Semantic understanding and automated evidence extraction method and system for medical data complaint text

Abstract

The present invention relates to the field of automated evidence extraction and discloses a method and system for semantic understanding and automated evidence extraction of medical appeal texts. The method comprises: constructing a three-level encoder for medical data appeal texts, performing semantic analysis and feature extraction at the character level, entity level, and document level to obtain multi-granularity semantic information, fusing the multi-granularity semantic information, and establishing a multi-granularity semantic fusion structure model; constructing a dynamic spatiotemporal evidence graph based on the trained multi-granularity semantic fusion structure to dynamically adjust edge weights; and verifying and eliminating contradictions in multimodal evidence through intelligent mining of key evidence chains.

Description

Technical Field

[0001] The present invention relates to the technical field of automated evidence extraction, and in particular to a method and system for semantic understanding and automated evidence extraction of medical data complaint texts. Background Art

[0002] In the existing technology, medical data complaint texts usually contain complex medical terms, entity relationships and time series information. Traditional text processing methods may find it difficult to fully understand and analyze this information. In medical complaint scenarios, the complexity of the text includes but is not limited to multi-level semantics, time series contradictions and cross-modal conflicts, which puts higher requirements on the model. Traditional models have high misjudgment rates due to single-granularity analysis, static knowledge dependence and broken evidence chains. Existing technology cannot process complex texts. Traditional graph neural networks (GNNs) have fixed entity relationship weights. For example, the "drug-indication" edge weight is always 0.9, which cannot reflect the time decay effect, such as the risk of failure of postoperative antibiotics after use for more than 24 hours, and ignores drug metabolic kinetic parameters, such as the impact of half-life on association strength. Therefore, it is necessary to provide semantic understanding and automated evidence extraction methods and systems for medical data complaint texts. Summary of the Invention

[0003] The present invention aims to provide a method and system for semantic understanding and automated evidence extraction of medical data complaint texts. To address the above-mentioned technical problems, the present invention is implemented through the following technical solutions:

[0004] First, the semantic understanding and automated evidence extraction method for medical data complaint text includes the following steps:

[0005] A three-level encoder is constructed for medical data complaint texts, which performs semantic analysis and feature extraction at the character level, entity level, and document level to obtain multi-granular semantic information.

[0006] Fuse multi-granularity semantic information and establish a multi-granularity semantic fusion structure model;

[0007] Based on the multi-granularity semantic fusion structure obtained through training, a dynamic spatiotemporal evidence graph is constructed to dynamically adjust edge weights;

[0008] Through intelligent mining of key evidence chains, multimodal evidence can be verified and contradictions eliminated.

[0009] In a second aspect, the semantic understanding and automated evidence extraction system for medical data complaint text provided by the present invention specifically includes the following modules:

[0010] Data collection module: collects relevant data of medical data complaint texts;

[0011] Feature extraction module: performs semantic analysis and feature extraction on relevant data of medical data complaint text at the character level, entity level, and document level;

[0012] Semantic fusion module: fuses multi-granularity semantic information and establishes a multi-granularity semantic fusion structure model;

[0013] Dynamic evidence graph construction module: Based on the multi-granularity semantic fusion structure obtained through training, it constructs a dynamic spatiotemporal evidence graph and dynamically adjusts edge weights;

[0014] Evidence verification and optimization module: Through intelligent mining of key evidence chains, multimodal evidence is verified and contradictions are eliminated.

[0015] Beneficial effects of the present invention:

[0016] 1. Semantic analysis and feature extraction are performed using encoders at three different granularities: character level, entity level, and document level. The entity-level encoder identifies medical entities, constructs a graph structure, and mines the relationship information between entities. The document-level encoder uses a hierarchical attention mechanism to highlight important sentences and information, capture the temporal characteristics of disease progression, and fuse multi-granularity features to more accurately reflect the semantics of the text than single-granularity feature extraction methods. The granular feature fusion mechanism integrates multi-level information from fine-grained character level to coarse-grained document level, reducing the risk of overfitting and thereby improving the model's generalization and stability. It accurately extracts features and performs semantic analysis on medical data complaint texts of different formats and expression habits. Through multi-granularity feature extraction and fusion, as well as encoding mechanisms designed specifically for different granularities, it better handles complex texts and improves the model's understanding of medical data complaint texts, thereby more accurately completing related tasks.

