Dynamic self-adaptive semantic analysis method for electronic medical record

Abstract

The invention discloses a dynamic self-adaptive semantic analysis method for an electronic medical record, and the method can recognize composite medical terms in the electronic medical record when the medical terms are subjected to mask operation, so as to avoid splitting the composite medical terms, so that the overall mask operation can be carried out on the composite medical terms, and the processing efficiency of the electronic medical record is improved. The integrity of medical terms is reserved, and the accuracy of processing the electronic medical records is improved; besides, the change trend of the examination indexes and the incidence relation between the change trend and the diagnosis conclusion are captured by calculating the first layer attention weight and the second layer attention weight of each examination index; the time sequence relation of the medical data is obtained to provide more intuitive and accurate data support for subsequent analysis of the state development condition of the target object associated with the electronic medical record.

Description

Technical Field

[0001] The present disclosure generally relates to the cross - field of medical artificial intelligence and natural language processing, and specifically relates to a dynamic adaptive semantic analysis method for electronic medical records. Background Art

[0002] With the wide application of electronic medical records, the scale and complexity of medical data are constantly increasing. Therefore, it is necessary to process medical data in order to perform structured electronic medical record, clinical decision - making support, cross - institutional medical data analysis, etc. based on the processed results.

[0003] Currently, a common method for processing medical data is to perform random masking and other operations on electronic medical records using general pre - trained models (such as BERT, CMT).

[0004] The above - mentioned processing method not only causes medical terms in electronic medical records to be split, but also fails to obtain the dynamic association between examination indicators and diagnostic conclusions, resulting in the inability to provide reliable data support for subsequent applications. Summary of the Invention

[0005] In view of the above - mentioned defects or deficiencies in the prior art, it is desirable to provide a dynamic adaptive semantic analysis method for electronic medical records, which can automatically perform network access security detection and improve the reliability of network access security detection.

[0006] The present application provides a dynamic adaptive semantic analysis method for electronic medical records, and the method includes: Identifying compound medical terms in the electronic medical record through a first model to obtain the compound medical terms of the electronic medical record, where the compound medical terms include at least two independent medical terms; Performing a masking operation on the compound medical terms as a whole to obtain masked medical terms; Inputting the electronic medical record into a second model for temporal feature extraction to obtain feature vectors of at least one examination indicator, where the feature vectors include test values and / or diagnostic conclusions of the examination indicator at different times; For each examination indicator, performing hierarchical attention processing according to the feature vector of the examination indicator to obtain a first - layer attention weight and a second - layer attention weight of the examination indicator. The first - layer attention weight is used to characterize the change trend of the test value corresponding to the examination indicator, and the second - layer attention weight is used to characterize the association relationship between the change trend and the diagnostic conclusion corresponding to the examination indicator; Analyzing the state development of the target object associated with the electronic medical record based on the masked medical terms, the first attention weight, and the second attention weight to obtain an analysis result, where the analysis result is used to characterize the state development of the target object.

[0007] In an optional embodiment, for each of the inspection indicators, hierarchical attention processing is performed according to the feature vector of the inspection indicator to obtain the first-layer attention weight and the second-layer attention weight of the inspection indicator, including: Determine the first-layer attention weight according to the feature vector of the inspection indicator and the first-layer attention parameter; Determine the second-layer attention weight according to the product of the first-layer attention weight and the feature vector of the inspection indicator and the second-layer attention parameter.

[0008] In an optional embodiment, the method for analyzing the state development of the target object associated with the electronic medical record based on the masked medical term, the first attention weight, and the second attention weight to obtain an analysis result, where the analysis result is used to characterize the state development of the target object, specifically includes: Perform state development analysis on the target object associated with the electronic medical record based on the masked medical term, the first attention weight, the second attention weight, and the text feature encoding vector to obtain the analysis result.

[0009] In an optional embodiment, the method further includes: Input the electronic medical record into a third model to extract text features in the electronic medical record; Perform feature encoding on the text features in the electronic medical record to obtain the text feature encoding vector.

[0010] In an optional embodiment, the method further includes the steps of training the second model and the third model: Obtain a plurality of first text features, a plurality of second text features, and a plurality of numerical features as positive samples, where the first text features are the features in the electronic medical record representing the diagnostic conclusions of the inspection indicators, and the second text features are the text features in the electronic medical record other than the first text features; Input the plurality of first text features and the plurality of numerical features into a first initial model for model training to obtain the second model; Input the plurality of second text features into a second initial model for model training to obtain the third model.

[0011] In an optional embodiment, the method further includes: Obtain verification text and verification numerical features, where the verification text is the text feature in the electronic medical record other than the diagnostic conclusion representing the inspection indicator; Input the verification numerical features into the second model to obtain verification numerical vectors; Input the verification text features into the third model to obtain verification text vectors; Calculate the cosine similarity loss between the verification numerical vectors and the verification text vectors; Determine the semantic alignment degree between the verification numerical vectors and the verification text vectors according to the cosine similarity loss; If the semantic alignment degree does not meet the first target requirement, adjust the model parameters of the second model and the third model for retraining until the semantic alignment degree meets the first target requirement; If it meets, output the second model and the third model.

[0012] In an optional embodiment, the method further includes the step of training the first model: Obtain a plurality of independent medical terms and the context information of each independent medical term from a medical knowledge base; Annotate each independent medical term based on the context information of each independent medical term to obtain an initial sample set; Convert the independent medical terms in each initial sample set into word embedding vectors; Input each word embedding vector into a third initial model for model training to obtain the first model.

[0013] In an optional embodiment, the method further includes: Obtain a verification medical term sample; Input the verification medical term sample into the first model to obtain a verification composite medical term; Input the verification composite medical term into a domain discriminator to obtain a first probability value, and determine whether the first probability value meets the second target requirement; If the first probability value does not meet the second target requirement, adjust the training parameters of the first initial model for retraining until the first probability value meets the second target requirement; If the first probability value meets the second target requirement, output the first model.

[0014] In an optional embodiment, the method further includes: Calculate a first adversarial training loss value of the first probability value, and determine whether the first adversarial training loss value meets the third target requirement; If the first adversarial training loss value does not meet the third target requirement, adjust the training parameters of the third initial model for retraining until the first adversarial training loss value meets the third target requirement; If the first adversarial training loss value meets the third target requirement, output the first model.

