Intelligent writing method and system for outpatient medical records based on deep learning

Abstract

The present disclosure provides a method and system for intelligently writing outpatient medical records based on deep learning, including the following steps: obtaining the patient's condition information and constructing a patient feature vector; obtaining a recommended expression based on the patient's feature vector and a deep learning model group to assist doctors in writing outpatient medical records; wherein the deep learning model group is composed of multiple deep learning models, and each deep learning model corresponds to each element in the patient's feature vector. The present disclosure uses deep learning models and knowledge graphs to analyze the patient's condition, quickly complete the writing of high-quality medical records, improve the speed and quality of medical record writing, reduce the workload of doctors in writing medical records, and improve the efficiency of outpatient diagnosis.

Description

Technical Field

[0001] The present invention belongs to the field of medical information technology, and specifically relates to a method and system for intelligently writing outpatient medical records based on deep learning. Background Art

[0002] The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute prior art.

[0003] Outpatient electronic medical records record the diagnosis and treatment activities performed by doctors in outpatient clinics and play a very important role in medical services. Since doctors are very busy when working in outpatient clinics, the quality of outpatient electronic medical records recorded by doctors is generally not high, and there are often missing and errors in the content. In order to facilitate the writing of outpatient electronic medical records, doctors generally maintain templates in advance for patients with various diseases. When writing outpatient electronic medical records, they independently select templates, import them, and then modify them. However, on the one hand, the templates maintained by doctors are not flexible enough, and the content often cannot meet the diverse needs of reality; on the other hand, when doctors make modifications based on templates, they often omit certain content that needs to be modified. The efficiency and quality of doctors writing outpatient electronic medical records cannot be guaranteed. Summary of the invention

[0004] In order to solve the above problems, the present invention proposes a method and system for intelligent writing of outpatient medical records based on deep learning, which intelligently assists doctors to flexibly reference templates in the template library and modify the content that needs to be changed in the template, so as to quickly complete high-quality outpatient electronic medical records.

[0005] According to some embodiments, the first solution of the present disclosure provides an outpatient medical record intelligent writing method based on deep learning, which adopts the following technical solutions:

[0006] A method and system for intelligently writing outpatient medical records based on deep learning, comprising the following steps:

[0007] Obtain the patient's condition information and construct the patient's feature vector;

[0008] Based on the patient feature vector and deep learning model group, the recommended statement is obtained to assist doctors in writing outpatient medical records;

[0009] Among them, the deep learning model group is composed of multiple deep learning models, and each deep learning model corresponds to each element in the patient feature vector.

[0010] As a further technical limitation, the standard outpatient medical record template library divides all outpatient patients in the historical data into m groups according to gender, age and disease diagnosis, and selects k different outpatient medical records from each group; each outpatient medical record includes the basic information of different patients as well as the chief complaint, current medical history, past medical history and allergy history in the outpatient medical record.

[0011] Furthermore, the patient feature vector is a row vector containing patient condition information.

[0012] Furthermore, the construction process of the deep learning model group is as follows:

[0013] Construct training sets, validation sets, and test sets based on patient medical records in the standard outpatient medical record template library;

[0014] A deep learning model was built in combination with the standard outpatient medical record template library to obtain a deep learning model group of patient feature vectors.

[0015] Furthermore, the training process of the deep learning model group is:

[0016] Collect the text content of each patient's outpatient medical record and convert it into vector data as a sample set;

[0017] A deep learning model group based on the bidirectional long short-term memory (LSTM) model was established. The patient feature vector was input and the patient's outpatient medical record text content information vector was output. The goal was to minimize the loss function. All the information vectors of each patient in the training set were input into the bidirectional LSTM model for training, and the parameters of the bidirectional LSTM model were determined.

[0018] As a further technical limitation, in the process of assisting the doctor in writing outpatient medical records, the patient feature vector is continuously updated, and the elements in the updated patient feature vector are matched with each deep learning model to obtain a recommended statement.

