Methods, devices and readable storage media for pre-filling medical information
By modeling patient diagnosis and treatment information as a non-Euclidean graph structure and using graph convolutional networks to extract semantic features, the problem of doctors manually filling in diagnosis and treatment information is solved, achieving efficient and accurate pre-filling of diagnosis and treatment information, reducing costs and improving flexibility.
Patent Information
- Application Number
- CN202211007679.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-08-22
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2042-08-22
AI Technical Summary
Manually filling in patient information during remote consultations is time-consuming and laborious for doctors, resulting in low efficiency. Furthermore, the template pre-filling method suffers from high design costs, maintenance difficulties, and insufficient flexibility.
Patient diagnosis and treatment information is modeled as a non-Euclidean graph structure. Semantic features are extracted using graph convolutional networks and matched with a pre-stored semantic representation library to achieve pre-filling of diagnosis and treatment information.
It improves the efficiency and accuracy of doctors filling in medical information, reduces labor costs and maintenance difficulties, and enables flexible filling of personalized information.
Smart Images

Figure CN115458102B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of information processing technology, and in particular to a method, apparatus and readable storage medium for pre-filling medical information. Background Technology
[0002] With the rapid development of digital and intelligent healthcare, telemedicine systems have emerged to accelerate the medical process for patients across regions and departments, making it more convenient, faster, and more economical for them. However, due to the electronic information requirements of various medical institutions, doctors need to fill in a large amount of patient-related medical information during remote consultations, which consumes a significant amount of their time and energy, making it laborious and prone to missed or incorrect diagnoses.
[0003] Therefore, how to assist doctors in quickly completing the information entry for diagnosis and treatment, and shift the focus of doctors' work back to the diagnosis and treatment itself, so as to achieve rapid and high-quality diagnosis and treatment, is an important problem that telemedicine systems urgently need to solve. Summary of the Invention
[0004] In view of this, this application provides a method, device and readable storage medium for pre-filling medical information, which solves the problem of reduced medical efficiency caused by doctors manually filling in a large amount of patient-related medical information in related technologies.
[0005] In a first aspect, embodiments of this application provide a method for pre-filling medical information, including:
[0006] Obtain the diagnosis and treatment information graph structure corresponding to the patient's diagnosis and treatment information. The diagnosis and treatment information graph structure includes edges representing the basic attributes in the diagnosis and treatment information and nodes connected to the edges that represent the corresponding values of the basic attributes.
[0007] The diagnostic information graph structure is transformed into an edge feature matrix, a node feature matrix, and an adjacency matrix corresponding to the diagnostic information graph structure.
[0008] The edge feature matrix, node feature matrix, and adjacency matrix are input into the graph convolutional network, and the first semantic feature that expresses the structure of the diagnostic information graph is output.
[0009] Based on the first semantic feature, the corresponding target diagnosis and treatment information is extracted from the pre-stored semantic expression library, and the patient's diagnosis and treatment information is pre-filled.
[0010] The diagnostic information pre-filling method according to the embodiments of this application may also have the following additional technical features:
[0011] In the above technical solution, the diagnostic information graph structure is a non-Euclidean graph structure.
[0012] In any of the above technical solutions, the transformation of the diagnosis and treatment information graph structure into an edge feature matrix, a node feature matrix, and an adjacency matrix corresponding to the diagnosis and treatment information graph structure includes: performing one-hot encoding on the edges and nodes in the diagnosis and treatment information graph structure respectively, and constructing the edge feature matrix, the node feature matrix, and the adjacency matrix corresponding to the diagnosis and treatment information graph structure, wherein the one-hot encoding lengths of the edge feature matrix and the node feature matrix are the same or different.
[0013] In any of the above technical solutions, before performing one-hot encoding on the edges and nodes in the diagnosis and treatment information graph structure, the method further includes: normalizing the diagnosis and treatment information, wherein the normalization process is to standardize the expression of the basic attributes of the diagnosis and treatment information and / or the values corresponding to the basic attributes.
[0014] In any of the above technical solutions, the graph convolutional network is used to: obtain the distributed representation vectors of all edges and all nodes in the diagnosis and treatment information graph structure based on the edge feature matrix, node feature matrix, and adjacency matrix; sum the distributed representation vectors of all edges to obtain the edge aggregated representation vector, and sum the distributed representation vectors of all nodes to obtain the node aggregated representation vector; concatenate the edge aggregated representation vector and the node aggregated representation vector to obtain the original distributed representation vector of the diagnosis and treatment information; and convert the original distributed representation vector into the first semantic feature corresponding to the diagnosis and treatment information graph structure through a feedforward neural network; wherein, the first semantic feature is the first distributed representation vector, and the first distributed representation vector is:
[0015] γ=ReLU(ReLU(ReLU(xW 0 )xW 1 W n )
[0016] γ is the first distributed vector, x is the original distributed representation vector, ReLU is the non-linear activation function, W is the learnable parameter of each layer of the feedforward neural network, and n is the number of layers of the feedforward neural network.