[0017] 2. Generate composite vectors through multi-granularity semantic fusion structure, integrate character-level, entity-level and document-level features, combine cosine similarity and drug metabolism dynamics, dynamically adjust edge weights, retain semantic associations, quantify time decay effects, and achieve a comprehensive semantic representation of medical events. Introduce the Boolean value of the NCCN guideline recommendation path, integrate the authority of clinical guidelines into the construction of the evidence chain, and realize multi-dimensional correlation analysis, temporal logic verification and guideline-driven reasoning of medical data through dynamic spatiotemporal evidence graphs and key evidence chain mining. It provides technical support for building a reliable and explainable medical evidence system and realizes high-precision and high-efficiency automated evidence extraction in medical data complaint scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0019] Figure 1 This is a flowchart of the steps of the method for semantic understanding and automated evidence extraction of medical data complaint text provided in Example 1 of the present invention;

[0020] Figure 2 It is a structural diagram of the semantic understanding and automated evidence extraction system for medical data complaint text provided in Example 2 of the present invention. DETAILED DESCRIPTION

[0021] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work should fall within the scope of protection of the present invention.

[0022] Example 1

[0023] like Figure 1 As shown, the method for semantic understanding and automatic evidence extraction of medical data complaint text provided by the embodiment of the present invention specifically includes the following steps:

[0024] Step 1: Construct a three-level encoder for medical data complaint text, perform semantic analysis and feature extraction at the character level, entity level, and document level, and fuse multi-granularity semantic information to establish a multi-granularity semantic fusion structure model;

[0025] In step one:

[0026] First, the specific process of constructing a character-level encoder for medical data complaint text is as follows:

[0027] The collected medical complaint text data was convolved using a hybrid convolutional neural network (Hybrid-CNN). The parameters of the one-dimensional convolution kernel were adjusted and a maximum pooling operation was performed to optimize the extraction of character-level features.

[0028] It should be noted that the hybrid convolutional neural network refers to a model design method that combines different types of convolution operations or structures. It aims to improve model performance, enhance feature extraction capabilities, and optimize computational efficiency by integrating multiple convolution methods. The one-dimensional convolution kernel represents a filter or weight matrix used to perform convolution operations on one-dimensional data. Maximum pooling means selecting the maximum value within a local area of ​​a given feature map, usually a rectangular window, as the output.

[0029] Specifically, a hybrid convolutional neural network is used to capture the word-formation features of medical terms such as drug names. Medical terms have specific word-formation patterns, such as "enteric-coated aspirin tablets" or "cefixime capsules." By analyzing character-level features, the meaning of medical data complaint text can be better understood.

[0030] Specifically, the medical data complaint text is character-encoded to obtain the encoding vector corresponding to each character. The input text has a characters, and the encoding vector dimension of each character is d. Therefore, the input data can be represented as an a×d matrix X;

[0031] Based on capturing the features of characters of different lengths, w one-dimensional convolution kernels of different sizes are used. The sizes of the one-dimensional convolution kernels are preset as: k1, k2, k3, ..., k w ; For each size k i One-dimensional convolution kernel, where i∈{1,2,3,...,w}, specifically represents a k i ×d-sized matrix W i , where d is the dimension of the character encoding vector;

[0032] For example, one-dimensional convolution kernels of different sizes are used, for example, k1=3, k2=4, k3=5, where k1, k2, and k3 are the sizes of the one-dimensional convolution kernels, to perform a convolution operation on the character sequence. Each one-dimensional convolution kernel captures character features of different sizes. For example, a one-dimensional convolution kernel of size 3 captures the common drug name suffix feature "-pilin".