[0019] Furthermore, the updating process of the patient feature vector is:

[0020] For structured features, find the structured data corresponding to the structured features in combination with the hospital information system and update the patient feature vector;

[0021] For unstructured features, the feature names are split based on the outpatient medical records of the patient confirmed by the doctor, and the feature descriptions corresponding to all unstructured features in the knowledge graph Q are extracted from the split feature names;

[0022] Specifically, the knowledge graph Q: unstructured feature -> feature name -> related data -> feature description. An unstructured feature contains multiple feature names. For each feature name, the related data refers to the text part that describes the feature, and the feature description is the various descriptions of the feature in the related data.

[0023] According to some embodiments, the second solution of the present disclosure provides an outpatient medical record intelligent writing system based on deep learning, which adopts the following technical solutions:

[0024] An intelligent outpatient medical record writing system based on deep learning, including:

[0025] An acquisition module is configured to acquire the patient's condition information and construct a patient feature vector;

[0026] The writing module is configured to obtain a recommendation statement based on the patient feature vector and the deep learning model group to assist doctors in writing outpatient medical records;

[0027] Among them, the deep learning model group is composed of multiple deep learning models, and each deep learning model corresponds to each element in the patient feature vector.

[0028] According to some embodiments, a third solution of the present disclosure provides a computer-readable storage medium, which adopts the following technical solution:

[0029] A computer-readable storage medium stores a program thereon, which, when executed by a processor, implements the steps in the outpatient medical record intelligent writing method based on deep learning as described in the first aspect of the present disclosure.

[0030] According to some embodiments, a fourth solution of the present disclosure provides an electronic device, which adopts the following technical solution:

[0031] An electronic device comprises a memory, a processor and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the method for intelligently writing outpatient medical records based on deep learning as described in the first aspect of the present disclosure are implemented.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] The present invention adopts deep learning models and knowledge graphs to analyze the patient's condition, quickly complete the writing of high-quality medical records, improve the speed and quality of medical record writing, reduce the workload of doctors in writing medical records, and improve the efficiency of outpatient diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The accompanying drawings constituting a part of the present disclosure are used to provide a further understanding of the present disclosure. The illustrative embodiments of the present disclosure and their descriptions are used to explain the present disclosure and do not constitute an improper limitation on the present disclosure.

[0035] Figure 1 It is a flow chart of the outpatient medical record intelligent writing method based on deep learning in the first embodiment of the present disclosure;

[0036] Figure 2is a structural diagram of a neural network model of a patient feature vector in the first embodiment of the present disclosure;

[0037] Figure 3 It is a structural block diagram of the outpatient medical record intelligent writing system based on deep learning in the second embodiment of the present disclosure. DETAILED DESCRIPTION

[0038] The present disclosure is further described below in conjunction with the accompanying drawings and embodiments.

[0039] It should be noted that the following detailed descriptions are all illustrative and are intended to provide further explanation of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present disclosure belongs.

[0040] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0041] In the absence of conflict, the embodiments in the present disclosure and the features in the embodiments may be combined with each other.

[0042] Embodiment 1

[0043] Embodiment 1 of the present disclosure introduces a method for intelligently writing outpatient medical records based on deep learning.

[0044] like Figure 1 The method for intelligently writing outpatient medical records based on deep learning is shown, comprising the following steps:

[0045] Step S01: construct a standard outpatient medical record template library H;

[0046] Step S02: construct patient feature vector Tu;

[0047] Step S03: constructing a deep learning model group corresponding to the patient's feature vector;

[0048] Step S04: assisting doctors to quickly complete the writing of outpatient medical records;

[0049] Step S05: Update the patient feature vector Tu;

[0050] Step S06: Update the recommendation statement based on the patient feature vector Tu.

[0051] As one or more implementation modes, in step S01, a standard outpatient medical record template library H is constructed, which contains high-quality historical outpatient medical record data. The content of each outpatient medical record covers the basic information of different patients (gender, age, main diagnosis, etc.), as well as the vectorized representation of the chief complaint, current medical history, past medical history, allergy history, etc. in the outpatient medical record.

