Ophthalmic disease information extraction method and auxiliary diagnosis device based on electronic medical record

Abstract

The present application relates to an electronic medical record-based ophthalmic disease information extraction method and an auxiliary diagnosis device. The method comprises: differentiating and extracting negative descriptions and positive descriptions from each segment of text in the electronic medical record of the ophthalmic disease; respectively performing semantic coding on the positive descriptions and the negative descriptions to obtain a positive description static semantic representation vector and a negative description static semantic representation vector, and performing joint coding to obtain a dynamic semantic representation vector; taking the dynamic semantic representation vector and the label embedding vector of each disease as input, extracting information related to different ophthalmic diseases, and outputting an ophthalmic disease-related information representation vector. The device takes the ophthalmic disease-related information representation vector as the input of the neural network and outputs a disease prediction probability. The present application takes the electronic medical record data as input and outputs whether it is affected by various ophthalmic diseases, which can achieve the purposes of ophthalmic disease information extraction and ophthalmic disease auxiliary diagnosis.

Description

Technical Field

[0001] This invention belongs to the fields of information technology and medical technology, and specifically relates to a method for extracting ophthalmic disease information and an auxiliary diagnostic device based on electronic medical records. Background Technology

[0002] The eye is one of the most important sensory organs in the human body, and eye health is a crucial component of overall health. In recent years, artificial intelligence technologies such as machine learning have developed rapidly and have been applied in numerous fields with significant results. Simultaneously, the widespread adoption of electronic medical record systems in ophthalmology and advancements in other medical information technologies have laid a solid foundation for leveraging artificial intelligence in the ophthalmology industry. Against this backdrop, developing an auxiliary diagnostic method for ophthalmic diseases based on electronic medical records can reduce the occurrence of missed diagnoses and misdiagnoses by doctors.

[0003] From the perspective of the input data upon which the technology relies, existing methods can be mainly divided into three categories: image-based, clinical text-based, and numerical data-based, with the vast majority of related methods falling into the first category. Image-based methods typically transform the diagnostic aid problem into an image classification problem, generally employing deep learning methods such as convolutional neural networks as the image classification model. Images such as fundus retinal images or optical coherence tomography (OCT) images serve as model input, and the model outputs the presence or absence of a relevant disease. Clinical text-based methods typically perform overall embedding encoding on the clinical text, or perform local embedding encoding on different segments of the clinical text and then concatenate them to obtain a feature vector representing the case. This vector is then used as input to a machine learning classification model to obtain the disease diagnosis result. Numerical data-based methods usually require feature engineering. First, various feature values ​​relevant to disease prediction are extracted from image examination reports and clinical text records to form a feature vector. This feature vector is then used as input to a prediction model to obtain the disease diagnosis result.

[0004] The existing methods have the following problems: (1) Image-based methods have the problem of limited scope of disease-aided diagnosis. Such methods rely solely on image data, which limits their diagnostic scope to diseases that can be represented by the image. For example, they cannot diagnose ocular surface diseases such as pterygium based solely on fundus retinal image data.

[0005] (2) There are generally two approaches to methods based on clinical text data: one is to directly encode the entire clinical text, and the other is to encode each segment of the clinical text locally and then concatenate them. Both approaches have certain problems. The first approach, which directly encodes the entire clinical text into a fixed-length vector, leads to information loss and the inability to highlight important information. The second approach, which encodes each segment of the text locally into vectors and then concatenates them, results in a high-dimensional vector after concatenation, containing a large amount of redundant information in the representation of the clinical text, and also makes the prediction model prone to overfitting on the training set. In addition, the simple concatenation operation makes the representation of each segment of the text lack interaction, and cannot effectively represent the text information.

[0006] (3) Methods based on numerical data have the problem of heavy manual tasks. These methods first require the participation of experts to select features or design feature variables, and then require heavy and time-consuming feature engineering, making them difficult to be widely used in real medical scenarios. Summary of the Invention

[0007] This invention constructs an ophthalmic disease information extraction method and an ophthalmic disease auxiliary diagnosis device (or system) based on ophthalmic electronic medical record data. It takes the electronic medical record data of a single case as input and outputs whether the patient suffers from various ophthalmic diseases, thereby achieving the purpose of ophthalmic disease information extraction and ophthalmic disease auxiliary diagnosis.

