Network representation learning method across medical data sources

Abstract

The present disclosure proposes a network representation learning method across medical data sources, including: S1, generating medical network data including a source network and a target network; S2, randomly sampling a set number of nodes from the source network and the target network; S3, obtaining an L-layer neural network, and calculating the structural features and expression features of the source network and the target network for each layer, and calculating the distance loss between the network features of the source network and the target network; S4, obtaining the output of the source network in the L-layer neural network, and calculating the loss value according to the classification loss and the distance loss, and updating the parameters of the algorithm according to the back propagation algorithm; S5, repeating steps S2-S4 until the entire algorithm converges, so that the accuracy of the algorithm for disease classification does not increase within multiple iterations. The present disclosure takes into account the problem of inconsistent data distribution between different hospital data sources, and has a broad application space by extracting the structural information of the network and the node attribute information, and minimizing the feature distance to compensate for the information loss.

Description

Technical Field

[0001] The present disclosure belongs to the field of medical technology, and in particular, relates to a network representation learning method across medical data sources. Background Art

[0002] As deep learning technology matures, it is becoming possible to apply it in medical scenarios. By modeling patient cases, deep learning technology can help doctors infer the patient's condition to promote the development of the medical industry. In medical scenarios, our research objects often include patients, symptoms, diseases, and drugs, and there are complex logical relationships between them. This relationship is difficult to describe directly through a simple deep fully connected network. At the same time, considering that in existing deep learning technologies, the network embedding representation learning algorithm can intuitively represent the connection relationship between objects and has strong reasoning ability. Therefore, we can use the technology of network embedding representation learning algorithm to model and solve it from the perspective of the network. Specifically, we can regard each object involved as a node in a network. If there are edges between nodes, it means that there is a certain relationship between them. For example, the relationship between nodes in the network can be the treatment relationship between diseases and drugs, the manifestation relationship between diseases and symptoms, and so on. By abstracting the complex medical relationship into a network, and then extracting information from the network through the network embedding expression algorithm, the role of inferring a condition can be achieved.

[0003] Existing network embedding representation algorithms can be mainly divided into two categories: one is the direct representation learning algorithm. Given a target network, the direct representation algorithm directly optimizes the expression vector of each node through the node attributes and network relationships, such as DeepWalk and Node2vec. The other is the inductive representation learning algorithm. The inductive representation algorithm often learns a mapping function. As long as the attributes of the input node and its neighbors are given, the expression vector of the node can be inferred through the mapping function, such as GCN, GraphSAGE and GAT.

[0004] In real medical scenarios, medical data often comes from different hospitals. This leads to inconsistent distribution of patient data in different hospitals. For example, for a cold, for people in the south, it may be a cold caused by the body being cold, while for people in the north, it may be a hot cold caused by indoor heating. For another example, for the same disease, there may be multiple drugs with the same pathology that can be used to treat it, and the medication habits of different hospitals may also be different. It is precisely because of these real reasons that the data distribution between different hospitals is different. This leads to the fact that if you directly use data from multiple different medical data sources without considering the inconsistency of their actual distribution, the knowledge learned is likely to be biased when you want to use as much data as possible to help the model learn.

[0005] However, existing algorithms cannot solve the problem of network representation learning across medical data sources. Specifically:

[0006] (1) For the direct inference algorithm, since the direct inference algorithm directly optimizes the node expression vectors in the network composed of a certain hospital data, it cannot directly infer the expression vectors of the nodes in the new network composed of another hospital data. Therefore, the direct inference algorithm does not have any available knowledge to be used for network learning across medical data sources.

[0007] (2) For inductive algorithms, although they consider learning a mapping function between node attributes and structural information when modeling, which can naturally perform network inference across medical data sources, inductive algorithms do not take into account that the data distribution between networks is different. The patterns or knowledge summarized from the medical network of one hospital may not be well applicable to the medical network of another hospital. Therefore, inductive algorithms also have certain defects in the problem of network representation learning across medical data sources.

[0008] Therefore, the existing technology needs to be improved.

[0009] The above background technology content is only used to help understand the present disclosure, and does not represent an admission or recognition that any of the content mentioned is part of the common knowledge relative to the present disclosure. Summary of the invention

[0010] To solve the above technical problems, the present disclosure proposes a network representation learning method across medical data sources.