[0052] Specifically, all outpatients in the historical data are divided into m groups according to gender, age, and disease diagnosis (for example, male, 20-29 years old, and patients with a primary diagnosis of hypertension stage I are considered as one group), and k different outpatient medical records are selected from each group (these medical records are sorted according to the number of historical usages 1, 2, ..., k), and each outpatient medical record includes the chief complaint, current medical history, past medical history, allergy history, etc. These outpatient medical records constitute the standard outpatient medical record template library H.

[0053] Assume that each outpatient medical record contains h items, for example, item 1 is the chief complaint, item 2 is the current medical history, item 3 is the past medical history...

[0054] The definition of “different outpatient medical records” is that the similarity between every two outpatient medical record texts calculated using the Levenshtein distance is less than a certain threshold, such as 0.4.

[0055] For the outpatient medical record text content in the standard outpatient medical record template library H, word segmentation technology (for example, Jieba word segmentation) and word vector technology (for example, word2vec, BERT, etc.) are used to represent the words in the text content such as chief complaint, current medical history, past history, allergy history, etc. with word vectors, thereby completing the vectorized representation of the text content such as chief complaint, current medical history, past history, allergy history, etc.

[0056] As one or more implementations, in step S02, for each patient u, a patient feature vector Tu is constructed, and each dimension in the vector includes the content of the patient feature. For example, a patient A, male, 29 years old, mainly diagnosed with hypertension grade I, symptoms of dizziness for 3 days, pulse of 70 beats / min, blood pressure of 110 / 80..., then the patient feature vector is

[0057] Tu=[A, male, 29 years old, hypertension stage I, dizziness for 3 days, 70 beats / min, 110 / 80,...]

[0058] As one or more implementation modes, in step S03, for each of the m groups in step S01, a corresponding deep learning model group is constructed (for the i-th group, its corresponding deep learning model group is called Mi).

[0059] The specific steps to build Mi are as follows:

[0060] Step S301: Establish training set, validation set and test set;

[0061] For the k patients in the i-th group in the standard outpatient medical record template library H, the vectors of 60%, 20%, and 20% of the outpatient medical record text information of the patients are respectively divided into the training set, the validation set, and the test set.

[0062] Step S302: Structure of deep learning model group Mi:

[0063] In the standard outpatient medical record template library H, the patient's outpatient medical record text contains h items of content, for example, item 1 is the chief complaint, item 2 is the current medical history, item 3 is the past medical history...

[0064] The deep learning model group Mi includes h-1 deep learning models, which are denoted as M(i,1), M(i,2), ..., M(i,h-1). Among them, the input of the j-th model M(i,j) is the text information vector of the j-th content in the h-th content of the outpatient medical record text, and the output is the recommendation of the j+1-th content in the h-th content of the outpatient medical record text. The recommendation is selected and recommended from the j+1-th content of the outpatient medical record text of all patients in the i-th group of the outpatient medical record text in the standard outpatient medical record template library H (denoted as B(j+1), which contains k options), that is, recommendation is made from k options.

[0065] The structure of the deep learning model M(i,j) (e.g. Figure 2 As shown) as follows:

[0066] The text information vector of the jth item in the h item of the outpatient medical record text (which may be multiple word vectors, denoted as aj) is input into the model M(i,j). aj first enters the bidirectional long short-term memory recurrent neural network model, and the vectors output by each word vector are subjected to the "max pooling technology" to obtain bj. The kernel vector output by the bidirectional long short-term memory recurrent neural network model is denoted as cj, and bj and cj are combined to form Lj, that is, Lj = [bj, cj].

