cross-attention mechanism and adopting a progressive feature fusion strategy to achieve further deep fusion and optimization at the feature level.

[0097] Previously, cross-attention has achieved good results in feature fusion tasks. Its main purpose is to use the self-attention mechanism to obtain context information from the query (Q) features that enhance the key-value (KV) features, thereby refining the final representation.

[0098] Specifically, S5 may include:

[0099] The cross-attention spatio-temporal feature fusion module converts the patient's global collaborative spatial features of this patient into query feature Q, and converts the patient's spatio-temporal fusion features of this patient into key feature K and value feature V, obtaining the patient's deep spatio-temporal fusion features

[0100] In the embodiment of the present invention, taking the patient's global collaborative spatial features as Q and the patient's spatio-temporal fusion features as K-V, the cross-attention mechanism is used to explicitly incorporate the patient's global spatial context collaborative information into the patient's spatio-temporal fusion features, realizing deep spatio-temporal semantic alignment to enrich the patient's personalized feature expression.

[0101] See Figure 9 Understand that first, the patient's global collaborative spatial features and the patient's spatio-temporal fusion features are respectively converted into corresponding Q, K, and V through multiple linear projection layers, expressed as:

[0102]

[0103] Among them, W Q , W K , W V respectively represent the mapping matrices for converting the features into the query, key, and value spaces. So far, the cross-attention (Cross-Attention, abbreviated as Cro_Attn) for feature fusion is calculated as follows:

[0104]

[0105]

[0106] d k is the embedding dimension of features in the query, key, and value spaces. is the finally learned patient's deep spatio-temporal fusion feature.

[0107] S6. Based on each patient's deep spatio-temporal fusion feature, obtain the inspection prediction vector of the corresponding training sequence data; and based on the inspection prediction vectors of all patients and the corresponding training labels, perform model training to obtain a trained medical inspection recommendation model for recommending the required inspections for patient sequence data.

[0108] Among them, the obtaining of the inspection prediction vector of the corresponding training sequence data based on each patient's deep spatio-temporal fusion feature includes:

[0109] For each patient's deep spatio-temporal fusion feature, use a scoring function and the patient's deep spatio-temporal fusion feature to obtain the inspection prediction vector of the patient's training sequence data, where the scoring function is expressed as:

[0110]

[0111] where the inspection (i.e., the given candidate inspection entity item) e ∈ ε xam ; ε xam represents the set of all inspection items; represents the global collaborative space feature for the inspection item; is the patient's deep spatio-temporal fusion feature of patient u; the inspection prediction vector represents the score vector of patient u for each candidate inspection item, where the higher the score of a certain inspection, the higher the probability that it is predicted as a recommended inspection.

[0112] Therefore, through the above processing, each patient's training sequence data can obtain a corresponding inspection prediction vector.

[0113] It can be understood that as the training sequence data of each patient, its training label represents the inspection that actually occurs at the next moment of the training sequence data, that is, the true value, while the inspection prediction vector represents the predicted value of the training sequence data.

[0114] Therefore, the performing of model training based on the inspection prediction vectors of all patients and the corresponding training labels to obtain a trained medical inspection recommendation model includes:

[0115] Based on the inspection prediction vectors of each patient and the corresponding training labels, use cross-loss for model training to obtain a trained medical inspection recommendation model.

[0116] Among them, the loss function is expressed as:

[0117]

[0118] Among them, y ue represents the one-hot encoded vector of the real inspection items for the next interaction of patient u, i.e., the training label; θ represents all model parameters; ||·||2 represents the L2 norm; λ is used to control the regularization strength.

[0119] When the loss function reaches convergence, the training ends and a trained medical examination recommendation model is obtained. The trained medical examination recommendation model can be used to recommend appropriate examination items for any patient sequence data.

