CN109065174B - Medical record theme acquisition method and device considering similarity constraint - Google Patents

Abstract

The present invention provides a medical record subject acquisition method and apparatus that considers similar constraints. The method includes: calculating the similarity between any two medical record documents in the initial medical record, and obtaining a similarity-constrained medical record set composed of a plurality of medical record documents whose similarity is greater than or equal to a similarity threshold; Each medical record document in the medical record collection is sequentially input into a preset LDA model, and the document-topic distribution and the topic-word distribution of each medical record document are derived through the preset LDA model. It can be seen that, in this embodiment, the thinking process of determining the medical record text can be well simulated by the doctor's diagnosis and treatment process, which is beneficial to obtain the accuracy of the subject.

Description

Method and device for obtaining medical record subject considering similar constraints

technical field

The invention relates to the technical field of data mining, and in particular, to a method and device for obtaining medical record subjects considering similar constraints.

Background technique

At present, topic models are mostly used in the evolution analysis of online public opinion topics in the field of online social media, which is beneficial to effectively monitor the changes of online public opinion according to the distribution of network topics in different time periods, and even actively guide its development direction. In addition, the topic model also has a small number of applications in the field of clinical diagnosis and treatment. The purpose is to analyze the diagnosis and treatment rules between disease-medication and disease-symptom in medical record documents. The analysis process includes: inputting each medical record document as an independent sample. A large number of training, to get the final topic analysis results.

However, in the process of realizing the solution of the present invention, the inventors found that: on the one hand, due to the similarity in the development of the disease between two patients with the same disease, the diagnosis plan made by a doctor will be affected by the diagnosis of similar patients who have previously treated them. impact of the program. On the other hand, there are individual differences between two patients, such as constitution, gender, age, disease stage, etc., so doctors will issue different diagnosis and treatment plans according to different patients. In the actual diagnosis and treatment process, there may be two patients with similar physical conditions and diseases, and their diagnosis and treatment plans also have similar parts. For example, diabetic patients will suffer from multiple diabetic complications at the same time, but the diagnosis and treatment plan and disease development of the same complications should be similar.

SUMMARY OF THE INVENTION

Aiming at the deficiencies in the prior art, the present invention provides a method and device for obtaining a medical record subject that considers similar constraints, so as to solve the technical problems existing in the related art.

In a first aspect, an embodiment of the present invention provides a method for obtaining a medical record subject that considers similar constraints, and the method includes:

Calculate the similarity between any two medical record documents in the initial medical record, and obtain a similarity-constrained medical record set composed of multiple medical record documents whose similarity is greater than or equal to the similarity threshold;

Each medical record document in the similarity-constrained medical record collection is input into a preset LDA model in turn, and the document-topic distribution and the topic-word distribution of each medical record document are derived by the preset LDA model.

Optionally, calculating the similarity between any two medical record documents in the initial medical record includes:

Obtain multiple similarity calculation factors of medical records and the weight values of each similarity calculation factor;

Calculate the numerical value of each similarity calculation factor for any two medical record documents respectively;

The similarity of any two medical record documents is calculated according to the numerical value of each similarity calculation factor and the weight value of each similarity calculation factor.

Optionally, the similarity calculation factors include: distances of gender attributes, distances of segments to which age belongs, and distances of diagnosis results.

Optionally, deriving the document-topic distribution and topic-word distribution of each medical record document by using the preset LDA model includes:

Randomly assign subject number z to each word in each medical record document in the similarity-constrained medical record collection;

重新采样主题，使得到的新主题满足GibbsSampling收敛；Rescan the set of similarity constrained medical records, for each word according to

Resampling topics so that the new topics obtained satisfy GibbsSampling convergence;

Count the topic-word co-occurrence frequency matrix in the corpus to obtain document-topic distribution and topic-word distribution.

