Processing Method, Device, Storage Medium and Electronic Device for Electronic Medical Records

Abstract

The present disclosure relates to a method, apparatus, storage medium, and electronic device for processing electronic medical records. The processing method includes: identifying target features corresponding to each medical visit record in a first electronic medical record, where the first electronic medical record includes at least two medical visit records; generating a feature matrix corresponding to the first electronic medical record based on the target features corresponding to each medical visit record, where the feature matrix includes multiple high-dimensional multi-hot vectors; generating a medical event matrix corresponding to the first electronic medical record based on the target features corresponding to each medical visit record, where the medical event matrix includes multiple low-dimensional dense vectors; and generating a first patient representation vector corresponding to the first electronic medical record according to the feature matrix and the medical event matrix, where the first patient representation vector is used to represent the medical record information in the first electronic medical record. Thus, the patient representation vector obtained through training by the above method can more accurately reflect the medical record information in the patient's electronic medical record.

Description

Technical Field

[0001] The present disclosure relates to the field of data processing, and specifically, to a method, apparatus, storage medium, and electronic device for processing electronic medical records. Background Art

[0002] The electronic medical record data of patients has become an important basic resource for the application of medical artificial intelligence. With the continuous accumulation of data and the continuous improvement of quality, mining the implicit information in the medical records based on the electronic medical records provides strong data support for medical researchers. The multiple visit records of patients can be abstracted into serialized data with temporal characteristics, which can be applied to prediction tasks of various clinical outcomes such as disease diagnosis, assisted treatment, examinations, and tests, and has guiding and practical significance for clinical medical research and practice. However, due to the characteristics of real-time, temporal, high-dimensional, sparse, and multi-noisy of patients' electronic medical record data, these heterogeneous characteristics make it difficult for the existing feature extraction methods used in machine learning to accurately process. Summary of the Invention

[0003] The purpose of the present disclosure is to provide a method, apparatus, storage medium, and electronic device for processing electronic medical records to solve the technical problem that the vector representation of electronic medical records in related technologies is inaccurate.

[0004] To achieve the above purpose, the present disclosure provides a method for processing electronic medical records, the method comprising:

[0005] Identifying target features corresponding to each visit record in a first electronic medical record, where the first electronic medical record includes at least two visit records;

[0006] Generating a feature matrix corresponding to the first electronic medical record based on the target features corresponding to each visit record, where the feature matrix includes a plurality of high-dimensional multi-hot vectors;

[0007] Generating a medical event matrix corresponding to the first electronic medical record based on the target features corresponding to each visit record, where the medical event matrix includes a plurality of low-dimensional dense vectors;

[0008] Generating a first patient representation vector corresponding to the first electronic medical record according to the feature matrix and the medical event matrix, where the first patient representation vector is used to represent the medical record information in the first electronic medical record.

[0009] Optionally, the method further comprises:

[0010] Decoding the first patient representation vector based on a neural network model to generate a first visit sequence;

[0011] Based on the general task learning objective, perform deep learning on the first visit sequence and the feature matrix to train the neural network model;

[0012] Optionally, the method further includes:

[0013] Obtain the new visit record of the patient;

[0014] According to the new visit record, obtain the corresponding new patient representation vector;

[0015] According to the neural network model, predict the first patient representation vector to generate a predicted patient representation vector;

[0016] Based on the domain task learning objective, perform deep learning on the new patient representation vector and the preset patient representation vector to train the neural network model.

[0017] Optionally, the method further includes:

[0018] Input the second electronic medical record into the neural network model;

[0019] Based on the neural network model, identify the second electronic medical record and generate a second patient representation vector corresponding to the second electronic medical record.

[0020] Optionally, the generating the feature matrix corresponding to the first electronic medical record based on the target features corresponding to each visit record includes:

[0021] Identify at least one dimensional feature corresponding to at least one target feature corresponding to each visit record;

[0022] For each visit record, based on the at least one target feature and the at least one dimensional feature of the visit record, generate a high-dimensional multi-hot vector corresponding to the visit record, so as to obtain multiple high-dimensional multi-hot vectors corresponding to each visit record;

[0023] According to the multiple high-dimensional multi-hot vectors, generate the feature matrix corresponding to the electronic medical record.

[0024] Optionally, the method further includes:

[0025] Obtain a preset feature term set, where the preset feature term set includes the mapping relationship between multiple initial target features and multiple initial dimensional features;

[0026] According to the preset feature term set, perform data cleaning on the target features corresponding to each visit record according to a preset rule to obtain the target features corresponding to each visit record after cleaning.

[0027] Optionally, for each visit record, generating a high-dimensional multi-hot vector corresponding to the visit record based on the at least one target feature and the at least one dimensional feature of the visit record includes:

[0028] Generating a dimension allocation table based on at least one first dimensional feature of a first visit record, where the dimension allocation table includes a mapping relationship between the at least one dimensional feature corresponding to the first visit record and the allocated dimension size;

[0029] Encoding a first target feature in the first visit record according to the dimension allocation table to obtain a high-dimensional multi-hot vector corresponding to the first visit record, where the first visit record is any visit record in the first electronic medical record.

[0030] Optionally, the generating a medical event matrix corresponding to the first electronic medical record based on the target features corresponding to the respective visit records includes:

[0031] Generating a patient knowledge graph corresponding to the first electronic medical record based on the target features corresponding to the respective visit records;

[0032] Obtaining a corresponding high-dimensional sparse vector according to the patient knowledge graph, where the high-dimensional sparse vector is used to represent the dimensional relationship of the target features corresponding to the respective visit records in the patient knowledge graph;

[0033] Encoding the high-dimensional sparse vector into a low-dimensional dense vector according to the hierarchical relationship of the target features corresponding to the respective visit records in the patient knowledge graph;

[0034] Generating a medical event matrix corresponding to the first electronic medical record according to the low-dimensional dense vector.

[0035] Optionally, the generating a patient knowledge graph based on the target features includes:

[0036] Obtaining at least one conceptual feature corresponding to the first electronic medical record according to the target features corresponding to the respective visit records;

[0037] Generating a patient knowledge graph corresponding to the first electronic medical record based on the hierarchical relationship between the target features and the at least one conceptual feature.

[0038] Optionally, the generating a first patient representation vector corresponding to the first electronic medical record according to the feature matrix and the medical event matrix includes:

[0039] Performing a linear operation on the feature matrix and the medical event matrix to obtain a low-dimensional visit vector corresponding to the first electronic medical record;

[0040] Based on the neural network model, generate the first patient representation vector corresponding to the first electronic medical record according to the low-dimensional medical visit vector.

