Dynamic Adaptive Semantic Analysis Method for Electronic Medical Records

Abstract

The present application discloses a dynamic adaptive semantic analysis method for electronic medical records. The method identifies compound medical terms in the electronic medical record when performing masking operations on medical terms to avoid splitting the compound medical terms. In this way, the compound medical terms can be masked as a whole, thereby retaining the integrity of the medical terms and improving the accuracy of electronic medical record processing. In addition, the present application captures the changing trend of the examination indicators and the correlation between the changing trend and the diagnostic conclusion by calculating the first-level attention weight and the second-level attention weight of each examination indicator, and obtains the temporal relationship of the medical data to provide more intuitive and accurate data support for the subsequent analysis of the status development of the target object associated with the electronic medical record.

Description

Technical Field

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

[0002] With the widespread use of electronic medical records, the scale and complexity of medical data continue to increase, so medical data needs to be processed to structure electronic medical records, provide clinical decision support, and conduct cross-institutional medical data analysis based on the processed results.

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

[0004] The above processing method not only causes the medical terms in the electronic medical record 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 desired to provide a dynamic adaptive semantic analysis method for electronic medical records that 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, the method comprising:

[0007] Performing compound medical term recognition on the electronic medical record using the first model to obtain compound medical terms of the electronic medical record, wherein the compound medical terms include at least two independent medical terms;

[0008] performing a masking operation on the compound medical term as a whole to obtain a masked medical term;

[0009] Inputting the electronic medical record into a second model to perform time series feature extraction to obtain a feature vector of at least one examination indicator, wherein the feature vector includes the test value and / or diagnosis conclusion of the examination indicator at different times;

[0010] For each of the inspection indicators, hierarchical attention processing is performed according to the characteristic vector of the inspection indicator to obtain a first-layer attention weight and a second-layer attention weight of the inspection indicator, where the first-layer attention weight is used to characterize a change trend of the test value corresponding to the inspection indicator, and the second-layer attention weight is used to characterize a correlation between the change trend and the diagnostic conclusion corresponding to the inspection indicator;

[0011] Based on the masked medical term, the first attention weight and the second attention weight, a status development analysis is performed on the target object associated with the electronic medical record to obtain an analysis result, which is used to characterize the status development of the target object.

[0012] In an optional embodiment, for each inspection indicator, performing hierarchical attention processing according to the characteristic vector of the inspection indicator to obtain the first-layer attention weight and the second-layer attention weight of the inspection indicator includes:

[0013] Determining the first-layer attention weight according to the characteristic vector of the inspection indicator and the first-layer attention parameter;

[0014] The second-layer attention weight is determined according to the product of the first-layer attention weight and the characteristic vector of the inspection indicator and the second-layer attention parameter.

[0015] In an optional embodiment, the state development analysis of the target object associated with the electronic medical record is performed based on the masked medical term, the first attention weight, and the second attention weight to obtain an analysis result, which is used to characterize the state development of the target object, specifically including:

[0016] Based on the masked medical term, the first attention weight, the second attention weight and the text feature encoding vector, a state development analysis is performed on the target object associated with the electronic medical record to obtain the analysis result.

[0017] In an optional embodiment, the method further includes:

[0018] Inputting the electronic medical record into a third model to extract text features in the electronic medical record;

[0019] Feature encoding is performed on the text features in the electronic medical record to obtain the text feature encoding vector.

[0020] In an optional embodiment, the method further comprises the steps of training the second model and the third model:

[0021] Acquire multiple first text features, multiple second text features, and multiple numerical features as positive samples, wherein the first text features are features representing the diagnosis conclusions of examination indicators in the electronic medical record, and the second text features are text features in the electronic medical record other than the first text features;

[0022] 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;

[0023] The plurality of second text features are input into a second initial model for model training to obtain the third model.

[0024] In an optional embodiment, the method further includes:

[0025] Acquire verification text and verification numerical features, wherein the verification text is a text feature in the electronic medical record other than the diagnostic conclusion representing the examination indicators;

[0026] Inputting the verification numerical feature into the second model to obtain a verification numerical vector;

[0027] Inputting the verification text feature into the third model to obtain a verification text vector;

[0028] Calculating the cosine similarity loss between the verification numerical vector and the verification text vector;

[0029] Determining a degree of semantic alignment between the verification numerical vector and the verification text vector based on the cosine similarity loss;

[0030] If the semantic alignment degree does not meet the first target requirement, adjusting the model parameters of the second model and the third model for retraining until the semantic alignment degree meets the first target requirement;

[0031] If so, the second model and the third model are output.

[0032] In an optional embodiment, the method further comprises the step of training the first model:

[0033] Acquire multiple independent medical terms and context information of each independent medical term from a medical knowledge base;

[0034] Annotating each of the independent medical terms based on context information of each of the independent medical terms to obtain an initial sample set;

[0035] Converting the independent medical terms in each of the initial sample sets into word embedding vectors;

[0036] Each of the word embedding vectors is input into a third initial model for model training to obtain the first model.

[0037] In an optional embodiment, the method further includes:

[0038] Obtain samples of validated medical terminology;

[0039] Inputting the verification medical term sample into the first model to obtain a verification compound medical term;

[0040] Inputting the verification compound medical term into a domain discriminator to obtain a first probability value, and determining whether the first probability value meets a second target requirement;

[0041] If the first probability value does not meet the second target requirement, adjusting the training parameters of the first initial model and retraining until the first probability value meets the second target requirement;

[0042] If the first probability value meets the second target requirement, the first model is output.

[0043] In an optional embodiment, the method further includes:

[0044] Calculating a first adversarial training loss value for the first probability value, and determining whether the first adversarial training loss value meets a third objective requirement;

[0045] If the first adversarial training loss value does not meet the third target requirement, adjusting the training parameters of the third initial model for retraining until the first adversarial training loss value meets the third target requirement;

[0046] If the first adversarial training loss value meets the third target requirement, the first model is output.