Optionally, the preset LDA model includes:

The similarity constraint of any two medical record documents is represented by the topic distribution distance dis(θr ^m , θr ⁿ ), and the formula is:

where θr ^m = {θ _m,1 ,θ _m,2 ,...,θ _m,Lm }, indicating that each medical record file includes L _m disease course records; θ _{m, Lm} indicates the subject of the L _mth course record; d (θm _,Lm ,θn _,Ln ) is expressed as the Euclidean distance between the subject vectors of the two disease courses;

The preset LDA model also includes a Gibbs-EM iteration function, which is:

代表相似性约束病历集合中主题为k的单词i的数量。

represents the number of words i with subject k in the similarity-constrained medical record collection.

In a second aspect, an embodiment of the present invention provides an apparatus for obtaining a medical record subject that considers similar constraints, and the apparatus includes:

a medical record collection acquisition module, configured to calculate the similarity between any two medical record documents in the initial medical record, and obtain a similarity-constrained medical record collection composed of a plurality of medical record documents whose similarity is greater than or equal to a similarity threshold;

The topic distribution derivation module is used for sequentially inputting each medical record document in the similarity constrained medical record collection into a preset LDA model, and deriving the document-topic distribution and topic-word distribution of each medical record document through the preset LDA model.

Optionally, the medical record collection acquisition module includes:

a weight value obtaining unit, used for obtaining a plurality of similarity calculation factors of the medical record and the weight value of each similarity calculation factor;

The factor data calculation unit is used to calculate the numerical value of each similarity calculation factor for any two medical record documents respectively;

The similarity calculation unit is configured to calculate the similarity of any two medical record documents according to the numerical value of each similarity calculation factor and the weight value of each similarity calculation factor.

Optionally, the similarity calculation factors include: distances of gender attributes, distances of segments to which age belongs, and distances of diagnosis results.

Optionally, the topic distribution derivation module includes:

a subject numbering unit, for randomly assigning subject number z to each word in each medical record document in the similarity-constrained medical record collection;

重新采样主题，使得到的新主题满足GibbsSampling收敛；A topic iteration unit for rescanning the similarity-constrained medical record collection, for each word according to

Resampling topics so that the new topics obtained satisfy GibbsSampling convergence;

The topic distribution computing unit is used to count the topic-word co-occurrence frequency matrix in the corpus to obtain document-topic distribution and topic-word distribution.

Optionally, the preset LDA model includes:

The similarity constraint of any two medical record documents is represented by the topic distribution distance dis(θr ^m , θr ⁿ ), and the formula is:

where θr ^m = {θ _m,1 ,θ _m,2 ,...,θ _m,Lm }, indicating that each medical record file includes L _m disease course records; θ _{m, Lm} indicates the subject of the L _mth course record; d (θm _,Lm ,θn _,Ln ) is expressed as the Euclidean distance between the subject vectors of the two disease courses;

The preset LDA model also includes a Gibbs-EM iteration function, which is:

代表相似性约束病历集合中主题为k的单词i的数量。

represents the number of words i with subject k in the similarity-constrained medical record collection.

It can be seen from the above technical solutions that in the embodiment of the present invention, by calculating the similarity of two medical record documents, a plurality of medical record documents greater than or equal to the similarity threshold can be selected from the initial medical record, and the subsequent process is composed of multiple medical record documents. Similarity constrained medical record collections as subject analysis documents. It can be seen that, in this embodiment, the thinking process of determining the medical record text can be well simulated by the doctor's diagnosis and treatment process, which is beneficial to obtain the accuracy of the subject.

Description of drawings

In order to explain the embodiments of the present invention or the technical solutions in the prior art more clearly, the following briefly introduces the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only These are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained from these drawings without creative efforts.