[0033] For example, the medical data complaint text "The patient felt unwell after taking aspirin" is encoded at the character level to obtain a character sequence. A one-dimensional convolution kernel of size 3 is used for convolution to simplify the character encoding so that each character corresponds to a simple numerical vector representation. The character encoding vector obtained after encoding the character sequence is shown in Table 1:

[0034] character Encoded vector Suffer from [0.1,0.2,0.3] Writer [0.4,0.5,0.6] Clothes [0.7,0.8,0.9] use [0.1,0.1,0.1] Ah [0.2,0.2,0.2] manage [0.3,0.3,0.3] match [0.4,0.4,0.4] Forest [0.5,0.5,0.5] back [0.6,0.6,0.6] out [0.7,0.7,0.7] now [0.8,0.8,0.8] No [0.9,0.9,0.9] suitable [1.0,1.0,1.0]

[0035] Table 1 Statistics of encoding vectors corresponding to characters

[0036] Using a one-dimensional convolution kernel of size 3, the weight matrix W of the one-dimensional convolution kernel is: Slide the one-dimensional convolutional kernel over the character sequence, and for each time, take the encoding vectors of 3 consecutive characters and perform a convolution operation with the one-dimensional convolutional kernel;

[0037] For each one-dimensional convolutional kernel of size k i W i , perform a convolution operation on the input matrix X. The one-dimensional convolutional kernel slides over X, and for each time, take the encoding vectors of k i consecutive characters and perform matrix multiplication with W i , add a bias term b i , and perform a non-linear transformation through the activation function GELU function. Set the one-dimensional convolutional kernel W i to perform a convolution operation at the j-th position, and the resulting convolution result c i,j is: where represents the sub-matrix of the input matrix X from the j-th row to the (j + k) i -1-th row, and W i T represents the transpose matrix of the one-dimensional convolutional kernel W i . For each one-dimensional convolutional kernel of size k i , after the convolution operation, a convolution result vector C [[ID=第32行]] i with a length of n + k i +1 is obtained;

[0038] It should be noted that for each one-dimensional convolutional kernel, in addition to the weight matrix, i.e., the convolutional kernel itself, a corresponding bias term is set to increase the degree of freedom of the model, enabling the model to flexibly fit the training data;

[0039] Exemplarily, when the one-dimensional convolutional kernel acts on the first 3 characters "患", "者", "服", the convolution result vector is calculated as: [0.1, 0.2, 0.3] * 0.1 + [0.4, 0.5, 0.6] * 0.4 + [0.7, 0.8, 0.9] * 0.7 = [0.66, 0.78, 0.9];

[0040] The one-dimensional convolutional kernel slides over the entire character sequence. After the convolution operation, the convolution results are shown in Table 2:

[0041] Convolution position Convolution result vector 1 [0.66,0.78,0.9] 2 [0.5,0.6,0.7] 3 [0.4,0.5,0.6] 4 [0.3,0.4,0.5] 5 [0.2,0.3,0.4] 6 [0.1,0.2,0.3] 7 [0.05,0.1,0.15] 8 [0.02,0.03,0.04] 9 [0.01,0.02,0.03] 10 [0.005,0.01,0.015] 11 [0.002,0.003,0.004]

[0042] Table 2 Convolution result vectors corresponding to convolution positions

[0043] Perform a max pooling operation on the convolution results to extract the most important features output by each one-dimensional convolutional kernel. Use a pooling window of size N (N is a positive integer and N is greater than or equal to 2). The pooling window slides over the convolution results, and for each time, take the maximum value of each dimension within the window as the output;

[0044] It should be noted that the pooling window size is preset by the professional and technical personnel of the present invention based on historical experience and can be adjusted according to specific circumstances;

[0045] For example, a pooling window of size 2 is used. When the pooling window is applied to the first two convolution results [0.66, 0.78, 0.9] and [0.5, 0.6, 0.7], the maximum value of each dimension is taken to obtain the pooling result vector: [max(0.66, 0.5), max(0.78, 0.6), max(0.9, 0.7)] = [0.66, 0.78, 0.9]. The pooling window continues to slide, and the convolution result vectors calculated based on Table 2 are sequentially max-pooled to obtain all pooling result vectors, as shown in Table 3:

[0046] Pooling location Pooling result vector 1 [0.66,0.78,0.9] 2 [0.4,0.5,0.6] 3 [0.2,0.3,0.4] 4 [0.05,0.1,0.15] 5 [0.01,0.02,0.03]

[0047] Table 3 Maximum pooling result vector statistics

[0048] It should be noted that through the maximum pooling operation, the most important features in each pooling window are extracted from the convolution result, reducing the dimension of the data while retaining key information. The obtained pooling result is used for subsequent feature splicing, and finally a character-level feature representation is obtained;

[0049] Specifically, based on the obtained encoding vector, multiple one-dimensional convolution kernels of different sizes are used for convolution operation, and the results of each one-dimensional convolution kernel are subjected to maximum pooling operation. All pooled features are concatenated to capture richer character-level features, and the character-level feature representation C is obtained, whose dimension is recorded as d. C ;

[0050] The second specific process of constructing an entity-level encoder for medical data complaint text is as follows:

[0051] Use named entity recognition (NER) technology to identify medical entities from medical data complaint text, including but not limited to: drug entity D, disease entity I, and symptom entity S;

[0052] It should be noted that named entity recognition (NER) technology represents an important task in natural language processing, which aims to identify and classify specific types of entities from text;

[0053] The identified entities are used as nodes, and edges between entities are constructed based on the semantic information in the text, forming a graph structure G = (V, E), where V is the node set and E is the edge set;

[0054] For example, if the text mentions that "Drug A can treat Disease B", then an edge from Drug A to Disease B is added to the graph;

[0055] Use the gated graph neural network Gated-GNN to propagate information and update the feature representation of the node. i , its feature update formula is: in, is node v i The feature representation at time t is: Represents the update gate, which is used to control the old features and new features The fusion ratio, ⊙ represents the element-by-element multiplication;

[0056] It should be noted that the Gated-GNN is a neural network model specifically designed for processing graph-structured data. In the medical data complaint text processing scenario, the text contains various medical entities, such as drugs, diseases, and symptoms. The complex relationships between entities can be abstracted as a graph structure, with nodes representing entities and edges representing relationships between entities. By processing this graph-structured data with the Gated-GNN, we can reason about the relationships between medical entities.

[0057] Based on the gated graph neural network, the feature representation of each entity is obtained, which includes the relationship information between entities. All entity features are averaged and pooled to obtain the entity-level feature representation E entity , whose dimension is denoted as d E ;

[0058] The third specific process of building a document-level encoder for medical data complaint text is as follows:

[0059] Based on the obtained medical data complaint text, the hierarchical attention mechanism is used to highlight important sentences and information, capture the characteristics of the disease progression, obtain the time sequence, and obtain the time series information of the disease progression.

[0060] It should be noted that the Hierarchical Attention mechanism is a model architecture specifically designed for processing long documents or sequence data. It captures important information in the text by applying the attention mechanism at different levels and assigning different weights to different parts.

[0061] Specifically, the medical data complaint text is divided into x sentences, and each sentence is encoded using the pre-trained language model BERT; the attention mechanism is applied at the sentence level, and the formula Calculate the importance weight α of each sentence x , where exp represents the exponential function, u q It means sentence s qThe feature representation of It represents the sentence feature representation u q The transpose of v s Represents the attention vector of the sentence;

[0062] Apply the attention mechanism at the document level, combine the attention results at the sentence level, and calculate the feature representation of the entire document. Get the document-level feature representation D doc , whose dimension is denoted as d D ;

[0063] Fourthly, the specific process of fusing multi-granularity semantic information and establishing a multi-granularity semantic fusion structure is as follows:

[0064] Based on the obtained character-level feature representation C and entity-level feature representation E entity and document-level feature representation D doc , a multi-granularity semantic fusion structure T is generated through a multi-granularity feature fusion mechanism sem :T sem =[B char ,B entity ,B doc ]∈R a×d×3 , where represents the number of characters, i.e. the length of the medical data complaint text, R represents the set of real numbers, and B char It represents the feature matrix of character-level feature representation C, B entity It represents the entity-level feature representation E entity The characteristic matrix, B doc It represents the document-level feature representation D doc The characteristic matrix of

[0065] It should be noted that the multi-granularity feature fusion mechanism refers to a technology used to integrate features at different levels in deep learning models. By combining multi-level information from fine-grained to coarse-grained, it improves the model's understanding of input data and task performance. The semantic information of the multi-granularity semantic fusion structure allows the model to learn richer feature representations, better adapt to various changes when facing text data of different types and fields, reduce the risk of overfitting, and thus improve the model's generalization ability and stability.