[0017] In any of the above technical solutions, the distributed representation vectors of all edges and all nodes in the diagnosis and treatment information graph structure are obtained based on the edge feature matrix, node feature matrix, and adjacency matrix. This includes: encoding the latent vector of the l-th layer corresponding to the node pair into a latent vector of the (l+1)-th layer as the message vector participating in the computation in the graph convolutional network. Here, the node pair is any one of multiple node pairs corresponding to the central node of the diagnosis and treatment information graph structure. The node pair includes the central node, the adjacent nodes of the central node, and the edge between the central node and the adjacent nodes. The message vectors of the multiple node pairs are aggregated to obtain aggregated features. The aggregated features are updated to the central node to obtain the distributed vector representation of the central node at the (l+1)-th layer.
[0018] In any of the above technical solutions, based on the first semantic feature, the corresponding target diagnosis and treatment information is extracted from the pre-stored semantic expression library, and the patient's diagnosis and treatment information is pre-filled, including: calculating the similarity scores between the second semantic features corresponding to multiple pre-stored patient graph structures in the semantic expression library and the first semantic feature, respectively, to obtain multiple similarity scores; performing standardized calculation on the multiple similarity scores, and generating a target score vector based on the standardized multiple similarity scores, wherein the components in the target score vector are the standardized similarity scores, and the second semantic feature is used to express the pre-stored patient graph structure; determining the target diagnosis and treatment information based on a preset threshold and the target score vector, and pre-filling the patient's diagnosis and treatment information based on the target diagnosis and treatment information.
[0019] In any of the above technical solutions, the target diagnosis and treatment information is determined according to a preset threshold and a target scoring vector, and the patient's diagnosis and treatment information is pre-filled according to the target diagnosis and treatment information, including: determining the top N similarity scores that are greater than or equal to the preset threshold in the target scoring vector, where N is a positive integer; determining the pre-stored patient index account corresponding to the top N similarity scores; searching for the target diagnosis and treatment information corresponding to the pre-stored patient index account, and pre-filling the patient's diagnosis and treatment information according to the target diagnosis and treatment information.
[0020] Secondly, embodiments of this application provide a diagnostic information pre-filling device, comprising:
[0021] The graph structure modeling module is used to obtain the diagnosis and treatment information graph structure corresponding to the patient's diagnosis and treatment information. The diagnosis and treatment information graph structure includes edges representing the basic attributes in the diagnosis and treatment information and nodes connected to the edges that represent the corresponding values of the basic attributes.
[0022] The graph structure transformation module is used to transform the diagnosis and treatment information graph structure into an edge feature matrix, a node feature matrix, and an adjacency matrix corresponding to the diagnosis and treatment information graph structure.
[0023] The semantic feature learning module is used to input the edge feature matrix, node feature matrix and adjacency matrix into the graph convolutional network and output the first semantic feature that expresses the structure of the diagnosis and treatment information graph.
[0024] The pre-filling module is used to extract the corresponding target diagnosis and treatment information from the pre-stored semantic expression library based on the first semantic feature, and pre-fill the patient's diagnosis and treatment information.
[0025] Thirdly, embodiments of this application provide a readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the method as described in the first aspect.
[0026] This application provides a method, apparatus, and readable storage medium for pre-filling medical information. By acquiring the medical information graph structure corresponding to the medical information related to the current patient, the medical information graph structure is transformed into an edge feature matrix, a node feature matrix, and an adjacency matrix corresponding to the medical information graph structure. Then, the edge feature matrix, node feature matrix, and adjacency matrix are input into a graph convolutional network, outputting a first semantic feature corresponding to the medical information graph structure. This first semantic feature is compared with the semantic features of other existing patient graph structures in a pre-stored semantic expression library to obtain the target medical information to be pre-filled. This achieves comprehensive and fine-grained capture of the potential semantic features of patient medical information, improving the efficiency of doctors filling in medical information and enhancing the accuracy and completeness of the medical information.
[0027] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description
[0028] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:
[0029] Figure 1 A flowchart illustrating the diagnostic information pre-filling method according to an embodiment of this application is shown;
[0030] Figure 2 This paper illustrates a schematic diagram of the overall architecture for pre-filling diagnostic and treatment information according to an embodiment of this application.
[0031] Figure 3 This illustration shows a schematic diagram of the encoding of the diagnostic information graph structure according to an embodiment of this application;
[0032] Figure 4 A schematic diagram of an MCN graph convolutional network according to an embodiment of this application is shown;
[0033] Figure 5 A schematic block diagram of a medical information pre-filling device according to an embodiment of this application is shown;
[0034] Figure 6 A schematic block diagram of an electronic device according to an embodiment of this application is shown. Detailed Implementation
[0035] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.
[0036] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.
[0037] Filling out patient medical information constitutes a significant portion of a doctor's workload during remote consultations. It is time-consuming, labor-intensive, and prone to errors, severely reducing the doctor's original clinical work and impacting the efficiency of diagnosis and treatment.
[0038] In related technologies, the template pre-filling method is usually used. That is, the template information that has been researched in advance is entered into the telemedicine system, and the corresponding template is selected when entering business information. The patient's medical information is then added to or modified based on the template.