[0120] The embodiment of the present invention provides a method for constructing a medical examination recommendation model based on progressive spatiotemporal feature fusion, aiming to solve the performance bottleneck of existing medical examination recommendation methods when modeling patient time series data and heterogeneous data. Specifically:

[0121] 1. Medical examination recommendation aims to use the heterogeneous data and their temporal relationships in the patient's historical interaction to model, evaluate the patient's current health status, and recommend appropriate medical examination items for the next stage. However, the current related methods have limited performance when processing data with both heterogeneous and temporal characteristics. To this end, the present invention proposes a method for constructing a medical examination recommendation model based on progressive spatiotemporal feature fusion. In the face of the heterogeneous and temporal characteristics of the data, the method of the present invention specifically designs a powerful graph encoder and sequence encoder. Through the progressive spatiotemporal feature fusion strategy, the advantages of the graph encoder and the sequence encoder are cleverly combined. With the help of KANsformer and cross-attention mechanism, spatiotemporal features are progressively integrated at the project and patient levels, respectively, and the spatiotemporal representation in the patient interaction data is fully integrated, thereby effectively utilizing the heterogeneous and temporal information of the data, realizing comprehensive modeling of patient data, and accurate patient health status assessment, thereby further improving the accuracy of the recommendation results.

[0122] 2. In the heterogeneous data of patient historical interactions, the existence of noisy data cannot be ignored. For example, some historical medical record interactions from a long time ago may have a misleading impact on the assessment of the patient's current health status. However, existing methods still have deficiencies in denoising heterogeneous data and adaptively improving the efficiency of downstream recommendation tasks. To make full use of the heterogeneous data in patient interaction information, the present invention starts from the perspective of collaborative filtering and globally models the collaborative features among patients. To this end, the present invention connects all the patient interaction entity data and their personal information entities. Since a large number of user interactions have common entities, a huge global collaborative heterogeneous graph can ultimately be constructed. However, the global collaborative heterogeneous graph contains a large number of different types of entities and complex dependencies between entities. In fact, only a partial subset is truly relevant to the downstream recommendation task, while other knowledge relationships will become noise, potentially misleading the model's understanding and assessment of the patient's health status, and denoising processing is required. To effectively remove the noisy data in the global collaborative heterogeneous graph, the present invention designs a task-adaptive denoising strategy based on the diffusion model. This strategy utilizes the excellent denoising and reconstruction capabilities of the diffusion model, combined with an explicit loss feedback mechanism, to directly generate graph structure data that is helpful for the downstream examination recommendation task, thereby providing a solid foundation for the graph neural network RGAT to extract high-quality spatial features.

[0123] 3. According to the characteristics of medical examination recommendation data, it is also necessary to mine the heterogeneous information between data while modeling the patient sequence data. However, existing sequence encoders have limited performance in modeling the above complex sequences. To fully mine the temporal information in patient interaction data, the present invention considers arranging the interaction data into sequence data in chronological order and modeling it using a sequence encoder. To this end, inspired by the powerful capabilities demonstrated by KANs in learning complex function mappings, the present invention integrates KAN into the Transformer framework to construct a KANsformer sequence encoder module. Compared with the traditional Transformer model, KANsformer not only demonstrates more excellent performance in capturing the temporal features of complex sequence data but also effectively realizes the implicit preliminary fusion of the patient's spatio-temporal features at the item level.

[0124] In the second aspect, the embodiments of the present invention also provide a medical examination recommendation method, as Figure 10 shown, the medical examination recommendation method includes:

[0125] S100, obtaining the patient sequence data to be measured;

[0126] S200, inputting the patient sequence data to be measured into a pre-constructed medical examination recommendation model to obtain the recommended examination;

[0127] Among them, the pre-constructed medical examination recommendation model is obtained according to the construction method of the medical examination recommendation model based on progressive spatio-temporal feature fusion described in the first aspect.

[0128] For the specific processing process of this medical examination recommendation model, please refer to the relevant content of the first aspect, which will not be elaborated here.

[0129] It can be understood that in the process of training the medical examination recommendation model in the embodiments of the present invention, a large amount of patient sequence data can be collected to construct a training set, which can basically cover all the characteristic situations of the patient sequence data, so that the trained medical examination recommendation model is more general and universal. In this way, when facing a patient sequence data to be measured, it can be directly input into the trained medical examination recommendation model, and the recommended examination items can be directly output.