[0066] Based on the obtained multi-granularity semantic fusion structure T sem , capture the word formation rules of drug names, build a dynamic knowledge graph of "drug → indication → contraindication", and establish a temporal causal chain of "drug use → examination → symptom worsening" to build a complete chain of evidence;

[0067] The technical solution of the embodiment of the present invention is: semantic analysis and feature extraction are performed through encoders at three different granularities: character level, entity level and document level, to comprehensively capture the information in the medical data complaint text from different angles, analyze the word formation features of medical terms such as drug names, and better understand the meaning of medical terms; the entity level encoder identifies medical entities and constructs a graph structure to mine the relationship information between entities; the document level encoder highlights important sentences and information through a hierarchical attention mechanism, captures the temporal characteristics of the disease course evolution, and fuses multi-granularity features, so that the model can learn richer and more comprehensive feature representations, which is more accurate than the single-granularity feature extraction method. The granular feature fusion mechanism integrates multi-level information from fine-grained character level to coarse-grained document level, enabling the model to better adapt to text data of different types and fields. It uses rich semantic information for learning and reasoning, reducing the risk of overfitting, thereby improving the model's generalization and stability. It accurately extracts features and performs semantic analysis on medical data complaint texts of different formats and expression habits. Through multi-granular feature extraction and fusion, as well as encoding mechanisms designed specifically for different granularities, it better handles complex texts and improves the model's understanding of medical data complaint texts, thereby completing related tasks more accurately.

[0068] Example 2

[0069] like Figure 1 As shown, the method for semantic understanding and automatic evidence extraction of medical data complaint text provided by the embodiment of the present invention specifically includes the following steps:

[0070] Step 2: Based on the multi-granularity semantic fusion structure obtained through training, a dynamic spatiotemporal evidence graph is constructed to dynamically adjust edge weights. Through intelligent mining of key evidence chains, multimodal evidence is verified and contradictions are eliminated.

[0071] In step 2:

[0072] First, the specific process of dynamically constructing the spatiotemporal evidence graph based on the multi-granularity semantic fusion structure obtained through training is as follows:

[0073] Based on the obtained multi-granularity semantic fusion structure, a 256-dimensional composite vector is constructed and the weighted calculation formula g is used. i =β1F char +β2F entity +β3F doc Get the semantic composite vector g i , where F char Represents the character-level feature vector, F entity represents the entity-level feature vector, F docIt represents the document-level feature vector, and β1, β2, and β3 represent the preset weight coefficients, with values ​​of 0.4, 0.3, and 0.3 respectively;

[0074] It should be noted that

[0075] For example, the vector of the event “ceftriaxone 2 g used on postoperative day 3” contains drug name, surgery time window, and administration route features;

[0076] Based on the calculated semantic composite vector g i and g j , through the formula Calculate the cosine similarity s between semantic composite vectors sem , where g i and g j All are semantic composite vectors;

[0077] The cosine similarity s sem Compared with the preset strong correlation threshold, if the cosine similarity s sem If the value is greater than or equal to the preset strong correlation threshold, the semantic composite vector g i and g j The corresponding medical data complaint text is marked as strongly associated;

[0078] Based on drug metabolism kinetics, the half-life of the drug is obtained by the formula Calculate the time difference weight w time , where e represents the natural constant, λ represents the preset drug time decay factor, and t i It represents the actual half-life of the drug, t j It represents the theoretical half-life of the drug;