[0015] In an alternative embodiment, the method further includes: Inputting the verification text feature and the verification numerical feature into a domain discriminator respectively to obtain a second probability value and a third probability value, and determining whether the second probability value and the third probability value meet a fourth target requirement; If the second probability value and the third probability value do not meet the fourth target requirement, adjusting the model parameters of the first initial model and the second initial model to retrain until the second probability value and the third probability value meet the fourth target requirement; If the second probability value and the third probability value meet the fourth target requirement, outputting the second model and the third model.

[0016] In an alternative embodiment, the method further includes: Calculating a second adversarial training loss value and a third adversarial training loss value of the second probability value and the third probability value, and determining whether the second adversarial training loss value and the third adversarial training loss value meet a fifth target requirement; If the second adversarial training loss value and the third adversarial training loss value do not meet the fifth target requirement, adjusting the model parameters of the first initial model and the second initial model to retrain until the second adversarial training loss value and the third adversarial training loss value meet the fifth target requirement; If the second adversarial training loss value and the third adversarial training loss value meet the fifth target requirement, outputting the second model and the third model.

[0017] A dynamic adaptive semantic analysis method for electronic medical records provided by the present application takes into account the problem that currently, a pre-trained model is usually used to perform random masking and other processing on electronic medical records, resulting in the splitting of compound medical terms in the electronic medical records and the inability to obtain the dynamic association between inspection indicators and diagnostic conclusions, thus unable to provide reliable data support for subsequent applications. When performing a masking operation on medical terms, the present application will identify compound medical terms in the electronic medical records to avoid splitting the compound medical terms, so that an overall masking operation can be performed on the compound medical terms, retaining the integrity of the medical terms and improving the accuracy of processing the electronic medical records; in addition, the present application captures the change trend of inspection indicators and the association relationship between the change trend and the diagnostic conclusion by calculating the first-layer attention weight and the second-layer attention weight of each inspection indicator, and obtains the time series relationship of medical data, providing more intuitive and accurate data support for analyzing the state development of the target object associated with the electronic medical record subsequently. Description of the Drawings

[0018] Other features, objectives, and advantages of the present application will become more apparent by reading the detailed description of the non-limiting embodiments with reference to the following drawings: Figure 1 A scenario diagram of the application of a dynamic adaptive semantic analysis method for electronic medical records provided by the present application; Figure 2 A flowchart of a dynamic adaptive semantic analysis method for electronic medical records provided by the present application; Figure 3 A step flowchart of a dynamic adaptive semantic analysis method for electronic medical records provided by the present application; Figure 4 A step flowchart of another dynamic adaptive semantic analysis method for electronic medical records provided by the present application; Figure 5 A step flowchart of another dynamic adaptive semantic analysis method for electronic medical records provided by the present application; Figure 6 A step flowchart of a dynamic adaptive semantic analysis method for electronic medical records provided by the present application; Figure 7 A step flowchart of another dynamic adaptive semantic analysis method for electronic medical records provided by the present application; Figure 8 A step flowchart of another dynamic adaptive semantic analysis method for electronic medical records provided by the present application; Figure 9 A step flowchart of another dynamic adaptive semantic analysis method for electronic medical records provided by the present application; Figure 10 A step flowchart of another dynamic adaptive semantic analysis method for electronic medical records provided by the present application; Figure 11 A step flowchart of another dynamic adaptive semantic analysis method for electronic medical records provided by the present application; Figure 12 A step flowchart of another dynamic adaptive semantic analysis method for electronic medical records provided by the present application; Figure 13 A step flowchart of another dynamic adaptive semantic analysis method for electronic medical records provided by the present application; Figure 14 A step flowchart of another dynamic adaptive semantic analysis method for electronic medical records provided by the present application; Figure 15 A step flowchart of another dynamic adaptive semantic analysis method for electronic medical records provided by the present application; Figure 16 A step flowchart of another dynamic adaptive semantic analysis method for electronic medical records provided by the present application. Detailed implementation manners

[0019] The present application will be further described in detail below with reference to the accompanying drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the related invention, rather than limiting the invention. In addition, it should be noted that, for the sake of description, only the parts related to the invention are shown in the drawings.

[0020] It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. The present application will be described in detail below with reference to the drawings and embodiments.

[0021] Please refer to Figure 1 , Figure 1 , which is an application scenario diagram of a dynamic adaptive semantic analysis method for electronic medical records provided by the present application. The application scenario diagram includes a terminal device 100 and a service device 200. The user uploads the electronic medical record of patient A to the service device 200 through the terminal device 100. The first model set on the service device 200 is used to identify the composite medical terms in the electronic medical record of patient A and perform a masking operation; then the second model is used to extract the temporal features of the electronic medical record to obtain the feature vectors of multiple inspection indicators in the electronic medical record; finally, hierarchical attention processing is performed on the feature vectors of the multiple inspection indicators to obtain the first-layer attention weights and the second-layer attention weights of each inspection indicator, so as to analyze the state development of patient A through the masked medical terms, the first-layer attention weights, and the second-layer attention weights after the above processing, and obtain the state development of patient A. Among them, the terminal device 100 is, for example, a smart phone, a tablet, a notebook computer, a desktop computer, a smart watch, etc. The service device 200 is, for example, a single server, a server cluster, etc.

[0022] Next, the detection method provided by the present application will be described.

[0023] As Figure 2 shown, Figure 2 , which is a step flowchart of a dynamic adaptive semantic analysis method for electronic medical records provided by the present application. The method includes: Step S20, identifying medical terms in the electronic medical record through a first model to obtain a medical term set of the electronic medical record. The medical term set includes composite medical terms, and the composite medical terms include at least two independent medical terms; Among them, the first model is a neural network model carried on a service device for medical term recognition. This neural network model can be obtained by training with a medical knowledge base. Exemplarily, the first model is a named entity recognition (NER) model. This first model can not only recognize independent medical terms, such as white blood cells, hypertension, heart rate, blood pressure, hemoglobin; but also recognize compound medical terms. Among them, a compound medical term is a medical term that includes at least two independent medical terms, which replaces long descriptive words, achieving the unity of conciseness and professionalism. It not only conforms to the multi-source characteristics of medical terms but also facilitates accurate expression in cross-disciplinary communication. For example, glycated hemoglobin, infectious pneumonia, white blood cell count, etc.