[0008] This invention does not rely on any external knowledge sources or knowledge bases. It starts directly from the patient's case itself, using the patient's electronic ophthalmology medical record as input to extract ophthalmic disease information and output various possible ophthalmic diseases the patient may have. This invention can be divided into two parts: an electronic medical record information extraction model and a disease diagnosis model. For the information extraction model, this invention solves the technical problem of distinguishing between positive and negative descriptions in electronic medical records. For the disease diagnosis model, this invention solves the technical problems related to end-to-end disease diagnosis models, enabling semantic representation of each segment of the electronic medical record and further achieving dynamic joint representation between semantic representations.

[0009] The technical solution adopted in this invention is as follows: A method for extracting ophthalmic disease information based on electronic medical records includes the following steps: The text in the electronic medical record of ophthalmology disease is divided into negative and positive descriptions and extracted to obtain a set of positive descriptions and a set of negative descriptions. Semantic encoding is performed on the positive descriptions in the positive description set and the negative descriptions in the negative description set to obtain static semantic representation vectors for positive and negative descriptions. The dynamic semantic representation vector is obtained by jointly encoding the static semantic representation vector of positive description and the static semantic representation vector of negative description. Taking dynamic semantic representation vectors and label embedding vectors of various diseases as input, the label attention layer is used to extract information related to different ophthalmic diseases and output ophthalmic disease-related information representation vectors.

[0010] Furthermore, the process of distinguishing and extracting negative and positive descriptions from each segment of text in the electronic medical record for ophthalmological diseases includes: Based on a combination of BiLSTM-CRF model and rule matching, negative and positive descriptions are extracted from each segment of text in electronic medical records to obtain sets of positive and negative descriptions.

[0011] Furthermore, the semantic encoding of positive descriptions in the positive description set and negative descriptions in the negative description set is based on the MC-BERT model for semantic encoding of positive and negative descriptions, including: Each positive description in the set of positive descriptions is used as input to MC-BERT to obtain a set of positive description representation vectors. Similarly, each negative description in the set of negative descriptions is used as input to MC-BERT to obtain a set of negative description representation vectors. Perform a transpose operation on each positive and negative description vector to obtain the static semantic representation vector matrix of positive descriptions. and negative description static semantic representation vector matrix .

[0012] Further, the joint encoding of the positive description static semantic representation vector and the negative description static semantic representation vector to obtain the dynamic semantic representation vector includes: A multi-head self-attention mechanism is used to model the static semantic representation vector matrix of positive descriptions. The relationship between the positive descriptions and the dynamic semantic representation vector matrix are obtained. ; A soft attention mechanism is used to model the static semantic representation vector matrix of positive descriptions. and negative description static semantic representation vector matrix The dependency relationship between them and Joint encoding is performed to obtain the interactive dynamic semantic representation vector matrix.

[0013] Furthermore, the extraction of information related to different ophthalmic diseases using the label attention layer includes: Will and By splicing The matrix is ​​then represented by ophthalmic disease label vectors. L and As input to the label attention layer, the output is a vector matrix representing ophthalmic disease-related information. ,in It is the representation vector of information related to ophthalmic disease i extracted from its input by the label attention mechanism; the model parameters involved in this step are further updated through gradient backpropagation during the training process of the ophthalmic disease auxiliary diagnosis task.