[0039] While the template pre-filling method saves doctors' information entry time and alleviates their workload to some extent, it still has the following problems: (1) Template design cost. Different hospitals and departments have different requirements for diagnosis and treatment content and different entry methods. The telemedicine system needs to design and configure corresponding templates for each implementation unit, which results in high labor costs. (2) Template version maintenance. As the system accumulates operating time, a large number of templates will be generated. At the same time, due to the diversity of templates, if the template content is modified or updated, the implementation workload will be extremely heavy and the maintenance cost will be extremely high. (3) Flexibility. With the template pre-filling method, only a single configured template can be selected, and it mainly provides information at the framework level. It lacks filling based on personalized information and is not flexible enough.
[0040] This application models patient treatment information as a medical information graph structure, extracts semantic representations of the medical information graph structure through a graph convolutional network, obtains matching pre-filled information based on the semantic representation, and then backfills the information, thereby realizing the pre-filling of treatment forms. This application is data-driven, dynamically matching information according to the current patient treatment information, which not only improves the efficiency of doctors in filling out treatment information but also enhances the accuracy and completeness of the information, greatly reducing the workload of doctors and allowing them to devote more energy to actual medical treatment.
[0041] The following description, in conjunction with the accompanying drawings, details the diagnostic information pre-filling method, diagnostic information pre-filling device, and readable storage medium provided in this application through specific embodiments and application scenarios.
[0042] This application provides a method for pre-filling medical information, such as... Figure 1 As shown, the method includes:
[0043] Step 101: Obtain the diagnosis and treatment information graph structure corresponding to the patient's diagnosis and treatment information. The diagnosis and treatment information graph structure includes edges representing basic attributes in the diagnosis and treatment information and nodes connected to the edges that represent the corresponding values of the basic attributes.
[0044] Step 102: Transform the diagnosis and treatment information graph structure into an edge feature matrix, a node feature matrix, and an adjacency matrix corresponding to the diagnosis and treatment information graph structure;
[0045] Step 103: Input the edge feature matrix, node feature matrix and adjacency matrix into the graph convolutional network, and output the first semantic feature that expresses the structure of the diagnosis and treatment information graph;
[0046] Step 104: Based on the first semantic feature, extract the corresponding target diagnosis and treatment information from the pre-stored semantic expression library and pre-fill the patient's diagnosis and treatment information.
[0047] In this embodiment, the system first acquires relevant patient information, such as basic patient information, symptoms, and signs. Diagnostic information can be obtained through a doctor's verbal description or from forms or reports that reflect medical information. This information is then modeled as a medical information graph structure (i.e., a non-Euclidean graph structure) according to business rules. A graph convolutional network is then used to output a distributed representation of this medical information graph structure, yielding its first semantic feature. This allows for comprehensive and fine-grained capture of the latent semantic features of the patient's medical information. Next, the first semantic feature is compared with the semantic features of other existing patient graph structures in a pre-stored semantic representation library to derive the target medical information. Finally, the target medical information is retrieved and returned to the front end for text pre-filling of the patient's medical information.
[0048] This application proposes a method for pre-filling medical information based on graph convolutional networks. Driven by existing patient medical information, the method models structured form data into a medical information graph structure, outputs its semantic features through a graph convolutional network, and then obtains matching pre-filled information by comparing it with existing patient medical information. This information is then backfilled, thereby achieving pre-filling of medical forms, improving the efficiency of doctors filling in medical information, and enhancing the accuracy and completeness of the information.
[0049] Compared to template pre-filling methods in related technologies, the diagnostic information pre-filling method of this application embodiment does not require designing and configuring corresponding templates for each implementation unit, reducing labor and maintenance costs, realizing personalized information filling, and improving the flexibility of information filling.
[0050] Furthermore, the medical information graph structure in this application embodiment can be a non-Euclidean structure, which can represent more related information in medical information.
[0051] In an exemplary embodiment, the overall architecture of the pre-population of diagnostic and treatment information in this application embodiment is as follows: Figure 2 As shown, specifically, the method for pre-filling diagnostic and treatment information in this application embodiment includes:
[0052] I. Patient visit data modeling.
[0053] This stage involves modeling the patient's structured diagnostic and treatment information into a non-Euclidean graph structure. The patient's displayed semantic representation is constructed using a graph approach, primarily comprising three parts:
[0054] (1) Presentation Layer Side Encoding. First, collect the medical information related to the patient's medical form to be filled, such as the patient's basic information, symptoms, signs, etc. Model this key-value information based on attribute-value pairs as a medical information graph structure. The medical information graph structure can be a non-Euclidean graph structure. The basic attributes in the medical information are represented as edges in the medical information graph structure. The basic attributes include at least one of name, signs, and symptoms. Signs include weight, age, and height, etc. The values corresponding to the basic attributes in the medical information are represented as nodes in the medical information graph structure. For example, if the basic attribute is name, the value corresponding to the basic attribute is Zhang San. If the basic attribute is weight, the value corresponding to the basic attribute is 60Kg. The mathematical expression of the medical information graph structure is: G = (V, E), where V is the set of nodes representing the basic attributes, V = {v1, v2, v3, ..., v...} |v| Let |V| be the total number of nodes, and E be the set of edges representing basic attributes, E = {e1, e2, e3, ..., e}. |E|}, where |E| is the total number of edges. For example... Figure 3As shown, the middle node represents the patient ID node, each connection represents a basic attribute, and each branch node represents the value of the basic attribute, thus modeling the correspondence between patient attributes and values.