[0041] The second part of the embodiments of the present disclosure provides a processing device for electronic medical records, and the device includes:

[0042] An identification module, configured to identify target features corresponding to medical visit records in a first electronic medical record, where the first electronic medical record includes at least two medical visit records;

[0043] A first generation module, configured to generate a feature matrix corresponding to the first electronic medical record based on the target features corresponding to the respective medical visit records, where the feature matrix includes a plurality of high-dimensional multi-hot vectors;

[0044] A second generation module, configured to generate a medical event matrix corresponding to the first electronic medical record based on the target features corresponding to the respective medical visit records, where the medical event matrix includes a plurality of low-dimensional dense vectors;

[0045] A third generation module, configured to generate a first patient representation vector corresponding to the first electronic medical record according to the feature matrix and the medical event matrix, where the first patient representation vector is used to represent the medical record information in the first electronic medical record.

[0046] The third part of the embodiments of the present disclosure provides a non-transitory computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of the method described in any item of the first part are implemented.

[0047] The fourth part of the embodiments of the present disclosure provides an electronic device, including:

[0048] A memory, on which a computer program is stored;

[0049] A processor, configured to execute the computer program in the memory to implement the steps of the method described in any item of the first part.

[0050] Through the above technical solutions, at least the following beneficial technical effects can be achieved:

[0051] Identify the target features corresponding to each medical visit record in the first electronic medical record. The first electronic medical record includes at least two medical visit records. Based on the target features corresponding to each medical visit record, generate a feature matrix corresponding to the first electronic medical record. The feature matrix includes multiple high-dimensional multi-hot vectors. Based on the target features corresponding to each medical visit record, generate a medical event matrix corresponding to the first electronic medical record. The medical event matrix includes multiple low-dimensional dense vectors. According to the feature matrix and the medical event matrix, generate a first patient representation vector corresponding to the first electronic medical record. The first patient representation vector is used to represent the medical record information in the first electronic medical record. Thus, through the above method, the electronic medical record information corresponding to the patient is transformed into a unified and general patient representation vector by using the feature matrix and the medical event matrix, so that the obtained patient representation vector can accurately express the semantic information in the electronic medical record, and the patient representation vector can more accurately reflect the medical record features in the patient's electronic medical record. And use a general task training model and a domain task training model to train the neural network model, so that this solution does not depend on a specific task, has good generality and robustness, and the obtained patient representation vector can well handle the problems of sparse and noisy electronic medical record data.

[0052] Other features and advantages of the present disclosure will be described in detail in the subsequent specific implementation section. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The drawings are used to provide a further understanding of the present disclosure, and constitute a part of the specification. Together with the following specific implementation, they are used to explain the present disclosure, but do not constitute a limitation to the present disclosure. In the drawings:

[0054] Figure 1 is a flowchart of a method for processing an electronic medical record according to an exemplary embodiment.

[0055] Figure 2 is a flowchart of a method for generating a feature matrix according to an exemplary embodiment.

[0056] Figure 3 is a flowchart of a method for generating a medical event matrix according to an exemplary embodiment.

[0057] Figure 4 is a schematic structural diagram of a patient knowledge graph according to an exemplary embodiment.

[0058] Figure 5 is a flowchart of another method for generating a medical event matrix according to an exemplary embodiment.

[0059] Figure 6 is a flowchart of a method for training a neural network model according to an exemplary embodiment.

[0060] Figure 7 is a schematic flowchart of a training method for another neural network model shown according to an exemplary embodiment.

[0061] Figure 8 is a block diagram of a processing device for an electronic medical record shown according to an exemplary embodiment.

[0062] Figure 9 is a block diagram of an electronic device shown according to an exemplary embodiment.

[0063] Figure 10 is a block diagram of another electronic device shown according to an exemplary embodiment. Detailed implementation manners

[0064] The following will describe in detail the specific implementation manners of the present disclosure with reference to the accompanying drawings. It should be understood that the specific implementation manners described herein are only used to illustrate and explain the present disclosure, and are not used to limit the present disclosure.

[0065] Existing patient representation methods are all by-products of prediction tasks, targeting some specific tasks rather than general and universal methods. When different tasks need to be adapted, these methods are not applicable. In the prior art, there are mainly the following two vector representation methods for electronic medical records:

[0066] One-hot or multi-hot encoding, where the information of a patient's one-time visit is corresponding to multiple diagnosis codes, each code is represented by a one-hot vector, and then multiple codes are concatenated into a vector or matrix to represent the information of this visit. This encoding method is simple and efficient, with a single data type, but the disadvantage is the lack of semantic information.

[0067] Using the neural network model RNN (Recurrent Neural Network) model for temporal encoding. This method uses the RNN model to model the temporal characteristics of the patient's visit information and generates a low-dimensional vector representation of the patient with temporal characteristics, which has certain semantic information. However, due to the characteristics of the patient's electronic medical record data such as real-time, temporal, high-dimensional, sparse, and multi-noisy, these heterogeneous characteristics will affect the semantic representation, making it difficult for the feature extraction methods used in existing machine learning to accurately process. The heterogeneous characteristics of electronic medical records will affect the semantics of the vector representation obtained through machine learning, resulting in inaccurate or even lack of semantic information in the vector representation corresponding to the semantic information, and making the semantic representation of the obtained patient representation vector inaccurate, thus making the vector representation method of electronic medical records less effective.

[0068] To solve the technical problem of inaccurate vector representation of electronic medical records in the prior art, the embodiments of the present disclosure provide a processing method for electronic medical records. Figure 1FIG. 1 is a flowchart of a method for processing an electronic medical record according to an exemplary embodiment. Figure 1 As shown, the method for processing the electronic medical record includes the following steps:

[0069] Step S11, identifying target features corresponding to each medical record in a first electronic medical record, where the first electronic medical record includes at least two medical records.

[0070] It should be noted that in the electronic medical record database corresponding to each patient, one patient corresponds to one electronic medical record, and the medical records are arranged in the order of the medical records based on the number of visits and the time of the visits. The electronic medical record records at least one medical record, and the medical record includes the description information of the patient, the patient's medical history information, the patient's medication information, the patient's surgical information, the patient's personal identity information, etc. recorded by the doctor based on the patient's condition description. In the electronic medical record database, there is only one medical record of the medical record, and the corresponding patient medical event is accidental and sudden. For the comprehensive study of the corresponding condition, it is impossible to identify the regular characteristics of universality and accuracy, and it has no guiding and practical significance for clinical medical research. Therefore, in this embodiment, it is necessary to first screen the electronic medical records in the electronic medical record database, obtain the electronic medical records with more than two medical records as the first electronic medical record, and identify the target features corresponding to each medical record in the first electronic medical record.