1 is a schematic flowchart of a method for obtaining a medical record subject that considers similar constraints according to an embodiment of the present invention;

Fig. 2 is the disease course record in the medical record file;

Figure 3 shows the distribution of the number of diabetic complications in male patients;

Figure 4 shows the distribution of the number of diabetic complications in female patients;

5 is a schematic diagram of the relationship between the number of topics and the similarity constraint indication SIM when the similarity thresholds are 0.5 and 0.6 respectively;

6 is a schematic diagram of the relationship between the number of topics and the similarity constraint indication SIM when the similarity thresholds are 0.7 and 0.8 respectively;

Figure 7 shows the relationship between the number of topics and the interaction information;

8 to 10 are block diagrams of an apparatus for obtaining a medical record subject that considers similar constraints according to an embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

FIG. 1 is a schematic flowchart of a method for obtaining a medical record subject that considers similar constraints according to an embodiment of the present invention. Referring to Figure 1, a medical record subject acquisition method that considers similar constraints includes:

101. Calculate the similarity between any two medical record documents in the initial medical record, and obtain a similarity-constrained medical record set composed of multiple medical record documents whose similarity is greater than or equal to a similarity threshold;

102. Input each medical record document in the similarity-constrained medical record collection into a preset LDA model in turn, and derive document-topic distribution and topic-word distribution of each medical record document by using the preset LDA model.

The steps of the method for obtaining the subject of medical records considering similar constraints will be described in detail below with reference to the accompanying drawings and embodiments.

First, 101 is introduced, the steps of calculating the similarity between any two medical record documents in the initial medical record, and obtaining a similarity-constrained medical record set composed of multiple medical record documents whose similarity is greater than or equal to the similarity threshold.

In the process of hospitalization, patients will generate various test records, such as admission records, discharge records, disease course records, and consultation records. If the similarity between detection records is directly calculated, the calculation amount will be greatly increased. For convenience of description, in this embodiment, the detection record before processing is referred to as the initial medical record.

In order to reduce the amount of calculation, only the similarity of the admission diagnosis part in the initial medical record is considered in this embodiment. In one embodiment, the similarity is to calculate the distance between any two initial medical records, and the construction of the medical record similarity constraint can be understood as collecting a medical record set whose distance between the two is less than a certain threshold.

In practical applications, the initial medical record will also include multiple complications of a certain condition, such as diabetes, which can lead to multiple complications, as shown in Table 1.

Table 1 Examples of complications in patients with diabetes

Analysis of Table 1 shows that patients of different age groups have differences in the characteristics of diabetes and its complications; in addition, patients of different age groups have different tolerance to drugs, which leads to differences in clinical diagnosis and treatment. Therefore, the basic information of the patient needs to be considered when calculating the similarity of the medical record documents. In this embodiment, the patient's name and age are included in the similarity calculation factor of the medical record document.

In one embodiment, the distance of gender attributes between the same gender is set to 1, and the distance of gender attributes between different genders is set to 0, as shown in the following formula:

Among them, sex _i and sex _j represent the gender of two different people.

In one embodiment, age is divided into 4 age groups according to the international population age structure, which are: teenagers, 0-17 years old, denoted as 1; youth, 18-45 years old, denoted as 2; middle-aged, 18-45 years old, denoted as 2; 45 years old, expressed as 3; old age, more than 59 years old, expressed as 4. In this way, the present embodiment can calculate the distance between the age groups to which two patients belong, as expressed by the following formula:

Among them, age _i and age _j represent the ages of two different people, and flag _i and flag _j represent the segments to which different ages belong. In addition, the closer the segments to which the two ages belong, the smaller the distance, and the farther the segments belong, the larger the distance.

Considering that the discrete textual description is used in the initial medical records, in this embodiment, the Jaccard distance is used to calculate the distance between the diagnosis results in different initial medical records, as shown in the following formula:

Among them, dia _i , dia _j represent the Boolean vector space of discharge diagnosis of medical record i and medical record j, and a large number of diseases between diabetes complications are considered in this paper.

For example: dia _i ={123}, dia _j ={234}, dia _i ∩dia _j ={2,3}; dia _i ∪dia _j ={1,2,3,4}, then d(dia _i , dia _j )=2/4=0.5.