[0047] In an optional embodiment, the method further includes:

[0048] 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;

[0049] 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 and retraining until the second probability value and the third probability value meet the fourth target requirement;

[0050] If the second probability value and the third probability value meet the fourth target requirement, the second model and the third model are output.

[0051] In an optional embodiment, the method further includes:

[0052] Calculating a second adversarial training loss value and a third adversarial training loss value for 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 objective requirement;

[0053] 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;

[0054] If the second adversarial training loss value and the third adversarial training loss value meet the fifth objective requirement, the second model and the third model are output.

[0055] The present application provides a dynamic adaptive semantic analysis method for electronic medical records. Considering that pre-trained models are usually used to perform random masking and other processing on electronic medical records, the compound medical terms in the electronic medical records are split, and the dynamic association between the examination indicators and the diagnostic conclusions cannot be obtained, resulting in the problem of being unable to provide reliable data support for subsequent applications. The present application will identify the compound medical terms in the electronic medical record when performing masking operations on medical terms to avoid splitting the compound medical terms. In this way, the compound medical terms can be masked as a whole, the integrity of the medical terms is retained, and the accuracy of electronic medical record processing is improved. In addition, the present application captures the changing trend of the examination indicators and the correlation between the changing trend and the diagnostic conclusion by calculating the first-layer attention weight and the second-layer attention weight of each examination indicator, and obtains the temporal relationship of the medical data to provide more intuitive and accurate data support for the subsequent analysis of the status development of the target object associated with the electronic medical record. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Other features, objects and advantages of the present application will become more apparent upon reading the detailed description of non-limiting embodiments made with reference to the following drawings:

[0057] Figure 1 This is an application scenario diagram of a dynamic adaptive semantic analysis method for electronic medical records provided in this application;

[0058] Figure 2 A flowchart of a dynamic adaptive semantic analysis method for electronic medical records provided in this application;

[0059] Figure 3 A flowchart of the steps of a dynamic adaptive semantic analysis method for electronic medical records provided in this application;

[0060] Figure 4 A flowchart of the steps of another method for dynamic adaptive semantic analysis of electronic medical records provided by this application;

[0061] Figure 5 A flowchart of the steps of another dynamic adaptive semantic analysis method for electronic medical records provided in this application;

[0062] Figure 6 A flowchart of the steps of a dynamic adaptive semantic analysis method for electronic medical records provided in this application;

[0063] Figure 7 A flowchart of the steps of another method for dynamic adaptive semantic analysis of electronic medical records provided by this application;

[0064] Figure 8 A flowchart of the steps of another dynamic adaptive semantic analysis method for electronic medical records provided in this application;

[0065] Figure 9 A flowchart of the steps of another method for dynamic adaptive semantic analysis of electronic medical records provided by this application;

[0066] Figure 10 A flowchart of the steps of another dynamic adaptive semantic analysis method for electronic medical records provided in this application;

[0067] Figure 11 A flowchart of the steps of another dynamic adaptive semantic analysis method for electronic medical records provided in this application;

[0068] Figure 12 A flowchart of the steps of another method for dynamic adaptive semantic analysis of electronic medical records provided by this application;

[0069] Figure 13 A flowchart of the steps of another dynamic adaptive semantic analysis method for electronic medical records provided in this application;

[0070] Figure 14 A flowchart of the steps of another dynamic adaptive semantic analysis method for electronic medical records provided in this application;

[0071] Figure 15 A flowchart of the steps of another dynamic adaptive semantic analysis method for electronic medical records provided in this application;

[0072] Figure 16 This is a flowchart of the steps of another dynamic adaptive semantic analysis method for electronic medical records provided in this application. DETAILED DESCRIPTION

[0073] The present application will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the relevant invention and are not intended to limit the invention. It should also be noted that, for ease of description, only portions relevant to the invention are shown in the accompanying drawings.

[0074] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0075] Please refer to Figure 1 , Figure 1 This application scenario diagram illustrates a dynamic adaptive semantic analysis method for electronic medical records provided herein. The application scenario diagram includes a terminal device 100 and a service device 200. A user uploads patient A's electronic medical record to service device 200 via terminal device 100. A first model configured on service device 200 identifies and masks the complex medical terms in patient A's electronic medical record. A second model then extracts temporal features from the electronic medical record to obtain feature vectors for multiple examination indicators. Finally, layered attention processing is performed on the feature vectors for the multiple examination indicators to obtain first-layer and second-layer attention weights for each examination indicator. The masked medical terms, first-layer and second-layer attention weights processed in this manner are then used to analyze the progression of patient A's condition, thereby obtaining the progression of patient A's condition. The terminal device 100 may be, for example, a smartphone, tablet, laptop, desktop computer, or smartwatch. The service device 200 may be, for example, a single server or a server cluster.

[0076] The following describes the detection method provided by this application.

[0077] like Figure 2 As shown, Figure 2 This application provides a flowchart of a method for dynamic adaptive semantic analysis of electronic medical records, which includes:

[0078] Step S20: performing medical term recognition on the electronic medical record using the first model to obtain a medical term set of the electronic medical record, where the medical term set includes a compound medical term, and the compound medical term includes at least two independent medical terms;

[0079] The first model is a neural network model installed on the service device for medical terminology recognition. This neural network model can be trained using a medical knowledge base. Exemplarily, the first model is a network entity recognition (NER) model. This model can not only recognize independent medical terms such as white blood cells, hypertension, heart rate, blood pressure, and hemoglobin, but also compound medical terms. Compound medical terms are medical terms composed of at least two independent medical terms. They replace lengthy descriptive phrases, achieving a balance between simplicity and professionalism. This not only reflects the multi-source nature of medical terminology, but also facilitates precise expression in interdisciplinary communication. Examples include glycated hemoglobin, infectious pneumonia, and white blood cell count.