[0130] In the above recommendation process, it is not necessary to go through the process of constructing the global collaborative heterogeneous graph again, but directly use the global collaborative heterogeneous graph obtained during the training process. Moreover, it is not necessary to output the heterogeneous subgraph by training the global collaborative heterogeneous graph through the guiding diffusion model, but directly use the diffusion model with the parameters solidified after training to output the corresponding heterogeneous subgraph, that is, use the heterogeneous subgraph output during the training stage. For the global collaborative spatial features, the set h of the item global collaborative spatial features during the training stage can also be directly used I spa For the patient global collaborative spatial features corresponding to the patient sequence data to be measured, referring to the collaborative filtering idea, the patient representation most similar to its sequence data can be matched in the training set

[0131] Then use the set of item global collaborative spatial features Obtain the re-represented patient sequence data, and then use the KANsformer sequence encoder module to process the re-represented patient sequence data to obtain the patient spatio-temporal fusion features After that, use the cross-attention spatio-temporal feature fusion module for processing to obtain the patient deep spatio-temporal fusion features

[0132] Finally, use the scoring function and the patient deep spatio-temporal fusion features of this patient to obtain the examination prediction vector of the training sequence data of this patient The examination with the highest value among them is the recommended examination.

[0133] The above specific process can be understood in combination with the relevant content of the first aspect, and no more detailed description will be given here.

[0134] The medical examination recommendation method provided by the embodiments of the present invention can obtain more accurate recommended examination items by using a pre-trained medical examination recommendation model, and has a relatively high recommendation efficiency.

[0135] The following presents the experiments of the present invention:

[0136] Table 1 shows the performance of the model proposed in the present invention and the baseline model on the test set. To exclude random interference and verify the reliability of the results, the present invention conducted 5 repeated experiments and proved its superiority compared with the strongest baseline model based on significance tests, where the p-value < 0.05, indicating that the improvement is statistically significant.

[0137] Table 1 Comparative experiment results

[0138]

[0139]

[0140] The baseline models involved in the experiment include: GRU-based sequential recommendation model (GRU4Rec), Neural Attentive Recommendation Machine (NARM), Convolutional Sequence Embedding Recommendation Model (Caser), Self-Attention based Sequential Model (SASRec), Sequential Recommendation with Bidirectional Encoder Representations from Transformer (BERT4Rec), Graph Contextualized Self-Attention Network for Session-based Recommendation (GCSAN), Self-Supervised Learning for Sequential Recommendation (S3-Rec), Low-rank decomposed self-attention networks (LightSANs), Diffusion Recommender Model (DiffRec), Symptom-based Set-to-set Small and Safe Drug Recommendation (4SDrug). The Progressive SpatioTemporal feature Fusion model for medical examination Recommendation (PSTFRec) represents the model of the embodiment of the present invention.

[0141] As shown in Table 1, the present invention uses two evaluation metrics widely used in the recommendation field, Hit@K and NDCG@K, to evaluate the performance of all methods, where K is set to 5 and 10 respectively. Hit@K measures whether the target item is included in the top K recommendation results of the model, while NDCG@K evaluates the sorting quality of the recommendation list. The higher the values of these two metrics, the better the performance of the model.

[0142] The experimental results show that, compared with 10 benchmark models, the proposed progressive spatio-temporal feature fusion model PSTFRec of the present invention has obtained the highest evaluation index and demonstrated the best recommendation performance.

[0143] It should be noted that in the description of this specification, the description with reference to terms such as "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. In addition, those skilled in the art can combine and combine the different embodiments or examples described in this specification.