[0071] It is understandable that each medical record corresponding to the electronic medical record is a record of the patient's condition recorded by the doctor based on the patient's description, including medical record features, punctuation marks, modal particles, conjunctions, etc. When conducting clinical medical research, doctors need to read a large number of electronic medical records and filter out medical record features in the electronic medical records. In order to improve the efficiency of the consultation, it is necessary to eliminate interference items in the medical records to facilitate the doctor to view the data. Therefore, it is necessary to extract the medical record features in the medical records to obtain the target features corresponding to each medical record. It is understandable that the target features include multiple medical record features corresponding to the patient. For example, the categories of medical record features can include: identity features such as patient height and age, symptom features such as cough and fever, examination features such as CT examination and blood routine, medication features such as cephalosporin cough syrup, and surgical features such as appendectomy and orthopedic surgery. Optionally, according to prior medical knowledge, the medical record features that may appear in the electronic medical record are summarized and sorted to obtain a medical feature knowledge table, and then according to the medical feature knowledge table, each medical record in the electronic medical record is identified to obtain the target features corresponding to each medical record.

[0072] Step S12: generating a feature matrix corresponding to the first electronic medical record based on the target features corresponding to each medical record, wherein the feature matrix includes a plurality of high-dimensional multi-hot vectors.

[0073] According to the corresponding target features in each medical record, using an encoding algorithm, each medical record is encoded into multiple high-dimensional multi-hot vectors. Each medical record corresponds to a high-dimensional multi-hot vector, and they are sorted according to the chronological order corresponding to the medical records, so as to obtain the feature matrix corresponding to the first electronic medical record, and this feature matrix can be used to represent the target features of the first electronic medical record in each medical record.

[0074] Figure 2 It is a schematic flowchart of a method for generating a feature matrix proposed according to an exemplary embodiment. Refer to Figure 2 The above step S12 may include:

[0075] Step S121, identify at least one dimensional feature corresponding to at least one target feature corresponding to each medical record.

[0076] It can be understood that the target feature is a set of multiple medical record features, and the dimensional feature is the medical record attribute corresponding to each medical record feature in the target feature. Optionally, the dimensional feature may be an identity feature, a diagnosis feature, a symptom feature, an examination feature, a medication feature, a surgical feature, etc. Each medical record records at least one target feature corresponding to the patient, and then the dimensional features corresponding to the target features are identified to obtain at least one dimensional feature. Optionally, obtain multiple initial target features that may appear in the electronic medical record, then identify the multiple initial dimensional features corresponding to the multiple initial target features, and establish a medical feature knowledge table, which includes the mapping relationship between the multiple initial target features and the multiple initial dimensional features. According to the medical feature knowledge table, determine the multiple dimensional features corresponding to each medical record feature in the target feature. For example, the dimensional features corresponding to the target features of cough, fever, and runny nose are symptom features, the dimensional features corresponding to the target features of height, weight, age, and name are identity features, and the dimensional features corresponding to the target features of amoxicillin capsules, cephalosporins, and cough syrup are medication features.

[0077] Optionally, after the above step S121, the processing method further includes:

[0078] Obtain a preset feature term set, which includes the mapping relationship between multiple initial target features and multiple initial dimensional features.

[0079] According to the preset feature term set, perform data cleaning on the target features corresponding to each medical record according to a preset rule, so as to obtain the target features corresponding to each medical record after cleaning.

[0080] It should be noted that the medical records in the electronic medical records are the medical record information recorded according to the doctor's description of the patient's current visit. Due to the doctor's description habits, the same target feature may appear multiple times in the same medical record, multiple target features with the same meaning may appear, target features with different meanings may use the same term, the target feature description is incomplete and cannot be recognized, and the target feature has no reference value. Therefore, it is necessary to perform data cleaning on the obtained multiple target features.

[0081] For example, in this embodiment, according to the corresponding medical knowledge and medical description habits, a mapping relationship between the initial target feature and the initial dimension feature is established to obtain the preset feature term set. The preset feature term set also includes the corresponding relationship between the target feature and its alternative names, abbreviations, general names, and other multiple description methods. By identifying the corresponding description methods, the corresponding target feature and dimension feature can be returned in the medical record, so that the target features with the same meaning but different descriptions are aligned. The target features that cannot be recognized according to the preset feature term set are deleted, so as to screen out the target features that need to be encoded. For example, the doctor's hand-drawn diagram, the patient's electrocardiogram, etc. cannot be used to represent the patient vector because their forms of expression are relatively special. In the data cleaning stage, the corresponding unnecessary target features are directly deleted. It can be understood that when the doctor makes a record, due to a slip of the pen or writing habits, the described target feature may be unclear or inaccurate. For example, descriptions such as stomach disease, eye disease, heart disease, cold, etc. are habitual descriptions, and the expressed target features are not accurate and need to be data-cleaned to obtain accurate target features. For example, mark the target features that cannot be recognized according to the preset feature term set and do not belong to the above unnecessary target features that need to be deleted, and send out the corresponding prompt information to remind the relevant personnel to revise and supplement.

[0082] Step S122, for each medical record, generate a high-dimensional multi-hot vector corresponding to the medical record based on at least one target feature and at least one dimension feature of the medical record, so as to obtain multiple high-dimensional multi-hot vectors corresponding to each medical record.

[0083] It can be understood that the high-dimensional multi-hot vectors corresponding to each medical record are used to represent at least one target feature included in the medical record and at least one dimensional feature corresponding to each target feature. For multiple medical records in the electronic medical record, multiple corresponding high-dimensional multi-hot vectors are determined. By way of example, one medical record corresponds to one high-dimensional multi-hot vector, and the lengths of all the high-dimensional multi-hot vectors are the same. According to the medical feature knowledge table in the above embodiment, positions and sizes are assigned to each initial dimensional feature in the initial high-dimensional multi-hot vector, and then at least one target feature and at least one dimensional feature in the medical record are compared, and the target feature corresponding to the dimensional feature is recorded at the corresponding position in the initial high-dimensional multi-hot vector, so as to obtain the high-dimensional multi-hot vector. The above steps are repeated for each medical record, so as to obtain multiple high-dimensional multi-hot vectors corresponding to multiple medical records.