[0014] An auxiliary diagnostic device for ophthalmic diseases based on electronic medical records, comprising: The medical record information extraction module is used to distinguish and extract negative and positive descriptions from each segment of text in the electronic medical record of ophthalmology diseases, and obtain a set of positive descriptions and a set of negative descriptions. The text semantic representation module is used to perform semantic encoding on the positive descriptions in the positive description set and the negative descriptions in the negative description set, respectively, to obtain the static semantic representation vectors of the positive descriptions and the static semantic representation vectors of the negative descriptions. The semantic representation interaction module is used to jointly encode the static semantic representation vectors of positive and negative descriptions to obtain dynamic semantic representation vectors. The disease information extraction module takes dynamic semantic representation vectors and label embedding vectors of various diseases as inputs, uses the label attention layer to extract information related to different diseases, and outputs ophthalmic disease-related information representation vectors. The disease prediction module uses vectors representing ophthalmic diseases as input to a neural network and outputs disease prediction probabilities.

[0015] Furthermore, the disease prediction module employs a set of feedforward neural network models, taking an ophthalmic disease-related information representation vector matrix as input, and outputting the probability of the existence of each ophthalmic disease.

[0016] Furthermore, the number of feedforward neural network models in the set of feedforward neural network models is the same as the number of target ophthalmic disease categories, and the i-th feedforward neural network model is represented by a vector matrix with disease-related information. In Given the input, output the predicted probability of the i-th eye disease. Each feedforward neural network has two layers of neurons, and a drop-out mechanism is added between the first and second layers of neurons to enhance the generalization ability of the model. A ReLU activation function is added after the first layer of neurons to increase the non-linear mapping ability of the model.

[0017] Key aspects of this invention include: 1. This invention addresses the characteristics of ophthalmic electronic medical records by employing a positive-negative description differentiation encoding method for each segment of text description in the medical record. It proposes a feasible and effective method for extracting positive-negative description differentiation and uses the biomedical pre-trained model MC-BERT as the encoder to obtain static encoded representations.

[0018] 2. This invention considers that the meaning of each text description should have different meanings in different contexts, and that there are dependencies between the meanings of each text description. After encoding each description using MC-BERT to obtain a static encoded representation, Multi-head Attention and Soft Attention are used to enable interaction between the static encoded representations, thereby obtaining a dynamic semantic encoded representation. This invention takes into account the overall context of the electronic medical record, models the dependencies between each text description, and obtains a more accurate semantic representation of each text description in the ophthalmology electronic medical record.

[0019] 3. Considering that different diseases require different information, this invention, based on the dynamic semantic encoding representation of each segment of the electronic medical record text, adopts a Label-wise Attention mechanism to further extract relevant information for each disease and remove noise information to obtain a disease-related information representation vector.

[0020] The beneficial effects of this invention are as follows: 1. This invention uses ophthalmic electronic medical record text as input, and has the advantages of wide disease prediction range, high applicability, and low usage requirements.

[0021] 2. This invention takes into account the significant differences between positive and negative descriptions, as well as the excessive noise in negative descriptions, and proposes a distinguishing encoding representation method. This method differentiates and extracts positive and negative descriptions and encodes them separately, thereby obtaining a more accurate representation vector and better utilizing the information contained in each segment of the electronic medical record.

[0022] 3. The semantic encoding method, semantic encoding interaction method, and disease-related information extraction method proposed in this invention can extract disease-related information and remove redundant information based on obtaining accurate dynamic encoding representation. This not only enables better vector representation of ophthalmological electronic medical records but also allows for better extraction of information important to each disease from the vector representation. Attached Figure Description

[0023] Figure 1 This is a flowchart of the ophthalmic disease information extraction method based on electronic medical records of the present invention.

[0024] Figure 2This is a schematic diagram illustrating the extraction process to differentiate between positive and negative responses.

[0025] Figure 3 This is a schematic diagram of text semantic representation.

[0026] Figure 4 This is a semantic representation interaction diagram.

[0027] Figure 5 This is a diagram illustrating disease information extraction.

[0028] Figure 6 This is a schematic diagram of the module structure of the ophthalmic disease auxiliary diagnostic device based on electronic medical records of the present invention. Detailed Implementation

[0029] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0030] In one embodiment of the present invention, a method for extracting ophthalmic disease information based on electronic medical records is provided. The process of this method is as follows: Figure 1 As shown, the process includes steps such as medical record information extraction, text semantic representation, semantic representation interaction, and disease information extraction.