[0080] The electronic medical record may be the electronic medical record of a single patient uploaded by the terminal device, or may be the electronic medical record of multiple patients uploaded by the terminal device, and this application does not limit this. When the terminal device uploads the electronic medical records of multiple patients, each electronic medical record may include a unique patient identifier so that the electronic medical records of different patients can be processed separately to obtain medical analysis data specific to each patient.

[0081] This application uses the first model to identify medical terms in electronic medical records. It can not only identify independent medical terms, but also accurately identify compound medical terms in electronic medical records to avoid splitting compound medical terms and causing inaccurate medical analysis data to be provided.

[0082] Step S30, performing a masking operation on the medical term set as a whole to obtain masked medical terms;

[0083] Masking refers to replacing an entire medical term with a mask. Since the first model can recognize compound medical terms, masking can be performed on compound medical terms as a whole, ensuring semantic integrity and resolving the issue of prior art in medical term recognition that can damage semantic integrity.

[0084] For example, the content of the input electronic medical record includes the following parts:

[0085] A 56-year-old male patient presented with chest pain for 3 hours. He had a 10-year history of hypertension and irregular medication use. Three hours prior, the patient experienced chest pain without apparent cause. The pain was located in the precordial area, felt compressive, and accompanied by profuse sweating, and persisted unrelieved. He had a history of hypertension, with a peak blood pressure of 180 / 110 mmHg, and irregular use of antihypertensive medication. He denied any history of diabetes or coronary artery disease. He also denied any family history of cardiovascular disease. Physical examination revealed a temperature of 36.8°C, a pulse of 100 beats / minute, respiratory rate of 20 breaths / minute, and a blood pressure of 160 / 100 mmHg. Auscultation of the heart and lungs revealed no abnormalities. An electrocardiogram revealed sinus rhythm with ST-segment elevation of 0.2-0.3 mV in leads V1-V6. A preliminary diagnosis of acute myocardial infarction was made. He was immediately treated with 300 mg of chewed aspirin, 300 mg of clopidogrel orally, and 4000 IU of low-molecular-weight heparin subcutaneously. The patient was admitted to the CCU, where further coronary angiography revealed complete occlusion of the proximal left anterior descending artery. PCI was immediately performed, successfully opening the artery. Postoperatively, antiplatelet, lipid-lowering, and vasodilation therapy were administered. The patient's condition gradually stabilized, with relief of chest pain and stable vital signs.

[0086] Medical terminology recognition is performed using the first model, and the medical term set obtained is:

[0087] Chest pain, hypertension, coronary heart disease, acute myocardial infarction;

[0088] Performing a global masking operation on the medical term collection yields the following output:

[0089] A 56-year-old male patient presented with complaints of [mask] for 3 hours. He had a 10-year history of [mask] and did not take medication regularly. Three hours prior, he developed [mask] without apparent cause. The pain was located in the precordial area, was compressive, and accompanied by profuse sweating, and persisted without relief. He had a history of [mask], with a maximum blood pressure of 180 / 110 mmHg, and did not take antihypertensive medication regularly. He denied diabetes or a history of [mask]. He also denied a family history of cardiovascular disease. Physical examination: temperature 36.8°C, pulse 100 beats / min, respirations 20 breaths / min, blood pressure 160 / 100 mmHg. Auscultation of the heart and lungs revealed no abnormalities. Electrocardiogram: sinus rhythm, ST-segment elevation of 0.2-0.3 mV in leads V1-V6. Preliminary diagnosis: [mask]. The patient was immediately given 300 mg of chewable aspirin, 300 mg of clopidogrel orally, and 4000 IU of low-molecular-weight heparin subcutaneously. The patient was admitted to the CCU, where further coronary angiography revealed complete occlusion of the proximal left anterior descending artery. PCI was immediately performed, successfully recanalizing the vessel. Postoperatively, antiplatelet, lipid-lowering, and vasodilator therapy were administered. The patient's condition gradually stabilized, [MASK] resolved, and his vital signs remained stable.

[0090] Step S40: Input the electronic medical record into the second model to extract time series features, and obtain a feature vector of at least one examination indicator, where the feature vector includes the value of the examination indicator at different times and / or the diagnosis conclusion;

[0091] The second model is a neural network model installed on the server for extracting time series features. The neural network model can be trained based on a recurrent neural network (LSTM or GRU).

[0092] Time series features refer to characteristics of an electronic medical record that change over time. These can include the values of examination indicators at different times or the diagnostic conclusions of examination indicators at different times. For example, time series features include a white blood cell count of 12.5×10^9 / L, a blood pressure of 160 / 100 mmHg, and an ST segment elevation of 0.2-0.3 mV in leads V1-V6.

[0093] Extracting time series features from time series features refers to vectorizing the values of the examination indicators and / or the diagnosis conclusions. Since an electronic medical record may contain multiple examination indicators, the second model can extract the feature vector of at least one examination indicator.

[0094] In addition, the present application may be to extract the time series features of the inspection indicators at different times, and then sort the feature vectors of the same inspection indicator in time sequence such as week, month, and year to obtain a set of feature vectors corresponding to each inspection indicator.

[0095] Exemplarily, the electronic medical record input into the second model includes the following parts:

[0096] 2024-02-15 09:00: White blood cell count 12.5X10^9 / L, diagnosed as infectious pneumonia.

[0097] 2024-02-15 10:00: Heart rate 100 beats / min, blood pressure 160 / 100 mmHg.

[0098] 2024-02-15 11:00: ECG shows: Neural rhythm, ST segment elevation of 0.2-0.3mV in leads V1-V6.