[0144] The above are only the preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A construction method of a medical examination recommendation model based on progressive spatio-temporal feature fusion, characterized in that include: Acquire a number of patient sequence data; wherein the patient sequence data represents the diseases, symptoms and examinations of the patients in chronological order and has associated items, including the age and gender of the patients; Determine an examination as a training label in each patient sequence data; use the data before the training label as the patient's training sequence data, thereby obtaining a training set consisting of the training sequence data of all patients; wherein the training sequence data of each patient retains associated items; A global collaborative heterogeneous graph of the training set is constructed using a heterogeneous graph aggregation module, and the global collaborative heterogeneous graph is denoised using a task adaptive denoising strategy based on a diffusion model to generate a heterogeneous subgraph that is beneficial to downstream examination recommendation tasks; the heterogeneous subgraph is processed using a graph encoder to obtain global collaborative space features, including patient global collaborative space features and item global collaborative space features; the items include diseases, symptoms, examinations, age and gender; Based on the global collaborative spatial features of all items, the re-represented patient training sequence data representations are obtained through position encoding, and the KANsformer sequence encoder module is used to process each re-represented patient training sequence data representation to obtain the corresponding patient spatiotemporal fusion features; wherein the KANsformer sequence encoder module is obtained by integrating KANs into the Transformer framework; For each patient, the corresponding patient spatiotemporal fusion features and the patient's global collaborative spatial features are processed using the cross-attention spatiotemporal feature fusion module to obtain the patient's deep spatiotemporal fusion features; Based on the deep spatiotemporal fusion features of each patient, the examination prediction vector of the corresponding training sequence data is obtained; and model training is performed based on the examination prediction vectors and corresponding training labels of all patients to obtain a trained medical examination recommendation model, which is used to recommend the required examinations for the patient sequence data.

2. The construction method of the medical examination recommendation model based on progressive spatio-temporal feature fusion according to claim 1, wherein The global collaborative heterogeneous graph of the training set constructed by the heterogeneous graph aggregation module is expressed as: Among them, represents a global collaborative heterogeneous graph, which contains six types of entity nodes, namely patient nodes, disease nodes, symptom nodes, examination nodes, age nodes, and gender nodes. Among them, disease nodes, symptom nodes, examination nodes, age nodes, and gender nodes are project nodes. Patient nodes are only connected to project nodes, and project nodes are only connected to patient nodes. Patient nodes are connected to each other through their common nodes; u represents a patient; U represents the set of patients; r represents a relational edge; R represents the set of relational edges; i represents a project. represents a disease symptom examination δ xam The overall set of these three project entities; represents the set of ages; represents the set of genders; ∪ represents the union.

3. The method for constructing a medical examination recommendation model based on progressive spatio-temporal feature fusion according to claim 1 or 2, characterized in that The method of denoising the global collaborative heterogeneous graph by using a task adaptive denoising strategy based on a diffusion model to generate a heterogeneous subgraph that is beneficial to downstream inspection and recommendation tasks includes: By introducing a task adaptation function during the training process of the diffusion model to guide the diffusion model to generate heterogeneous subgraphs that are beneficial to downstream inspection recommendation tasks, which can be expressed by the formula: Among them, L rec and L diffsion represent the loss functions of the recommendation task and the diffusion model respectively; L d ′ iffsion is the loss function of the diffusion model processed by the task adaptive denoising strategy; represents the difference between the current training Epoch and the last Epoch of the global collaborative heterogeneous graph; represents the average change of L rec in the first δ epochs, where δ > 1 and ε < 1.

4. The construction method of the medical examination recommendation model based on progressive spatio-temporal feature fusion according to claim 2, wherein, The using a graph encoder to process the heterogeneous subgraph to obtain a global collaborative spatial feature includes: The relation-aware graph attention network RGAT is used as the graph encoder, and its message aggregation mechanism is used to aggregate knowledge to obtain entity representations at each layer; wherein the message aggregation mechanism is expressed as follows: Among them, the heterogeneous subgraph is represented by ; during the knowledge aggregation process, N m represents all adjacent entity sets connected by entity m through relationship r m,n in the heterogeneous subgraph; the embeddings of entities m and n are respectively represented as α(m, r m,n , n) represents the attention correlation of entities and relationships estimated during the knowledge aggregation process, which is used to capture the unique semantics of the relationship between entities m and n; the Τ in the upper right corner represents the transpose operation; W represents a learnable weight matrix; LeakyReLU(·) represents the LeakyReLU activation function; Drop(·) represents the Dropout function; Norm(·) represents the normalization process; exp(·) represents the natural exponential function; || represents the concatenation operation; represents the entity representation obtained by fusing entity m at the l g -th layer of RGAT; The entity representations of each layer are integrated as the global collaborative space feature; the global collaborative space feature is expressed as: Among them, represents the global collaborative space feature of any entity m in the heterogeneous subgraph, including the patient global collaborative space feature, denoted as and the project global collaborative space feature, denoted as d is the embedding dimension, is the set of patient global collaborative space features, is the set of project global collaborative space features; L g represents the number of layers of RGAT.