[0011] According to one aspect of the embodiment of the present disclosure, a network representation learning method across medical data sources is disclosed, including:

[0012] S1, generating medical network data including a source network and a target network, wherein the source network is generated by medical records of a certain hospital, and the target network is generated by medical records of another hospital different from the hospital, and the medical network data includes medical record information of patients, and constructs network relationships between symptoms, diseases, drugs, and diagnostic methods;

[0013] S2, randomly sampling a set number of nodes from the source network and the target network respectively, where the number of collected nodes is related to the degree of the medical network;

[0014] S3, obtaining an L-layer neural network from step S2, and calculating the structural features and expression features of the source network and the target network for each layer, and calculating the distance loss between the network features of the source network and the target network;

[0015] S4, obtains the output of the source network in the L-layer neural network from S3, calculates the loss value based on the classification loss and distance loss, and updates the algorithm parameters according to the back propagation algorithm;

[0016] S5, repeating steps S2-S4 until the entire algorithm converges, so that the accuracy of the algorithm for disease classification does not increase within multiple iterations.

[0017] In another embodiment of the network representation learning method across medical data sources disclosed in the present invention, the step S3 obtains an L-layer neural network from step S2, and calculates the structural features and expression features of the source network and the target network for each layer, and calculates the distance loss between the network features of the source network and the target network, including:

[0018] S30, inputting the node features of the source network and the target network into the neural network of the L layer;

[0019] S31, in each layer of the L-layer neural network, the node feature expression vector of each network passes through a message routing module to obtain the structural feature, and the structural feature passes through a message aggregation module to obtain a new expression feature vector of the current node;

[0020] S32, calculating the distance loss value between the node features of the current layer from the source network and the target network through the network alignment module across the medical data sources;

[0021] S33, repeat steps S31 to S32 for L times to obtain the node feature vectors of the final source network and the target network and the L-layer accumulated structural feature distance loss and expression feature distance loss.

[0022] In another embodiment of the network representation learning method across medical data sources disclosed in the present invention, in step S31, in each layer of the L-layer neural network, the node feature expression vector of each network is passed through a message routing module to obtain a structural feature, and the structural feature is passed through a message aggregation module to obtain a new expression feature vector of the current node, including:

[0023] The message routing module of each layer is represented as:

[0024] ;

[0025] ;

[0026] In the formula, For Node In an L-layer neural network The structural feature vector of the layer, is the first layer in the L-layer neural network The expression feature vectors of the source network and the target network of the layer, the expression feature vector of the 0th layer is composed of the original feature vector of the node express, For the The parameter matrix involved in the message routing module of the layer, For the The parameter matrix involved in the message routing module of the layer, is the activation function, is the direct concatenation operation of two vectors, For Node The set of directly connected neighbors, For Node Transmit to Node The message weight;

[0027] The message aggregation module of each layer is expressed as:

[0028] ;

[0029] ;

[0030] In the formula, and is the parameter matrix involved in the message aggregation module, is a vector representing the node aggregation level.

[0031] In another embodiment of the network representation learning method across medical data sources disclosed in the present invention, the step S32, calculating the distance loss value between the node features of the current layer from the source network and the target network through the network alignment module across medical data sources, includes:

[0032] The structural feature distance of each layer is:

[0033] ;

[0034] In the formula, is the structural feature vector of the source network and the target network and The distribution of is a distance function used to calculate the structural feature vector and The expected distance;

[0035] The expression feature distance loss of each layer is:

[0036] ;

[0037] In the formula, Express feature vectors for nodes in the source and target networks and The distribution of is a distance function used to calculate the node expression feature vector and expected distance.

[0038] In another embodiment of the network representation learning method across medical data sources disclosed in the present invention, the step S33, repeating steps S31 to S32 for L times, obtaining the node feature vectors of the final source network and the target network and the L-layer accumulated structural feature distance loss and expression feature distance loss includes:

[0039] The accumulated structural feature distance loss of the L layer is:

[0040] ;

[0041] The accumulated expression feature distance loss of layer L is:

[0042] .

[0043] Compared with the prior art, the present invention has the following advantages:

[0044] The network representation learning method across medical data sources disclosed in this paper takes into account the problem of inconsistent data distribution between data sources of different hospitals. By minimizing the feature distance to compensate for the information loss caused by the inconsistency, it can make full use of multi-source medical data and has broad application space in the medical field. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A flowchart of an embodiment of the network representation learning method across medical data sources proposed in the present disclosure;

[0046] Figure 2 A flowchart of another embodiment of the network representation learning method across medical data sources proposed in the present disclosure. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solution and advantages of the embodiments of the present disclosure clearer, the technical solution in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, not all of the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present disclosure.