[0055] (2) Sparse high-dimensional encoding. The medical information graph structure is converted into a sparse high-dimensional numerical representation through one-hot encoding (i.e., one-hot encoding). That is, the edge feature matrix FE, the node feature matrix FV, and the adjacency matrix X corresponding to the diagnosis and treatment information graph structure are constructed respectively, which are used as the input of the subsequent graph convolutional network.
[0056] Furthermore, since edges and nodes express different semantics—edges express role semantics, while nodes express the semantics of the role holder—therefore, as... Figure 3 As shown, this application encodes edges and nodes separately, constructing their respective semantic spaces. The mathematical expression corresponding to the one-hot encoding of a node is v. * ∈(0,1) |V| The mathematical expression for the one-hot encoding of an edge is e * ∈(0,1) |E| The encoding lengths of edges and nodes can be the same or different, but it is preferable to use different encoding lengths to effectively distinguish between edges and nodes.
[0057] like Figure 3 As shown, there is an optional step in this stage: the normalization of medical information. The diversity of terminology in patient medical information can cause ambiguity in the system data. To improve system usability, reduce maintenance complexity, and enhance deep learning efficiency, a normalization operation can be added before one-hot encoding to standardize the magnitude of terms in the medical information, thereby ensuring the accuracy of subsequent similarity scoring. Normalization involves standardizing the basic attributes and / or their corresponding values in the medical information. For example, a patient's symptom is right upper limb pain. Doctor A writes "right arm pain," Doctor B writes "right upper arm pain," Doctor C writes "right arm pain," and Doctor D writes "right arm ache," etc. If these descriptions can be standardized to all be written as "right upper limb pain," the amount of data will be reduced, and accuracy will be improved.
[0058] (3) Database-side encoding. The pre-stored patient-related structured medical information in the pre-stored semantic representation library is modeled as a pre-stored patient graph structure. The basic attributes of the pre-stored patient-related structured medical information are the edges in the pre-stored patient graph structure, and the corresponding values are the nodes in the pre-stored patient graph structure. Database-side encoding is incremental encoding and does not need to be built in real-time with each form request. It is recommended to build it at startup and then according to the scheduled task time.
[0059] II. Distributed Representation.
[0060] This stage involves inputting the adjacency matrix X, node feature matrix FV, and edge feature matrix FE corresponding to the diagnostic information graph structure output from stage one into the graph convolutional network. Through distributed representation learning, the high-dimensional sparse vector is mapped to a low-dimensional space, and the low-dimensional dense numerical vector representation (i.e., the first semantic feature) corresponding to the diagnostic information graph structure is output. This vector can approximately express the potential hidden semantics contained in the diagnostic information graph structure in the high-dimensional algebraic space.
[0061] In one example, the medical information graph structure is non-Euclidean, and the graph convolutional network is an efficient distributed representation learning model designed for non-Euclidean graph structures based on Graph Convolutional Networks (GCN). Figure 4 As shown, this graph convolutional network mainly includes three steps: distributed representation learning of graph nodes and edges, aggregation of nodes and edges, and feature computation.
[0062] (1) Distributed representation learning of graph nodes and edges. This step extracts the distributed representation of nodes and edges in the diagnosis and treatment information graph structure based on the GCN principle. The graph convolutional network includes message passing units M and aggregation units. The update unit U and the message passing unit M are responsible for updating the hidden vectors of the l-th layer of the diagnosis and treatment information graph structure corresponding to the node pairs {v, u, vu} in the diagnosis and treatment information graph structure. The message is encoded as a (Dimn, 1)-dimensional vector, where v, u ∈ V, vu ∈ E, v is the center node, u is the neighboring node of the center node, and the node pairs are determined based on the edge feature matrix, the node feature matrix, and the adjacency matrix. Aggregation Unit The update unit U is responsible for calculating the aggregated features of all first-order adjacent nodes of the central node. The update unit U updates the central node with the calculated aggregated features, thus giving the central node the semantic features inherent in its adjacent nodes and corresponding edges. Unlike GCN, which can only perform distributed representation of nodes, this application incorporates edges E into the design of the message passing unit M. This allows for distributed representation of both nodes and edges, further integrating more semantics. After calculating the message vectors of all layers of the diagnostic information graph structure, the distributed representation vectors of all edges and all nodes are finally obtained.
[0063] The specific calculation formula is as follows:
[0064] 1) The mathematical definition of the message passing unit M is:
[0065]
[0066]
[0067]
[0068]
[0069] Among them, u i Let vu be the i-th adjacent node of the central node v. i Let be the edge between the center node and the i-th adjacent node. The message vector features of the i-th node pair The vector representation of the central node at layer l. Let be the vector representation of the i-th adjacent node at layer l. Let be the vector representation of the edge between the central node and the i-th adjacent node at layer l, p be the transformation function of the node in the diagnostic information graph structure, ω be the transformation function of the edge in the diagnostic information graph structure, W1 be the learnable parameter matrix of the node, responsible for transforming the node dimension, and W2 be the learnable parameter matrix of the edge, responsible for transforming the edge dimension. `dimv` is the square root degree matrix, responsible for data normalization; `·` represents the vector inner product operation. The `sum` function sums the transformed vectors of the center node and its adjacent nodes, resulting in a (Dimv, 1)-dimensional node feature vector. The `concat` function concatenates the node feature vector with the transformed feature vectors of the edges, resulting in a (Dimv + dime, 1)-dimensional feature vector for the node pair. `linear` is a linear layer that further extracts features from the node pair feature vectors output in the previous step, and performs non-linear processing on the linear extraction operation of the node pair feature vectors using the `leakyRelu` activation function.