[0084] Optionally, step S122 above may further include:

[0085] Generate a dimension allocation table based on at least one first dimensional feature of the first medical record. The dimension allocation table includes the mapping relationship between at least one dimensional feature corresponding to the first medical record and the allocated dimension size.

[0086] Encode the target features in the first medical record according to the dimension allocation table to obtain the high-dimensional multi-hot vector corresponding to the first medical record, where the first medical record is any medical record in the first electronic medical record.

[0087] By way of example, to make it more convenient for the machine to recognize, the target features in each medical record identified through the above steps can be numerically encoded according to a certain rule. Optionally, in this embodiment, the International Classification of Diseases (ICD) can be used to encode the target features in the medical record. For example, the first medical record of the patient in the electronic medical record is "the patient has symptoms such as allergic cough, chest tightness, and sneezing", and the identified target features are: "allergic cough", "chest tightness", "sneezing". The character representation corresponding to the first medical record obtained through ICD encoding is "vist (visit) 1: R05.X51 R06.003 R06.751". By obtaining the character representation of each medical record, the character representation of "vist1: 558.9 477.9 401.9 247.9 530.8; vist2: 278.0 584.9 995.91 518.81; vist3: v58.61 428.0 780.2 786.50" is obtained to represent the electronic medical record corresponding to the patient.

[0088] By recognizing the first character representation corresponding to the above first medical record, at least one first-dimensional feature corresponding to the target feature is determined. According to the medical feature knowledge table in the above embodiments, the number of target features included in each dimensional feature in the doctor feature knowledge table is determined, and a dimensional size is assigned to the dimensional feature. For example, the following assignment table can be referred to:

[0089]

[0090]

[0091] Based on the above assignment table, a position storage relationship is established for all target features included in each dimensional feature. For the target feature corresponding to the first medical record, when the first target feature included in the first-dimensional feature is recognized, the first target feature is then stored at the corresponding position of the first-dimensional feature. For the target feature that does not exist in the first-dimensional feature corresponding to the first medical record, the storage state is "none", thereby establishing a dimensional assignment table. For example, the medical record of the patient is "The patient began to have symptoms such as coughing and expectorating without obvious cause 10 days ago. At that time, the patient received an injection (antibiotics, 30 ml) at an individual clinic for 3 days, and there was no obvious improvement. In recent days, the above symptoms have worsened. Therefore, for further diagnosis and treatment, the patient came to the outpatient department of our hospital today. The outpatient department admitted the patient to the hospital with bronchopneumonia. During the course of the disease, the patient had poor sleep and normal bowel and bladder movements". The target features recognized through the above steps are "coughing, expectorating, antibiotics, bronchopneumonia, poor sleep, normal bowel movement, normal urine output". Among them, the dimensional features corresponding to "coughing, expectorating, poor sleep, normal bowel movement, normal urine output" are symptom features, the dimensional feature corresponding to "antibiotics" is the medication feature, and the dimensional feature corresponding to "bronchopneumonia" is the diagnosis feature. According to the above assignment table, the storage positions of each target feature in the dimensional feature are determined, thereby establishing the following dimensional assignment table:

[0092] According to the above dimension allocation table, encode the target features corresponding to the first medical record to obtain a high-dimensional multi-hot vector corresponding to the first medical record. For example, in this embodiment, the multi-hot algorithm can be used for encoding, and the target features are sequentially allocated to the high-dimensional multi-hot vector according to the categories of the dimension features. For example, the high-dimensional multi-hot vector is sorted in segments in the order of identity features, diagnostic features, symptom features, examination features, verification features, medication features, and surgical features. According to the number of dimension features included in the medical feature knowledge table, determine the dimension size of each dimension feature, and allocate the code length according to the dimension size. Sort the medical record feature categories under the same dimension feature category according to the preset sorting method, and use "0" and "1" to represent the non-existence and existence of the corresponding medical record features at the corresponding positions under the category of the dimension feature respectively. By identifying whether the corresponding medical record features exist at the corresponding positions of the dimension features in the first medical record, generate a string of numbers containing "0" and "1", so as to obtain the high-dimensional multi-hot vector corresponding to the first medical record.

[0093] Step S123, generate a feature matrix corresponding to the electronic medical record according to multiple high-dimensional multi-hot vectors.

[0094] After obtaining multiple high-dimensional multi-hot vectors corresponding to multiple medical records through the above steps, arrange the multiple high-dimensional multi-hot vectors in the order of the medical records, so as to obtain a feature matrix corresponding to the electronic medical record.

[0095] Step S13, generate a medical event matrix corresponding to the first electronic medical record based on the target features corresponding to each medical record. The medical event matrix includes multiple low-dimensional dense vectors.

[0096] It should be noted that the target features include multiple medical record features, and there are internal corresponding relationships between the medical record features. For example, the dimension feature corresponding to "acute coronary syndrome" is the diagnostic feature, and the medical record features manifested at the patient level are symptom features such as "shortness of breath after exertion", "palpitation", "palpitation", "chest pain", "shortness of breath", etc. Query according to the above medical feature knowledge table to determine one or more superior medical record features and one or more inferior medical record features corresponding to each medical record feature in the target features, and determine the subordination relationship between the medical record features. By identifying the subordination relationship between the medical record features in the target features, encode each medical record feature into a low-dimensional dense vector according to the preset algorithm, and generate a medical event matrix corresponding to the first electronic medical record through the obtained multiple low-dimensional dense vectors.

[0097] Figure 3 is a schematic flowchart of a method for generating a medical event matrix according to an exemplary embodiment. See Figure 3 The above step S13 may include:

[0098] Step S131: Generate a patient knowledge graph corresponding to the first electronic medical record based on the target features corresponding to each medical record.

[0099] It can be understood that during the process of medical treatment, doctors usually make a diagnosis based on a comprehensive judgment of multiple medical record features such as the patient's corresponding symptom conditions, medication conditions, surgical history, etc., to determine the disease that the patient is most likely to have. Therefore, when building a patient knowledge graph, it is necessary to establish the corresponding relationships between the various medical record features in the target features to facilitate doctors to make an accurate diagnosis based on the corresponding relationships. In this embodiment, based on medical knowledge, the corresponding relationships between the various medical record features can be established in the above-mentioned medical feature recognition table, and the patient knowledge graph can be generated by querying the medical feature recognition table.

[0100] Optionally, the above step S131 may include:

[0101] Obtain at least one conceptual feature corresponding to the first electronic medical record according to the target features corresponding to each medical record.