[0024] The electronic medical record can be the electronic medical record of a certain patient uploaded by a terminal device, or the electronic medical records of multiple patients uploaded by a terminal device. This application does not limit this. When the electronic medical records uploaded by the terminal device are those of multiple patients, each electronic medical record can include a unique identifier of the patient, so as to separately process the electronic medical records of different patients and obtain medical analysis data for each patient.

[0025] This application uses the first model to perform medical term recognition on the electronic medical record, which can not only recognize independent medical terms but also accurately recognize compound medical terms in the electronic medical record, so as to avoid the situation where the splitting of compound medical terms leads to inaccurate medical analysis data provided.

[0026] Step S30: Perform an overall masking operation on the medical term set to obtain masked medical terms; Among them, the masking operation refers to the technique of replacing the entire medical term with Mask. Since the above first model can recognize compound medical terms, the overall masking operation can be performed on the compound medical terms, ensuring the integrity of semantics and solving the problem of damaging the semantic integrity in the recognition of medical terms in the prior art.

[0027] Exemplarily, the content in the input electronic medical record includes the following parts: Patient, male, 56 years old, complained of chest pain for 3 hours. He had a history of hypertension for 10 years and did not take medicine regularly. The patient developed chest pain without obvious cause 3 hours ago. The pain was located in the precordial area, was compressive, accompanied by profuse sweating, and did not relieve continuously. He had a history of hypertension, with the highest blood pressure reaching 180 / 110 mmHg, and did not take antihypertensive medicine regularly. Denied a history of diabetes and coronary heart disease. Denied a family history of cardiovascular diseases. Physical examination: body temperature 36.8°C, pulse 100 beats per minute, respiratory rate 20 breaths per minute, blood pressure 160 / 100 mmHg. No abnormalities were heard on auscultation of the heart and lungs. Electrocardiogram showed: sinus rhythm, ST segment elevation of 0.2 - 0.3 mV in leads V1 - V6. Preliminary diagnosis: acute myocardial infarction. Immediately gave 300 mg of aspirin to chew, 300 mg of clopidogrel orally, and 4000 IU of low molecular weight heparin by subcutaneous injection. The patient was admitted to the CCU ward and further underwent coronary angiography, which found that the proximal segment of the left anterior descending artery was completely occluded. Immediately underwent PCI and successfully opened the blood vessel. After the operation, antiplatelet, lipid-lowering, vasodilating and other treatments were given. The patient's condition gradually stabilized, chest pain relieved, and vital signs were stable.

[0028] Medical term recognition is performed through the first model, and the medical term set obtained is: Chest pain, hypertension, coronary heart disease, acute myocardial infarction; Perform an overall masking operation on the medical term set, and the following output is obtained: Patient, male, 56 years old, complained of [MASK] for 3 hours. He had a history of [MASK] for 10 years and did not take medicine regularly. The patient developed [MASK] without obvious cause 3 hours ago. The pain was located in the precordial area, was compressive, accompanied by profuse sweating, and did not relieve continuously. He had a history of [MASK], with the highest blood pressure reaching 180 / 110 mmHg, and did not take antihypertensive medicine regularly. Denied a history of diabetes and [MASK]. Denied a family history of cardiovascular diseases. Physical examination: body temperature 36.8°C, pulse 100 beats per minute, respiratory rate 20 breaths per minute, blood pressure 160 / 100 mmHg. No abnormalities were heard on auscultation of the heart and lungs. Electrocardiogram showed: sinus rhythm, ST segment elevation of 0.2 - 0.3 mV in leads V1 - V6. Preliminary diagnosis: [MASK]. Immediately gave 300 mg of aspirin to chew, 300 mg of clopidogrel orally, and 4000 IU of low molecular weight heparin by subcutaneous injection. The patient was admitted to the CCU ward and further underwent coronary angiography, which found that the proximal segment of the left anterior descending artery was completely occluded. Immediately underwent PCI and successfully opened the blood vessel. After the operation, antiplatelet, lipid-lowering, vasodilating and other treatments were given. The patient's condition gradually stabilized, [MASK] relieved, and vital signs were stable.

[0029] Step S40, input the electronic medical record into the second model for temporal feature extraction to obtain the feature vectors of at least one examination index, where the feature vectors include the values of the examination index at different times and / or the diagnosis conclusion; Among them, the second model is a neural network model carried on the service device for extracting time series features. This neural network model can be obtained through training based on a recurrent neural network (LSTM or GRU).

[0030] Time series features refer to some features in the electronic medical record that have variable quantities over time. They can include the values of examination indicators at different times, or the diagnostic conclusions of examination indicators at different times. Exemplarily, the time series features are the count value of white blood cell count 12.5×10^9 / L, blood pressure value 160 / 100 mmHg, ST segment elevation value of V1-V6 leads 0.2-0.3 mV, etc.

[0031] Performing time series feature extraction on time series features means vectorizing the values and / or diagnostic conclusions of examination indicators. Since there can be multiple examination indicators in an electronic medical record, the second model can at least extract the feature vectors of one examination indicator.

[0032] In addition, in this application, after extracting the time series features of the examination indicators at different times, the feature vectors of the same examination indicator can be sorted according to time sequences such as weeks, months, and years to obtain the feature vector set corresponding to each examination indicator.

[0033] Exemplarily, the electronic medical record input into the second model includes the following parts: 2024-02-15 09:00: White blood cell count 12.5X10^9 / L, diagnosed with infectious pneumonia.

[0034] 2024-02-15 10:00: Heart rate 100 beats / min, blood pressure 160 / 100 mmHg.

[0035] 2024-02-15 11:00: Electrocardiogram shows: Sinus rhythm, ST segment elevation of V1-V6 leads 0.2-0.3mV.

[0036] 2024-02-18 09:00: White blood cell count 15.0X10^9 / L, diagnosed with aggravated infection.