[0048] The following is a more detailed description of a network representation learning method across medical data sources provided by the present disclosure in conjunction with the accompanying drawings and embodiments.

[0049] Figure 1 A flowchart of an embodiment of the network representation learning method across medical data sources proposed in the present disclosure, such as Figure 1 As shown, the network representation learning method across medical data sources:

[0050] S1, generating medical network data including a source network and a target network, wherein the source network is generated by medical records of a certain hospital, and the target network is generated by medical records of another hospital different from the hospital, and the medical network data includes medical record information of patients, and constructs network relationships between symptoms, diseases, drugs, and diagnostic methods;

[0051] S2, randomly sample a set number of nodes from the source network and the target network respectively. The number of collected nodes is related to the degree of the medical network. Taking disease diagnosis as an example, it is necessary to iterate L times from each disease node to collect its neighboring nodes, and use the node attributes corresponding to the collected nodes as the input data of the algorithm;

[0052] S3, obtaining an L-layer neural network from step S2, and calculating the structural features and expression features of the source network and the target network for each layer, and calculating the distance loss between the network features of the source network and the target network;

[0053] S4, obtains the output of the source network in the L-layer neural network from S3, calculates the loss value based on the classification loss and distance loss, and updates the algorithm parameters according to the back propagation algorithm;

[0054] S5, repeating steps S2-S4 until the entire algorithm converges, so that the accuracy of the algorithm for disease classification does not increase within multiple iterations.

[0055] Figure 2A flowchart of another embodiment of the network representation learning method across medical data sources proposed in the present disclosure, such as Figure 2 As shown, in step S3, a neural network with L layers is obtained from step S2, and the structural features and expression features of the source network and the target network are calculated for each layer, and the distance loss between the network features of the source network and the target network is calculated, including:

[0056] S30, inputting the node features of the source network and the target network into the neural network of the L layer;

[0057] S31, in each layer of the L-layer neural network, the node feature expression vector of each network passes through a message routing module to obtain the structural feature, and the structural feature passes through a message aggregation module to obtain a new expression feature vector of the current node;

[0058] S32, calculating the distance loss value between the node features of the current layer from the source network and the target network through the network alignment module across the medical data sources;

[0059] S33, repeat steps S31 to S32 for L times to obtain the node feature vectors of the final source network and the target network and the L-layer accumulated structural feature distance loss and expression feature distance loss.

[0060] In step S31, in each layer of the L-layer neural network, the node feature expression vector of each network is passed through a message routing module to obtain a structural feature, and the structural feature is passed through a message aggregation module to obtain a new expression feature vector of the current node, including:

[0061] The message routing module of each layer is represented as:

[0062] ;

[0063] ;

[0064] In the formula, For Node In an L-layer neural network The structural feature vector of the layer, is the first layer in the L-layer neural network The expression feature vectors of the source network and the target network of the layer, the expression feature vector of the 0th layer is composed of the original feature vector of the node express, For the The parameter matrix involved in the message routing module of the layer, For the The parameter matrix involved in the message routing module of the layer, is the activation function, is the direct concatenation operation of two vectors, For Node The set of directly connected neighbors, For Node Transmit to Node The message weight;

[0065] The message aggregation module of each layer is expressed as:

[0066] ;

[0067] ;

[0068] In the formula, and is the parameter matrix involved in the message aggregation module, is a vector representing the node aggregation level.

[0069] The step S32, calculating the distance loss value between the node features of the source network and the target network at the current layer through the network alignment module across the medical data source, includes:

[0070] The structural feature distance of each layer is:

[0071] ;

[0072] In the formula, is the structural feature vector of the source network and the target network and The distribution of is a distance function used to calculate the structural feature vector and The expected distance;

[0073] The expression feature distance loss of each layer is:

[0074] ;

[0075] In the formula, Express feature vectors for nodes in the source and target networks and The distribution of is a distance function used to calculate the node expression feature vector and expected distance.

[0076] The step S33, repeating steps S31 to S32 for L times, obtaining the node feature vectors of the final source network and the target network and the L-layer accumulated structural feature distance loss and expression feature distance loss includes:

[0077] The accumulated structural feature distance loss of the L layer is:

[0078] ;

[0079] The accumulated expression feature distance loss of layer L is:

[0080] .

