Recommendation method and system for multi-channel information fusion scheme

Abstract

The application provides a recommendation method and system of a multi-channel information fusion scheme, comprising the following steps: S1, collecting relevant data; the data comprises medical guideline data, historical case data and patient physical index data; S2, constructing a medical guideline hypergraph to obtain hypergraph node representation; S3, processing the relevant data, scoring different schemes and outputting the scoring results. The application fully integrates a data-driven method and a decision rule, uses a more flexible coding mode, models the conditions and results of the rules in the authoritative and widely used medical guidelines in the form of distance in the representation space, models the information from three channels of the medical guidelines, the historical cases and the patient physical indexes at the same time, fully considers the embedded representation of the patient under different dimensions, and improves the accuracy and interpretability of the recommendation.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical decision-making, and more specifically, to a method and system for recommending multi-channel information fusion schemes. Background Technology

[0002] Intelligent medical decision-making is a crucial tool in the healthcare field, enabling medical professionals to develop more rapid and comprehensive treatment plans for patients. Integrating data-driven approaches with the knowledge gained from medical guidelines is essential for effective treatment decisions. On one hand, data-driven methods can learn the underlying knowledge behind patient data based on its complex structure and rich training data. On the other hand, medical guidelines, a set of authoritative and widely used decision-making rules summarized by medical experts, can guide data-driven methods in extracting and modeling key information. For example, attributes from medical guidelines often hold greater significance when training data-driven models. This integrated approach, utilizing both data-driven methods and professional expertise, promises to improve the accuracy and personalization of treatment decisions.

[0003] Existing knowledge-integrated healthcare recommendation methods typically take a coarse approach to combining data-driven and rule-based approaches, heavily relying on data quality. They often treat healthcare rules as a constraint on recommendations or learn rules using large-scale data, thus potentially encountering the following problems:

[0004] 1. Poor generalization ability of decision rules. This leads to the failure of constraints in decision rules; in addition, due to the relatively small amount of data to match decision rules, the ability to learn decision rules using data-driven methods is relatively weak, so the coding of decision rules needs to be improved.

[0005] 2. They neglect the relationships behind the decision rules. They often treat the rules as a whole for learning, ignoring the connections between the smallest unit constraints within the rules; this makes it difficult to model the key relationships between the conditions and outcomes of the rules.

[0006] These issues lead to an imbalance in the combination of data-driven methods and decision rule-based knowledge-driven methods, which limits the performance of existing methods.

[0007] Chinese patent document CN108877921B discloses a medical intelligent triage method and system. This medical intelligent triage method includes: extracting disease-related symptoms and signs from patient information as candidate factor information; based on the candidate factor information, obtaining various diseases and treatment measures related to the symptoms from medical literature as identification knowledge information; matching the identification knowledge information with the candidate factor information; repeating the above steps until the department to which the disease belongs is determined or the patient information has been extracted and matched, and returning the triage result. However, this patent document mainly provides a method and system for selecting the appropriate department for a patient, while this invention scores different options, with higher scores indicating higher relevance, and recommends the highest-scoring option to the doctor, shortening the doctor's selection time. This is fundamentally different from the problem solved by this invention patent. Summary of the Invention

[0008] In view of the deficiencies in the existing technology, the purpose of this invention is to provide a recommended method and system for multi-channel information fusion schemes.

[0009] A method for recommending a multi-channel information fusion scheme according to the present invention includes:

[0010] Step S1: Collect relevant data;

[0011] The data includes medical guideline data, historical case data, and patient physical indicator data;

[0012] Step S2: Construct a medical guideline hypergraph and obtain hypergraph node representations;

[0013] Step S3: Process the relevant data, score the different schemes, and output the scoring results.

[0014] Preferably, step S2 includes using the patient's physical indicator data as a condition node in the hypergraph and using the scheme corresponding to the patient's physical indicator data as a result node in the hypergraph; the condition node and the result node constitute a decision rule.

[0015] Preferably, in step S2, the condition node is defined as a. i The resulting node is t i SuperMap

[0016] in, Let G represent the set of all condition nodes and all result nodes, where condition nodes and result nodes belonging to the same decision rule are connected by edges; ε represents the set of all edges in the hypergraph G; for each decision rule r, its associated node set is... Let represent the set of nodes for all decision rules, and

[0017] Preferably, step S2 includes the following sub-steps:

[0018] Step S2.1: Process the condition nodes and result nodes;

[0019] Step S2.2: Model the condition nodes and result nodes and complete the training of node representations.

[0020] Preferably, step S3 includes the following sub-steps:

[0021] Step S3.1: Process patient physical indicator data;

[0022] Step S3.2: Perform adaptive multi-channel information fusion.

[0023] Preferably, step S2.1 includes the following sub-steps:

[0024] Step S2.1.1: For the set Each node x in i Perform random initialization to obtain the initial node embedding representation x. i ∈R K Where K represents the vector dimension, and node x i Represents condition node a i Or result node t i , and a i ∈R K ,t i ∈R K ;

[0025] Step S2.1.2: Use GCN to capture the medical guideline hypergraph structure, along with each node x i The structural contributions of adjacent nodes are:

[0026]

[0027] in, It is node x i All neighboring nodes, w ik Representative node x i and adjacent node x k The weight of the edge between them, d i and d k These are nodes x i and x k The degree, Θ T It is the transpose of the weight parameter matrix of a trainable GCN;

[0028] Step S2.1.3: Strengthen the connections between condition nodes, and add t jAs a query in the attention mechanism, Key=Value=N is constructed. j ,and:

[0029]

[0030]

[0031]

[0032] Among them, a k For condition node a k Embedded representation, It is a set of condition nodes; Represents condition node a i The set of neighboring nodes, The embedding representations of the nodes form a vector matrix N. j , Represents the vector matrix N j The transpose of t′; softmax(·) indicates the use of the softmax normalization function, t′ j For the result node t after the attention mechanism is updated j The embedding representation; mean(·) represents taking the mean; It is condition node a i The embedded representation incorporates information from second-order neighbor nodes;

[0033] The embedded representation of the result node is obtained in the same way.