[0037] 2024-02-18 10:00: Heart rate 95 beats / min, blood pressure 150 / 95 mmHg.

[0038] 2024-02-18 11:00: Electrocardiogram shows: Sinus rhythm, ST segment elevation of V1-V6 leads 0.1-0.2 mV.

[0039] The obtained output is: h_1 = [12.5, NaN, NaN, NaN, NaN] h_2 = [NaN, 100.0, 160, 100, NaN] h_3 = [NaN, NaN, NaN, NaN, 0.3] h_4 = [15.0, NaN, NaN, NaN, NaN] h_5 = [NaN, 95.0, 150, 95, NaN] h_6 = [NaN, NaN, NaN, NaN, 0.2] Sorting the time series feature vectors gives: h_11 = [12.5, Infectious pneumonia, 15, Worsening of infection] h_21 = [100.0, 95] h_31 = [160 / 100, 150 / 95] h_41 = [Arrhythmia, Sinus rhythm] h_51 = [0.2 - 0.3, 0.1 - 0.2] By extracting the time series features in the electronic medical record, the present application can capture the dynamic changes of each time series feature in the time series, providing data support for subsequent hierarchical attention processing.

[0040] Step S50: For each inspection index, perform hierarchical attention processing according to the feature vector of the inspection index to obtain the first - layer attention weight and the second - layer attention weight of the inspection index. The first - layer attention weight is used to characterize the change trend of the test value corresponding to the inspection index, and the second - layer attention weight is used to characterize the correlation between the change trend and the diagnostic conclusion corresponding to the inspection index. Among them, after extracting the time series features of the electronic medical record according to the above - mentioned second model, feature vectors of at least one inspection index are obtained. Since only a single feature vector is obtained by time series feature extraction, it cannot reflect the deeper correlation and logical relationship of each inspection index. Therefore, the present application will perform hierarchical attention processing on each extracted inspection index to capture the change trend of the inspection index over time, as well as the correlation between the change trend of the inspection index and the diagnostic conclusion corresponding to the inspection index, so as to provide more intuitive and more conducive - to - analysis medical data for users.

[0041] Hierarchical attention processing refers to calculating the first-layer attention weights and the second-layer attention weights for the feature vectors of each inspection index, so as to characterize the change trend of the value corresponding to the inspection index through the first-layer attention weights, and to characterize the correlation between the change trend and the diagnostic conclusion corresponding to the inspection index through the second-layer attention weights.

[0042] Exemplarily, the present application can extract temporal features in the following manner:

[0043] z is the time-series data; h is the time-series feature vector; θ RNN is the model parameter.

[0044] Exemplarily, the input time-series feature vector is: h_11 = [12.5, Infectious pneumonia, 15, Worsening infection] h_21 = [100.0, 95] h_31 = [160 / 100, 150 / 95] h_41 = [Arrhythmia, Sinus rhythm] h_51 = [0.2 - 0.3, 0.1 - 0.2] The output first-layer attention weight a1 is: a1_1 = 0.8 a1_2 = 0.7 a1_3 = 0.6 a1_4 = 0.9 a1_5 = 0.85 The output second-layer attention weight a2 is: a2_1 = 0.75 a2_2 = 0.65 a2_3 = 0.55 a2_4 = 0.85 a2_5 = 0.8 Exemplarily, as Figure 3 shown, Figure 3 is the schematic diagram of the hierarchical temporal attention architecture provided by the present application: Input: Medical test indicators and diagnostic conclusions: Represents the input data, including the test indicators of the patient (such as white blood cell count, blood glucose level, etc.) and diagnostic conclusions (such as infection, diabetes, etc.).

[0045] Hierarchical temporal attention architecture: The core architecture for processing input data and extracting temporal correlations.

[0046] Temporal Feature Extraction Layer: Extract temporal features from the input data and represent the test indicators and diagnostic conclusions at each time point as vectors.

[0047] First Layer of Attention Mechanism: Capture the temporal correlations between test indicators, such as the trend of changes in white blood cell count.

[0048] Second Layer of Attention Mechanism: Capture the dynamic correlations between test indicators and diagnostic conclusions, such as the logical relationship between changes in white blood cell count and the development of the infection course.

[0049] Output Layer: Temporal Correlation Result: Output the result of temporal correlation to provide support for clinical decision-making.

[0050] Clinical Decision Support System: Finally, apply the temporal correlation result to the clinical decision support system to help doctors make more accurate diagnoses.

[0051] Step S60, analyze the status development of the target object associated with the electronic medical record based on the masked medical terms, the first-layer attention weights, and the second-layer attention weights, and obtain an analysis result, where the analysis result is used to characterize the status development of the target object.

[0052] Among them, after obtaining the masked medical terms, the first-layer attention weights, and the second-layer attention weights according to the above operations, the user can analyze the status development of the target object associated with the electronic medical record based on these data to obtain an analysis result.

[0053] Exemplarily, according to the masked medical terms, it is determined that the user has symptoms of chest pain and is initially diagnosed with acute myocardial infarction. The user determines that the change in the white blood cell count of the target object over time is continuously increasing according to a1_1, and determines that pneumonia is worsening according to a2_1 = 0.75.

[0054] In an optional embodiment, as Figure 4 shown, Figure 4 Steps for obtaining the first-layer attention weights and the second-layer attention weights provided by an exemplary embodiment of the present application: Step S301, determine the first-layer attention weights according to the feature vectors of the inspection indicators and the first-layer attention parameters; Among them, the first-layer attention weights of the present application can be obtained through the following method:

[0055] Among them, h is the temporal feature vector; a1 is the first-layer attention weight, and θ att1 is the first-layer attention parameter.

[0056] Step S302: Determine the second-layer attention weight according to the product of the first-layer attention weight and the feature vector of the inspection index and the second-layer attention parameter.

[0057] Among them, the second-layer attention weight can be obtained by the following method in this application:

[0058] Among them, h is the time-series feature vector; a1 is the first-layer attention weight, and θ att2 is the first-layer attention parameter; a2 is the first-layer attention weight, represents element-wise multiplication.