[0070] 2) Aggregation unit The mathematical definition is:
[0071]
[0072] in, For the aggregation feature of the central node, the node neighbor function ne(v) returns all first-order adjacent nodes of the central node v. Aggregation Unit The output is the aggregated feature obtained by adding the results of the calculations of all first-order adjacent nodes and their corresponding edges of the central node through the message passing unit M.
[0073] 3) The mathematical definition of the update unit U is:
[0074]
[0075] in, The distributed vector representation of the central node at layer l. The distributed vector representation of the central node at level l+1 is given. The update unit U adds the distributed vector representation of the central node v at the previous level to the distributed vector representations of all its corresponding first-order neighbor nodes to obtain the distributed vector representation of the next level.
[0076] (2) Aggregation of nodes and edges. For example... Figure 4 As shown, this step is divided into two parts:
[0077] 1) Sum the distributed representation vectors of all edges and all nodes obtained in stage (1) through the MCN graph convolutional network, that is, perform node aggregation and edge aggregation respectively to obtain a unified node aggregation representation vector and a unified edge aggregation representation vector.
[0078] 2) The unified node aggregation representation vector and the unified edge aggregation representation vector are concatenated in the zeroth dimension to obtain the original distributed representation vector of the diagnosis and treatment information.
[0079] (3) Feature Calculation. This step is responsible for generating the final distributed vector representation of the diagnostic information graph structure (i.e., the first semantic feature). For example... Figure 4 As shown, through automatic feature extraction by an n-layer feedforward neural network, the original distributed representation vector output by stage (2) is transformed into a low-dimensional dense first semantic feature. This first semantic feature integrates the semantic features of the nodes and edges of the entire diagnostic information graph structure, and can approximately express the latent semantics of the diagnostic information graph structure. The first semantic feature is also the first distributed representation vector, and the calculation formula for the first distributed representation vector γ is as follows:
[0080] γ=ReLU(ReLU(ReLU(xW 0 )xW 1 W n )
[0081] Where x is the original distributed representation vector, ReLU is the non-linear activation function, W is the learnable parameter of each layer of the feedforward neural network, and n is the number of layers of the feedforward neural network.
[0082] III. Similarity Calculation. This step involves calculating the similarity score between the first semantic feature corresponding to the patient's medical information and the second semantic feature of other pre-stored patient graph structures in the pre-stored semantic representation library. The top N pre-stored patient medical information with the highest similarity scores (i.e., the top N pre-stored patient medical information with similarity scores greater than or equal to a preset threshold) are used as input for the next stage. For similarity score calculation, numerous methods exist, including cosine similarity, Chebyshev distance, Mahalanobis distance, Euclidean distance, and inner product methods. Since the mapping space of distributed representation is a linear space, based on practical testing, this embodiment uses the computationally simple yet efficient inner product method as an example for similarity calculation. Specifically, the first semantic feature is the first distributed representation vector γ, and the second semantic feature is the second distributed representation vector S. First, γ is multiplied by all known S values to obtain similarity scores, which are then arranged into a (Dim|γ|, 1)-dimensional vector. Then, the vector is standardized using the softmax function. The standardized similarity scores are arranged in descending order to generate the target score vector. The components of the target score vector are the standardized similarity scores. Based on the actual situation, the top N similarity scores are determined from the target score vector. The specific calculation formula for the top N similarity scores is as follows:
[0083] score = rank(softmax(S·γ))
[0084] Where score is the similarity score, γ is the first distributed representation vector, S is the matrix composed of the second distributed representation vectors with shape (n, Dimγ[0]), and n is the number of the second distributed representation vectors. "·" is the inner product operation. The softmax function is the normalization function (i.e., the standardization function), which maps the values of each component of the vector to the interval [0, 1], and the sum of each component is 1. The rank function is the sorting function, which sorts the components from largest to smallest score.
[0085] IV. Pre-filling. This step involves retrieving pre-stored patient medical information to complete pre-filling. First, the second semantic features corresponding to the top N similarity scores from the previous stage are determined, and the pre-stored patient index accounts corresponding to these second semantic features are selected. Then, based on the pre-stored patient index accounts, the target medical information (i.e., the pre-stored patient medical information) is found and returned to the presentation layer as reference data for writing the current patient medical information. Doctors can directly modify based on this reference data or selectively copy the reference. The entire process can be updated in real-time as the doctor fills in the patient medical information.
[0086] In this embodiment, patient medical information is modeled as a non-Euclidean graph structure in a non-Euclidean space. A graph convolutional network is used to calculate the similarity of the edge matrix, node matrix, and adjacency matrix representing the non-Euclidean graph structure using distributed vectors to obtain matching pre-filled information. This pre-filled information is then backfilled, thus achieving pre-filling of the medical form. This method is data-driven, dynamically matching information based on the current patient information. This not only improves the efficiency of doctors filling in medical information but also enhances the accuracy and completeness of the information, significantly reducing the workload of doctors and allowing them to focus more on actual medical treatment.