[0067] The vector Lj is mapped to the exponential vector ρ representing the similarity between aj and all k j+1th items in B(j+1) by using the logistic regression method, that is, Where c1 and c2 are parameter tensors in the logistic regression relationship, and the training in step S03 below is required to determine the specific values. The dimension of ρ is the number of all possible j+1 items in B(j+1) (i.e., k), and the mth element ρ(m) in the vector ρ represents the probability of selecting the mth item in B(j+1). The cross entropy of the 0-1 probability distribution q of the j+1th item result corresponding to ρ and aj in the sample library H is used as the loss function of the model M(i,j).

[0068] Step S303: training to obtain the parameters of model M(i,j):

[0069] The information vector of the text content of the outpatient medical record of each patient in the i-th group of the training set established in step S301 is trained through the entire process in step S302 with the minimization of the loss function as the goal. Training all the information vectors of each patient in the training set once according to the process of step S302 is called training once on the training set.

[0070] The entire training process is:

[0071] Step S30301: After training on the training set for every 100 times, save the obtained model once;

[0072] Step S30302: After saving the model once, use the model to perform a validation on the validation set, obtain the sum of the loss functions of all patients in the validation set and save it;

[0073] Step S30303: After 100 verifications, select the parameters corresponding to the model with the smallest sum of verification loss functions as the final parameters of model M(i,j).

[0074] As one or more implementation modes, in step S04, based on the deep learning model group Mi, the doctor is assisted to quickly complete the writing of the outpatient medical record content; the specific steps are:

[0075] Step S401: for a patient u, according to the method in step S01, group them according to their gender, age, and diagnosis, and assume that they belong to the i-th group. The chief complaint content in the first medical record content in the i-th group in the template library H is recommended to the doctor. The doctor directly adopts or adopts it after modification based on the recommended content, confirms the content of the chief complaint, outputs the content of the chief complaint, and records the content of the chief complaint as z1;

[0076] Step S402: Update the feature vector Tu of patient u (see step S05 for specific methods). Input the chief complaint z1 confirmed in step S401 into the model M(i,1). Model M(i,1) outputs the recommended current medical history, and the content of the recommended current medical history is recorded as o(i,1). Based on Tu, update o(i,1) (see step S06 for specific methods), and recommend the updated o(i,1) to the doctor. Based on o(i,1), the doctor directly adopts or adopts it after modification, confirms the content of the current medical history, and records the current medical history as z2.

[0077] Step S403: Repeat the method in step S402 (ie, replace z1 with z2 to obtain o(i, 2) and z3, and so on) until the assistant doctor completes the complete outpatient medical record content.

[0078] As one or more implementation modes, in step S05, the patient feature vector Tu is updated. Taking the operation in step S402 as an example, the specific steps are:

[0079] Step S501: For the structured features contained in Tu, find the corresponding structured data in the hospital information system, and update the data into the feature vector Tu.

[0080] Step S502: For the unstructured features contained in Tu, z1 (i.e., the main complaint) is split into small sentences, that is, it is divided according to colons, commas, semicolons, and periods. Each small segment is called a small sentence. These small sentences are added to the set W in the order in which they appear.

[0081] Step S503: Starting from the first sentence in W, check each sentence in W whether it contains negative characters or words in the knowledge graph N, and exclude the sentences containing these negative characters or words from W.

[0082] Specifically, knowledge graph N: negative words and vocabulary; in the description of text data, there are many descriptions with negative meanings, for example, "no abdominal pain", where "no" is a negative word. Knowledge graph N records the negative words and vocabulary in the medical record text.

[0083] Step S504: For each sentence in W, check whether it contains the feature descriptions corresponding to all the unstructured features in the knowledge graph Q. If there are corresponding feature descriptions, extract these feature descriptions, and then update these feature descriptions to the features recorded in the feature vector Tu. For example, if a sentence in W is "fever for 3 days" (this description belongs to a feature description of the unstructured feature "symptoms" and the feature name is "fever"), update the content of the feature "symptoms" in Tu to "fever for 3 days".