[0059] The first-layer attention weight is used to capture the time-series correlation between inspection indexes, such as the change trend of white blood cell count; the second-layer attention weight further captures the correlation between inspection indexes and diagnostic conclusions, such as the logical relationship between the change of white blood cell count and the development of the pneumonia infection course. Hierarchical attention processing can more accurately understand the time-series relationship of medical data, so as to provide more accurate and intuitive data support for clinical decision-making.

[0060] In another optional embodiment, the analysis result can also be obtained by the following method in this application: Perform a state development analysis on the target object associated with the electronic medical record based on the masked medical term, the first attention weight, the second attention weight, and the text feature encoding vector to obtain the analysis result.

[0061] Among them, when performing the state development analysis on the target object above, only the majority values of the inspection indexes and some text features (diagnostic conclusions) are feature-encoded (that is, time-series feature extraction), and other text features in the electronic medical record are not feature-encoded, resulting in some text features and numerical features not being in the same semantic space, which cannot provide a data basis for cross-modal contrast learning. Therefore, this application also performs a state development analysis on the target object associated with the electronic medical record through the text feature vector to obtain a more complete and accurate analysis result.

[0062] It should be noted here that the steps of feature-encoding the text features in the electronic medical record to obtain the text feature encoding vector include the following: Input the electronic medical record into the third model for text feature encoding to obtain the text feature encoding vector.

[0063] Among them, the third model is a neural network model used for text feature encoding and installed on the service device. This neural network model can be a BERT model and a multi-layer perceptron (MLP).

[0064] This application can perform text feature encoding in the following way:

[0065] Among them, f t is the text feature vector, and is the model parameter.

[0066] For example, Figure 5 as shown, in an optional embodiment, the present application further provides a step of training a second model and a third model: Step S401, obtain a plurality of first text features, a plurality of second text features, and a plurality of numerical features as positive samples, where the first text features are the features representing the diagnostic conclusions of examination indicators in the electronic medical record, and the second text features are the text features in the electronic medical record other than the first text features; Among them, the first text features are, for example, infectious pneumonia, myocardial infarction, myocarditis, hypertension, etc. The second text features are, for example, white blood cell count, heart rate, electrocardiogram indication, blood pressure, ST segment elevation in leads V1-V6, etc.

[0067] The first text features, the second text features, and the numerical features can be obtained from the historical electronic medical records stored in the memory, or can be obtained by accessing different hospital systems. The present application does not limit this. Using the data obtained from different systems as samples for model training can enable the trained model to maintain good performance when facing new data from different institutions or fields.

[0068] Step S402, input the plurality of first text features and the plurality of numerical features into a first initial model for model training to obtain a second model; Among them, the first initial model is, for example, a BERT model.

[0069] Step S403, input the plurality of second text features into a second initial model for model training to obtain a third model.

[0070] Among them, the second initial model is, for example, a multi-layer perceptron.

[0071] It should be noted here that model training is a conventional technical means and will not be elaborated here.

[0072] For example, Figure 6 as shown, in an optional embodiment, the present application further provides a step of training a second model and a third model: Step S501, obtain verification text features and verification numerical features, where the verification text features are the text features in the electronic medical record other than the diagnostic conclusions representing the examination indicators; Among them, during the above model training, the text features in the obtained electronic medical records except for the diagnostic conclusions representing inspection indicators can be divided into a training set and a validation set according to a ratio, so as to perform model training through the above steps and perform model validation through the following steps. Similarly, the numerical features obtained during model training can also be divided into a training set and a validation set according to a ratio.

[0073] Step S502: Input the validation numerical features into the second model to obtain a validation numerical vector; Step S503: Input the validation text features into the third model to obtain a validation text vector; Step S504: Calculate the cosine similarity loss between the validation numerical vector and the validation text vector; Among them, the present application can calculate the cosine similarity loss through the following method: First, calculate the cosine similarity between the validation numerical vector and the validation text vector:

[0074] Among them, ft is the text feature vector; fv is the numerical feature vector.

[0075] The fv numerical feature vector can be obtained through the following formula:

[0076] Among them, fv is the numerical feature vector, and MLP represents a multi-layer perceptron. are the MLP parameters.

[0077] Next, calculate the cosine similarity loss between the two:

[0078] Among them, Lcos is the cosine similarity loss.

[0079] Step S505: Determine the semantic alignment degree between the validation numerical vector and the validation text vector according to the cosine similarity loss; if the semantic alignment degree does not meet the first target requirement, execute step S506; if the semantic alignment degree meets the first target requirement, execute step S507; Step S506: Adjust the model parameters of the first initial model and the second initial model for retraining until the semantic alignment degree meets the first target requirement; Step S507: Output the second model and the third model.

[0080] Among them, the first objective requirement is used to limit the value of the cosine similarity loss. Then, in this application, it can be determined that the semantic alignment degree between the verification numerical vector and the verification text vector is low when the calculated cosine similarity loss is greater than the preset threshold, and it can be determined that the semantic alignment degree between the verification numerical vector and the verification text vector is high when the cosine similarity loss is less than the preset threshold. This application does not limit the preset threshold, which can be set according to the specific application environment. Exemplarily, the first objective requirement is that the cosine similarity loss is equal to 0.05.

[0081] In this application, by continuously adjusting the model parameters of the first initial model and the second initial model, the feature vectors obtained from the first initial model and the second initial model are aligned to minimize the cosine similarity loss, so that the text features and numerical features are closer in the semantic space. This optimization mechanism can improve the accuracy of multimodal data fusion to help the second model and the third model obtained by training better understand the semantic relationship between different modal data.

[0082] Exemplarily, the verification text feature vector ft: ft = [0.5, 0.6, 0.7, 0.8, 0.9] The verification numerical feature vector fv: fv = [0.4, 0.5, 0.6, 0.7, 0.8] The cosine similarity Sim(ft, fv) = 0.95 The cosine similarity loss Lcos = 0.05 The calculated cosine similarity loss is equal to 0.05, indicating that the semantic alignment degree has met the first objective requirement, and the second model and the third model can be output.

[0083] As Figure 7 shown, it is a schematic diagram of the cross-modal contrast learning module: Input: Medical text and numerical data: It represents the input multimodal data, including medical text descriptions (such as medical record texts) and numerical data (such as test indicators, imaging features, etc.).