[0099] 2024-02-18 09:00: White blood cell count 15.0X10^9 / L, diagnosed as worsening infection.

[0100] 2024-02-18 10:00: Heart rate 95 beats / min, blood pressure 150 / 95 mmHg.

[0101] 2024-02-18 11:00: ECG showed: sinus rhythm, ST segment elevation of 0.1-0.2 mV in leads V1-V6.

[0102] The output obtained is:

[0103] h_1 = [12.5, NaN, NaN, NaN, NaN]

[0104] h_2 = [NaN, 100.0, 160, 100, NaN]

[0105] h_3 = [NaN, NaN, NaN, NaN, 0.3]

[0106] h_4 = [15.0, NaN, NaN, NaN, NaN]

[0107] h_5 = [NaN, 95.0, 150, 95, NaN]

[0108] h_6 = [NaN, NaN, NaN, NaN, 0.2]

[0109] Sorting the time series feature vectors to obtain:

[0110] h_11 = [12.5, infectious pneumonia, 15, worsening infection]

[0111] h_21 = [ 100.0, 95]

[0112] h_31 = [160 / 100, 150 / 95]

[0113] h_41= [Black rhythm, sinus rhythm]

[0114] h_51 = [0.2-0.3, 0.1-0.2]

[0115] This application can capture the dynamic changes of each time series feature in the time series by extracting the time series features in the electronic medical records, and provide data support for subsequent hierarchical attention processing.

[0116] Step S50: For each inspection indicator, hierarchical attention processing is performed based on the characteristic vector of the inspection indicator to obtain the first-layer attention weight and the second-layer attention weight of the inspection indicator. The first-layer attention weight is used to represent the change trend of the test value corresponding to the inspection indicator, and the second-layer attention weight is used to represent the correlation between the change trend and the diagnostic conclusion corresponding to the inspection indicator.

[0117] Among them, after extracting the time series features of the electronic medical record according to the second model, a feature vector of at least one examination indicator is obtained. Since the time series feature extraction only obtains a single feature vector, it cannot reflect the deeper correlation and logical relationship between the examination indicators. Therefore, this application will perform layered attention processing on each extracted examination indicator to capture the changing trend of the examination indicator over time, as well as the correlation between the changing trend of the examination indicator and the corresponding diagnostic conclusion of the examination indicator, thereby providing users with more intuitive and easier to analyze medical data.

[0118] Hierarchical attention processing refers to calculating the first-layer attention weight and second-layer attention weight for the characteristic vector of each inspection indicator, so as to represent the changing trend of the numerical value corresponding to the inspection indicator through the first-layer attention weight, and the correlation between the changing trend and the diagnostic conclusion corresponding to the inspection indicator through the second-layer attention weight.

[0119] For example, the present application may extract time series features in the following manner:

[0120]

[0121] z is the time series data; h is the time series feature vector; θ RNN are model parameters.

[0122] For example, the input time series feature vector is:

[0123] h_11 = [12.5, infectious pneumonia, 15, worsening infection]

[0124] h_21 = [ 100.0, 95]

[0125] h_31 = [160 / 100, 150 / 95]

[0126] h_41= [Black rhythm, sinus rhythm]

[0127] h_51 = [0.2-0.3, 0.1-0.2]

[0128] The first layer attention weight a1 of the output is:

[0129] a1_1 = 0.8

[0130] a1_2 = 0.7

[0131] a1_3 = 0.6

[0132] a1_4 = 0.9

[0133] a1_5 = 0.85

[0134] The output second layer attention weight a2 is:

[0135] a2_1 = 0.75

[0136] a2_2 = 0.65

[0137] a2_3 = 0.55

[0138] a2_4 = 0.85

[0139] a2_5 = 0.8

[0140] For example, Figure 3 As shown, Figure 3 Schematic diagram of the layered temporal attention architecture provided for this application:

[0141] Input: Medical test indicators and diagnosis conclusions: Represents input data, including the patient's test indicators (such as white blood cell count, blood sugar level, etc.) and diagnosis conclusions (such as infection, diabetes, etc.).

[0142] Hierarchical Temporal Attention Architecture: The core architecture for processing input data and extracting temporal correlations.

[0143] Time series feature extraction layer: extracts time series features from the input data and represents the inspection indicators and diagnostic conclusions at each time point as vectors.

[0144] The first-layer attention mechanism captures the temporal correlation between test indicators, such as the changing trend of white blood cell count.

[0145] The second-level attention mechanism captures the dynamic relationship between test indicators and diagnostic conclusions, such as the logical relationship between changes in white blood cell count and the progression of infection.

[0146] Output layer: Time series association results: Output the results of time series association to provide support for clinical decision-making.

[0147] Clinical decision support system: Eventually, the time series correlation results will be applied to the clinical decision support system to help doctors make more accurate diagnoses.

[0148] Step S60 , analyzing the status development of the target object associated with the electronic medical record based on the masked medical terminology, the first-layer attention weight, and the second-layer attention weight to obtain an analysis result, which is used to characterize the status development of the target object.

[0149] 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 perform a status development analysis on the target object associated with the electronic medical record based on these data to obtain the analysis results.

[0150] For example, the user is determined to have chest pain symptoms based on the masked medical terminology and is initially diagnosed with acute myocardial infarction. The user determines based on a1_1 that the target subject's white blood cell count is increasing over time, and determines based on a2_1= 0.75 that pneumonia is worsening.

[0151] In an optional embodiment, if Figure 4 As shown, Figure 4 The steps for obtaining the first-layer attention weight and the second-layer attention weight provided in an exemplary embodiment of the present application are as follows:

[0152] Step S301, determining the first-layer attention weight according to the characteristic vector of the inspection index and the first-layer attention parameter;

[0153] Among them, this application can obtain the first-level attention weight through the following method:

[0154]

[0155] Among them, h is the time series feature vector; a1 is the first layer attention weight, θ att1 is the first layer attention parameter.