[0031] In the medical record information extraction step, a combination of BiLSTM-CRF model and rule matching is used to distinguish and extract negative and positive descriptions from each segment of text in the electronic medical record (including descriptions of present illness, past medical history, and various examination results). This results in two sets of descriptions: positive and negative. Positive and negative refer to the presence or absence of a symptom or sign; a positive description indicates the presence of a certain symptom or sign, while a negative description indicates the opposite.

[0032] In the text semantic representation step, the extracted positive and negative descriptions are grouped and semantically encoded based on the MC-BERT model to obtain the semantic representation vector of each text segment.

[0033] In the semantic representation interaction step, the semantic representation vectors obtained above are used as inputs to the semantic representation interaction layer, enabling the semantic representations to interact with each other, modeling their dependencies, completing the dynamic joint representation between semantic representations, and outputting a set of dynamic semantic representation vectors.

[0034] Then, in the disease information extraction step, the label embedding vector of each disease and the dynamic semantic representation vector output by the semantic representation interaction layer are used as the input of the label-wise attention layer to extract information related to different diseases from the semantic representation vector and output the disease-related information representation vector.

[0035] 1. Medical record information extraction This method uses ophthalmology electronic medical records as input. From the perspective of a single case, the ophthalmology electronic medical record contains multiple text descriptions, including present medical history, past medical history, text descriptions of examination results for various parts of the eye, and descriptions of auxiliary examination results. Each text description can be divided into positive and negative categories. Considering that positive and negative text descriptions differ significantly in descriptive characteristics, information density, and predictive value for disease, and that negative descriptions contain more noise, this invention distinguishes and encodes positive and negative descriptions.

[0036] This method first uses a combination of deep learning information extraction models and rule matching to distinguish between positive and negative descriptions in each segment of text in electronic medical records. The extraction is performed on comma-separated clauses. The process for distinguishing between positive and negative descriptions is as follows: Figure 2 As shown. Specifically, based on the analysis and statistics of a large number of negative descriptions in each text segment, a series of matching rules are designed to extract negative descriptions with obvious patterns. Then, for negative descriptions with less obvious patterns, a BiLSTM-CRF deep learning model is trained and used to distinguish them using sequence labeling. Before inputting the text description data into BiLSTM-CRF, the input text needs to be preprocessed, such as word segmentation. This method first performs word statistical analysis on ophthalmology electronic medical record data, adding domain-specific vocabulary to a general lexicon to form a new lexicon, and then segments the input data based on JIEBA word segmentation and this lexicon.

[0037] Long Short-Term Memory (LSTM) networks are a type of recurrent neural network (RNN). Compared to ordinary RNNs, they can learn long-term dependencies more effectively, making them suitable for processing sequential inputs such as text and audio, and widely used in natural language processing. BiLSTM refers to a bidirectional long short-term memory network. Conditional Random Fields (CRFs) are discriminative probabilistic models, a type of random field, commonly used for labeling or analyzing sequential data, such as natural language text or biological sequences. The BiLSTM-CRF model is a classic sequence labeling model, often used in information extraction. The model first uses BiLSTM to read text characters sequentially from both left and right directions. Then, BiLSTM outputs the probability distribution of the sequence labeling result for each character. This probability distribution is used as input to the CRF model, which outputs the sequence labeling result for each character. Based on the sequence labeling results, information description of the text is extracted.

[0038] 2. Text Semantic Representation After the medical record information extraction step, this method obtained a set of negative descriptions and a set of positive descriptions. In the text semantic representation step, this method uses the MC-BERT model to perform semantic representation encoding on each segment of positive and negative descriptions, such as... Figure 3 As shown.