[0084] Text tower: Text encoder: Encodes medical text data into text feature vectors, usually using a pre-trained language model (such as BERT) for encoding.

[0085] Numerical tower: Numerical encoder: Encodes numerical data into numerical feature vectors, and neural networks or other encoding methods can be used.

[0086] Cross-modal alignment module: Maps text feature vectors and numerical feature vectors to the same semantic space for contrast learning.

[0087] Loss function: Cosine similarity loss, which calculates the cosine similarity between the text feature vector and the numerical feature vector, and optimizes the alignment effect through the loss function.

[0088] Output: Optimized cross-modal representation: Output the optimized cross-modal representation for subsequent multi-modal data analysis.

[0089] Multi-modal data analysis system: Apply the optimized cross-modal representation to the multi-modal data analysis system, such as clinical decision support or disease prediction.

[0090] In yet another alternative embodiment, as Figure 8 shown, the present application also provides a step of training the second model and the third model: Step S601, input the verification text features and verification numerical features into the domain discriminator respectively, obtain the second probability value and the third probability value, and determine whether the second probability value and the third probability value meet the fourth target requirement; if the second probability value and the third probability value do not meet the fourth target requirement, then execute step S602; if the second probability value and the third probability value meet the fourth target requirement, then execute step S603; Step S602, adjust the model parameters of the first initial model and the second initial model and retrain until the second probability value and the third probability value meet the fourth target requirement; Step S603, output the second model and the third model.

[0091] Among them, when performing the above model training, the text features in the obtained electronic medical records except for the diagnostic conclusions representing the examination indicators can be divided into a training set and a verification set according to a ratio, so as to perform model training through the above steps and perform model verification through the following steps. Similarly, the numerical features obtained during model training can also be divided into a training set and a verification set according to a ratio.

[0092] The domain discriminator is, for example:

[0093] Among them, is the output of the domain discriminator, indicating the probability that the input data belongs to a specific domain (or institution) under the condition of the given feature representation f. θ D is the discriminator parameter.

[0094] The fourth target requirement is used to limit the range of the second probability value and the third probability value. For the second probability value and the third probability value, the fourth target requirement can be different or the same. The present application does not limit this, and it can be set according to the specific application scenario. Exemplarily, the fourth target requirement is, for example, that the output probability value is 0.5.

[0095] Adjusting parameters for model training is a conventional technical means for model training, and this application does not limit it. For example Figure 9 As shown, this application obtains a cross-institutional domain adaptive model through cross-modal comparison learning, which is used to weaken the ability of the model to distinguish which specific institution the data comes from, and obtain a clinical decision support result. Thus, the technical effect of improving the domain adaptability of the model is achieved.

[0096] Exemplarily: the text feature representation f = [0.5, 0.6, 0.7, 0.8, 0.9], and the domain discrimination probability p(d∣f): = 0.5, indicating insensitivity to the domain, that is, the second probability value meets the fourth target requirement, and the second model can be output.

[0097] In another optional embodiment, as Figure 10 shown, this application also provides a step of training a second model and a third model: Step S701, calculate the second adversarial training loss value and the third adversarial training loss value of the second probability value and the third probability value, and determine whether the second adversarial training loss value and the third adversarial training loss value meet the fifth target requirement; if the second adversarial training loss value and the third adversarial training loss value do not meet the fifth target requirement, then execute step S702; if the second adversarial training loss value and the third adversarial training loss value meet the fifth target requirement, then step S703; Among them, adversarial training is used to train electronic medical record data from different institutions as different domains, so that the model can automatically adapt to the medical data environment of different institutions, and further improve the adaptive ability of the model.

[0098] The fifth target requirement is used to limit the range of the adversarial loss value, which can be set according to the specific application scenario and is not limited here. Exemplarily, the fifth target requirement is that the adversarial training loss value is 0.1.

[0099] This application can calculate the adversarial training loss through the following formula:

[0100] where L adv is the adversarial training loss, is the output of the domain discriminator, which is used to measure the performance of the model in adversarial training.

[0101] Step S702, adjust the model parameters of the first initial model and the second initial model and retrain until the second adversarial training loss value and the third adversarial training loss value meet the fifth target requirement; Step S703, output the second model and the third model.

[0102] Among them, adjusting parameters to adjust model training is a conventional technical means for model training, and this application does not limit it. This application intervenes in the process of model training by setting adversarial training losses, enabling the model to learn more stable and generalizable feature representations, so that when facing new data from different institutions or fields, it can maintain good performance. This training method can improve the adaptability of the model in cross-institutional medical data analysis.

[0103] Exemplarily, the text feature representation f = [0.5, 0.6, 0.7, 0.8, 0.9]; p(d∣f): = 0.5; L adv = 0.1. It means being insensitive to the domain, that is, the adversarial training loss value meets the fifth target requirement, and the second model can be output.

[0104] Here, it should be noted that when training the second model and the third model, the model training can be intervened by any one of the above semantic alignment degree, probability value, and anti-training loss value to achieve the technical effect of improving the performance of the model; any two can also be selected for model training intervention to achieve the technical effect of improving the performance of the model; or the model training can be intervened in the above three ways in sequence to achieve the technical effect of improving the performance of the model, and this application does not limit it.

[0105] In an alternative embodiment, as Figure 11 shown, Figure 11 The steps for training the first model provided by an exemplary embodiment of this application are as follows: Step S801, obtain a plurality of independent medical terms, a plurality of compound medical terms, and the context information of each compound medical term from the medical knowledge base; Among them, the medical knowledge base includes a large number of independent medical terms, compound medical terms, and the context information of each compound medical term. It can provide reliable training samples for model training to obtain a model with better performance.

[0106] Step S802, label each compound medical term according to the context information of each compound medical term to obtain intermediate medical terms, and use the set including each intermediate medical term and each independent medical term as the initial sample set; Among them, since a compound medical term can include at least two independent medical terms and can be combined with prefixes, suffixes, etc. to obtain new medical terms, if recognized by an ordinary model, the compound medical term is easily split into multiple independent medical terms, which will lead to incorrect identification of electronic medical records and further affect the subsequent analysis of the status development.

[0107] Therefore, the accurate recognition of compound medical terms can improve the accuracy of analyzing the status development of the target object.