It should be noted that in this embodiment, only the similarity calculation factors are considered including: the distance of the gender attribute, the distance of the age to which the segment belongs, and the distance of the diagnosis result. When the application scene of the text topic acquisition method changes, The specific composition of the similarity calculation factor can also be adjusted accordingly, and the adjusted solution also falls within the protection scope of the present application.

After determining the similarity calculation factors, set the weight adjustment adjustment parameters μ ₁ , μ ₂ , μ ₃ respectively, and calculate the similarity between any two initial medical records, as shown in the following formula:

sim(T _i ,T _j )=μ ₁ *d(sex _i ,sex _j )+μ ₂ *d(age _i ,age _j )+μ ₃ *d(dia _i ,dia _j )

(3)

μ ₁ +μ ₂ +μ ₃ =1 (4)

0≤μ ₁ ,μ ₂ ,μ ₃ ≤1 (5)

Finally, compare the similarity with the similarity threshold τ, screen out multiple initial medical records whose similarity is greater than or equal to the similarity threshold, and obtain a similarity-constrained medical record set composed of multiple initial medical records, denoted as D={( T _i ,T _j )|i,j∈[1,M]}.

Next, introduce 102, the steps of sequentially inputting each medical record document in the similarity constrained medical record collection into a preset LDA model, and deriving the document-topic distribution and topic-word distribution of each medical record document through the preset LDA model.

In this embodiment, the preset LDA model is improved on the basis of the existing LDA model. In order to facilitate the technical person to better understand the preset LDA model, first describe the basic principle of the LDA model:

Latent Dirichlet Allocation (LDA) is a topic model whose purpose is to find document topics, including a three-layer structure of documents, topics and words, and each document has a probability distribution related to its own topic, while the document Medium words are sampled from different topic distributions, as shown in Equation (6) below:

∑p(word|document)=∑p(word|topic)*p(topic|document) (6)

The LDA model is used to model the medical record documents. The total number of medical record documents M is set. There are Nm clinical description words in the mth medical record document. Each word is represented as ω _m,n . According to the existing bag of words model (bag of words) ) represent documents and words as document-topic distribution and topic-word distribution. In the medical record text, the subject can be understood as a general term for clinical care methods such as medication, observation, symptoms, surgery, etc. Each medical record text is a multinomial distribution of multiple topics, that is, each medical record text is a combination of multiple steps in the clinical nursing process. made.

In the related art, the steps of generating medical record text by the LDA model are shown in Table 2.

Understandably, since each topic is a multinomial distribution of multiple words, each clinical care step contains multiple clinical practices, and both the document-topic distribution and the topic-word distribution conform to the Dirichlet parameters α and β. Prior distribution, so the LDA model can well simulate the thinking process of doctors making medical record texts in the process of diagnosis and treatment.

并根据

计算主题-单词分布和文档-主题分布。实际上，计算过程中可以直接推导出主题-单词分布和文档-主题分布，而无需计算

Based on the above analysis, the purpose of the LDA model reasoning is to calculate the unknown parameters in the LDA model through the current test document set

and according to

Compute topic-word distribution and document-topic distribution. In fact, the topic-word distribution and document-topic distribution can be directly derived during the calculation process without computing

In practical application, the parameter inference algorithm of LDA model includes two kinds of Gibbs sampling and EM variation. Two methods are described below.

First, the core idea of Gibbs Sampling is the Markov Monte Carlo (MCMC) method, which only changes the parameter value of one dimension in each iteration process until it converges and outputs the parameter value to be estimated. According to the Dirichlet parameter estimation, the inference can be obtained:

表示文档-主题分布，

表示主题-单词分布，

表示单词

分布为k的概率，i为一个数据对(m,n),表示第m个文档中的第n个词。in:

represents the document-topic distribution,

represents the topic-word distribution,

represent words

The probability that the distribution is k, i is a data pair (m, n), representing the nth word in the mth document.