[0034]

[0035]

[0036]

[0037] in, Represents condition node a j The set of neighboring nodes, N j It is its embedded representation, Indicates matrix transpose; It is the result node t i The set of neighboring nodes; t k Represents the result node t in the set k Embedded representation; a′ j It is condition node a j Embedded representation after attention mechanism and normalization It is the result node t i Vector representation of;

[0038] Step S2.1.4: Based on GCN and the path-based attention mechanism, the initial representations of the condition nodes and result nodes are updated through a weighted combination:

[0039]

[0040] Where α and β are hyperparameters, using Representation of condition node embedding Or result node embedding representation

[0041] Preferably, step S2.2 includes the following sub-steps:

[0042] Step S2.2.1: Based on contrastive learning, model the nodes, taking condition nodes and result nodes belonging to the same decision rule subgraph as positive samples, and nodes not belonging to the same decision rule as negative samples; the type of nodes in the negative samples is not limited.

[0043]

[0044]

[0045] Among them, the set of positive sample pairs is The set of negative sample pairs is For positive sample pairs, Negative sample pairs;

[0046] Step S2.2.2: Calculate node x using the dot product. i and x j distance Training node representations using the cross-entropy loss function:

[0047]

[0048]

[0049] in, For node x i The transpose of the embedded representation. Representing the dot product of vectors, we obtain the distance. e i For positive or negative sample pairs, y i For the corresponding label; when e i When y is a positive sample pair i When e is 1 i When y is a negative sample pair i =0; For e i The distance between the included nodes;

[0050] Step S2.2.3: Represent the decision rule r embedding as the mean of the embeddings of all condition nodes within the rule subset, by narrowing down r and t. r Distance, update the loss function:

[0051]

[0052]

[0053] Among them, t r For the result node of rule r, the embedding representation r of decision rule r is obtained by calculation. Transpose it.

[0054] Preferably, step S3.1 includes the following sub-steps:

[0055] Step S3.1.1: For each patient p i Define the patient's physical indicator data as v j , For patient p i All body metrics data v j The set; Matching with all nodes in V yields the condition node a, which is related to the decision rule. j The set M i M i The representations of the condition nodes are weighted and fused to obtain the vector representation of the medical guideline data channel. and:

[0056]

[0057]

[0058]

[0059] Step S3.1.2: For each patient p in the memory set m All exist

[0060] Among them, t m Equivalent to It is patient p m The treatment plan is matched with V to obtain the result node embedding; s′ i,m Indicates patient p i and patient p m The cosine similarity is used to standardize the similarity, making the patient p i With candidate set S i The sum of the similarities between patients is 1, thus obtaining patient p. i With patient p mSimilarity s i,m ,and:

[0061]

[0062]

[0063] Among them, S i It is the set of K patients with the highest similarity in the memory set, where K is adjusted according to the dataset. The treatment plans of the patients with the smallest score differences are combined to obtain p. i Neighbor passage indicates

[0064]

[0065] Step S3.1.3: Preprocess the patient's physical indicator data. Based on data-driven methods and combined with the DPCNN model, obtain the representation of the patient's physical indicator data channels.

[0066]

[0067] Here, preproc(·) represents the preprocessing operation and makes custom adjustments based on the real data; concat(·) represents concatenating the patient's attribute vectors into a matrix, which is then passed through a fully connected layer fc1 to obtain the patient's embedded representation.

[0068] Preferably, step S3.2 includes the following sub-steps:

[0069] Step S3.2.1: Based on the fully connected layer fc2, adaptively fuse the representations of the three channels: medical guideline data, historical case data, and patient physical indicator data.

[0070]

[0071] Step S3.2.2: Based on the fully connected layer fc3, connect the patient p i The fusion vector representation is mapped to a probability vector, and the input vector is mapped to the range (0,1) through the Sigmoid activation function σ to obtain the patient p. i And the predicted probability vector of all treatment options y i,j For patient p i Treatment plan t j The true label, For the corresponding probability values, train the model using the cross-entropy loss function:

[0072]

[0073]

[0074] A recommendation system based on a multi-channel information fusion scheme provided by the present invention includes:

[0075] Module M1: Collects relevant data;

[0076] The data includes medical guideline data, historical case data, and patient physical indicator data;

[0077] Module M2: Construct a medical guideline hypergraph and obtain hypergraph node representations;

[0078] Module M3: Processes relevant data, scores different schemes, and outputs the scoring results.

[0079] Preferably, module M2 includes using patient physical indicator data as condition nodes in a hypergraph and using the solutions corresponding to the patient physical indicator data as result nodes in the hypergraph; the condition nodes and the result nodes constitute decision rules.

[0080] Preferably, in module M2, the condition node is defined as a. i The resulting node is t i SuperMap

[0081] in, Let G represent the set of all condition nodes and all result nodes, where condition nodes and result nodes belonging to the same decision rule are connected by edges; ε represents the set of all edges in the hypergraph G; for each decision rule r, its associated node set is... Let represent the set of nodes for all decision rules, and

[0082] Preferably, module M2 includes the following sub-modules:

[0083] Module M2.1: Processes condition nodes and result nodes;

[0084] Module M2.2: Models condition nodes and result nodes and completes the training of node representations.

[0085] Preferably, module M3 includes the following sub-modules:

[0086] Module M3.1: Processes patient physical indicator data;

[0087] Module M3.2: Performs adaptive multi-channel information fusion.