[0087] Specifically, the method in this application embodiment performs graph modeling and graph representation learning on existing patient diagnosis and treatment data to obtain its low-dimensional dense distributed representation vector in a linear space. Through similarity calculation, the top N optimal pre-filled information are selected. Furthermore, this application embodiment is a real-time dynamic method, calculating and updating in real time based on the content filled in by the doctor.
[0088] The MCN algorithm in this application is an original graph neural network algorithm based on GCN. By fusing graph structure features and semantic features of nodes and edges in the graph, the algorithm can effectively capture the potential semantic features of the entire patient diagnosis and treatment graph globally and locally.
[0089] As a specific implementation of the above-mentioned method for pre-filling medical information, this application provides a device for pre-filling medical information. Figure 5 As shown, the diagnostic information pre-filling device 500 includes: a graph structure modeling module 501, a graph structure transformation module 502, a semantic feature learning module 503, and a pre-filling module 504.
[0090] Among them, the graph structure modeling module 501 is used to obtain the diagnosis and treatment information graph structure corresponding to the patient's diagnosis and treatment information. The unit diagnosis and treatment information graph structure includes edges representing the basic attributes in the unit diagnosis and treatment information and nodes connected to the unit edges that represent the corresponding values of the unit's basic attributes.
[0091] Graph structure transformation module 502 is used to transform the diagnosis and treatment information graph structure into an edge feature matrix, a node feature matrix, and an adjacency matrix corresponding to the diagnosis and treatment information graph structure.
[0092] The semantic feature learning module 503 is used to input the edge feature matrix, node feature matrix and adjacency matrix into the graph convolutional network and output the first semantic feature that expresses the structure of the diagnosis and treatment information graph.
[0093] The pre-filling module 504 is used to extract the corresponding target diagnosis and treatment information from the pre-stored semantic expression library based on the first semantic feature, and pre-fill the patient's diagnosis and treatment information.
[0094] In this embodiment, firstly, medical information related to the current patient's medical record is acquired, such as basic patient information, symptoms, and signs. This information is then modeled into a medical information graph structure according to business rules. Next, a graph convolutional network is used to output a distributed representation of this medical information graph structure, obtaining its first semantic feature. This allows for comprehensive and fine-grained capture of the latent semantic features of the patient's medical information. Then, the first semantic feature is compared with the semantic features of other existing patient graph structures in a pre-stored semantic representation library to derive the target medical information. Finally, the target medical information is retrieved and returned to the front end for text pre-filling of the patient's medical information.
[0095] This application proposes a method for pre-filling medical information based on graph convolutional networks. Driven by existing patient medical information, the method models structured medical information as a graph structure in non-Euclidean space, learns its semantic features through graph convolutional networks, obtains matching pre-filled information by comparing it with existing patient medical information, and then backfills the information, thereby realizing the pre-filling of medical forms, improving the efficiency of doctors in filling in medical information, and enhancing the accuracy and completeness of the medical information.
[0096] Furthermore, the diagnostic information graph structure is a non-Euclidean graph structure.
[0097] Furthermore, the graph structure transformation module 502 is specifically used to: perform one-hot encoding on the edges and nodes in the diagnosis and treatment information graph structure, and construct the edge feature matrix, the node feature matrix, and the adjacency matrix corresponding to the diagnosis and treatment information graph structure, respectively.
[0098] Furthermore, the medical information pre-filling device 500 also includes a normalization processing module for normalizing the medical information.
[0099] Furthermore, the graph convolutional network is used to: obtain the distributed representation vectors of all edges and all nodes in the diagnostic information graph structure based on the edge feature matrix, node feature matrix, and adjacency matrix; sum the distributed representation vectors of all edges to obtain the edge aggregate representation vector, and sum the distributed representation vectors of all nodes to obtain the node aggregate representation vector; concatenate the edge aggregate representation vector and the node aggregate representation vector to obtain the original distributed representation vector of the diagnostic information; and convert the original distributed representation vector into the first semantic feature corresponding to the diagnostic information graph structure through a feedforward neural network; wherein, the first semantic feature is the first distributed representation vector, and the first distributed representation vector is:
[0100] γ=ReLU(ReLU(ReLU(xW 0 )xW 1 W n )
[0101] γ is the first distributed vector, x is the original distributed representation vector, ReLU is the non-linear activation function, W is the learnable parameter of each layer of the feedforward neural network, and n is the number of layers of the feedforward neural network.
[0102] Furthermore, the graph convolutional network includes a message passing unit, an aggregation unit, and an update unit. The message passing unit encodes the latent vector of the l-th layer corresponding to a node pair into a latent vector of the (l+1)-th layer, which serves as the message vector in the graph convolutional network algorithm. The node pair is any one of multiple node pairs corresponding to the central node of the diagnostic information graph structure, and each node pair includes the central node, its neighboring nodes, and the edges between the central node and its neighboring nodes. The aggregation unit aggregates the message vectors of the multiple node pairs to obtain aggregated features. The update unit updates the aggregated features to the central node so that the central node possesses the semantic features of its corresponding neighboring nodes and edges.