[0081] For those skilled in the art, it is obvious that the embodiments of the present disclosure are not limited to the details of the above exemplary embodiments, and the embodiments of the present disclosure can be implemented in other specific forms without departing from the spirit or basic features of the embodiments of the present disclosure. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive, and the scope of the embodiments of the present disclosure is limited by the attached claims rather than the above description, so it is intended to include all changes that fall within the meaning and scope of the equivalent elements of the claims in the embodiments of the present disclosure. Any figure mark in the claims should not be regarded as limiting the claims involved. In addition, it is obvious that the word "including" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units, modules or devices stated in the system, device or terminal claims can also be implemented by the same unit, module or device through software or hardware. The words first, second, etc. are used to indicate names, and do not indicate any specific order.

[0082] Finally, it should be noted that the above implementation modes are only used to illustrate the technical solutions of the embodiments of the present disclosure and are not intended to limit them. Although the embodiments of the present disclosure are described in detail with reference to the above preferred implementation modes, those skilled in the art should understand that the technical solutions of the embodiments of the present disclosure may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the embodiments of the present disclosure.

Claims

1. A network representation learning method across medical data sources, characterized in that: include: S1, generating medical network data including a source network and a target network, wherein the source network is generated by medical records of a certain hospital, and the target network is generated by medical records of another hospital different from the hospital, and the medical network data includes medical record information of patients, and constructs network relationships between symptoms, diseases, drugs, and diagnostic methods; S2, randomly sampling a set number of nodes from the source network and the target network respectively, where the number of collected nodes is related to the degree of the medical network; S3, obtaining an L-layer neural network from step S2, and calculating the structural features and expression features of the source network and the target network for each layer, and calculating the distance loss between the network features of the source network and the target network; S4, obtains the output of the source network in the L-layer neural network from S3, calculates the loss value based on the classification loss and distance loss, and updates the algorithm parameters according to the back propagation algorithm; S5, repeating steps S2-S4 until the entire algorithm converges, so that the accuracy of the algorithm for disease classification does not increase within multiple iterations.

2. The network representation learning method across medical data sources according to claim 1, characterized in that: In step S3, a neural network with L layers is obtained from step S2, and the structural features and expression features of the source network and the target network are calculated for each layer, and the distance loss between the network features of the source network and the target network is calculated, including: S30, inputting the node features of the source network and the target network into the neural network of the L layer; S31, in each layer of the L-layer neural network, the node feature expression vector of each network passes through a message routing module to obtain the structural feature, and the structural feature passes through a message aggregation module to obtain a new expression feature vector of the current node; S32, calculating the distance loss value between the node features of the current layer from the source network and the target network through the network alignment module across the medical data sources; S33, repeat steps S31 to S32 for L times to obtain the node feature vectors of the final source network and the target network and the L-layer accumulated structural feature distance loss and expression feature distance loss.

3. The network representation learning method across medical data sources according to claim 2, characterized in that: In step S31, in each layer of the L-layer neural network, the node feature expression vector of each network is passed through a message routing module to obtain a structural feature, and the structural feature is passed through a message aggregation module to obtain a new expression feature vector of the current node, including: The message routing module of each layer is represented as: ； ； In the formula, For Node In an L-layer neural network The structural feature vector of the layer, is the first layer in the L-layer neural network The expression feature vectors of the source network and the target network of the layer, the expression feature vector of the 0th layer is composed of the original feature vector of the node express, For the The parameter matrix involved in the message routing module of the layer, For the The parameter matrix involved in the message routing module of the layer, is the activation function, is the direct concatenation operation of two vectors, For Node The set of directly connected neighbors, For Node Transmit to Node The message weight; The message aggregation module of each layer is expressed as: ； ； In the formula, and is the parameter matrix involved in the message aggregation module, is a vector representing the node aggregation level.

4. The network representation learning method across medical data sources according to claim 3, characterized in that: The step S32, calculating the distance loss value between the node features of the source network and the target network at the current layer through the network alignment module across the medical data source, includes: The structural feature distance of each layer is: ； In the formula, is the structural feature vector of the source network and the target network and The distribution of is a distance function used to calculate the structural feature vector and The expected distance; The expression feature distance loss of each layer is: ； In the formula, Express feature vectors for nodes in the source and target networks and The distribution of is a distance function used to calculate the node expression feature vector and expected distance.

5. The network representation learning method across medical data sources according to claim 4, characterized in that: The step S33, repeating steps S31 to S32 for L times, obtaining the node feature vectors of the final source network and the target network and the L-layer accumulated structural feature distance loss and expression feature distance loss includes: The accumulated structural feature distance loss of the L layer is: ； The accumulated expression feature distance loss of layer L is: 。

Images