[0088] Preferably, module M2.1 includes the following sub-modules:

[0089] Module M2.1.1: For sets Each node x in i Perform random initialization to obtain the initial node embedding representation x. i ∈R K Where K represents the vector dimension, and node x i Represents condition node a i Or result node t i , and a i ∈R K ,t i ∈R K ;

[0090] Module M2.1.2: Uses GCN to capture the hypergraph structure of medical guidelines, with each node x i The structural contributions of adjacent nodes are:

[0091]

[0092] in, It is node x i All neighboring nodes, w ik Representative node x i and adjacent node x k The weight of the edge between them, d i and d k These are nodes x i and x k The degree, Θ T It is the transpose of the weight parameter matrix of a trainable GCN;

[0093] Module M2.1.3: Enhances the connections between condition nodes, including t j As a query in the attention mechanism, Key=Value=N is constructed. j ,and:

[0094]

[0095]

[0096]

[0097] Among them, a k For condition node a k Embedded representation, It is a set of condition nodes; Represents condition node a i The set of neighboring nodes, The embedding representations of the nodes form a vector matrix N. j , Represents the vector matrix N jThe transpose of t′; softmax(·) indicates the use of the softmax normalization function, t′ j For the result node t after the attention mechanism is updated j The embedding representation; mean(·) represents taking the mean; It is condition node a i The embedded representation incorporates information from second-order neighbor nodes;

[0098] The embedded representation of the result node is obtained in the same way.

[0099]

[0100]

[0101]

[0102] in, Represents condition node a j The set of neighboring nodes, N j It is its embedded representation, Indicates matrix transpose; It is the result node t i The set of neighboring nodes; t k Represents the result node t in the set k Embedded representation; a′ j It is condition node a j Embedded representation after attention mechanism and normalization It is the result node t i Vector representation of;

[0103] Module M2.1.4: Based on GCN and path-based attention mechanisms, the initial representations of condition nodes and result nodes are updated through weighted combination:

[0104]

[0105] Where α and β are hyperparameters, using Representation of condition node embedding Or result node embedding representation

[0106] Preferably, module M2.2 includes the following sub-modules:

[0107] Module M2.2.1: Based on contrastive learning, node modeling uses conditional nodes and result nodes belonging to the same decision rule subgraph as positive samples, and nodes not belonging to the same decision rule as negative samples; the type of nodes in the negative samples is unrestricted.

[0108]

[0109]

[0110] Among them, the set of positive sample pairs is The set of negative sample pairs is For positive sample pairs, Negative sample pairs;

[0111] Module M2.2.2: Calculate node x using the dot product method. i and x j distance Training node representations using the cross-entropy loss function:

[0112]

[0113]

[0114] in, For node x i The transpose of the embedded representation. Representing the dot product of vectors, we obtain the distance. e i For positive or negative sample pairs, y i For the corresponding label; when e i When y is a positive sample pair i When e is 1 i When y is a negative sample pair i =0; For e i The distance between the included nodes;

[0115] Module M2.2.3: Represents the embedding of decision rule r as the mean of the embeddings of all condition nodes within the rule subset, by narrowing down r and t. r Distance, update the loss function:

[0116]

[0117]

[0118] Among them, t r For the result node of rule r, the embedding representation r of decision rule r is obtained by calculation. Transpose it.

[0119] Preferably, module M3.1 includes the following sub-modules:

[0120] Module M3.1.1: For each patient p i Define the patient's physical indicator data as v j, For patient p i All body metrics data v j The set; Matching with all nodes in V yields the condition node a, which is related to the decision rule. j The set M i M i The representations of the condition nodes are weighted and fused to obtain the vector representation of the medical guideline data channel. and:

[0121]

[0122] W i ={…,a j ,…};

[0123]

[0124] Module M3.1.2: For each patient p in the memory set m All exist

[0125] Among them, t m Equivalent to It is patient p m The treatment plan is matched with V to obtain the result node embedding; s′ i,m Indicates patient p i and patient p m The cosine similarity is used to standardize the similarity, making the patient p i With candidate set S i The sum of the similarities between patients is 1, thus obtaining patient p. i With patient p m Similarity s i,m ,and:

[0126]

[0127]

[0128] Among them, S i It is the set of K patients with the highest similarity in the memory set, where K is adjusted according to the dataset. The treatment plans of the patients with the smallest score differences are combined to obtain p. i Neighbor passage indicates

[0129]

[0130] Module M3.1.3: Preprocesses patient physical indicator data, using a data-driven approach and incorporating a DPCNN model to obtain channel representations of the patient physical indicator data.

[0131]

[0132] Here, preproc(·) represents the preprocessing operation and makes custom adjustments based on the real data; concat(·) represents concatenating the patient's attribute vectors into a matrix, which is then passed through a fully connected layer fc1 to obtain the patient's embedded representation.

[0133] Preferably, module M3.2 includes the following sub-modules:

[0134] Module M3.2.1: Based on the fully connected layer fc2, it adaptively fuses the representations of three channels: medical guideline data, historical case data, and patient physical indicator data.

[0135]

[0136] Module M3.2.2: Based on the fully connected layer fc3, the patient p i The fusion vector representation is mapped to a probability vector, and the input vector is mapped to the range (0,1) through the Sigmoid activation function σ to obtain the patient p. i And the predicted probability vector of all treatment options y i,j For patient p i Treatment plan t j The true label, For the corresponding probability values, train the model using the cross-entropy loss function:

[0137]

[0138]

[0139] Compared with the prior art, the present invention has the following beneficial effects:

[0140] 1. This invention fully integrates data-driven methods and decision rules, and uses a more flexible encoding method to model the conditions and results of rules in authoritative and widely used medical guidelines in the form of distance in the representation space; at the same time, it models information from three channels: medical guidelines, historical cases and patient physical indicators, and fully considers the patient's embedded representation under different dimensions, thereby improving the accuracy and interpretability of recommendations.