[0102] Generate a patient knowledge graph corresponding to the first electronic medical record based on the hierarchical relationship between the target feature and the at least one conceptual feature.

[0103] It can be understood that each medical record feature corresponds to a corresponding conceptual feature. For example, the conceptual features corresponding to aspirin, cephalosporin, antibiotics, etc. are medication features, and the conceptual features corresponding to PCI (percutaneous coronary intervention) and CABG (Coronary Artery Bypass Graft) are surgical features; based on the input habits of doctors when describing medical records, some medical record features can be recognized as corresponding conceptual features. For example, for symptom features, when doctors describe symptoms, they will append the word "symptom" after describing the specific symptoms to summarize the corresponding symptoms, and the conceptual feature "symptom" corresponding to medical record features such as "cough, runny nose, sneeze" can be extracted through field recognition; however, for some medical record features, doctors usually do not introduce the corresponding conceptual features when recording. For example, for surgical features such as PCI and CABG, when doctors input, there is usually no conceptual feature such as "surgery" to describe the medical record. Therefore, it is necessary to search and supplement the corresponding conceptual features according to the above-mentioned medical feature knowledge table to determine the corresponding relationship between the target feature and the conceptual feature.

[0104] Figure 4 It is a schematic structural diagram of a patient knowledge graph proposed according to an exemplary embodiment. Refer to Figure 4, it can be understood that the same-level concept features can have a corresponding relationship with the concept features at a higher level. For example, surgical features and medication features both belong to treatment features, coronary heart disease and angina pectoris both belong to disease features, and blood routine and urine routine both belong to examination features. Therefore, it is necessary to search for and supplement the concept features at each level by querying medical feature knowledge and determine the corresponding relationship between each medical record feature and the concept feature. It should be understood that referring to Figure 4 as shown, the corresponding relationship between the medical record features and the concept features includes the mutual relationship between each medical record feature and other medical record features, the mutual relationship between each medical record feature and each concept feature, and the mutual relationship between each concept feature and other concept features, and then establish a Figure 4 as shown patient knowledge graph.

[0105] Step S132, according to the patient knowledge graph, obtain the corresponding high-dimensional sparse vector, which is used to represent the dimensional relationship of the target features corresponding to each medical record in the patient knowledge graph.

[0106] Step S133, according to the hierarchical relationship of the target features corresponding to each medical record in the patient knowledge graph, encode the high-dimensional sparse vector into a low-dimensional dense vector.

[0107] Step S134, according to the low-dimensional dense vector, generate the medical event matrix corresponding to the first electronic medical record.

[0108] Figure 5 is a flowchart showing another method for generating a medical event matrix according to an exemplary embodiment. Referring to Figure 5 as shown, according to the patient knowledge graph obtained in the above steps, use the one-hot algorithm to encode each medical record feature or concept feature into a high-dimensional sparse vector e i , and then use the attention layer to obtain the membership relationship and semantic information of the high-dimensional sparse vector in the patient knowledge graph, so as to encode the high-dimensional sparse vector into a low-dimensional dense vector gi. Exemplarily, the encoding can be performed according to the following calculation formula:

[0109]

[0110]

[0111] is the node set of the patient knowledge graph, represents other nodes related to the i node in the patient knowledge graph, α ij represents calculating g i when e j vector's attention weight, and α ij ∈R + . α ijThe calculation method of attention weight is as follows:

[0112]

[0113]

[0114] w a is e i and e j The weight matrix connected, and w a ∈R l×2m , l×2m represents the real number dimension, that is, a matrix with l rows and 2m columns, b a ∈R l is the bias vector, u a ∈R l is the weight vector for generating scalar, represents the transpose of u a , e k represents all node vectors related to the patient in the patient knowledge graph.

[0115] Step S14, generate the first patient representation vector corresponding to the first electronic medical record according to the feature matrix and the medical event matrix, and the first patient representation vector is used to represent the medical record features in the first electronic medical record.

[0116] It can be understood that the feature matrix obtained through the above steps is used to characterize the target features existing in the electronic medical record, and the medical event matrix is used to characterize the corresponding relationship between each feature in the electronic medical record. By combining the feature matrix and the medical event matrix, the first patient representation vector used to characterize the medical record information in the first electronic medical record is generated. By viewing the first patient representation vector, the doctor can read the corresponding symptom features and the corresponding relationship between each symptom in the first electronic medical record corresponding to the patient, and then diagnose and study the medical events of the patient.

[0117] Optionally, the above step S14 may include:

[0118] Perform a linear operation on the feature matrix and the medical event matrix to obtain a low-dimensional visit vector corresponding to the first electronic medical record.

[0119] Based on the neural network model, generate the first patient representation vector corresponding to the first electronic medical record according to the low-dimensional visit vector.

[0120] Exemplarily, perform matrix multiplication on the medical event matrix extracted from the patient knowledge graph and the feature matrix, and obtain the low-dimensional visit vector X corresponding to the patient's visit record through non-linear operation. Then send X into the neural network model RNN (Recurrent Neural Network), and thus generate the corresponding patient representation vector P. Exemplarily, the calculation can be performed according to the following formula:

[0121] x1, x2, x3, …, x i = tanh([g1, g2, g3, …, g i ·[v1, v2, v3, …, v i )

[0122] h1, h2, h3, …, h i = RNN(x1, x2, x3, …, x i ; θ)

[0123] p = h i

[0124] θ represents the RNN model parameters, and h i represents the patient representation vector corresponding to each visit record of the patient.

[0125] Through the above technical solution, the target features corresponding to each visit record in the first electronic medical record are identified. The first electronic medical record includes at least two visit records. Based on the target features corresponding to each visit record, a feature matrix corresponding to the first electronic medical record is generated. The feature matrix includes multiple high-dimensional multi-hot vectors. Based on the target features corresponding to each visit record, a medical event matrix corresponding to the first electronic medical record is generated. The medical event matrix includes multiple low-dimensional dense vectors. According to the feature matrix and the medical event matrix, a first patient representation vector corresponding to the first electronic medical record is generated. The first patient representation vector is used to represent the medical record information in the first electronic medical record. Thus, through the above method, the electronic medical record information corresponding to the patient is converted into a unified and general patient representation vector by using the feature matrix and the medical event matrix, so that the obtained patient representation vector can accurately express the semantic information in the electronic medical record and can more accurately reflect the medical record features in the patient's electronic medical record. And a general task training model and a domain task training model are used to train the neural network model, so that this solution does not depend on a specific task, has good generality and robustness, and the obtained patient representation vector can well handle the problems of data sparsity and noise in the electronic medical record.