[0156] Step S302: Determine the second-layer attention weight according to the product of the first-layer attention weight and the characteristic vector of the inspection index and the second-layer attention parameter.

[0157] Among them, this application can obtain the second-level attention weight through the following method:

[0158]

[0159] Among them, h is the time series feature vector; a1 is the first layer attention weight, θ att2 is the first layer attention parameter; a2 is the first layer attention weight, Represents element-wise multiplication.

[0160] The first layer of attention weights captures temporal relationships between examination indicators, such as the changing trend of white blood cell counts. The second layer of attention weights further captures the correlation between examination indicators and diagnostic conclusions, such as the logical relationship between white blood cell count changes and the progression of pneumonia infection. Layered attention processing can more accurately understand the temporal relationships in medical data, providing more accurate and intuitive data support for clinical decision-making.

[0161] In another optional embodiment, the present application may also obtain analysis results according to the following method:

[0162] Based on masked medical terms, first attention weight, second attention weight and text feature encoding vector, the state development of the target object associated with the electronic medical record is analyzed to obtain the analysis results.

[0163] Among them, since the above-mentioned status development analysis of the target object only performed feature encoding (that is, time series feature extraction) on the majority of the examination indicators and some text features (diagnostic conclusions), and did not perform feature encoding on other text features in the electronic medical record, some text features and numerical features are not in the same semantic space, and cannot provide a data basis for cross-modal comparative learning. Therefore, this application also uses text feature vectors to perform status development analysis on the target object associated with the electronic medical record to obtain more complete and accurate analysis results.

[0164] It should be noted here that the steps for encoding text features in electronic medical records to obtain text feature encoding vectors include the following:

[0165] The electronic medical record is input into the third model for text feature encoding to obtain a text feature encoding vector.

[0166] The third model is a neural network model installed on the server for text feature encoding. The neural network model can be a BERT model or a multi-layer perceptron (MLP).

[0167] This application can be used to encode text features in the following ways:

[0168]

[0169] Among them, f t is the text feature vector, are model parameters.

[0170] like Figure 5 As shown, in an optional embodiment, the present application further provides a step of training the second model and the third model:

[0171] Step S401: Acquire multiple first text features, multiple second text features, and multiple numerical features as positive samples, where the first text features are features in the electronic medical record that represent the diagnosis conclusion of the examination indicators, and the second text features are text features in the electronic medical record other than the first text features.

[0172] The first text feature may be, for example, infectious pneumonia, myocardial infarction, myocarditis, hypertension, etc. The second text feature may be, for example, white blood cell count, heart rate, electrocardiogram, blood pressure, ST segment elevation in leads V1-V6, etc.

[0173] The first text feature, the second text feature, and the numerical feature may be obtained from historical electronic medical records stored in a memory, or may be obtained from different hospital systems by accessing different hospital systems, and this application is not limited thereto. Using data obtained from different systems as samples for model training enables the trained model to maintain good performance when faced with new data from different institutions or fields.

[0174] Step S402: inputting the plurality of first text features and the plurality of numerical features into the first initial model for model training to obtain a second model;

[0175] The first initial model is, for example, a BERT model.

[0176] Step S403: Input the plurality of second text features into the second initial model for model training to obtain a third model.

[0177] The second initial model is, for example, a multi-layer perceptron.

[0178] It should be noted here that model training is a conventional technical means and will not be elaborated here.

[0179] like Figure 6 As shown, in an optional embodiment, the present application further provides a step of training the second model and the third model:

[0180] Step S501: obtaining verification text features and verification numerical features, where the verification text features are text features in the electronic medical record other than the diagnostic conclusions representing the examination indicators;

[0181] During the model training described above, the text features obtained from the electronic medical records, excluding the diagnostic conclusions representing the examination indicators, can be divided into a training set and a validation set in proportion, so as to train the model through the above steps and validate the model through the following steps. Similarly, the numerical features obtained during model training can also be divided into a training set and a validation set in proportion.

[0182] Step S502, inputting the verification numerical feature into the second model to obtain a verification numerical vector;

[0183] Step S503: input the verification text features into the third model to obtain a verification text vector;

[0184] Step S504, calculating the cosine similarity loss between the verification numerical vector and the verification text vector;

[0185] Among them, the present application can calculate the cosine similarity loss by the following method:

[0186] First, calculate the cosine similarity between the validation numerical vector and the validation text vector:

[0187]

[0188] Among them, ft is the text feature vector; fv is the numerical feature vector.

[0189] The fv numerical eigenvector can be obtained by the following formula:

[0190]

[0191] Where fv is a numerical feature vector, MLP represents a multi-layer perceptron, are the MLP parameters.

[0192] Next, calculate the cosine similarity loss between the two:

[0193]

[0194] Among them, Lcos is the cosine similarity loss.

[0195] Step S505: determining the degree of semantic alignment between the verification numerical vector and the verification text vector based on the cosine similarity loss; if the degree of semantic alignment does not meet the first target requirement, executing step S506; if the degree of semantic alignment meets the first target requirement, executing step S507;

[0196] Step S506: Adjust the model parameters of the first initial model and the second initial model and retrain them until the semantic alignment degree meets the first target requirement;

[0197] Step S507: output the second model and the third model.

[0198] Among them, the first target requirement is used to limit the value of the cosine similarity loss. Then, this application can be to determine that the semantic alignment between the verification numerical vector and the verification text vector is low when the calculated cosine similarity loss is greater than a preset threshold, and to determine that the semantic alignment 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 and can be set according to the specific application environment. For example, the first target requirement is that the cosine similarity loss is equal to 0.05.