[0141] 2. This invention can score different solutions, with the score representing the degree of correlation between the solutions, thus shortening the time doctors spend choosing a solution. By optimizing the representation of medical rules, it overcomes the problem of poor generalization ability of decision rules in existing methods, ensuring that the rules can be effectively applied in various situations, solving the problem of insufficient matching data for decision rules, and providing a richer source of information for data-driven methods.

[0142] 3. The medical guideline hypergraph constructed in this invention can creatively propose a decision rule encoding method by using the key connections between the condition nodes and result nodes corresponding to medical rules and employing comparative learning.

[0143] 4. This invention adopts a data-driven approach to directly model patient physical indicators, thereby effectively covering complex patient situations with weak rule connections; it adaptively fuses information from three channels—medical guidelines, historical similar cases, and the patient's own physical indicators—skillfully balancing the relationship between data-driven and rule-driven approaches, thus improving the overall performance of the recommendation system.

[0144] Other beneficial effects of the present invention will be explained in detail through the introduction of specific technical features and technical solutions in specific embodiments. Those skilled in the art should be able to understand the beneficial technical effects brought about by these technical features and technical solutions through the introduction of these technical features and technical solutions. Attached Figure Description

[0145] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0146] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0147] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.

[0148] Reference Figure 1 As shown, a medical guideline hypergraph is first constructed based on the decision rules. The numerical conditions of the physical indicators and the corresponding treatment plans are regarded as condition nodes and result nodes in the hypergraph, respectively. Positive sample pairs and negative sample pairs are constructed based on the graph. The hypergraph node representation is obtained by contrastive learning, and the representation learning of medical rules is carried out.

[0149] Subsequently, information from three channels—medical guidelines, historical cases, and patient physical indicators—was modeled and adaptively fused into multi-channel information. The treatment plan was then scored, with higher scores indicating greater relevance. This allowed doctors to be recommended treatment plans for patients, saving them time.

[0150] This invention integrates a data-driven approach with decision rules from clinical manuals. It not only captures potential knowledge from patient data through data-driven methods, but also constructs a rule learning method with strong generalization capabilities, thereby helping medical professionals to develop more rational treatment plans for patients.

[0151] Specifically, a recommended method for a multi-channel information fusion scheme includes the following steps:

[0152] Step 1: Construct the medical guideline hypergraph G. Several "conditions" and one "outcome" constitute a decision rule, such as rule G.

[0153] The "conditions" are a1 and a2, which are medical indicator restrictions, such as "er≥1", and the "outcome" is t. i For example, a treatment plan, such as "ec-t". "Condition" and "outcome" are defined as two types of nodes in the hypergraph G, and the hypergraph is constructed based on their relationships. Among them, set It contains all condition nodes and result nodes; condition nodes and result nodes belonging to the same rule are connected by edges, and the set ε contains all edges in the hypergraph G; for each rule r, its associated node set is... Rule representation set The set of nodes that describes all rules, i.e.

[0154] Step 2: For Each node x i Perform random initialization to obtain the initial node embedding representation x. i ∈r K Node x i It can represent condition node a i Or result node t i That is, there exists a i ∈R K ,t i ∈R K , where K is the dimension of the embedding representation.

[0155] Step 3: Use GCN to capture the hypergraph structure of the medical guidelines, for each node x i The structural contribution of directly adjacent nodes is taken into account, and the specific calculation formula is as follows:

[0156]

[0157] in, It is node x i All neighboring nodes, w ik Representative node x i and adjacent node x k The weight of the edge between them, d i and d k It is node x i and x k The degree, Θ T It is the transpose of the weight parameter matrix of a trainable GCN.

[0158] Step 4: In the hypergraph, condition nodes and result nodes belonging to the same rule form a rule subgraph. The same result node may exist in different subgraphs, thus creating second-order paths that connect condition nodes in different subgraphs, such as...

[0159] Among them, the result node t j Acting as a bridge to enhance the connection between conditional nodes; inspired by the self-attention mechanism, t j Treat it as a Query in the attention mechanism, and construct Key=Value=N j Taking the update of the embedded representation of condition nodes as an example, the following formula applies:

[0160]

[0161]

[0162]

[0163] Among them, a k For condition node a k Embedded representation. Represents the result node t j The set of neighboring nodes, since there are edges connecting the condition nodes and the result nodes in the hypergraph, therefore It is a set of condition nodes. Similarly, Represents condition node a i The set of neighboring nodes. The embedding representations of the nodes form a vector matrix N. j , Represents the vector matrix N j The transpose of t′. softmax(·) indicates the use of the softmax normalization function, t′ j For the result node t after the attention mechanism is updated j The embedded representation. mean(·) represents taking the mean. It is condition node a i The embedding representation incorporates information from neighboring nodes updated using the attention mechanism, which means it incorporates information from second-order neighboring nodes. The embedding representation of the resulting node can be obtained using the same method.

[0164] Step 5: Using GCN and a path-based attention mechanism, the initial representations of condition nodes and result nodes can be updated using the following weighted combination, where α and β are hyperparameters:

[0165]

[0166] Step 6: After obtaining the node representation, use contrastive learning to model the nodes. Conditional nodes and treatment nodes belonging to the same rule subgraph are considered as positive samples, and nodes that do not belong to the same rule are considered as negative samples. The type of nodes in the negative samples is not restricted.