[0084] Specifically, the knowledge graph Q: unstructured feature -> feature name -> related data -> feature description. A certain unstructured feature contains multiple feature names. For each feature name, the related data refers to the text part that describes the feature, and the feature description is the various descriptions of the feature in the related data. For example, the feature "symptoms" includes feature names such as "fever" and "vomiting". The related data of the feature name "fever" is the chief complaint or current medical history of the outpatient medical record, and its feature description includes "fever 38 degrees", "fever", "high fever does not subside", "fever for 3 days", etc.

[0085] As one or more implementation modes, in step S06, the recommendation expression is updated based on the patient feature vector Tu. Taking the updating of o(i,1) in step S402 as an example, the specific steps are as follows:

[0086] Step S601: divide o(i,1) into small sentences, that is, divide it according to colon, comma, semicolon, and period, each small segment is called a small sentence, and add these small sentences in the order they appear in sequence to the set X. Starting from the first small sentence in X, for each small sentence in X, check whether it contains negative words or phrases in the knowledge graph N, and exclude the small sentences containing these negative words or phrases from X.

[0087] Step S602: Use the knowledge graph Q to find the unstructured features that may be described in o(i,1), match all feature descriptions contained in the knowledge graph Q with the clauses in the set X, and find the features described in o(i,1) and the descriptions of these features. For example, if o(i,1) may contain feature descriptions corresponding to the feature names in the unstructured feature "symptoms", then match all feature descriptions of all feature names in the feature "symptoms" in the knowledge graph Q with the clauses in the set X (for example, the set X contains the clause "fever") to find the feature names described in o(i,1) and the descriptions of these features (for example, the matched feature name is "fever" and the description is "fever").

[0088] Step S603: If Tu contains the feature found in step S602, update the feature description in Tu to o(i,1). If the description of the feature "symptom" in Tu is "fever for 3 days", and in step S602, o(i,1) contains the description of the feature "symptom" "fever", then update the description "fever" in o(i,1) to "fever for 3 days".

[0089] This embodiment uses deep learning models and knowledge graphs to analyze the patient's condition, quickly complete the writing of high-quality medical records, improve the speed and quality of medical record writing, reduce the workload of doctors in writing medical records, and improve the efficiency of outpatient diagnosis.

[0090] Embodiment 2

[0091] Embodiment 2 of the present disclosure introduces an intelligent outpatient medical record writing system based on deep learning.

[0092] like Figure 3 The intelligent outpatient medical record writing system based on deep learning shown in the figure includes:

[0093] An acquisition module is configured to acquire the patient's condition information and construct a patient feature vector;

[0094] The writing module is configured to obtain a recommendation statement based on the patient feature vector and the deep learning model group to assist doctors in writing outpatient medical records;

[0095] Among them, the deep learning model group is composed of multiple deep learning models, and each deep learning model corresponds to each element in the patient feature vector.

[0096] The detailed steps are the same as the deep learning-based intelligent writing method for outpatient medical records provided in Example 1, and will not be repeated here.

[0097] Embodiment 3

[0098] A third embodiment of the present disclosure provides a computer-readable storage medium.

[0099] A computer-readable storage medium stores a program thereon, which, when executed by a processor, implements the steps in the outpatient medical record intelligent writing method based on deep learning as described in the first embodiment of the present disclosure.

[0100] The detailed steps are the same as the deep learning-based intelligent writing method for outpatient medical records provided in Example 1, and will not be repeated here.

[0101] Embodiment 4

[0102] A fourth embodiment of the present disclosure provides an electronic device.

[0103] An electronic device comprises a memory, a processor and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the method for intelligently writing outpatient medical records based on deep learning as described in the first embodiment of the present disclosure are implemented.

[0104] The detailed steps are the same as the deep learning-based intelligent writing method for outpatient medical records provided in Example 1, and will not be repeated here.