[0039] MC-BERT is a Chinese biomedical pre-trained language model, a type of BERT-based pre-trained language model. BERT is a Transformer-based pre-trained language model with powerful text representation capabilities. The overall training framework of BERT consists of two stages: pre-training and fine-tuning. In the pre-training stage, the model performs unsupervised pre-training tasks on massive corpora. In the fine-tuning stage, the parameters obtained in the pre-training stage are used to further train the model on downstream supervised tasks. MC-BERT optimizes the pre-training tasks for the biomedical domain based on the BERT model and is pre-trained on a large biomedical corpus. Compared to pre-trained language models trained on general corpora, it is better suited for downstream biomedical tasks. This method first uses a mask pre-training task to further pre-train the MC-BERT model on an ophthalmology electronic medical record corpus. Then, the model is used as the base part of the disease-assisted diagnosis model to obtain the semantic representation of the text description. The parameters of the model will be further updated through gradient backpropagation during the training process of the disease-assisted diagnosis task.

[0040] Let the two sets of input text descriptions be the positive description sets. and negative description set ,in Indicates the first i A positive description, Indicates the first i There are 1 negative description, totaling 1 m A positive description and n A negative description. Then, in this step, [the following will be done] Each description in As input to MC-BERT, a set of results is obtained. d 3D positive description vector ,Will Each description in Also used as input to MC-BERT, a set of results is obtained. d 3D negative description vector By transposing each positive and negative representation vector, we can obtain the positive description semantic representation vector matrix. , , and the semantic representation vector matrix of negative description , The calculation formula is described below.

[0041] 3. Semantic Representation Interaction After the semantic representation step, this method obtains a set of positive representation vectors. and a set of negative representation vectors Considering that the meaning of each text description should have different meanings in different contexts, and that the meanings of each text description are dependent on each other, the meaning represented by a text description should be considered in conjunction with the semantic representations of other text descriptions. The representation of that text description should interact with the semantic representations of other text descriptions and be jointly encoded. For the semantic representation of a certain text description, the semantic encoding representation vector obtained only based on its description itself without considering the other text descriptions is called the static semantic representation vector, obtained through the aforementioned semantic representation steps. and That is, the static semantic representation vector matrix.

[0042] In the semantic representation interaction step, this invention employs a multi-head self-attention mechanism and a soft attention mechanism to obtain two sets of dynamic semantic representation vectors. Specifically, as follows... Figure 4 As shown, this invention employs a multi-head self-attention mechanism layer to model the positive description static semantic representation vector matrix. The relationship between the positive and negative relationships is used to obtain a dynamic positive description representation. This layer uses a positive description static representation vector matrix. As input, output a dynamic semantic representation vector matrix of positive descriptions. Furthermore, in this step, the present invention employs a soft attention mechanism layer to model the positive description static semantic representation vector matrix. and negative description static semantic representation vector matrix The dependency relationship between them and Perform joint encoding, this layer with and Input and output interactive dynamic semantic representation vector matrix and The dimensions are respectively equivalent to .

[0043] Multi-head self-attention is a special type of multi-head attention operation. It consists of multiple independently performed attention operations, with the input being a query vector matrix. Query ) 、 Key vector matrix ( Key ) and value vector matrix ( Value Three matrices, each of which can be viewed as a set of... d Composed of dimensional row vectors. As described in formula (8), the multi-head attention mechanism uses a linear mapping method to... Query, Key and Value Mapped to different subspaces (head), the row vector dimension is changed from... d They are respectively mapped to and Then in each subspace, the corresponding Query, Key and Value The attention mechanism described in formula (9) is executed in parallel and outputs are generated. After the operations in each subspace are completed, the outputs from different subspaces are merged, and the merged vector is linearly mapped to finally obtain the output of the multi-head attention mechanism. The multi-head attention mechanism can dynamically jointly encode the input, and can pay attention to the information of different representation subspaces at different positions in the input, capturing richer features. Compared with the multi-head attention mechanism, the special feature of the multi-head self-attention mechanism is that the query vector matrix in its input ( Query ) 、 Key vector matrix ( Key ) and value vector matrix ( Value All three matrices are the same matrix.

[0044] The soft attention mechanism also operates on a query vector matrix ( Query ) 、 Key vector matrix ( Key ) and value vector matrix ( Value The three matrices are used as input, and then the Attention operation described in formula (9) will be performed. This method uses the positive description vector matrix. As a soft attention mechanism Query Describing the vector matrix in a negative form As Key and Value This allows for the extraction of valuable information from negative descriptions that is relevant to positive descriptions, and the output is a dynamic semantic representation vector matrix. .