[0167] Among them, the set of positive sample pairs is The set of negative sample pairs is The formula is defined as follows:

[0168]

[0169]

[0170] Step 7: Calculate node x using the dot product. i and x j distance in For node x i The transpose of the embedded representation. Representing the dot product of vectors, we obtain the distance. e i For positive or negative sample pairs, y i For the corresponding label, when e i When y is a positive sample pair i When e is 1 i When y is a negative sample pair i =0, For e i The distances between the included nodes are used. The node representations are trained using the cross-entropy loss function, as shown in the following formula:

[0171]

[0172]

[0173] Step 8: For rule r, its embedding representation can be calculated as the mean of the embeddings of all condition nodes within the rule subset, where tr For the result node of rule r, Let the transpose of the embedding representation r be used, and the loss function be updated accordingly, further narrowing the distance between the rule embedding and the treatment plan, as follows:

[0174]

[0175]

[0176] Step 9: For each patient p i Its physical indicators are expressed in the form of "attribute-value"; for example, "er value = 0.1", expressed as v. j This indicates that all of the patient's physical indicators constitute a set. Bundle Matching all nodes in V yields a set M of condition nodes related to the rule. i M i The representations of the condition nodes are weighted and fused to obtain the vector representation of the medical guide channel. The formula is as follows:

[0177]

[0178] M i ={…,a j ,…};

[0179]

[0180] Step 10: For each patient p in the memory set m All exist

[0181] Where t m Equivalent to It is patient p m The treatment plan is matched with V to obtain the result node embedding; s′ i,m Indicates patient p i and patient p m The cosine similarity is used to standardize the similarity, making the patient p i With candidate set S i The sum of the similarities between patients is 1, thus obtaining patient p. i With patient p m Similarity s i,m The similarity calculation formula is as follows:

[0182]

[0183]

[0184] Among them, Si It is the set of K patients with the highest similarity in the memory set, where K is adjusted according to the dataset; the treatment plans of the most similar patients are combined to obtain p. i Neighbor passage indicates

[0185]

[0186] Step 11: Preprocess the patient's physical indicator information. Since the patient's physical indicator data contains multiple attributes with different data types, differentiated preprocessing is required. `preproc(·)` represents a preprocessing operation, such as mapping discrete data to one-hot vectors, or mapping continuous data to vectors of the same shape through a fully connected layer. `preproc(·)` can be customized based on the actual data. `concat(·)` concatenates the patient's attribute vectors into a matrix, which is then passed through a fully connected layer `fc1` to obtain the patient's embedded representation. This step is purely data-driven, and here it combines a DPCNN model to obtain the representation of the patient's information channels. The formula is as follows:

[0187]

[0188] Step 12: Adaptively fuse the representations of the three channels using the fully connected layer fc2:

[0189]

[0190] Step 13: For patient p i The fusion vector representation is mapped to a probability vector through a fully connected layer fc3, and then the input vector is mapped to the range (0,1) through a sigmoid activation function σ to obtain the patient p. i And the predicted probability vector of all treatment options y i,j For patient p i Treatment plan t j The true label, Given the corresponding probability values, the model is trained using the cross-entropy loss function:

[0191]

[0192]

[0193] This invention fully integrates data-driven methods and decision rules, employing a more flexible encoding approach to model the conditions and outcomes of rules from authoritative and widely used medical guidelines in a distance-based representation space. Simultaneously, it models information from three channels: medical guidelines, historical cases, and patient physical indicators, fully considering the patient's embedded representation across different dimensions, thereby improving the accuracy and interpretability of recommendations.

[0194] The above are basic embodiments of the present invention. The technical solution of the present invention will be further described below through a preferred embodiment.

[0195] Example 1

[0196] Suppose we have a patient attributes and treatment plan table as shown in Table 1, and medical rules as shown in Table 2:

[0197] Table 1 Patient Attributes and Treatment Plans

[0198] <![CDATA[p1]]> <![CDATA[v1]]> <![CDATA[v3]]> <![CDATA[t2]]> <![CDATA[p2]]> <![CDATA[v2]]> <![CDATA[v4]]> <![CDATA[t3]]> <![CDATA[p3]]> <![CDATA[v2]]> None <![CDATA[t3]]> <![CDATA[p i ]]> <![CDATA[v1]]> <![CDATA[v4]]> <![CDATA[t2]]>

[0199] Table 2. Medical Guideline Rules and Their Inferential Relationships

[0200]

[0201] Based on this information, the recommended treatment plan includes the following steps:

[0202] Step 1: Construct a hypergraph G. For each rule in Table 2, construct condition nodes and result nodes. The mapping relationship of the defined condition nodes is shown in Table 3. For example, according to the defined relationship, for rule r1, the condition node is a1, the result node is t1, and there is an edge connecting them.

[0203] Table 3. Mapping Relationship between Rule Subconditions and Hypergraph Condition Nodes

[0204] <![CDATA[A1<v1]]> <![CDATA[a1]]> <![CDATA[A1=v1]]> <![CDATA[a2]]> <![CDATA[v1<A1<v2]]> <![CDATA[a3]]> <![CDATA[A1≥v2]]> <![CDATA[a4]]> <![CDATA[A2≤v3]]> <![CDATA[a5]]> <![CDATA[v3<A2<v4]]> <![CDATA[a6]]> <![CDATA[A2≥v4]]> <![CDATA[a7]]>

[0205] Step 2: Randomly initialize the embedding representation of all nodes in the hypergraph G, such that the node representation of rule r1 is a1,t1.