[0199] This application continuously adjusts the model parameters of the first and second initial models to minimize the cosine similarity loss and align the feature vectors obtained by the first and second initial models, making the text features and numerical features closer in the semantic space. This optimization mechanism can improve the accuracy of multimodal data fusion and help the trained second and third models better understand the semantic relationships between data of different modalities.

[0200] For example, verify the text feature vector ft:

[0201] ft = [0.5, 0.6, 0.7, 0.8, 0.9]

[0202] Verify the numerical feature vector fv:

[0203] fv = [0.4, 0.5, 0.6, 0.7, 0.8]

[0204] Cosine similarity Sim(ft, fv) = 0.95

[0205] Cosine similarity loss Lcos = 0.05

[0206] The calculated cosine similarity loss is equal to 0.05, indicating that the degree of semantic alignment has met the first objective requirement and the second and third models can be output.

[0207] like Figure 7 As shown in the figure, it is a schematic diagram of the cross-modal contrastive learning module:

[0208] Input: Medical text and numerical data: Represents multimodal input data, including medical text descriptions (such as medical records) and numerical data (such as test indicators, image features, etc.).

[0209] Text Tower: Text Encoder: Encodes medical text data into text feature vectors, typically using a pre-trained language model such as BERT.

[0210] Numeric Tower: Numeric Encoder: Encodes numerical data into numerical feature vectors, which can use neural networks or other encoding methods.

[0211] Cross-modal alignment module: maps textual feature vectors and numerical feature vectors into the same semantic space for contrastive learning.

[0212] Loss function: Cosine similarity loss, calculates the cosine similarity between the text feature vector and the numerical feature vector, and optimizes the alignment effect through the loss function.

[0213] Output: Optimized cross-modal representation: Output optimized cross-modal representation for subsequent multimodal data analysis.

[0214] Multimodal Data Analysis Systems: Apply the optimized cross-modal representations to multimodal data analysis systems, such as clinical decision support or disease prediction.

[0215] In another optional embodiment, Figure 8 As shown, the present application also provides a step of training the second model and the third model:

[0216] Step S601: Input the verification text feature and the verification numerical feature 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, execute step S602; if the second probability value and the third probability value meet the fourth target requirement, execute step S603;

[0217] 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;

[0218] Step S603: output the second model and the third model.

[0219] During the model training described above, the text features obtained from the electronic medical records, excluding the diagnostic conclusions representing the examination indicators, can be divided into a training set and a validation set in proportion, so as to train the model through the above steps and validate the model through the following steps. Similarly, the numerical features obtained during model training can also be divided into a training set and a validation set in proportion.

[0220] The domain discriminator is, for example:

[0221]

[0222] in, θ is the output of the domain discriminator, which indicates the probability that the input data belongs to a specific domain (or organization) under the given feature representation f. D are the discriminator parameters.

[0223] The fourth target requirement is used to limit the range of the second and third probability values. The fourth target requirement can be different or the same for the second and third probability values. This application does not limit this and can be set according to the specific application scenario. For example, the fourth target requirement is, for example, that the output probability value is 0.5.

[0224] Adjusting parameters to perform model training is a conventional technical means of model training, and this application does not limit this. Figure 9 As shown, this application obtains a cross-institutional domain-adaptive model through cross-modal comparative learning, which is used to weaken the model's ability to distinguish which specific institution the data comes from and obtain clinical decision support results. This achieves the technical effect of improving the model's domain adaptability.

[0225] For example, 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 indicate that the model is insensitive to the domain. That is, the second probability value meets the fourth objective requirement, and the second model can be output.

[0226] In another optional embodiment, Figure 10 As shown, the present application also provides a step of training the second model and the third model:

[0227] 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 execute step S703;

[0228] Among them, adversarial training is used to train electronic medical record data from different institutions as different fields, so that the model can automatically adapt to the medical data environment of different institutions, further improving the model's adaptability.

[0229] The fifth objective requirement is used to limit the range of the adversarial loss value. It can be set according to the specific application scenario and is not limited here. An exemplary fifth objective requirement is that the adversarial training loss value is 0.1.

[0230] This application can calculate the adversarial training loss using the following formula:

[0231]

[0232] Among them, L adv is the adversarial training loss, is the output of the domain discriminator and is used to measure the performance of the model in adversarial training.

[0233] 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;

[0234] Step S703: output the second model and the third model.

[0235] Among them, adjusting parameters to perform model training is a conventional technical means of model training, and this application does not limit this. This application intervenes in the model training process by setting an adversarial training loss, so that the model can learn more stable and generalized feature representations, thereby maintaining good performance when facing new data from different institutions or fields. This training method can improve the adaptability of the model in cross-institutional medical data analysis.

[0236] For example, the text feature representation f = [0.5, 0.6, 0.7, 0.8, 0.9]; p(d|f): = 0.5; L adv = 0.1. This indicates that the adversarial training loss value meets the fifth objective requirement and the second model can be output.

[0237] It should be noted here that when training the second model and the third model, model training intervention can be performed through any of the above-mentioned semantic alignment degree, probability value and anti-training loss value to achieve the technical effect of improving the performance of the model; any two of them can be selected for model training intervention to achieve the technical effect of improving the performance of the model; model training intervention can also be performed in sequence through the above three methods to achieve the technical effect of improving the performance of the model, and this application does not limit this.

[0238] In an alternative embodiment, Figure 11 As shown, Figure 11 The steps of training the first model provided in an exemplary embodiment of the present application are as follows:

[0239] Step S801, obtaining a plurality of independent medical terms, a plurality of compound medical terms, and context information of each compound medical term from a medical knowledge base;

[0240] The medical knowledge base includes numerous independent medical terms, compound medical terms, and contextual information for each compound medical term, which can provide reliable training samples for model training and thus obtain a better-performing model.

