CN111292818A - Query reconstruction method for electronic medical record description - Google Patents

Abstract

A query reconstruction method for electronic medical record description, which is a query reconstruction method for electronic medical record long text in clinical decision support. Query reconstruction in information retrieval refers to the process of automatically processing the information input by the user to form a new query expression. In the medical field, medical literature retrieval is an important application of clinical decision support. It uses electronic medical record texts as query input to obtain needed information from massive medical texts. However, the description of electronic medical records is very complex, and it is necessary to perform query reconstruction in order to obtain an effective retrieval effect. In this regard, the present invention describes the long text of electronic medical records with redundant information, uses sub-query segmentation and screening and query quality prediction technology, selects the sub-query with the highest query quality to replace the original query, and predicts the query intent to reconstruct the query, thereby improving the performance of the query. retrieval efficiency.

Description

A Query Reconstruction Method for Electronic Medical Record Description

technical field

The present invention relates to the field of text retrieval, in particular to the processing of query in text retrieval.

Background technique

Information retrieval is the process of finding the information needed by users from unstructured large-scale data, and it is an effective method to obtain key information in massive data. As early as the last century, medical experts considered the use of data and models to assist clinical decision-making, and thus proposed a clinical decision support system. The clinical decision support system is a medical information technology application system, including the use of information retrieval technology to predict diagnosis, that is, the patient description in the medical record is used as a query to find relevant medical documents to assist decision-making. Through this method, the clinical decision support system can effectively mine the deep data in medical treatment, improve the efficiency of medical service, and speed up the process of medical informatization.

With the development of medical and health services and the advancement of science and technology, the level of technology and informatization in the medical industry is constantly improving. As a very active application branch of clinical decision support systems, the diagnosis decision support system has been a research and development support system at home and abroad. Application hotspots. A diagnosis decision support system refers to a computer application system that can provide auxiliary support for doctors in the process of diagnosis and decision-making. A large number of studies have shown that the diagnosis decision support system can effectively solve the problem of limited knowledge of clinicians in the process of disease diagnosis, reduce human negligence, relatively reduce medical expenses, and provide guarantee for medical quality.

To better study medical text retrieval techniques, the Text Retrieval Conference TREC proposed the Clinical Decision Support (CDS) task in 2014. In this task, the description of the electronic medical record is given as the input query, and the contestant searches the existing medical document collection to return the most relevant documents to the query, which are provided to the doctor to assist in judging the real needs of the patient. The electronic medical record used as a query is stored in the form of free text. The medical record data comes from the ICU clinical database MIMIC-III, and the target document set as the retrieval library comes from the American biomedical and life science full-text database PubMed Central.

Judging from the existing research results of the TREC CDS task, the main work of document retrieval focuses on the processing of original query sentences. The mainstream method is query expansion based on keywords, and the query processing work for long texts of electronic medical records is very rare. . The electronic medical record text is highly ambiguous, with a lot of redundancy and unclear semantics. Its processing method is undoubtedly a difficult point in the task of medical document retrieval.

SUMMARY OF THE INVENTION

Judging from the existing research results of the TREC CDS task, the main work of document retrieval focuses on the processing of original query sentences. The mainstream method is query expansion based on keywords, and the query processing work for long texts of electronic medical records is very rare. . The length of a common query is several to a dozen query words, while the average length of any electronic medical record description in clinical decision-making is 50-200 query words. Most commercial and academic search engines are not ideal when dealing with such long queries. This means that it leaves the task of reducing the original query to a shorter query to the user. On the other hand, a clear query intent can be targeted for retrieval to improve retrieval efficiency.

Aiming at the above problems, the present invention aims to reconstruct the query statement, adopts the SVM classifier to obtain the query intent of the query statement, generates and filters the sub-queries of the electronic medical record, and then obtains the optimal sub-query by training the query quality prediction model. Subqueries, which are combined with query intents to generate refactored query statements.