[0206] Step 3: Use GCN to capture the hypergraph structure of the medical guidelines, for each node x i The structural contribution of directly adjacent nodes is taken into account, and the specific calculation formula is as follows:

[0207]

[0208] in, It is node x i All neighboring nodes, w ik Representative node xi and adjacent node x k The weight of the edge between them, d i and d k It is node x i and x k The degree, Θ, is a weight parameter of a trainable GCN, which yields the node representation x. GCN Taking rule r1 as an example, the nodes it contains are represented as follows:

[0209] Step 4: Enhance node connectivity in the hypergraph using a path-based attention mechanism, taking rules r1 and r4 as examples. Rule r1 has condition nodes a1 and result nodes t1, indicating the existence of a path. If the condition nodes of rule r4 include a3 and a6, and the result node is t1, then a path exists. If the result node is used as a bridge, then a path exists. Taking the update of the embedded representation of condition nodes as an example, the following formula is followed:

[0210]

[0211]

[0212]

[0213] Get the updated node embedding Update all condition nodes and result nodes in the graph using the same method.

[0214] Step 5: Using GCN and a path-based attention mechanism, the initial representations of condition nodes and result nodes can be updated using the following weighted combination, where α and β are hyperparameters:

[0215]

[0216] Then, using contrastive learning, condition nodes and result nodes belonging to the same rule are considered positive samples. For example, if the condition node of rule r1 is a1 and the result node is t1, then the positive sample pair is (a1, t1). If the condition nodes of rule r2 are a2 and a5 and the result node is t2, then the positive sample pairs are (a2, t2) and (a5, t2). If node a1 and result node t2 are not in the same rule, then a negative sample (a1, t2) is formed. And so on, negative sample pairs also include (a2, t1), (a5, t1), etc.

[0217] Step 6: Calculate the loss function, complete model training, and obtain the node embedding representation after training.

[0218] Step 7: Assume that patients p1-p3 in Table 1 represent patient information in the memory set. Using patient p... i The following example illustrates the multi-channel information fusion treatment plan recommendation process in this embodiment. The information of all patients in Table 1 is structured, and the non-empty attributes are converted into different "attribute name: attribute value" features, such as: p1:{A1:v1,A2:v3}.

[0219] Step 8: Patient p i Match the attribute pairs with the range of all condition nodes in the hypergraph, referring to Table 2, to obtain patient p. i The successfully matched attribute nodes are a2 and a7. Their embedding representations, a2 and a7, have already been obtained in the model trained using the previous hypergraph node embedding representation. A weighted average of a2 and a7 is then used to obtain the embedding representation of the patient in the regular channel.

[0220] Step 9: Similar to Step 8, for patients p1-p3 in Table 1, the embedding representation of the regular channels can be obtained, that is... Calculate patient p i Similarity to other patients:

[0221]

[0222]

[0223] Among them, S i It is the set of K patients with the highest similarity in the memory set, where K is adjusted according to the dataset; the treatment plans of the most similar patients are combined to obtain p. i Neighbor passage indicates

[0224]

[0225] Let s be obtained here i,m The values ​​are 0.5, 0.3, and 0.2 respectively, for s i,m The patients are sorted, and the K most similar patients are selected. In this embodiment, K is 3, and the embedding representation of the neighbor channels is...

[0226] Step 10: After preprocessing all patient data, it is directly input into the data-driven neural network model. This embodiment uses DPCNN to obtain the embedding representation of the data-driven channels.

[0227] Step 11: Perform adaptive fusion of the three channels to obtain the patient's embedded representation p iThe model is trained by obtaining the probability distribution vectors of patients and all treatment options through a fully connected layer, and then calculating the loss using the real treatment option t2.

[0228] Step 12: After the model training is completed, when recommending treatment plans for new patients, start from step 7 again and recommend treatment plans to doctors based on the probability distribution of treatment plans for new patients.

[0229] The medical guideline hypergraph constructed in this embodiment can creatively propose a decision rule encoding method by using contrastive learning to identify key connections between condition nodes and result nodes corresponding to medical rules. By adopting a data-driven approach, it directly models patient physical indicator information, thereby effectively covering complex patient situations where rule connections are weak. It adaptively fuses information from three channels: medical guidelines, historical similar cases, and the patient's own physical indicators, cleverly balancing the relationship between data-driven and rule-driven approaches, and improving the overall performance of the recommendation system.

[0230] The present invention also provides a recommendation system for a multi-channel information fusion scheme. The recommendation system for the multi-channel information fusion scheme can be implemented by executing the process steps of the recommendation method of the multi-channel information fusion scheme. That is, those skilled in the art can understand the recommendation method of the multi-channel information fusion scheme as a preferred embodiment of the recommendation system for the multi-channel information fusion scheme.

[0231] Specifically, a recommendation system based on a multi-channel information fusion scheme includes:

[0232] Module M1: Collects relevant data;

[0233] The data includes medical guideline data, historical case data, and patient physical indicator data;

[0234] Module M2: Construct a medical guideline hypergraph and obtain hypergraph node representations;

[0235] Module M3: Processes relevant data, scores different schemes, and outputs the scoring results.

[0236] The module M2 includes using patient physical index data as condition nodes in the hypergraph and using the solutions corresponding to the patient physical index data as result nodes in the hypergraph; the condition nodes and the result nodes constitute decision rules.

[0237] In module M2, the condition node is defined as a. i The resulting node is t i SuperMap

[0238] in, Let G represent the set of all condition nodes and all result nodes, where condition nodes and result nodes belonging to the same decision rule are connected by edges; ε represents the set of all edges in the hypergraph G; for each decision rule r, its associated node set is... Let represent the set of nodes for all decision rules, and

[0239] Module M2 includes the following sub-modules:

[0240] Module M2.1: Processes condition nodes and result nodes;

[0241] Module M2.2: Models condition nodes and result nodes and completes the training of node representations.

[0242] The module M3 includes the following sub-modules:

[0243] Module M3.1: Processes patient physical indicator data;

[0244] Module M3.2: Performs adaptive multi-channel information fusion.