[0103] Further, the pre-filling module 504 is specifically used for: calculating the similarity scores between the second semantic features corresponding to multiple pre-stored patient graph structures in the semantic expression library and the first semantic features, respectively, to obtain multiple similarity scores; standardizing the multiple similarity scores, and generating a target score vector based on the standardized multiple similarity scores, wherein the components in the target score vector represent the standardized similarity scores, and the second semantic features are used to express the pre-stored patient graph structures; determining the target diagnosis and treatment information based on a preset threshold and the target score vector, and pre-filling the patient's diagnosis and treatment information based on the target diagnosis and treatment information.
[0104] Furthermore, the pre-filling module 504 is specifically used for: determining the top N similarity scores that are greater than or equal to a preset threshold in the target scoring vector, where N is a positive integer; determining the pre-stored patient index accounts corresponding to the top N similarity scores; searching for the target diagnosis and treatment information corresponding to the pre-stored patient index accounts; and pre-filling the patient's diagnosis and treatment information according to the target diagnosis and treatment information.
[0105] The medical information pre-filling device 500 in this application embodiment can be an electronic device or a component of an electronic device, such as an integrated circuit or a chip. The electronic device can be a terminal or other devices besides a terminal. For example, the electronic device can be a tablet computer, laptop computer, handheld computer, robot, ultra-mobile personal computer (UMPC), netbook, or personal digital assistant (PDA), etc., and can also be a server, network attached storage (NAS), personal computer (PC), etc. This application embodiment does not specifically limit the specific implementation.
[0106] The diagnostic information pre-filling device 500 in this embodiment can be a device with an operating system. This operating system can be Android, iOS, or other possible operating systems; this embodiment does not specifically limit its use.
[0107] The diagnostic information pre-filling device 500 provided in this application embodiment can achieve... Figure 1 The various processes implemented in the method embodiment for pre-filling medical information will not be described again here to avoid repetition.
[0108] This application also provides an electronic device, such as... Figure 6 As shown, the electronic device 600 includes a processor 601 and a memory 602. The memory 602 stores a program or instruction that can run on the processor 601. When the program or instruction is executed by the processor 601, it implements the various steps of the above-described embodiment of the diagnostic information pre-filling method and can achieve the same technical effect. To avoid repetition, it will not be described again here.
[0109] It should be noted that the electronic devices in the embodiments of this application include the aforementioned mobile electronic devices and non-mobile electronic devices.
[0110] The memory 602 can be used to store software programs and various data. The memory 602 may primarily include a first storage area for storing programs or instructions and a second storage area for storing data. The first storage area may store the operating system, application programs or instructions required for at least one function (such as sound playback, image playback, etc.). Furthermore, the memory 602 may include volatile memory or non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct memory bus RAM (DRRAM). The memory 602 in this embodiment includes, but is not limited to, these and any other suitable types of memory.
[0111] Processor 601 may include one or more processing units; optionally, processor 601 integrates an application processor and a modem processor, wherein the application processor mainly handles operations involving the operating system, user interface, and applications, and the modem processor mainly handles wireless communication signals, such as a baseband processor. It is understood that the aforementioned modem processor may also not be integrated into processor 601.
[0112] This application also provides a readable storage medium storing a program or instructions. When the program or instructions are executed by a processor, they implement the various processes of the above-described diagnostic information pre-filling method embodiments and achieve the same technical effect. To avoid repetition, they will not be described again here.
[0113] This application also provides a chip, which includes a processor and a communication interface. The communication interface and the processor are coupled. The processor is used to run programs or instructions to implement the various processes of the above-described embodiments of the diagnostic information pre-filling method and can achieve the same technical effect. To avoid repetition, it will not be described again here.
[0114] It should be understood that the chip mentioned in the embodiments of this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.
[0115] This application also provides a computer program product, which is stored in a storage medium and executed by at least one processor to implement the various processes of the above-described diagnostic information pre-filling method embodiment, and can achieve the same technical effect. To avoid repetition, it will not be described again here.
[0116] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.
[0117] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.