[0045] All model parameters involved in this step will be further updated via gradient backpropagation during the training process of the ophthalmic disease auxiliary diagnosis task. The calculation formulas involved in this step are described below, where... Query, Key and Value Abbreviated as Q, K and V In formula (8) In formula (9) Representing the vector dimension, and related to the self-attention mechanism operation mentioned above. Value The row vector dimension of a matrix after being mapped to a subspace same.

[0046] In formula (9), Concat represents concatenation. This represents the result of the Attention operation performed in the i-th subspace. This represents the Attention operation described in formula (10).

[0047] 4. Disease Information Extraction After the semantic representation interaction step, two sets of dynamic semantic representation vectors are obtained, which contain the information required for disease prediction. However, different diseases require different information, meaning that different information has different levels of importance for different diseases. Some information may be important for one disease but useless or even noise for others. Therefore, in the disease information extraction step, this method uses a label-wise attention mechanism to extract relevant information from the input representation vectors for different diseases and filter out noise information.

[0048] Label-wise attention, also known as the label attention mechanism, uses labels to represent vector matrices. and feature representation vector matrix H For input, where Corresponding to the i The first label is used to indicate the first...i Each tag d 3D embedding representation vector. Label-wise Attention uses... L for Query , H for Key and Value Perform the Attention operation described in Formula 9 to output a vector matrix. ,in This is the information representation vector related to label i extracted from each feature representation vector.

[0049] As described in formulas (11) and (12), in this step, the semantic representation obtained in the interaction step is first... and By splicing The matrix is ​​then represented by ophthalmic disease label vectors. L and As input to the label attention layer, it outputs a set of vectors. ,in It is a representation vector of information related to ophthalmic disease i extracted from its input by the label attention mechanism, such as Figure 5 As shown. The model parameters involved in this step will be further updated through gradient backpropagation during the training process of the ophthalmic disease auxiliary diagnosis task. The calculation formulas involved are described below: in, d express Vector dimension in a matrix.

[0050] In one embodiment of the present invention, an auxiliary diagnostic device for ophthalmic diseases based on electronic medical records is provided, which uses the disease-related information representation vector obtained by the above method as the input of a neural network and outputs the disease prediction probability. Figure 6 As shown, the device includes: The medical record information extraction module is used to distinguish and extract negative and positive descriptions from each segment of text in the electronic medical record of ophthalmology diseases, and obtain a set of positive descriptions and a set of negative descriptions. The text semantic representation module is used to perform semantic encoding on the positive descriptions in the positive description set and the negative descriptions in the negative description set, respectively, to obtain the static semantic representation vectors of the positive descriptions and the static semantic representation vectors of the negative descriptions. The semantic representation interaction module is used to jointly encode the static semantic representation vectors of positive and negative descriptions to obtain dynamic semantic representation vectors. The disease information extraction module takes dynamic semantic representation vectors and label embedding vectors of various diseases as inputs, uses the label attention layer to extract information related to different diseases, and outputs ophthalmic disease-related information representation vectors. The disease prediction module uses vectors representing ophthalmic diseases as input to a neural network and outputs disease prediction probabilities.

[0051] The specific implementation process of the medical record information extraction module, text semantic representation module, semantic representation interaction module, and disease information extraction module can be found in the preceding description of the method of this invention.

[0052] After the disease information extraction step, a set of ophthalmic disease-related information representation vectors were obtained. In the disease prediction step, the disease prediction module uses a set of feed-forward neural network (FNN) models to represent vector matrices with ophthalmic disease-related information. As input, the output is the probability of the presence of each ophthalmic disease. Specifically, the number of models in this group is the same as the number of target ophthalmic disease categories in the auxiliary diagnostic system, and the i-th feedforward neural network model uses... Given the input, output the predicted probability of the i-th eye disease. Each feedforward neural network has two layers of neurons, with a drop-out mechanism added between the first and second layers to enhance the model's generalization ability. A ReLU activation function is added after the first layer to increase the model's non-linear mapping capability. During iterative training, the drop-out mechanism randomly selects some neurons with a certain probability to deactivate them, reducing the dependency between neurons and improving the model's robustness. The model parameters involved in this step are further updated through gradient backpropagation during the training process of the disease-assisted diagnosis task. The relevant calculation formulas are described below: in, , This represents the model parameters of the first layer of the i-th feedforward neural network. , This represents the model parameters of the second layer of the i-th feedforward neural network.