[0245] Module M2.1 includes the following sub-modules:

[0246] Module M2.1.1: For sets Each node x in i Perform random initialization to obtain the initial node embedding representation x. i ∈R K Where K represents the vector dimension, and node x i Represents condition node a i Or result node t i , and a i ∈RE K ,t i ∈R K ;

[0247] Module M2.1.2: Uses GCN to capture the hypergraph structure of medical guidelines, with each node x i The structural contributions of adjacent nodes are:

[0248]

[0249] in, It is node x i All neighboring nodes, w ik Representative node x i and adjacent node x k The weight of the edge between them, d i and d k These are nodes x i and xk The degree, Θ T It is the transpose of the weight parameter matrix of a trainable GCN;

[0250] Module M2.1.3: Enhances the connections between condition nodes, including t j As a query in the attention mechanism, Key=Value=N is constructed. j ,and:

[0251]

[0252]

[0253]

[0254] Among them, a k For condition node a k Embedded representation, It is a set of condition nodes; Represents condition node a i The set of neighboring nodes, The embedding representations of the nodes form a vector matrix N. j , Represents the vector matrix N j The transpose of t′; softmax(·) indicates the use of the softmax normalization function, t′ j For the result node t after the attention mechanism is updated j The embedding representation; mean(·) represents taking the mean; It is condition node a i The embedded representation incorporates information from second-order neighbor nodes;

[0255] The embedded representation of the resulting node is obtained in the same way.

[0256]

[0257]

[0258]

[0259] in, Represents condition node a j The set of neighboring nodes, N j It is its embedded representation, Indicates matrix transpose; It is the result node t i The set of neighboring nodes; t k Represents the result node t in the set k Embedded representation; a′ jIt is condition node a j Embedded representation after attention mechanism and normalization It is the result node t i Vector representation of;

[0260] Module M2.1.4: Based on GCN and path-based attention mechanisms, the initial representations of condition nodes and result nodes are updated through weighted combination:

[0261]

[0262] Where α and β are hyperparameters, using Representation of condition node embedding Or result node embedding representation

[0263] Module M2.2 includes the following sub-modules:

[0264] Module M2.2.1: Based on contrastive learning, node modeling uses conditional nodes and result nodes belonging to the same decision rule subgraph as positive samples, and nodes not belonging to the same decision rule as negative samples; the type of nodes in the negative samples is unrestricted.

[0265]

[0266]

[0267] Among them, the set of positive sample pairs is The set of negative sample pairs is For positive sample pairs, Negative sample pairs;

[0268] Module M2.2.2: Calculate node x using the dot product method. i and x j distance Training node representations using the cross-entropy loss function:

[0269]

[0270]

[0271] in, For node x i The transpose of the embedded representation. Representing the dot product of vectors, we obtain the distance. e i For positive or negative sample pairs, y i For the corresponding label; when e i When y is a positive sample pair iWhen e is 1 i When y is a negative sample pair i =0; For e i The distance between the included nodes;

[0272] Module M2.2.3: Represents the embedding of decision rule r as the mean of the embeddings of all condition nodes within the rule subset, by narrowing down r and t. r Distance, update the loss function:

[0273]

[0274]

[0275] Among them, t r For the result node of rule r, the embedding representation r of decision rule r is obtained by calculation. Transpose it.

[0276] The module M3.1 includes the following sub-modules:

[0277] Module M3.1.1: For each patient p i Define the patient's physical indicator data as v j , For patient p i All body metrics data v j The set; Matching with all nodes in V yields the condition node a, which is related to the decision rule. j The set M i M i The representations of the condition nodes are weighted and fused to obtain the vector representation of the medical guideline data channel. and:

[0278]

[0279] M i ={…,a j ,…};

[0280]

[0281] Module M3.1.2: For each patient p in the memory set m All exist

[0282] Among them, t m Equivalent to It is patient p m The treatment plan is matched with V to obtain the result node embedding; s′ i,m Indicates patient p iand patient p m The cosine similarity is used to standardize the similarity, making the patient p i With candidate set S i The sum of the similarities between patients is 1, thus obtaining patient p. i With patient p m Similarity s i,m ,and:

[0283]

[0284]

[0285] Among them, S i It is the set of K patients with the highest similarity in the memory set, where K is adjusted according to the dataset. The treatment plans of the patients with the smallest score differences are combined to obtain p. i Neighbor passage indicates

[0286]

[0287] Module M3.1.3: Preprocesses patient physical indicator data, using a data-driven approach and incorporating a DPCNN model to obtain channel representations of the patient physical indicator data.

[0288]

[0289] Here, preproc(·) represents the preprocessing operation and makes custom adjustments based on the real data; concat(·) represents concatenating the patient's attribute vectors into a matrix, which is then passed through a fully connected layer fc1 to obtain the patient's embedded representation.

[0290] The module M3.2 includes the following sub-modules:

[0291] Module M3.2.1: Based on the fully connected layer fc2, it adaptively fuses the representations of three channels: medical guideline data, historical case data, and patient physical indicator data.