Claims
1. A method for pre-filling medical information, characterized in that, include: Obtain a diagnosis and treatment information graph structure corresponding to the patient's diagnosis and treatment information. The diagnosis and treatment information graph structure includes edges representing basic attributes in the diagnosis and treatment information and nodes connected to the edges that represent the corresponding values of the basic attributes. One-hot encoding is performed on the edges and nodes in the diagnosis and treatment information graph structure to construct the edge feature matrix, the node feature matrix, and the adjacency matrix corresponding to the diagnosis and treatment information graph structure. The one-hot encoding lengths of the edge feature matrix and the node feature matrix are the same or different. The edge feature matrix, the node feature matrix, and the adjacency matrix are input into a graph convolutional network, which outputs a first semantic feature that expresses the structure of the diagnostic information graph. The graph convolutional network is used for: Based on the edge feature matrix, the node feature matrix, and the adjacency matrix, obtain the distributed representation vectors of all edges and all nodes in the diagnosis and treatment information graph structure; The distributed representation vectors of all edges are summed to obtain the edge aggregate representation vector, and the distributed representation vectors of all nodes are summed to obtain the node aggregate representation vector. The edge aggregation representation vector and the node aggregation representation vector are concatenated to obtain the original distributed representation vector of the diagnosis and treatment information. The original distributed representation vector is converted into the first semantic feature corresponding to the diagnostic information graph structure using a feedforward neural network; Wherein, the first semantic feature is a first distributed representation vector, and the first distributed representation vector is: Let be the first distributed representation vector. x Let be the original distributed representation vector. is a non-linear activation function, W is the learnable parameter of each layer of the feedforward neural network, and n is the number of layers of the feedforward neural network; Based on the first semantic feature, the corresponding target diagnosis and treatment information is extracted from the pre-stored semantic expression library, and the diagnosis and treatment information of the patient is pre-filled. The graph convolutional network includes a message passing unit, an aggregation unit, and an update unit. The message passing unit is used to encode the l-th hidden vector corresponding to the node pair into a (l+1)-th hidden vector as a message vector to be used in the graph convolutional network algorithm. The node pair is any one of multiple node pairs corresponding to the central node of the diagnosis and treatment information graph structure. The node pair includes the central node, the adjacent nodes of the central node, and the edge between the central node and the adjacent nodes. The specific calculation formula for the message transmission unit is as follows: Among them, u i Let vu be the i-th adjacent node of the central node v. i Let be the edge between the center node and the i-th adjacent node. The message vector features of the i-th node pair The central node is at the 1st l Vector representation of layers, For the i-th adjacent node in the i-th... l Vector representation of layers, The edge between the center node and the i-th adjacent node is at the 1st... l Vector representation of layers, For the transformation function of nodes in the diagnosis and treatment information graph structure, Let be the transformation function for the edges in the diagnostic information graph structure. This is the learnable parameter matrix for each node, responsible for transforming the node dimensions. This is the learnable parameter matrix for the edges, responsible for transforming the dimensions of the edges. The square root degree matrix is responsible for data normalization. " represents the vector dot product operation; The aggregation unit is used to aggregate the message vectors of multiple node pairs to obtain aggregation features; The update unit is used to update the aggregated features to the central node, so that the central node has the semantic features of its corresponding neighboring nodes and edges.
2. The method according to claim 1, characterized in that, The diagnostic information graph structure is a non-Euclidean graph structure.
3. The method according to claim 1, characterized in that, Before performing one-hot encoding on the edges and nodes in the diagnostic information graph structure, the method further includes: The diagnostic and treatment information is normalized, which involves standardizing the basic attributes and / or the values corresponding to the basic attributes of the diagnostic and treatment information.
4. The method according to any one of claims 1 to 3, characterized in that, The step of extracting corresponding target diagnostic and treatment information from a pre-stored semantic expression library based on the first semantic feature and pre-filling the patient's diagnostic and treatment information includes: The second semantic features corresponding to multiple pre-stored patient graph structures in the semantic expression library are compared with the first semantic features to calculate the similarity score, resulting in multiple similarity scores; The similarity scores are standardized and calculated, and a target score vector is generated based on the standardized similarity scores. The components of the target score vector are the standardized similarity scores, and the second semantic feature is used to express the pre-stored patient graph structure. Based on the preset threshold and the target scoring vector, target diagnostic and treatment information is determined, and the patient's diagnostic and treatment information is pre-filled based on the target diagnostic and treatment information.
5. The method according to claim 4, characterized in that, The step of determining target diagnostic and treatment information based on a preset threshold and the target scoring vector, and pre-filling the patient's diagnostic and treatment information based on the target diagnostic and treatment information, includes: In the target scoring vector, determine the top N similarity scores that are greater than or equal to a preset threshold, where N is a positive integer; Determine the pre-stored patient index accounts corresponding to the first N similarity scores; Locate the target medical information corresponding to the pre-stored patient index account, and pre-populate the patient's medical information based on the target medical information.
6. A medical information pre-filling device, characterized in that, For performing the diagnostic information pre-filling method as described in any one of claims 1 to 5, the diagnostic information pre-filling device comprises: The graph structure modeling module is used to obtain a diagnosis and treatment information graph structure corresponding to the patient's diagnosis and treatment information. The diagnosis and treatment information graph structure includes edges representing basic attributes in the diagnosis and treatment information and nodes connected to the edges that represent the corresponding values of the basic attributes. The graph structure conversion module is used to convert the diagnosis and treatment information graph structure into an edge feature matrix, a node feature matrix, and an adjacency matrix corresponding to the diagnosis and treatment information graph structure. The semantic feature learning module is used to input the edge feature matrix, the node feature matrix and the adjacency matrix into the graph convolutional network and output the first semantic feature that expresses the structure of the diagnosis and treatment information graph. The pre-filling module is used to extract the corresponding target diagnosis and treatment information from the pre-stored semantic expression library based on the first semantic feature, and pre-fill the diagnosis and treatment information of the patient.
7. A readable storage medium having a program or instructions stored thereon, characterized in that, When the program or instructions are executed by the processor, they implement the steps of the diagnostic information pre-filling method as described in any one of claims 1 to 5.
Citation Information
Patent Citations
Similar medical record retrieval method and system based on multi-modal medical record graph
CN113380360A
Artificial intelligence auxiliary diagnosis and treatment method and system, storage medium and electronic equipment
CN113972005A
Disease typing model training method, system and device and storage medium
CN114255878A