Since there are K topics in total, K iterations are required, and the training steps are shown in Table 3:

Second, the EM variational algorithm is to find suitable parameters to maximize the probability of the topic-word distribution observed in the text set, which is similar to the maximum likelihood estimation problem. The EM variational algorithm is divided into two iterative steps:

)求得近似后验概率分布q(θ,z|γ,

)。The variational E-step considers the difficulty in derivation of the posterior probability p(w|α,β) formula in the original step, and introduces variational parameters (γ,

) to obtain the approximate posterior probability distribution q(θ,z|γ,

).

β)。其中，先验狄利克雷分布参数(α,β)决定了主题-单词分布和文档-主题分布θ，w代表单词，z代表主题。The variational M-step maximizes the approximate function L(γ,

β). Among them, the prior Dirichlet distribution parameters (α, β) determine the topic-word distribution and document-topic distribution θ, w represents words, and z represents topics.

Since the iterative goal of the LDA model is to maximize the probability of word occurrence p(Z,W|α,β), this can effectively meet the data characteristics of diabetes course records, and at the same time, it will also lead to large differences in the subject distribution of similar medical records, resulting in Valid statistical analysis of medical records based on the subject distribution of medical records is not possible.

In order to establish a topic model that satisfies the constraint of similarity of medical records, this goal is achieved by changing the Gibbs sampling convergence condition strategy in this embodiment.

Considering that there will be multiple time-ordered medical records in each medical record at the same time, the similarity calculation of medical record documents should consider the similarity between different medical record sets in each medical record document, that is, the similarity constrained medical record set D. The document-topic distributions of the different course record collections were as similar as possible.

Let T _m represent the medical records numbered m, including L _m disease course records, and the subject set of the disease course records is represented as ^θrm ={θm _,1 ,θm _,2 ,...,θm _,Lm }. There are two medical record document subject sets θγ ^m , ^θrn , and the medical record similarity constraint can be calculated by using the mean distribution distance between the two subjects, as follows:

Among them, d(θm _,Lm ,θn _,Ln ) is expressed as the Euclidean distance between two disease and vectors, and the larger dis( ^θrm , ^θrn ) is, the lower the similarity is.

The maximum objective function can be modified as:

In this embodiment, the Gibbs-EM iterative method is used to deduce the LDA model, which modifies the document-topic distribution α _m to the normal distribution μ _m to obtain the preset LDA model:

Among them, μ _mk represents the probability that the medical record document m belongs to the topic k. Since μ _m is considered to obey the standard normal distribution, the improved maximum objective function is expressed as follows:

In addition, in this embodiment, the document topic distribution α _m is fixed first in the sampling process, then the Gibbs-EM iteration function expression is:

代表相似性约束病历集合中主题为k的单词i的数量，由于采用正态分布来代替原来的α，所以公式(14)可以用随机梯度下降方法进行推导，模型训练过程如表4：in,

Represents the number of words i with subject k in the similarity-constrained medical record set. Since the normal distribution is used to replace the original α, the formula (14) can be derived by the stochastic gradient descent method. The model training process is shown in Table 4:

After that, each medical record document in the similarity-constrained medical record collection is sequentially input into a preset LDA model, and the document-topic distribution and the topic-word distribution of each medical record document are deduced by the preset LDA model.

So far, on the basis of analyzing the influence of text mining on medical diagnosis and the modeling process and reasoning method of the underlying Dirichlet topic model, a preset LDA model based on similarity constraints of medical records is designed in the embodiment of the present invention. The preset LDA model not only considers the similarity constraints between different medical record documents, but also determines the medical text topic modeling goals, reasoning process and model-related metrics, so that the LDA model can clearly reflect the characteristics of each stage of diagnosis and treatment. The focus and the evolution process of the disease are conducive to improving the scientificity, effectiveness and accuracy of the subject mining of medical records.

The following adopts the LDA model and the preset LDA model of the present application (hereinafter referred to as Medical Record Similaritybased Latent Dirichlet Allocation, MRS-LDA) to compare experiments to illustrate the effectiveness of a method for obtaining a medical record subject provided by the embodiment of the present invention that considers similarity constraints and superiority.