[0292]

[0293] Module M3.2.2: Based on the fully connected layer fc3, the patient p i The fusion vector representation is mapped to a probability vector, and the input vector is mapped to the range (0,1) through the Sigmoid activation function σ to obtain the patient p. i and the predicted probability vector of all treatment options y i,j For patient p i Treatment plan t j The true label, For the corresponding probability values, train the model using the cross-entropy loss function:

[0294]

[0295]

[0296] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0297] In the description of this application, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0298] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for recommending multi-channel information fusion schemes, characterized in that, include: Step S1: Collect relevant data; The data includes medical guideline data, historical case data, and patient physical indicator data; Step S2: Construct a medical guideline hypergraph and obtain hypergraph node representations; Step S3: Process the relevant data, score the different solutions, and output the scoring results; Step S2 includes using the patient's physical index data as a condition node in the hypergraph and using the scheme corresponding to the patient's physical index data as a result node in the hypergraph; the condition node and the result node constitute a decision rule. Step S2 includes the following sub-steps: Step S2.1: Process the condition nodes and result nodes; Step S2.2: Model the condition nodes and result nodes and complete the training of node representations; Step S2.1 includes the following sub-steps: Step S2.1.1: For the set Each node in Perform random initialization to obtain the initial embedding representation of the nodes. ;in, Represents vector dimension, nodes Represents condition nodes or result node ,and , ; Step S2.1.2: Use Capture the hypergraph structure of medical guidelines, and associate it with each node. The structural contributions of adjacent nodes are: ； in, It is a node All neighboring nodes, Representative node and adjacent nodes The weight of the edges between them. and These are nodes and The degree, It is trainable Transpose of the weight parameter matrix; Step S2.1.3: Strengthen the connections between condition nodes, and... As a query in the attention mechanism, construct ,and: ； ； ； in, For condition nodes Embedded representation, It is a set of condition nodes; Representative condition node The set of neighboring nodes, The embedding representations of the nodes form a vector matrix. , Represents a vector matrix transpose; Indicates use Normalization function, The result node after attention mechanism update Embedded representation; This indicates taking the average value; It is a condition node The embedded representation incorporates information from second-order neighbor nodes; The embedded representation of the result node is obtained in the same way. : ； ； ； in, Representative condition node The set of neighboring nodes, It is its embedded representation, Indicates matrix transpose; It is the result node The set of neighboring nodes; Represents the result node in the set Embedded representation; It is a condition node Embedded representation after attention mechanism and normalization It is the result node Vector representation of; Step S2.1.4: According to In a path-based attention mechanism, the initial representations of condition nodes and result nodes are updated through a weighted combination: ； in, , These are hyperparameters, used Representation of condition node embedding Or result node embedding representation .

2. The recommendation method for a multi-channel information fusion scheme according to claim 1, characterized in that, In step S2, the condition node is defined as follows: The result node is SuperMap ; in, This represents the set of all condition nodes and all result nodes, where condition nodes and result nodes belonging to the same decision rule are connected by edges. Hypergraph The set of all edges in the set; for each decision rule Its related node set is ; Let represent the set of nodes for all decision rules, and .

3. The recommendation method for a multi-channel information fusion scheme according to claim 2, characterized in that, Step S3 includes the following sub-steps: Step S3.1: Process patient physical indicator data; Step S3.2: Perform adaptive multi-channel information fusion.

4. The recommendation method for a multi-channel information fusion scheme according to claim 1, characterized in that, Step S2.2 includes the following sub-steps: Step S2.2.1: Based on contrastive learning, model the nodes, taking condition nodes and result nodes belonging to the same decision rule subgraph as positive samples, and nodes not belonging to the same decision rule as negative samples; the type of nodes in the negative samples is not limited. ； ； Among them, the set of positive sample pairs is The set of negative sample pairs is ; For positive sample pairs, Negative sample pairs; Step S2.2.2: Calculate the nodes using the dot product method. and distance The node representation is trained using the cross-entropy loss function: ； ； in, For nodes The transpose of the embedded representation. Representing the dot product of vectors, we obtain the distance. Are they positive or negative sample pairs? For the corresponding tag; when When it is a positive sample pair, When it is 1, When it is a negative sample pair, =0; for The distance between the included nodes; Step S2.2.3: Decision rules Embedding is represented as the mean of the embeddings of all condition nodes within the rule subset, by narrowing down. and Distance, update the loss function: ； ； in, For rules The result node is used to calculate the decision rule. Embedded representation , Transpose it.

5. The recommendation method for a multi-channel information fusion scheme according to claim 3, characterized in that, Step S3.1 includes the following sub-steps: Step S3.1.1: For each patient Define patient physical indicator data as , For patients All body indicator data The set; and Match all nodes to obtain the condition nodes related to the decision rule. set ,Will The representations of the condition nodes are weighted and fused to obtain the vector representation of the medical guideline data channel. ,and: ； ； ； Step S3.1.2: For each patient in the memory set All exist ; in, Equivalent to The patient Treatment plan and The matching result node embedding; Indicates the patient and patients The cosine similarity is used to standardize the similarity, so that patients... With candidate set The sum of the similarities between patients is 1, thus obtaining the patient's... With patients Similarity ,and: ； ； in, It is the memory concentration with the highest similarity A collection of patients, Adjustments were made based on the dataset, combining the treatment plans of patients with the smallest score differences to obtain... Neighbor passage indicates : ； Step S3.1.3: Preprocess the patient's physical indicator data, based on data-driven methods and combined with... The model represents the patient's physical indicator data channels. : ； in, This indicates preprocessing operations, with custom adjustments made based on the actual data; This means concatenating the patient's attribute vectors into a matrix, which is then passed through a fully connected layer. This yields the patient's embedded representation.

6. The method for recommending a multi-channel information fusion scheme according to claim 3, characterized in that, Step S3.2 includes the following sub-steps: Step S3.2.1: Based on the fully connected layer The representations of three channels—medical guideline data, historical case data, and patient physical indicator data—are adaptively fused. ； Step S3.2.2: Based on the fully connected layer Patient The fusion vector representation is mapped to a probability vector and activated by the Sigmoid activation function. Mapping the input vector to the range (0,1) yields the patient's... And the predicted probability vector of all treatment options For patients Treatment plan The true label, For the corresponding probability values, train the model using the cross-entropy loss function: ； 。

Images