[0079] Based on the obtained time difference weight, the cosine similarity between the obtained semantic composite vectors, the time difference weight and the Boolean value of the recommended path of the latest version of the NCCN guidelines are multiplied to obtain the final edge weight w ij ;

[0080] The second specific process of verifying and eliminating contradictions in multimodal evidence through intelligent mining of key evidence chains is as follows:

[0081] Based on the multi-granularity semantic fusion structure, a unit time window is set to obtain medical events that occur within the same unit time window, and the cosine similarity between the semantic composite vectors obtained by the medical event analysis is obtained;

[0082] If there is a strong correlation between medical events A and B, set the time threshold specified in the guidelines, calculate the difference between the actual time of medical event A and the time threshold specified in the guidelines, and take the absolute value to obtain the time deviation;

[0083] The obtained time deviation is compared with the time threshold specified in the guidelines to obtain the time deviation degree;

[0084] Compare the obtained time deviation with the preset deviation threshold. If the time deviation is greater than or equal to the deviation threshold, it indicates a logical error and is marked as a logical error.

[0085] Based on the logical error obtained from medical event A, the logical error is fed back and a strongly related medical event B is obtained to determine whether a logical error occurs;

[0086] If a logical error is found in medical event B, the evidence corresponding to medical events A and B will be eliminated and an abnormal feedback will be given. If there is no logical error, the evidence corresponding to the strongly associated medical event B will be retained, the evidence corresponding to medical event A will be eliminated, and feedback will be given to verify medical event A.

[0087] For example, the text record "fever occurred on the 5th day of immunotherapy" is compared with the test data "CRP increase occurred on the 3rd day". If the time deviation of the text record "fever occurred on the 5th day of immunotherapy" is greater than the deviation threshold, a logical error is marked. If the time deviation of the test data "CRP increase occurred on the 3rd day" is less than the deviation threshold, there is no logical error, the evidence corresponding to the test data "CRP increase occurred on the 3rd day" is retained, the evidence corresponding to the text record "fever occurred on the 5th day of immunotherapy" is eliminated, and feedback is provided to verify the text record "fever occurred on the 5th day of immunotherapy".

[0088] The technical solution of the embodiment of the present invention is: generating a composite vector through a multi-granularity semantic fusion structure, integrating character-level, entity-level and document-level features, combining cosine similarity and drug metabolism dynamics, dynamically adjusting edge weights, retaining semantic associations, quantifying time decay effects, and realizing a comprehensive semantic representation of medical events. The Boolean value of the NCCN guideline recommendation path is introduced, and the authority of the clinical guidelines is integrated into the evidence chain construction. Through dynamic spatiotemporal evidence graphs and key evidence chain mining, multi-dimensional correlation analysis, temporal logic verification and guideline-driven reasoning of medical data are realized, providing technical support for building a reliable and explainable medical evidence system, and realizing high-precision and high-efficiency automated evidence extraction in medical data complaint scenarios.

[0089] Example 3

[0090] like Figure 2 As shown, the semantic understanding and automated evidence extraction system for medical data complaint text provided by the embodiment of the present invention specifically includes the following modules:

[0091] Data collection module: collects relevant data of medical data complaint texts;

[0092] Feature extraction module: performs semantic analysis and feature extraction on relevant data of medical data complaint text at the character level, entity level, and document level;

[0093] Semantic fusion module: fuses multi-granularity semantic information and establishes a multi-granularity semantic fusion structure model;

[0094] Dynamic evidence graph construction module: Based on the multi-granularity semantic fusion structure obtained through training, it constructs a dynamic spatiotemporal evidence graph and dynamically adjusts edge weights;

[0095] Evidence verification and optimization module: Through intelligent mining of key evidence chains, multimodal evidence is verified and contradictions are eliminated.