[0108] Therefore, when performing model training, medical terms with specific context relationships can be recognized without splitting and output as a whole medical term. Then, when training the model, it is necessary to train in combination with the context information of compound medical terms to obtain a model capable of recognizing compound medical terms. Then, after obtaining the compound medical terms and their context information through the medical knowledge base, the compound medical terms can be labeled and form a training set with independent medical terms for subsequent model training.

[0109] Step S803: Perform word embedding vector conversion processing on each initial sample set to obtain word embedding vectors. Among them, the word embedding vector conversion processing can map words to low-dimensional dense vectors, reduce the data dimension and reduce redundant information.

[0110] Step S804: Input each word embedding vector into the third initial model for model training to obtain the first model.

[0111] Among them, the third initial model is, for example, a named entity recognition (NER) model. Model training is a conventional technical means and will not be elaborated here.

[0112] As Figure 12 shown, when training the first model, the present application also performs the following operations: Step S901: Obtain verification medical term samples. Step S902: Input the verification medical term samples into the first model to obtain verified compound medical terms. Step S903: Input the verified compound medical terms into the domain discriminator to obtain a first probability value, and determine whether the first probability value meets the second target requirement; if the first probability value does not meet the second target requirement, then execute Step S904; if the first probability value meets the second target requirement, then execute Step S905. Step S904: Adjust the training parameters of the third initial model and retrain until the first probability value meets the second target requirement. Step S905: Output the first model.

[0113] As Figure 13 shown, when training the first model, the present application also performs the following operations: Step S1001: Calculate the first adversarial training loss value of the first probability value, and determine whether the first adversarial training loss value meets the third target requirement. If the first adversarial training loss value does not meet the third target requirement, execute Step S1002; if the first adversarial training loss value meets the third target requirement, execute Step S1003; Step S1002: Adjust the training parameters of the third initial model for retraining until the first adversarial training loss value meets the third target requirement; Step S1003: Output the first model.

[0114] The intervention in the training process of the first model is the same as that of the second and third models mentioned above, and will not be elaborated here.

[0115] Exemplarily combined with Figure 14 as shown, the training process of the first model is described as follows: Input: Electronic medical record data from different institutions: represents the electronic medical record data from different medical institutions, including the data of Institution A, Institution B, and Institution C.

[0116] Pre-trained language model: Input the data from different institutions into the pre-trained language model for feature extraction.

[0117] Feature extraction layer: Extract the feature representation of the input data, providing a basis for subsequent adversarial training.

[0118] Feature representation: The extracted feature vector is used for subsequent adversarial training.

[0119] Adversarial training module: Includes a generator and a discriminator, used to train the model to adapt to the data of different institutions.

[0120] Domain discriminator: The discriminator is used to predict the domain label of the feature representation (i.e., from which institution).

[0121] Domain label prediction: The discriminator outputs the domain label prediction result.

[0122] Generator optimization: The generator optimizes the feature representation through adversarial training to make it have better adaptability among different institutions.

[0123] Output: Domain-adaptive features: The feature representation optimized through adversarial training has cross-institutional adaptability.

[0124] Cross-institutional data analysis: Use the optimized features for cross-institutional medical data analysis, such as joint modeling or data sharing.

[0125] Exemplarily, in combination with Figure 15 Describe the whole process of inputting data into cross-modal contrast learning: Data input: Input medical record text x and time-series data z.

[0126] Medical term recognition: Identify medical terms through the first model.

[0127] Entity-level mask generation: Perform an overall masking operation on the set of medical terms to obtain masked medical terms.

[0128] Time-series feature extraction: Input the electronic medical record into the second model for time-series feature extraction to obtain the feature vectors of at least one examination index.

[0129] Hierarchical time-series attention: For each of the said examination indexes, perform hierarchical attention processing according to the feature vectors of the examination indexes.

[0130] Cross-modal contrast learning: Analyze the status development of the target object associated with the electronic medical record based on the masked medical terms, the first-layer attention weights, and the second-layer attention weights to obtain an analysis result.

[0131] Exemplarily, in combination with Figure 16 Describe the whole medical entity perception masking mechanism: Medical text input: Represent the original medical text input, for example: "The patient's glycated hemoglobin level is elevated, and blood sugar control is poor." Medical term recognition module: Identify compound medical terms and label medical terms.

[0132] Among them, identify compound medical terms: Identify compound medical terms from the input text (such as "glycated hemoglobin").

[0133] Label medical terms: Label the identified medical terms for subsequent processing. For example, label glycated hemoglobin as a compound medical term; Masking strategy generation module: Perform an overall masking operation on the identified medical terms to avoid term splitting. The overall masked medical terms represent performing an overall masking operation on the identified medical terms (such as replacing with [MASK]). Avoiding term splitting means ensuring that medical terms are not split and masked, and preserving semantic integrity.

[0134] Pre-trained language model: Learn context information to predict the masked medical terms. Learning context prediction mask means the model learns to predict the masked medical terms through context information.

[0135] Output prediction results: Verify whether the prediction results are consistent with the original medical terms. Verify the prediction results: Verify whether the prediction results of the model are accurate. Compare with the original terms: Compare the prediction results with the original medical terms to evaluate the performance of the model.

[0136] Verification and comparison: Finally, verify whether the prediction results are consistent with the original medical terms to complete the whole process.

[0137] A dynamic adaptive semantic analysis method for electronic medical records provided by the present application. Considering that currently, a pre-trained model is usually used to perform random masking and other processing on electronic medical records, resulting in the splitting of compound medical terms in the electronic medical records and the inability to obtain the dynamic association between inspection indicators and diagnosis conclusions, which leads to the problem of not being able to provide reliable data support for subsequent applications. When performing the masking operation on medical terms, the present application will identify the compound medical terms in the electronic medical records to avoid splitting the compound medical terms, so that the overall masking operation can be performed on the compound medical terms, retaining the integrity of the medical terms and improving the accuracy of processing the electronic medical records; in addition, the present application captures the change trend of the inspection indicators and the association relationship between the change trend and the diagnosis conclusion by calculating the first-layer attention weight and the second-layer attention weight of each inspection indicator, and obtains the time-series relationship of the medical data, providing more intuitive and accurate data support for analyzing the state development of the target object associated with the electronic medical record subsequently.