Among them, the initial medical records adopted the inpatient medical records of patients in the Department of Endocrinology, The First Affiliated Hospital of Anhui Medical University, including a total of 1294 inpatient records of diabetic patients from 2015 to 2017. display), consultation records and discharge records, etc. Among them, the ratio of the number of medical records of male and female patients was 648:646, which was roughly the same.

Referring to Figure 3 and Figure 4, among the diabetic patients admitted to the First Affiliated Hospital of Anhui Medical University, the number of complications suffered by patients of different ages and genders was significantly different according to the admission diagnosis. Among them, the number of diabetic complications that the elderly suffer from at the same time is much higher than that of other age groups. Middle-aged people mostly suffer from 3 to 5 complications at the same time. Young people will have diabetes, but there are no more complications. There are fewer children with diabetes.

In this embodiment, the gender, age, and admission diagnosis of the patient in the admission record are selected as the data basis for the similarity constraint calculation of the medical record, and the relevant subject analysis is performed by using the medical record of the doctor during the hospitalization of the patient. During the experiment, you can also do the following processing, including:

(1) Using the python crawler method, segment the text records of admission records, discharge records, course records, etc. from the medical record documents in HTML format of 1294 patients, and separate the required patient information, diagnosis results, and course record texts at the same time.

(2) Build a dictionary and stop word thesaurus. The research content of the present invention is some words related to medicine, such as symptoms, medicines, and treatment methods. In addition, the medical record text contains a large number of words that are irrelevant to this article. After counting the frequency of each word in the medical record, manually 12599 words were extracted as stop words and added to the stop word bank. At the same time, the disease name of ICD10 China was added as a supplementary feature to the dictionary.

(3) Use the jieba word segmentation in python as a word segmentation tool, and use the above dictionary and stop word library to perform word segmentation and stop word removal operations.

Considering that in the topic mining of medical record documents, the influence of the number of topics on the modeling of text topics and the difference in the number of similar medical records caused by different similarity thresholds, in this embodiment, the similarity threshold and the number of topics are used as adjustment parameters, and the similarity threshold of medical records is The value of τ ranges from 0.5 to 0.8, and the number of subjects is K=7, 10, 13, 15, 20, and 30. The PMI-Score and medical record similarity constraints of the model are calculated under the above parameters.

Referring to Figure 5 and Figure 6, the MRS-LDA model and the LDA model compare the similarity constraint results under different topic parameters and different degrees of similarity, where the abscissa is the number of topics K, and the ordinate is the similarity constraint index SIM. Comparative analysis of MRS-LDA model has obvious advantages in medical record similarity constraints. When the topic similarity threshold is the same, with the increase of the number of topics, the medical record similarity constraint has an insignificant decline, but the MRA-LDA model still has a greater advantage in the medical record similarity constraint index than the LDA model.

Referring to Figure 7, under different topic parameters and different similarity thresholds, the results of the MRS-LDA model and the LDA model cross-interaction information (PIM-Score) are compared, where the abscissa is the number of topics K, and the ordinate is the metric PIM-Score . When the number of subjects K=15, the MRS-LDA model outperforms the LDA model on the PIM-Score metric, and outperforms the LDA model when the medical record similarity threshold is 0.5.

Through comparative experiments, the MRS-LDA model has a good performance on the similarity constraint metric. Under the same medical record similarity threshold and number of subjects, the distance between the subject distributions of similar medical records obtained by the MRS-LDA model is smaller. , which can better describe the existing associations between similar medical records. That is to say, in the present invention, the constraint condition of similarity of medical records is added when constructing the objective function, which can make the subject distribution between similar medical records closer, and can be applied to the use scenario of subject mining of medical records, and the accuracy is high.

In a second aspect, an embodiment of the present invention provides an apparatus for obtaining a medical record subject that considers similar constraints. Referring to FIG. 8 , the apparatus includes:

A medical record collection acquisition module 801, configured to calculate the similarity between any two medical record documents in the initial medical record, and obtain a similarity-constrained medical record collection composed of a plurality of medical record documents whose similarity is greater than or equal to a similarity threshold;

The topic distribution derivation module 802 is configured to sequentially input each medical record document in the similarity constrained medical record set into a preset LDA model, and derive the document-topic distribution and topic-word distribution of each medical record document through the preset LDA model.