[0241] Step S802: annotate each compound medical term according to its context information to obtain intermediate medical terms, and use the set including each intermediate medical term and each independent medical term as the initial sample set;

[0242] Among them, since compound medical terms can include at least two independent medical terms, they can be combined with prefixes, suffixes and other words to obtain new medical terms. If they are identified through ordinary models, it is easy to split the compound medical terms into multiple independent medical terms, which will lead to errors in the recognition of electronic medical records, which will further affect the subsequent status development analysis.

[0243] Therefore, accurate recognition of compound medical terms can improve the accuracy of analysis of the target object's status development.

[0244] Therefore, during model training, medical terms with specific contextual relationships can be output as a single, integrated medical term, rather than being separated for recognition. This requires incorporating the contextual information of compound medical terms during model training to develop a model capable of recognizing compound medical terms. Once compound medical terms and their contextual information are obtained from the medical knowledge base, the compound medical terms can be annotated and combined with the independent medical terms to form a training set for subsequent model training.

[0245] Step S803, performing word embedding vector conversion processing on each initial sample set to obtain a word embedding vector;

[0246] Among them, the word embedding vector conversion process can map words into low-dimensional dense vectors, reduce data dimensions and reduce redundant information.

[0247] Step S804: Input each word embedding vector into the third initial model for model training to obtain a first model.

[0248] The third initial model is, for example, a network entity recognition (NER) model. Model training is a conventional technical means and will not be described in detail here.

[0249] like Figure 12 As shown, when performing the first model training, the application also performs the following operations:

[0250] Step S901, obtaining verification medical term samples;

[0251] Step S902: input the verification medical term sample into the first model to obtain the verification compound medical term;

[0252] Step S903: Verify the compound medical term input 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, execute step S904; if the first probability value meets the second target requirement, execute step S905;

[0253] Step S904: adjusting the training parameters of the third initial model and retraining until the first probability value meets the second target requirement;

[0254] Step S905: output the first model.

[0255] like Figure 13 As shown, when performing the first model training, the application also performs the following operations:

[0256] Step S1001, calculating a first adversarial training loss value of a first probability value, and determining whether the first adversarial training loss value meets a third target requirement; if the first adversarial training loss value does not meet the third target requirement, executing step S1002; if the first adversarial training loss value meets the third target requirement, executing step S1003;

[0257] Step S1002: Adjust the training parameters of the third initial model and retrain until the first adversarial training loss value meets the third target requirement;

[0258] Step S1003: output the first model.

[0259] The intervention in the training process of the first model is the same as that of the training of the second and third models, and will not be described here.

[0260] Exemplary combinations Figure 14 As shown, the training process of the first model is described:

[0261] Input: Electronic medical record data from different institutions:

[0262] Represents electronic medical record data from different medical institutions, including data from institution A, institution B, and institution C.

[0263] Pre-trained language model:

[0264] Input data from different institutions into the pre-trained language model for feature extraction.

[0265] Feature extraction layer:

[0266] Extract feature representations of input data to provide a basis for subsequent adversarial training.

[0267] Feature representation:

[0268] The extracted feature vector is used for subsequent adversarial training.

[0269] Adversarial training module:

[0270] Includes generators and discriminators for training models to adapt to data from different institutions.

[0271] Domain Discriminator:

[0272] The discriminator is used to predict the domain label (i.e., which institution the feature representation comes from).

[0273] Domain label prediction:

[0274] The discriminator outputs domain label prediction results.

[0275] Generator optimization:

[0276] The generator optimizes feature representation through adversarial training, making it more adaptable across different institutions.

[0277] Output: Domain Adaptive Features:

[0278] The feature representation optimized through adversarial training is adaptable across institutions.

[0279] Cross-institutional data analysis:

[0280] The optimized features are used for cross-institutional medical data analysis, such as joint modeling or data sharing.

[0281] For example, combined Figure 15 The entire process from data input to cross-modal contrastive learning is explained:

[0282] Data input: Input medical record text x and time series data z.

[0283] Medical term identification: Medical terms are identified through the first model.

[0284] Entity-level mask generation: Perform a masking operation on the medical term set as a whole to obtain masked medical terms.

[0285] Time series feature extraction: The electronic medical record is input into the second model for time series feature extraction to obtain a feature vector of at least one examination indicator.

[0286] Hierarchical temporal attention: For each of the inspection indicators, hierarchical attention processing is performed according to the characteristic vector of the inspection indicator.

[0287] Cross-modal contrastive learning: analyzing the state development of the target object associated with the electronic medical record based on the masked medical terminology, the first-layer attention weight, and the second-layer attention weight to obtain an analysis result.

[0288] For example, combined Figure 16 The entire medical entity-aware masking mechanism is described as follows:

[0289] Medical text input: Represents raw medical text input, for example: "The patient has elevated HbA1c levels and poor blood sugar control."

[0290] Medical terminology recognition module: identifies compound medical terms and labels them.

[0291] Among them, identifying compound medical terms: identifying compound medical terms (such as "glycated hemoglobin") from the input text.

[0292] Labeling medical terms: Labeling identified medical terms for subsequent processing, such as labeling glycated hemoglobin as a compound medical term;

[0293] Masking strategy generation module: This module performs overall masking on identified medical terms to prevent term splitting. The overall masked medical term representation performs overall masking on identified medical terms (e.g., replacing them with [MASK]). Avoiding term splitting ensures that medical terms are not split or masked, preserving their semantic integrity.

[0294] Pre-trained language model: Learns contextual information to predict masked medical terms. The Learning Contextual Prediction Masked Representation model learns to predict masked medical terms using contextual information.

[0295] Output prediction results: Verify whether the prediction results are consistent with the original medical terms. Verify prediction results: Verify whether the model's prediction results are accurate. Compare with original terms: Compare the prediction results with the original medical terms to evaluate model performance.