[0053] Experiments were conducted on electronic medical record datasets from real ophthalmology hospitals to test the proposed method and various multi-label text classification methods. The datasets covered ophthalmic diseases including cataracts, glaucoma, age-related macular degeneration, diabetic retinopathy, dry eye, pterygium, and vitreous opacity.

[0054] In the experiments, the parameters of the models involved in each method were first fine-tuned on the training set, and then the experimental results of each model on the test set were obtained. The evaluation metric used was the F1-score, and the following definitions were made for True Positive (TP), True Negative (TN), False Positive (FP), and False Negative (FN) (Table 1). After obtaining the experimental results of each model on the test set, the TP, TN, FP, and FN results were statistically analyzed, and the recall and precision were calculated. The F1-score was then calculated from the recall and precision, and the specific calculation formula is as follows: Table 1. Indicator Definitions Other methods used in the comparison include the BiLSTM model, BiLSTM-Attention model, textCNN model, and BERT model. These methods take as input a word sequence composed of concatenated sentences from the electronic medical record text and output a fixed-length vector as a representation of the electronic medical record text. This vector is then input into a feedforward neural network to obtain prediction results for each disease. Experimental results are shown in Table 2. The method proposed in this invention achieved the best performance, reaching an metric of 0.9362.

[0055] Table 2. Experimental Results In this invention, the MC-BERT used in the text semantic representation step can be replaced with other Transformer-based models; the multi-head attention mechanism used in the semantic representation interaction step can be replaced with other similar methods; and the feedforward neural network used in the disease prediction step can be replaced with other similar classifier models.

[0056] Based on the same inventive concept, another embodiment of the present invention provides a computer device (computer, server, smartphone, etc.) including a memory and a processor, wherein the memory stores a computer program configured to be executed by the processor, and the computer program includes instructions for performing the steps of the method of the present invention.

[0057] Based on the same inventive concept, another embodiment of the present invention provides a computer-readable storage medium (such as ROM / RAM, disk, optical disk) that stores a computer program, which, when executed by a computer, implements the various steps of the method of the present invention.

[0058] The specific embodiments of the present invention disclosed above are intended to help understand the content of the present invention and to implement it accordingly. Those skilled in the art will understand that various substitutions, changes, and modifications are possible without departing from the spirit and scope of the present invention. The present invention should not be limited to the content disclosed in the embodiments of this specification; the scope of protection of the present invention is defined by the claims.

Claims

1. A method for extracting ophthalmic disease information based on electronic medical records, characterized in that, Includes the following steps: The text in the electronic medical record of ophthalmology disease is divided into negative and positive descriptions and extracted to obtain a set of positive descriptions and a set of negative descriptions. Semantic encoding is performed on the positive descriptions in the positive description set and the negative descriptions in the negative description set to obtain static semantic representation vectors for positive and negative descriptions. The dynamic semantic representation vector is obtained by jointly encoding the static semantic representation vector of positive description and the static semantic representation vector of negative description. Taking dynamic semantic representation vectors and label embedding vectors of various diseases as input, the label attention layer is used to extract information related to different ophthalmic diseases and output ophthalmic disease-related information representation vectors. The joint encoding of the positive description static semantic representation vector and the negative description static semantic representation vector to obtain the dynamic semantic representation vector includes: A multi-head self-attention mechanism is used to model the static semantic representation vector matrix of positive descriptions. The relationship between the positive descriptions and the dynamic semantic representation vector matrix are obtained. ; A soft attention mechanism is used to model the static semantic representation vector matrix of positive descriptions. and negative description static semantic representation vector matrix The dependencies between them are described by positive vector matrices. As a soft attention mechanism Query Describing the vector matrix in a negative form As Key and Value ,right and Joint encoding is performed to obtain the interactive dynamic semantic representation vector matrix. .