Optionally, referring to FIG. 9 , the medical record collection acquisition module 801 includes:

a weight value obtaining unit 901, configured to obtain a plurality of similarity calculation factors of the medical record and the weight value of each similarity calculation factor;

A factor data calculation unit 902, configured to respectively calculate the numerical values of any two medical record documents with respect to each similarity calculation factor;

The similarity calculation unit 903 is configured to calculate the similarity of any two medical record documents according to the numerical value of each similarity calculation factor and the weight value of each similarity calculation factor.

Optionally, the similarity calculation factors include: distances of gender attributes, distances of segments to which age belongs, and distances of diagnosis results.

Optionally, referring to FIG. 10 , the topic distribution derivation module 802 includes:

Subject numbering unit 1001, for randomly assigning subject number z to each word in each medical record document in the similarity-constrained medical record collection;

重新采样主题，使得到的新主题满足GibbsSampling收敛；The topic iteration unit 1002 is used for rescanning the similarity-constrained medical record set, and for each word according to

Resampling topics so that the new topics obtained satisfy GibbsSampling convergence;

The topic distribution calculation unit 1003 is configured to count the topic-word co-occurrence frequency matrix in the corpus to obtain document-topic distribution and topic-word distribution.

Optionally, the preset LDA model includes:

The similarity constraint of any two medical record documents is represented by the topic distribution distance dis(θr ^m , θr ⁿ ), and the formula is:

where θr ^m = {θ _m,1 ,θ _m,2 ,...,θ _m,Lm }, indicating that each medical record file includes L _m disease course records; θ _{m, Lm} indicates the subject of the L _mth course record; d (θm _,Lm ,θn _,Ln ) is expressed as the Euclidean distance between the subject vectors of the two disease courses;

The preset LDA model also includes a Gibbs-EM iteration function, which is:

代表相似性约束病历集合中主题为k的单词i的数量。

represents the number of words i with subject k in the similarity-constrained medical record collection.

It should be noted that there is a one-to-one correspondence between the medical record subject acquisition device that considers similar constraints and the above method provided by the embodiment of the present invention, and the implementation details of the above method are also applicable to the above device. Detailed description.

In the description of the present invention, numerous specific details are set forth. It will be understood, however, that embodiments of the invention may be practiced without these specific details. In some instances, well-known methods, structures and techniques have not been shown in detail in order not to obscure an understanding of this description.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, but not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: The technical solutions described in the foregoing embodiments can still be modified, or some or all of the technical features thereof can be equivalently replaced; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the technical solutions of the embodiments of the present invention. The scope of the invention should be included in the scope of the claims and description of the present invention.

Claims (6)

1. A medical record subject acquisition method considering similar constraints, wherein the method comprises: Calculate the similarity between any two medical record documents in the initial medical record, and obtain a similarity-constrained medical record set composed of multiple medical record documents whose similarity is greater than or equal to the similarity threshold; Inputting each medical record document in the similarity-constrained medical record collection into a preset LDA model in turn, and deriving the document-topic distribution and topic-word distribution of each medical record document through the preset LDA model; Deriving the document-topic distribution and topic-word distribution of each medical record document through the preset LDA model includes: Randomly assign subject number z to each word in each medical record document in the similarity-constrained medical record collection; Rescan the set of similarity constrained medical records, for each word according to