[0126] Figure 6 is a schematic flowchart of a method for training a neural network model shown according to an exemplary embodiment. Refer to Figure 6 , after the above step S14, the processing method of the above electronic medical record may further include:

[0127] Step S15, based on the neural network model, decode the first patient representation vector to generate a first visit sequence.

[0128] Step S16, based on the general task learning objective, perform deep learning on the first visit sequence and the feature matrix to train the neural network model.

[0129] It is understandable that the first patient representation vector p is decoded through a neural network model to obtain the first visit sequence The first visit sequence is the patient visit information obtained through the neural network model. The first visit sequence is compared with the feature matrix V obtained through the first electronic medical record. If the first visit sequence is consistent with the patient visit information expressed by the feature matrix V, it means that the neural network model can accurately obtain the corresponding patient visit information based on the first electronic medical record. If the first visit sequence is not consistent with the feature matrix V, general task learning can be performed through the following formula to train the neural network model so that the neural network model can more accurately obtain the patient representation vector corresponding to the patient. The calculation formula is as follows:

[0130]

[0131]

[0132] n represents the number of electronic medical records corresponding to different patients in the collected samples, d represents the number of visits included in the electronic medical record, p represents the number corresponding to the patient electronic medical record information, p = 1 represents the electronic medical record information corresponding to the first patient, v represents the number of the visit record in the patient electronic medical record, and v = 1 represents the first visit record in the electronic medical record of the p-th patient.

[0133] Figure 7 is a schematic flowchart of another method for training a neural network model shown according to an exemplary embodiment. Refer to Figure 7 , after the above step S14, the above processing method may further include:

[0134] Step S21, obtaining the new visit record of the patient.

[0135] Step S22, obtaining the corresponding new patient representation vector according to the new visit record.

[0136] Step S23, predicting the first patient representation vector according to the neural network model to generate a predicted patient representation vector.

[0137] Step S24, performing deep learning on the new patient representation vector and the preset patient representation vector based on the domain task learning objective to train the neural network model.

[0138] It can be understood that the neural network model can perform deep learning based on different target tasks, thereby training the neural network model so that the neural network model is applicable to complex and variable patient electronic medical records and then making a more accurate patient vector representation; when diagnosing a disease, a doctor needs to predict the patient's condition based on the patient's electronic medical record. Therefore, in this embodiment, the patient representation vector corresponding to the patient is learned through the neural network model, and then the predicted patient representation vector at the t+1 moment is predicted for the patient. When the patient visits again, new visit records will be recorded in the electronic medical record corresponding to the patient. At this time, the new visit records are medical diagnoses made according to the patient's actual disease conditions; through the steps of the above embodiment, the new patient representation vector V corresponding to the new visit records is obtained. t+1 , the first patient representation vector P is used to generate the predicted patient representation vector at the t+1 moment of the patient through the neural network model. The predicted patient representation vector is compared with the new patient representation vector V t+1 . If the predicted patient representation vector and the new patient representation vector V t+1 express consistent patient visit information, it means that through the neural network model, the predicted patient representation vector corresponding to the patient can be accurately predicted. If the predicted patient representation vector and the new patient representation vector V t+1 cannot be consistent, domain task learning can be performed through the following formula to train the neural network model, so that the neural network model can more accurately predict the predicted patient representation vector corresponding to the patient. The calculation formula is as follows:

[0139]

[0140]

[0141] w P is the weight matrix connected by each vector in P, and b p is the bias vector related to P.

[0142] Optionally, the above processing method may further include:

[0143] Input the second electronic medical record into the neural network model.

[0144] Based on the neural network model, the second electronic medical record is recognized to generate the second patient representation vector corresponding to the second electronic medical record.

[0145] It can be understood that the and obtained through the above training method.Perform joint calculations, perform backpropagation to update the neural network model, and identify the obtained second electronic medical record based on the updated neural network model, so as to obtain the second patient representation vector corresponding to the second electronic medical record.

[0146] Figure 8 It is a block diagram of a processing device for electronic medical records shown according to an exemplary embodiment. Refer to Figure 8 The processing device 100 includes an identification module 110, a first generation module 120, a second generation module 130, and a third generation module 140.

[0147] The identification module 110 is configured to identify the target features corresponding to the medical records in the first electronic medical record, and the first electronic medical record includes at least two medical records.

[0148] The first generation module 120 is configured to generate a feature matrix corresponding to the first electronic medical record based on the target features corresponding to each medical record, and the feature matrix includes a plurality of high-dimensional multi-hot vectors.

[0149] The second generation module 130 is configured to generate a medical event matrix corresponding to the first electronic medical record based on the target features corresponding to each medical record, and the medical event matrix includes a plurality of low-dimensional dense vectors.

[0150] The third generation module 140 is configured to generate a first patient representation vector corresponding to the first electronic medical record according to the feature matrix and the medical event matrix, and the first patient representation vector is used to represent the medical record features in the first electronic medical record.

[0151] Optionally, the processing device 100 further includes:

[0152] The fourth generation module is configured to decode the first patient representation vector based on the neural network model to generate a first medical record sequence.

[0153] The first training module is configured to perform deep learning on the first medical record sequence and the feature matrix based on the general task learning objective to train the neural network model.

[0154] Optionally, the processing device 100 may further include:

[0155] The first acquisition module is configured to acquire the new medical records of the patient.

[0156] The second acquisition module is configured to acquire the corresponding new patient representation vector according to the new medical records.

[0157] The fifth generation module is configured to predict the first patient representation vector according to the neural network model to generate a predicted patient representation vector.

[0158] The second training module performs deep learning on the newly added patient representation vector and the preset patient representation vector based on the domain task learning objective to train the neural network model.

[0159] Optionally, the processing device 100 may further include:

[0160] An input module for inputting the second electronic medical record into the neural network model.

[0161] A sixth generation module for identifying the second electronic medical record based on the neural network model and generating a second patient representation vector corresponding to the second electronic medical record.

[0162] Optionally, the first generation module 120 may include:

[0163] An identification sub-module for identifying at least one dimensional feature corresponding to at least one target feature of each medical visit record;

[0164] A first generation sub-module for generating a high-dimensional multi-hot vector corresponding to each medical visit record based on the at least one target feature and at least one dimensional feature of the medical visit record, so as to obtain multiple high-dimensional multi-hot vectors corresponding to each medical visit record.

[0165] A second generation sub-module for generating a feature matrix corresponding to the electronic medical record according to the multiple high-dimensional multi-hot vectors.