5. The construction method of the medical examination recommendation model based on progressive spatio-temporal feature fusion according to claim 4, characterized in that, The patient training sequence data representation obtained by position encoding based on the global collaborative spatial features of all items includes: The set of project global collaborative space features is used as the initial representation of each patient training sequence entity; For the training sequence data of each patient, determine the position of the item data in the training sequence data of the patient in and obtain the representation of the patient's training sequence data that re-represents the patient by adding its position embedding representation p i to .

6. The method for constructing a medical examination recommendation model based on progressive spatio-temporal feature fusion according to claim 5, wherein, The KANsformer sequence encoder module uses the multi-head self-attention mechanism, which is the core component of the Transformer architecture. Its formula is as follows: Among them, represents the output of the l-th layer of the multi-head attention module based on KANs in KANsformer; s L s and N respectively represent the total number of multi-head self-attention blocks and the number of heads; A n represents the attention value of the n-th head; φ(·) is the single-layer EfficientKAN function; is the n-th Efficient KAN function corresponding to the query, key, and value in the attention mechanism, used for feature mapping transformation; where φ1 represents the first-layer Efficient KAN function and φ2 represents the second-layer EfficientKAN function; GELU represents the GELU activation function; represents the spatio-temporal fusion feature of the patient obtained by KANsformer; represents the output of the last layer of KANsformer, L s represents the last layer of KANsformer, and T in u T represents the length of the training sequence data of the patient.

7. The construction method of the medical examination recommendation model based on progressive spatio-temporal feature fusion according to claim 6, characterized in that, For each patient, the corresponding patient spatiotemporal fusion features and the patient's global collaborative spatial features are processed using a cross-attention spatiotemporal feature fusion module to obtain the patient's deep spatiotemporal fusion features, including: The cross-attention spatio-temporal feature fusion module converts the patient's global collaborative spatial feature into a query feature Q, and converts the patient's spatio-temporal fusion feature into a key feature K and a value feature V, obtaining the patient's deep spatio-temporal fusion feature 8. The construction method of the medical examination recommendation model based on progressive spatio-temporal feature fusion according to claim 7, characterized in that, The inspection prediction vector corresponding to the training sequence data obtained based on the depth spatio-temporal fusion features of each patient includes: For the depth spatio-temporal fusion features of each patient, using a scoring function and the depth spatio-temporal fusion features of this patient, obtain the inspection prediction vector of the training sequence data of this patient, where the scoring function is expressed as: Among them, check e ∈ ε xam ; ε xam represents the set of all inspection items; represents the global collaborative space feature for this inspection item; is the patient's deep spatio-temporal fusion feature for patient u; inspection prediction vector represents the score vector of patient u for each candidate inspection item.

9. The construction method of the medical examination recommendation model based on progressive spatio-temporal feature fusion according to claim 8, wherein The model training based on the inspection prediction vectors of all patients and the corresponding training labels to obtain a trained medical examination recommendation model includes: Based on the inspection prediction vectors of each patient and the corresponding training labels, use cross loss for model training to obtain a trained medical examination recommendation model.

10. A medical examination recommendation method, characterized in that, It includes: Obtain the patient sequence data to be measured; Input the patient sequence data to be measured into a pre-constructed medical examination recommendation model to obtain the recommended examination; wherein, the pre-constructed medical examination recommendation model is obtained according to the construction method of the medical examination recommendation model based on progressive spatio-temporal feature fusion described in any one of claims 1-9.

Images