Resampling topics so that the new topics obtained satisfy Gibbs Sampling convergence; where,

represent words

the probability that the distribution is k; Count the topic-word co-occurrence frequency matrix in the corpus to obtain document-topic distribution and topic-word distribution; The preset LDA model includes: The similarity constraint of any two medical record documents is represented by the topic distribution distance dis(θr ^m , θr ⁿ ), and the formula is:

where θr ^m = {θ _m,1 ,θ _m,2 ,...,θ _m,Lm }, indicating that each medical record file includes L _m disease course records; θ _{m, Lm} indicates the subject of the L _mth course record; d (θ m _{, Lm} ,θ ⁿ _{, Ln} ) is expressed as the Euclidean distance between the subject vectors of the two disease _courses ; The medical record file includes L _n disease course records; θ _n,Ln represents the subject of the L _nth disease course record; The preset LDA model also includes a Gibbs-EM iteration function, which is:

Represents the number of words i with subject k in the similarity-constrained medical record collection; prior Dirichlet distribution parameters (α, β); the total number of words in the medical record collection V; preset LDA model

_μmk represents the probability that medical record document m belongs to topic k. 2. The medical record subject acquisition method according to claim 1, wherein calculating the similarity between any two medical record documents in the initial medical record comprises: Obtain multiple similarity calculation factors of medical records and the weight values of each similarity calculation factor; Calculate the numerical value of each similarity calculation factor for any two medical record documents respectively; The similarity of any two medical record documents is calculated according to the numerical value of each similarity calculation factor and the weight value of each similarity calculation factor. 3 . The method for obtaining the subject of medical records according to claim 2 , wherein the similarity calculation factors include: distances of gender attributes, distances of segments to which age belongs, and distances of diagnosis results. 4 . 4. A medical record subject acquisition device considering similar constraints, characterized in that the device comprises: A medical record collection acquisition module, configured to calculate the similarity between any two medical record documents in the initial medical record, and obtain a similarity-constrained medical record collection composed of a plurality of medical record documents whose similarity is greater than or equal to a similarity threshold; The topic distribution derivation module is used for sequentially inputting each medical record document in the described similarity constrained medical record collection into a preset LDA model, and deriving the document-topic distribution and topic-word distribution of each medical record document through the preset LDA model; The topic distribution derivation module includes: a subject numbering unit, for randomly assigning subject number z to each word in each medical record document in the similarity-constrained medical record collection; A topic iteration unit for rescanning the similarity-constrained medical record collection, for each word according to

Resampling topics so that the new topics obtained satisfy GibbsSampling convergence; where,

represent words

the probability that the distribution is k; The topic distribution calculation unit is used to count the topic-word co-occurrence frequency matrix in the corpus to obtain document-topic distribution and topic-word distribution; The preset LDA model includes: The similarity constraint of any two medical record documents is represented by the topic distribution distance dis(θr ^m , θr ⁿ ), and the formula is:

where θr ^m = {θ _m,1 ,θ _m,2 ,...,θ _m,Lm }, indicating that each medical record file includes L _m disease course records; θ _{m, Lm} indicates the subject of the L _mth course record; d (θ m _{, Lm} ,θ ⁿ _{, Ln} ) is expressed as the Euclidean distance between the subject vectors of the two disease _courses ; The medical record file includes L _n disease course records; θ _n,Ln represents the subject of the L _nth disease course record; The preset LDA model also includes a Gibbs-EM iteration function, which is:

Represents the number of words i with subject k in the similarity-constrained medical record collection; prior Dirichlet distribution parameters (α, β); the total number of words in the medical record collection V; preset LDA model

_μmk represents the probability that medical record document m belongs to topic k. 5. The medical record subject acquisition device according to claim 4, wherein the medical record collection acquisition module comprises: a weight value obtaining unit, used for obtaining a plurality of similarity calculation factors of the medical record and the weight value of each similarity calculation factor; The factor data calculation unit is used to calculate the numerical value of each similarity calculation factor for any two medical record documents respectively; The similarity calculation unit is configured to calculate the similarity of any two medical record documents according to the numerical value of each similarity calculation factor and the weight value of each similarity calculation factor. 6 . The method and apparatus for obtaining the subject of medical records according to claim 5 , wherein the similarity calculation factors include: distances of gender attributes, distances of segments to which age belongs, and distances of diagnosis results. 7 .

Images