[0096] The above is a detailed description of an embodiment of the present invention, but the content is only a preferred embodiment of the present invention and cannot be considered to limit the scope of implementation of the present invention; the above formulas are all dimensionless and numerical calculations, and the formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field based on actual conditions and historical experience, and can be adjusted according to actual conditions; all equal changes and improvements made according to the scope of application of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A method for semantic understanding and automated evidence extraction of medical data complaint texts, characterized by: The following steps are involved: A three-level encoder is constructed for medical data complaint texts, which performs semantic analysis and feature extraction at the character level, entity level, and document level to obtain multi-granular semantic information. The method for obtaining character-level features is: The collected medical complaint text data was convolved using a hybrid convolutional neural network (Hybrid-CNN). The parameters of the one-dimensional convolution kernel were adjusted and a maximum pooling operation was performed to optimize the extraction of character-level features. The method for obtaining entity-level features is: Use named entity recognition (NER) technology to identify medical entities from medical data complaint texts, use the identified entities as nodes, build edges between entities based on the semantic information in the text, and calculate and update nodes to form a graph structure. ,in is a collection of nodes, is the edge set; The method for obtaining the update node is: Use the gated graph neural network Gated-GNN to propagate information and update the feature representation of the node. , its feature update formula is: ,in, is a node In the The characteristic representation of the moment, Represents the update gate, which is used to control the old features and new features The fusion ratio, It means element-by-element multiplication; The semantic information of multiple granularities is fused to obtain a semantic composite vector and a multi-granularity semantic fusion structure model is established; Based on the obtained multi-granularity semantic fusion structure model, a dynamic spatiotemporal evidence graph is constructed to dynamically adjust the edge weights to obtain the final edge weights. The method for obtaining the final edge weight is: Based on drug metabolic kinetics, the half-life of the drug is obtained, and the time difference weight is calculated by the formula; Based on the obtained time difference weight, the cosine similarity between the obtained semantic composite vectors, the time difference weight and the Boolean value of the recommended path of the latest version of the NCCN guidelines are multiplied to obtain the final edge weight; Through intelligent mining of key evidence chains, multimodal evidence can be verified and contradictions eliminated.

2. The method for semantic understanding and automated evidence extraction of medical data complaint text according to claim 1, characterized in that: The method for obtaining entity-level feature representation is: Based on the gated graph neural network, the feature representation of each entity is obtained, which includes the relationship information between entities. All entity features are averaged and pooled to obtain the entity-level feature representation, whose dimension is recorded as .

3. The method for semantic understanding and automated evidence extraction of medical data complaint text according to claim 1, characterized in that: The method for obtaining the semantic composite vector is: Based on the obtained multi-granularity semantic fusion structure, a 256-dimensional composite vector is constructed and the cosine similarity is calculated. The semantic composite vector is obtained through the weighted calculation formula.

4. The method for semantic understanding and automated evidence extraction of medical data complaint text according to claim 3, characterized in that: The method for obtaining the cosine similarity is: Based on the calculated semantic composite vector and , the cosine similarity between semantic composite vectors is calculated by the formula; The obtained cosine similarity is compared with the preset strong correlation threshold to determine whether there is a strong correlation.

5. The method for semantic understanding and automated evidence extraction of medical data complaint text according to claim 1, characterized in that: The specific process of eliminating contradictions is as follows: Based on the multi-granularity semantic fusion structure, a unit time window is set to obtain medical events that occur within the same unit time window, and the cosine similarity between the semantic composite vectors obtained by the medical event analysis is obtained; Based on the obtained logical errors, the logical errors are fed back, and medical events with strong correlations are obtained to determine whether logical errors occur and eliminate contradictions.

6. A semantic understanding and automated evidence extraction system for medical data complaint text, the extraction system being used to execute the extraction method according to any one of claims 1 to 5, characterized in that: include: Data collection module: collects relevant data of medical data complaint texts; Feature extraction module: performs semantic analysis and feature extraction on relevant data of medical data complaint text at the character level, entity level, and document level; Semantic fusion module: fuses multi-granularity semantic information and establishes a multi-granularity semantic fusion structure model; Dynamic evidence graph construction module: Based on the obtained multi-granularity semantic fusion structure model, a dynamic spatiotemporal evidence graph is constructed to dynamically adjust edge weights; Evidence verification and optimization module: Through intelligent mining of key evidence chains, multimodal evidence is verified and contradictions are eliminated.

Images