[0166] Optionally, the processing device 100 may further include:

[0167] A third acquisition module for acquiring a preset feature term set, which includes the mapping relationship between multiple initial target features and multiple initial dimensional features.

[0168] A data cleaning module for cleaning the target features corresponding to each medical visit record according to the preset feature term set and the preset rules to obtain the cleaned target features corresponding to each medical visit record.

[0169] Optionally, the first generation sub-module may be used to:

[0170] Generate a dimension allocation table based on at least one first dimensional feature of the first medical visit record, where the dimension allocation table includes the mapping relationship between at least one dimensional feature corresponding to the first medical visit record and the allocated dimension size.

[0171] Encode the first target feature in the first medical visit record according to the dimension allocation table to obtain a high-dimensional multi-hot vector corresponding to the first medical visit record, and the first medical visit record is any medical visit record in the first electronic medical record.

[0172] Optionally, the third generation module 130 may include:

[0173] The third generation sub-module is configured to generate a patient knowledge graph corresponding to the first electronic medical record based on the target features corresponding to each medical record.

[0174] The third acquisition sub-module is configured to obtain a corresponding high-dimensional sparse vector according to the patient knowledge graph, and the high-dimensional sparse vector is used to represent the dimensional relationship of the target features corresponding to each medical record in the patient knowledge graph.

[0175] The encoding sub-module is configured to encode the high-dimensional sparse vector into a low-dimensional dense vector according to the hierarchical relationship of the target features corresponding to each medical record in the patient knowledge graph.

[0176] The fourth generation sub-module is configured to generate a medical event matrix corresponding to the first electronic medical record according to the low-dimensional dense vector.

[0177] Optionally, the third generation sub-module is configured to:

[0178] Obtain at least one conceptual feature corresponding to the first electronic medical record according to the target features corresponding to each medical record.

[0179] Generate a patient knowledge graph corresponding to the first electronic medical record based on the hierarchical relationship between the target feature and at least one conceptual feature.

[0180] Optionally, the third generation module 140 is configured to:

[0181] Perform a linear operation on the feature matrix and the medical event matrix to obtain a low-dimensional medical record vector corresponding to the first electronic medical record.

[0182] Generate a first patient representation vector corresponding to the first electronic medical record based on the neural network model according to the low-dimensional medical record vector.

[0183] Regarding the device in the above embodiments, the specific manner in which each module performs operations has been described in detail in the embodiments related to the method, and will not be elaborated here.

[0184] Figure 9 It is a block diagram of an electronic device 900 shown according to an exemplary embodiment. As Figure 9 shown, the electronic device 900 may include: a processor 901, a memory 902. The electronic device 900 may further include one or more of a multimedia component 903, an input / output (I / O) interface 904, and a communication component 905.

[0185] Among them, the processor 901 is used to control the overall operation of the electronic device 900 to complete all or part of the steps in the above-mentioned method for processing electronic medical records. The memory 902 is used to store various types of data to support the operation of the electronic device 900. These data may include, for example, instructions for any application or method operating on the electronic device 900, as well as application-related data, such as contact data, sent and received messages, pictures, audio, video, and so on. The memory 902 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic memory, flash memory, magnetic disk, or optical disc. The multimedia component 903 may include a screen and an audio component. Among them, the screen can be, for example, a touch screen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signal can be further stored in the memory 902 or sent through the communication component 905. The audio component also includes at least one speaker for outputting audio signals. The I / O interface 904 provides an interface between the processor 901 and other interface modules, and the above-mentioned other interface modules can be a keyboard, a mouse, buttons, etc. These buttons can be virtual buttons or physical buttons. The communication component 905 is used for wired or wireless communication between the electronic device 900 and other devices. Wireless communication, such as Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, 4G, NB-IOT, eMTC, or other 5G, etc., or a combination of one or more of them is not limited here. Therefore, the corresponding communication component 905 may include: a Wi-Fi module, a Bluetooth module, an NFC module, and so on.

[0186] In an exemplary embodiment, the electronic device 900 can be implemented by one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components, and is used to execute the above-mentioned method for processing electronic medical records.

[0187] In another exemplary embodiment, a computer-readable storage medium including program instructions is further provided. When the program instructions are executed by a processor, the steps of the above-mentioned method for processing electronic medical records are implemented. For example, the computer-readable storage medium can be the above-mentioned memory 902 including program instructions, and the above-mentioned program instructions can be executed by the processor 901 of the electronic device 900 to complete the above-mentioned method for processing electronic medical records.

[0188] Figure 10 FIG. is a block diagram of another electronic device 1000 shown according to an exemplary embodiment. For example, the electronic device 1000 can be provided as a server. Refer to Figure 10 FIG., the electronic device 1000 includes a processor 1022, the number of which can be one or more, and a memory 1032 for storing computer programs executable by the processor 1022. The computer programs stored in the memory 1032 can include one or more modules each corresponding to a set of instructions. In addition, the processor 1022 can be configured to execute the computer program to execute the above-mentioned method for processing electronic medical records.

[0189] In addition, the electronic device 1000 can further include a power supply component 1026 and a communication component 1050. The power supply component 1026 can be configured to perform power management of the electronic device 1000, and the communication component 1050 can be configured to implement communication of the electronic device 1000, for example, wired or wireless communication. In addition, the electronic device 1000 can further include an input / output (I / O) interface 1058. The electronic device 1000 can operate based on an operating system stored in the memory 1032, such as Windows Server TM , Mac OS X TM , Unix TM , Linux TM and so on.

[0190] In another exemplary embodiment, a computer-readable storage medium including program instructions is further provided. When the program instructions are executed by a processor, the steps of the above-described method for processing electronic medical records are implemented. For example, the non-transitory computer-readable storage medium may be the above-described memory 1032 including program instructions, and the above program instructions may be executed by the processor 1022 of the electronic device 1000 to complete the above-described method for processing electronic medical records.

[0191] In another exemplary embodiment, a computer program product is further provided. The computer program product includes a computer program executable by a programmable device, and the computer program has a code portion for executing the above-described method for processing electronic medical records when executed by the programmable device.

[0192] The preferred embodiments of the present disclosure have been described in detail above in conjunction with the accompanying drawings. However, the present disclosure is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present disclosure, various simple modifications can be made to the technical solutions of the present disclosure, and these simple modifications all fall within the protection scope of the present disclosure.

[0193] In addition, it should be noted that, in the above specific embodiments, the various specific technical features described can be combined in any suitable manner without conflict.