[0138] The above description is only the preferred embodiments of the present application and the explanation of the applied technical principles. Those skilled in the art should understand that the scope of the invention involved in the present application is not limited to the technical solutions formed by the specific combination of the above technical features, and should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the inventive concept. For example, the technical solutions formed by mutually replacing the above features with the (but not limited to) technical features with similar functions disclosed in the present application.

Claims

1. A dynamic adaptive semantic analysis method for electronic medical records, characterized in that The method includes: Performing medical term recognition on the electronic medical record through a first model to obtain a medical term set of the electronic medical record, where the medical term set includes compound medical terms, and the compound medical terms include at least two independent medical terms; Performing an overall masking operation on the medical term set to obtain masked medical terms; Inputting the electronic medical record into a second model for temporal feature extraction to obtain feature vectors of at least one examination index, where the feature vectors include the values and / or diagnostic conclusions of the examination index at different times; For each of the examination indexes, performing hierarchical attention processing according to the feature vectors of the examination index to obtain a first-layer attention weight and a second-layer attention weight of the examination index, where the first-layer attention weight is used to characterize the change trend of the value corresponding to the examination index, and the second-layer attention weight is used to characterize the correlation between the change trend and the diagnostic conclusion corresponding to the examination index; Analyzing the status development of the target object associated with the electronic medical record based on the masked medical terms, the first-layer attention weight, and the second-layer attention weight to obtain an analysis result, where the analysis result is used to characterize the status development of the target object.

2. The method according to claim 1, wherein The step of, for each of the examination indexes, performing hierarchical attention processing according to the feature vectors of the examination index to obtain a first-layer attention weight and a second-layer attention weight of the examination index includes: Determining the first-layer attention weight according to the feature vector of the examination index and the first-layer attention parameter; Determining the second-layer attention weight according to the product of the first-layer attention weight and the feature vector of the examination index and the second-layer attention parameter.

3. The method according to claim 1, wherein The step of analyzing the status development of the target object associated with the electronic medical record based on the masked medical terms, the first-layer attention weight, and the second-layer attention weight to obtain an analysis result, where the analysis result is used to characterize the status development of the target object specifically includes: Analyzing the status development of the target object associated with the electronic medical record based on the masked medical terms, the first-layer attention weight, the second-layer attention weight, and the text feature encoding vector to obtain the analysis result.

4. The method according to claim 3, wherein The method further includes: Inputting the electronic medical record into a third model for text feature encoding to obtain the text feature encoding vector.

5. The method according to claim 4, wherein The method further includes the steps of training the second model and the third model: Obtaining a plurality of first text features, a plurality of second text features, and a plurality of numerical features as positive samples, where the first text features are the features in the electronic medical record representing the diagnostic conclusions of the examination indexes, and the second text features are the text features in the electronic medical record other than the first text features; Inputting the plurality of first text features and the plurality of numerical features into a first initial model for model training to obtain the second model; Inputting the plurality of second text features into a second initial model for model training to obtain the third model.

6. The method according to claim 5, wherein The method further includes: Obtain verification text features and verification numerical features, where the verification text features are text features in the electronic medical record other than the diagnosis conclusion representing inspection indicators; Input the verification numerical features into the second model to obtain a verification numerical vector; Input the verification text features into the third model to obtain a verification text vector; Calculate the cosine similarity loss between the verification numerical vector and the verification text vector; Determine the semantic alignment degree between the verification numerical vector and the verification text vector according to the cosine similarity loss; If the semantic alignment degree does not meet the first target requirement, adjust the model parameters of the first initial model and the second initial model for retraining until the semantic alignment degree meets the first target requirement; If it meets, output the second model and the third model.

7. The method according to claim 1, wherein The method further includes the step of training the first model: Obtain a plurality of independent medical terms, a plurality of compound medical terms, and the context information of each of the compound medical terms from a medical knowledge base; Annotate each of the compound medical terms according to the context information of each of the compound medical terms to obtain intermediate medical terms, and use the set including each of the intermediate medical terms and each of the independent medical terms as an initial sample set; Perform word embedding vector conversion processing on each of the initial sample sets to obtain word embedding vectors; Input each of the word embedding vectors into a third initial model for model training to obtain the first model.

8. The method according to claim 7, wherein The method further includes: Obtain a verification medical term sample; Input the verification medical term sample into the first model to obtain a verification compound medical term; Input the verification compound medical term into a domain discriminator to obtain a first probability value, and determine whether the first probability value meets the second target requirement; If the first probability value does not meet the second target requirement, adjust the training parameters of the third initial model for retraining until the first probability value meets the second target requirement; If the first probability value meets the second target requirement, output the first model.

9. The method according to claim 8, characterized in that The method further includes: Calculate the first adversarial training loss value of the first probability value, and determine whether the first adversarial training loss value meets the third target requirement; If the first adversarial training loss value does not meet the third target requirement, adjust the training parameters of the third initial model for retraining until the first adversarial training loss value meets the third target requirement; If the first adversarial training loss value meets the third target requirement, output the first model.

10. The method according to claim 6, characterized in that, The method further includes: Input the verification text features and verification numerical features into the domain discriminator respectively to obtain a second probability value and a third probability value, and determine whether the second probability value and the third probability value meet the fourth target requirement; If the second probability value and the third probability value do not meet the fourth target requirement, adjust the model parameters of the first initial model and the second initial model for retraining until the second probability value and the third probability value meet the fourth target requirement; If the second probability value and the third probability value meet the fourth target requirement, output the second model and the third model.

11. The method according to claim 10, characterized in that, The method further includes: Calculating a second adversarial training loss value and a third adversarial training loss value of the second probability value and the third probability value, and determining whether the second adversarial training loss value and the third adversarial training loss value meet a fifth target requirement; If the second adversarial training loss value and the third adversarial training loss value do not meet the fifth target requirement, adjusting the model parameters of the first initial model and the second initial model and retraining until the second adversarial training loss value and the third adversarial training loss value meet the fifth target requirement; If the second adversarial training loss value and the third adversarial training loss value meet the fifth target requirement, output the second model and the third model.

Images