[0296] Verification and comparison: Finally, the prediction results are verified to be consistent with the original medical terms, completing the entire process.

[0297] The present application provides a dynamic adaptive semantic analysis method for electronic medical records. Considering that pre-trained models are usually used to perform random masking and other processing on electronic medical records, the compound medical terms in the electronic medical records are split, and the dynamic association between the examination indicators and the diagnosis conclusion cannot be obtained, resulting in the problem of being unable to provide reliable data support for subsequent applications. The present application will identify the compound medical terms in the electronic medical record when performing masking operations on medical terms to avoid splitting the compound medical terms. In this way, the compound medical terms can be masked as a whole, the integrity of the medical terms is retained, and the accuracy of electronic medical record processing is improved. In addition, the present application captures the changing trend of the examination indicators and the correlation between the changing trend and the diagnosis conclusion by calculating the first-level attention weight and the second-level attention weight of each examination indicator, and obtains the temporal relationship of the medical data to provide more intuitive and accurate data support for the subsequent analysis of the status development of the target object associated with the electronic medical record.

[0298] The above description is merely a preferred embodiment of the present application and an illustration of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to the technical solutions formed by the specific combination of the above-mentioned technical features, but also encompasses other technical solutions formed by any combination of the above-mentioned technical features or their equivalents without departing from the inventive concept. For example, a technical solution formed by replacing the above-mentioned features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A dynamic adaptive semantic analysis method for electronic medical records, characterized in that: The method comprises: Performing medical term recognition on the electronic medical record using a first model to obtain a medical term set of the electronic medical record, wherein the medical term set includes a compound medical term, and the compound medical term includes at least two independent medical terms; Performing a masking operation on the medical term set as a whole to obtain masked medical terms; Inputting the electronic medical record into a second model to perform time series feature extraction to obtain a feature vector of at least one examination indicator, wherein the feature vector includes the value and / or diagnosis conclusion of the examination indicator at different times; For each of the inspection indicators, hierarchical attention processing is performed according to the characteristic vector of the inspection indicator to obtain a first-layer attention weight and a second-layer attention weight of the inspection indicator, where the first-layer attention weight is used to characterize a changing trend of the numerical value corresponding to the inspection indicator, and the second-layer attention weight is used to characterize a correlation between the changing trend and the diagnostic conclusion corresponding to the inspection indicator; Performing a status development analysis on the target object associated with the electronic medical record based on the masked medical term, the first-layer attention weight, and the second-layer attention weight to obtain an analysis result; The step of performing hierarchical attention processing on each inspection indicator according to the characteristic vector of the inspection indicator to obtain the first-layer attention weight and the second-layer attention weight of the inspection indicator includes: Determining the first-layer attention weight according to the characteristic vector of the inspection indicator and the first-layer attention parameter; The second-layer attention weight is determined according to the product of the first-layer attention weight and the characteristic vector of the inspection indicator and the second-layer attention parameter.

2. The method according to claim 1, characterized in that The step of performing a status development analysis on the target object associated with the electronic medical record based on the masked medical term, the first-layer attention weight, and the second-layer attention weight to obtain an analysis result, wherein the analysis result is used to characterize the status development of the target object, specifically including: Based on the masked medical term, the first layer attention weight, the second layer attention weight and the text feature encoding vector, the status development situation of the target object associated with the electronic medical record is analyzed to obtain the analysis result.

3. The method according to claim 2, characterized in that The method further comprises: The electronic medical record is input into a third model for text feature encoding to obtain the text feature encoding vector.

4. The method according to claim 3, characterized in that The method further comprises the steps of training the second model and the third model: Acquire multiple first text features, multiple second text features, and multiple numerical features as positive samples, wherein the first text features are features representing the diagnosis conclusions of examination indicators in the electronic medical record, and the second text features are 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; The plurality of second text features are input into a second initial model for model training to obtain the third model.

5. The method according to claim 4, characterized in that The method further comprises: Acquire verification text features and verification numerical features, wherein the verification text features are text features in the electronic medical record other than the diagnostic conclusion representing the examination indicators; Inputting the verification numerical feature into the second model to obtain a verification numerical vector; Inputting the verification text feature into the third model to obtain a verification text vector; Calculating the cosine similarity loss between the verification numerical vector and the verification text vector; Determining a degree of semantic alignment between the verification numerical vector and the verification text vector based on the cosine similarity loss; If the semantic alignment degree does not meet the first target requirement, adjusting 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 so, the second model and the third model are output.

6. The method according to claim 1, characterized in that The method further comprises the step of training the first model: Acquire multiple independent medical terms, multiple compound medical terms, and context information of each compound medical term from a medical knowledge base; Annotating each of the compound medical terms according to context information of each compound medical term to obtain intermediate medical terms, and using a 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 on each of the initial sample sets to obtain a word embedding vector; Each of the word embedding vectors is input into a third initial model for model training to obtain the first model.

7. The method according to claim 6, characterized in that The method further comprises: Obtain samples of validated medical terminology; Inputting the verification medical term sample into the first model to obtain a verification compound medical term; Inputting the verification compound medical term into a domain discriminator to obtain a first probability value, and determining whether the first probability value meets a second target requirement; If the first probability value does not meet the second target requirement, adjusting the training parameters of the third initial model and retraining until the first probability value meets the second target requirement; If the first probability value meets the second target requirement, the first model is output.

8. The method according to claim 7, characterized in that The method further comprises: Calculating a first adversarial training loss value for the first probability value, and determining whether the first adversarial training loss value meets a third objective requirement; If the first adversarial training loss value does not meet the third target requirement, adjusting 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, the first model is output.

9. The method according to claim 5, characterized in that The method further comprises: 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 and 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, the second model and the third model are output.

10. The method according to claim 9, characterized in that The method further comprises: 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 objective 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 objective requirement, the second model and the third model are output.

Images