2. The method for extracting ophthalmic disease information based on electronic medical records according to claim 1, characterized in that, The semantic encoding of positive descriptions in the positive description set and negative descriptions in the negative description set is based on the MC-BERT model for semantic encoding of positive and negative descriptions, including: Each positive description in the set of positive descriptions is used as input to MC-BERT to obtain a set of positive description representation vectors. Similarly, each negative description in the set of negative descriptions is used as input to MC-BERT to obtain a set of negative description representation vectors. Perform a transpose operation on each positive and negative description vector to obtain the static semantic representation vector matrix of positive descriptions. and negative description static semantic representation vector matrix .

3. The method for extracting ophthalmic disease information based on electronic medical records according to claim 1, characterized in that, The extraction of information related to different ophthalmic diseases using the label attention layer includes: Will and By splicing The matrix is ​​then represented by ophthalmic disease label vectors. L and As input to the label attention layer, the output is a vector matrix representing ophthalmic disease-related information. ,in It is the representation vector of information related to ophthalmic disease i extracted from its input by the label attention mechanism; the model parameters involved in this step are further updated through gradient backpropagation during the training process of the ophthalmic disease auxiliary diagnosis task.

4. An auxiliary diagnostic device for ophthalmic diseases based on electronic medical records, characterized in that, include: The medical record information extraction module is used to distinguish and extract negative and positive descriptions from each segment of text in the electronic medical record of ophthalmology diseases, and obtain a set of positive descriptions and a set of negative descriptions. The text semantic representation module is used to perform semantic encoding on the positive descriptions in the positive description set and the negative descriptions in the negative description set, respectively, to obtain the static semantic representation vectors of the positive descriptions and the static semantic representation vectors of the negative descriptions. The semantic representation interaction module is used to jointly encode the static semantic representation vectors of positive and negative descriptions to obtain dynamic semantic representation vectors. The disease information extraction module takes dynamic semantic representation vectors and label embedding vectors of various diseases as inputs, uses the label attention layer to extract information related to different diseases, and outputs ophthalmic disease-related information representation vectors. The disease prediction module is used to take vectors representing ophthalmic diseases as input to a neural network and output the disease prediction probability. The joint encoding of the positive description static semantic representation vector and the negative description static semantic representation vector to obtain the dynamic semantic representation vector includes: A multi-head self-attention mechanism is used to model the static semantic representation vector matrix of positive descriptions. The relationship between the positive descriptions and the dynamic semantic representation vector matrix are obtained. ; A soft attention mechanism is used to model the static semantic representation vector matrix of positive descriptions. and negative description static semantic representation vector matrix The dependencies between them are described by positive vector matrices. As a soft attention mechanism Query Describing the vector matrix in a negative form As Key and Value ,right and Joint encoding is performed to obtain the interactive dynamic semantic representation vector matrix. .

5. The ophthalmic disease auxiliary diagnostic device based on electronic medical records according to claim 4, characterized in that, The disease prediction module uses a set of feedforward neural network models, taking an ophthalmic disease-related information vector matrix as input, and outputting the probability of the existence of each ophthalmic disease.

6. The ophthalmic disease auxiliary diagnostic device based on electronic medical records according to claim 5, characterized in that, The number of feedforward neural network models in the set is the same as the number of target ophthalmic disease categories. The i-th feedforward neural network model is represented by a vector matrix with disease-related information. In Given the input, output the predicted probability of the i-th eye disease. Each feedforward neural network has two layers of neurons, and a drop-out mechanism is added between the first and second layers of neurons to enhance the generalization ability of the model. A ReLU activation function is added after the first layer of neurons to increase the non-linear mapping ability of the model.

7. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program including instructions for performing the method of any one of claims 1 to 3.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a computer, implements the method according to any one of claims 1 to 3.

Images