[0194] Furthermore, any combination can be made between the various different embodiments of the present disclosure as long as it does not violate the idea of the present disclosure, and it should also be regarded as the content disclosed by the present disclosure.

Claims

1. A method for processing electronic medical records, characterized in that, Including: Identifying target features corresponding to each medical visit record in a first electronic medical record, where the first electronic medical record includes at least two medical visit records; Generating a feature matrix corresponding to the first electronic medical record based on the target features corresponding to each medical visit record, where the feature matrix includes multiple high-dimensional multi-hot vectors; Generating a patient knowledge graph corresponding to the first electronic medical record based on the target features corresponding to each medical visit record; Obtaining a corresponding high-dimensional sparse vector according to the patient knowledge graph, where the high-dimensional sparse vector is used to represent the dimensional relationship of the target features corresponding to each medical visit record in the patient knowledge graph; Encoding the high-dimensional sparse vector into a low-dimensional dense vector according to the hierarchical relationship of the target features corresponding to each medical visit record in the patient knowledge graph; Wherein, encoding the high-dimensional sparse vector into a low-dimensional dense vector includes the following formula: , is the node set of the patient knowledge graph, j represents other nodes in the patient knowledge graph that are related to node i, represents the calculation when the attention weight of the vector, and , is the set of positive real numbers, is the low-dimensional dense vector of node i, is the high-dimensional sparse vector of node j, is the high-dimensional sparse vector of node i; Generating a medical event matrix corresponding to the first electronic medical record according to the low-dimensional dense vector; Generating a first patient representation vector corresponding to the first electronic medical record according to the feature matrix and the medical event matrix, where the first patient representation vector is used to represent the medical record information in the first electronic medical record.

2. The processing method according to claim 1, wherein The method further includes: Decoding the first patient representation vector based on a neural network model to generate a first medical visit sequence; Performing deep learning on the first medical visit sequence and the feature matrix based on a general task learning objective to train the neural network model.

3. The processing method according to claim 2, characterized in that, The method further includes: Obtaining a new medical visit record of a patient; Obtaining a corresponding new patient representation vector according to the new medical visit record; Predicting the first patient representation vector according to the neural network model to generate a predicted patient representation vector; Performing deep learning on the new patient representation vector and the predicted patient representation vector based on a domain task learning objective to train the neural network model.

4. The processing method according to claim 3, characterized in that, The method further includes: Inputting a second electronic medical record into the neural network model; Identifying the second electronic medical record based on the neural network model to generate a second patient representation vector corresponding to the second electronic medical record.

5. The processing method according to claim 1, characterized in that The generating the feature matrix corresponding to the first electronic medical record based on the target features corresponding to each medical visit record includes: Identifying at least one dimensional feature corresponding to at least one target feature corresponding to each medical visit record; For each medical visit record, generating a high-dimensional multi-hot vector corresponding to the medical visit record based on the at least one target feature and the at least one dimensional feature of the medical visit record, so as to obtain multiple high-dimensional multi-hot vectors corresponding to each medical visit record; Generating a feature matrix corresponding to the electronic medical record according to the multiple high-dimensional multi-hot vectors.

6. The processing method according to claim 5, wherein The method further includes: Obtaining a preset feature term set, where the preset feature term set includes mapping relationships between multiple initial target features and multiple initial dimensional features; Performing data cleaning on the target features corresponding to each medical visit record according to the preset feature term set according to a preset rule to obtain the target features corresponding to each medical visit record after cleaning.

7. The processing method according to claim 5, characterized in that For each visit record, generating a high-dimensional multi-hot vector corresponding to the visit record based on the at least one target feature and the at least one dimensional feature of the visit record, includes: Generating a dimension allocation table based on at least one first dimensional feature of a first visit record, where the dimension allocation table includes a mapping relationship between the at least one dimensional feature corresponding to the first visit record and the allocated dimension size; Encoding a first target feature in the first visit record according to the dimension allocation table to obtain a high-dimensional multi-hot vector corresponding to the first visit record, where the first visit record is any visit record in the first electronic medical record.

8. The processing method according to claim 1, wherein The generating a patient knowledge graph based on the target feature includes: Obtaining at least one conceptual feature corresponding to the first electronic medical record according to the target features corresponding to the respective visit records; Generating a patient knowledge graph corresponding to the first electronic medical record based on the hierarchical relationship between the target feature and the at least one conceptual feature.

9. The processing method according to claim 1, characterized in that The generating a first patient representation vector corresponding to the first electronic medical record according to the feature matrix and the medical event matrix includes: Performing a linear operation on the feature matrix and the medical event matrix to obtain a low-dimensional visit vector corresponding to the first electronic medical record; Generating the first patient representation vector corresponding to the first electronic medical record based on the low-dimensional visit vector according to a neural network model.

10. A processing device for electronic medical records, characterized in that, Includes: An identification module, configured to identify a target feature corresponding to a visit record in a first electronic medical record, where the first electronic medical record includes at least two visit records; A first generation module, configured to generate a feature matrix corresponding to the first electronic medical record based on the target features corresponding to the respective visit records, where the feature matrix includes a plurality of high-dimensional multi-hot vectors; A second generation module, configured to: Generate a patient knowledge graph corresponding to the first electronic medical record based on the target features corresponding to the respective visit records; Obtain a corresponding high-dimensional sparse vector according to the patient knowledge graph, where the high-dimensional sparse vector is used to characterize the dimensional relationship of the target features corresponding to the respective visit records in the patient knowledge graph; Encoding the high-dimensional sparse vector into a low-dimensional dense vector according to the hierarchical relationship of the target features corresponding to the respective visit records in the patient knowledge graph; Wherein, encoding the high-dimensional sparse vector into a low-dimensional dense vector includes the following formula: , is the set of nodes in the patient knowledge graph, j represents other nodes in the patient knowledge graph that are related to node i, represents the calculation when the attention weight of the vector, and , is the set of positive real numbers, is the low-dimensional dense vector of node i, is the high-dimensional sparse vector of node j, is the high-dimensional sparse vector of node i; Generating a medical event matrix corresponding to the first electronic medical record according to the low-dimensional dense vector; A third generation module, configured to generate a first patient representation vector corresponding to the first electronic medical record according to the feature matrix and the medical event matrix, where the first patient representation vector is used to represent the medical record information in the first electronic medical record.

11. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the steps of the method according to any one of claims 1-9.

12. An electronic device, characterized in that, Includes: A memory, on which a computer program is stored; A processor, configured to execute the computer program in the memory to implement the steps of the method according to any one of claims 1-9.

Images