[0105] The above description is only a preferred embodiment of the present disclosure and is not intended to limit the present disclosure. For those skilled in the art, the present disclosure may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A method for intelligently writing outpatient medical records based on deep learning, characterized in that: The following steps are involved: Obtain the patient's condition information and construct a patient feature vector; each dimension in the vector includes the content of the patient's characteristics; According to the patient feature vector and the deep learning model group, a recommended statement is obtained to assist doctors in writing outpatient medical records; in the process of assisting doctors in writing outpatient medical records, the patient feature vector is continuously updated, and the elements in the updated patient feature vector are matched with each deep learning model to obtain a recommended statement; the updating process of the patient feature vector is: For structured features, find the structured data corresponding to the structured features in combination with the hospital information system and update the patient feature vector; For unstructured features, the feature names are split in combination with the outpatient medical records of the patient confirmed by the doctor, and the feature descriptions corresponding to all unstructured features in the knowledge graph Q are extracted from the split feature names; wherein the knowledge graph Q is unstructured feature->feature name->related data->feature description, and a certain unstructured feature contains multiple feature names. For each feature name, the related data refers to the text part describing the feature, and the feature description is the various descriptions of the feature in the related data; Among them, the deep learning model group is composed of multiple deep learning models, and each deep learning model corresponds to each element in the patient feature vector.

2. The method for intelligently writing outpatient medical records based on deep learning as described in claim 1, characterized in that: The standard outpatient medical record template library divides all outpatient patients in the historical data into m groups according to gender, age and disease diagnosis, and selects k different outpatient medical records from each group; each outpatient medical record includes the basic information of different patients as well as the chief complaint, current medical history, past medical history and allergy history in the outpatient medical record.

3. The method for intelligently writing outpatient medical records based on deep learning as described in claim 2, characterized in that: The patient feature vector is a row vector containing patient condition information.

4. The method for intelligently writing outpatient medical records based on deep learning as described in claim 3, characterized in that: The construction process of the deep learning model group is as follows: Construct training sets, validation sets, and test sets based on patient medical records in the standard outpatient medical record template library; A deep learning model was built in combination with the standard outpatient medical record template library to obtain a deep learning model group of patient feature vectors.

5. The method for intelligently writing outpatient medical records based on deep learning as described in claim 4, characterized in that: The training process of the deep learning model group is: Collect the text content of each patient's outpatient medical record and convert it into vector data as a sample set; A deep learning model group based on the bidirectional LSTM model was established. The patient feature vector was input and the patient's outpatient medical record text content information vector was output. The goal was to minimize the loss function. All the information vectors of each patient in the training set were input into the bidirectional LSTM model for training, and the parameters of the bidirectional LSTM model were determined.

6. The outpatient medical record intelligent writing system based on deep learning is characterized by: include: An acquisition module is configured to acquire the patient's condition information and construct a patient feature vector; Each dimension in this vector includes the content of patient characteristics; The writing module is configured to obtain a recommended statement based on the patient feature vector and the deep learning model group to assist the doctor in writing the outpatient medical record; in the process of assisting the doctor in writing the outpatient medical record, the patient feature vector is continuously updated, and the elements in the updated patient feature vector are matched with each deep learning model to obtain the recommended statement; the updating process of the patient feature vector is: For structured features, find the structured data corresponding to the structured features in combination with the hospital information system and update the patient feature vector; For unstructured features, the feature names are split in combination with the outpatient medical records of the patient confirmed by the doctor, and the feature descriptions corresponding to all unstructured features in the knowledge graph Q are extracted from the split feature names; wherein the knowledge graph Q is unstructured feature->feature name->related data->feature description, and a certain unstructured feature contains multiple feature names. For each feature name, the related data refers to the text part describing the feature, and the feature description is the various descriptions of the feature in the related data; Among them, the deep learning model group is composed of multiple deep learning models, and each deep learning model corresponds to each element in the patient feature vector.

7. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the steps in the outpatient medical record intelligent writing method based on deep learning as described in any one of claims 1 to 5 are implemented.

8. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the outpatient medical record intelligent writing method based on deep learning as described in any one of claims 1-5 are implemented.

Images