In order to achieve the above object, the technical scheme provided by the present invention is:

The present invention provides a query reconstruction method for electronic medical record description, including:

Step 1. Preprocess the electronic medical record text and medical literature text in the data set;

Step 2, train the SVM classifier to predict the query intent of the electronic medical record text;

Step 3. Obtain all sub-queries of the electronic medical record text and perform preliminary pre-screening on them;

Step 4. Train the query quality prediction model, and select the optimal sub-query from the sub-queries output by the pre-screening in step 3;

Step 5: Combine the query intent obtained in Step 2 with the optimal sub-query output in Step 4 to obtain the final reconstructed query.

beneficial effect

The invention performs query reconstruction processing for the long description text of electronic medical records with redundant information, including prediction of query intention and reduction of original query based on query quality prediction technology. The invention analyzes the original query semantics, and trains a classifier to realize the judgment of the query intention. On the other hand, the present invention analyzes all sub-query sets of the original query, expresses each sub-query through a set of query quality indicators, and proposes an index reflecting query expansion performance for the first time, and on this basis, trains a query quality prediction model to obtain query quality The highest subquery replaces the original query.

The present invention conducts a query reconstruction experiment on the TREC CDS data set, and observes a significant improvement in performance, which also confirms that query reconstruction improves retrieval results. The query reconstruction method implemented for the long text of electronic medical records is of great significance for solving the problem of limited knowledge of clinicians in the process of disease diagnosis, reducing human negligence, relatively reducing medical expenses, and providing guarantee for medical quality.

Description of drawings

The accompanying drawings are further descriptions of the present invention and constitute a part of the specification, and are used to explain the present invention together with the following specific embodiments, but do not constitute a limitation of the present invention. In the attached image:

FIG. 1 is a schematic flowchart of a query reconstruction method;

Figure 2 is an example of a TREC query topic;

Figure 3 shows the result of processing electronic medical record text using the query reconstruction.

Detailed ways

The specific implementation process of the present invention, as shown in Figure 1, includes the following 5 steps:

Step 1. Preprocess the electronic medical record text and medical literature text in the data set;

Step 2, train the SVM classifier to predict the query intent of the electronic medical record text;

Step 3. Obtain all sub-queries of the electronic medical record text and perform preliminary pre-screening on them;

Step 4. Train the query quality prediction model, and select the optimal sub-query from the sub-queries output by the pre-screening in step 3;

Step 5: Combine the query intent obtained in Step 2 with the optimal sub-query output in Step 4 to obtain the final reconstructed query.

The individual steps are detailed below.

Step 1: Preprocess the EMR text and medical literature text in the dataset

As an example, the data set used is from the TREC CDS Track data set, including electronic medical record text and medical literature set. The electronic medical record text comes from the 90 query topics (Topics) defined by the TREC conference 2014-2016, and the medical record data comes from the ICU clinical database MIMIC-III. We select the description field in each topic as an original query. Figure 2 shows an example of a query topic (Topic). On the other hand, the medical literature collection is more than 730,000 medical-related documents from the American biomedical and life science text database PubMed Central.

In step 1, the electronic medical record text and medical document text need to be preprocessed, which specifically includes the following steps:

1.4. Extract plain text

This step is not required because the EHR text itself is plain text. The medical literature text is a web page file stored in XML format. It is necessary to remove the useless CSS and JS codes, and extract the required plain text data according to the XML tags, including the title, abstract, keywords and body parts of the literature. The subsequent literature text has a uniform format.

Provide the plain text of the electronic medical record and the extracted plain text of medical documents to step 1.2;

1.5. Remove stop words

Use the preprocessing vocabulary to remove stop words in plain text, including some words that do not contain semantic information, and words that are used too frequently.

After removing stop words, the result is provided to step 1.3;

1.6, restore part of speech

The integration of different parts of speech is restored to the root of the word. Words with the same meaning in English will have different tense changes, and these words will be restored by part of speech.

After restoring the part of speech, the text preprocessing work of step 1 is completed, and the preprocessed electronic medical record text is provided to steps 2 and 3, and the preprocessed medical document text is provided to step 4.

Step 2: Train the SVM three-classifier to predict the query intent of the electronic medical record text

Step 2 uses the preprocessed electronic medical record text obtained in step 1 as a training set to train the SVM classifier to judge the query intent, which specifically includes the following steps.

2.1. Label each electronic medical record text in the training set with three-category labels: if the electronic medical record text content belongs to Diagnosis, it is marked as 1; if the electronic medical record text content belongs to the treatment plan (Treatment), it is marked as 2; The text content belongs to the diagnostic testing method (Test), marked as 3. The annotated results are provided to step 2.2.

2.2. Train three classifiers.

The training of the three classifiers uses the existing support vector machine SVM algorithm, and the characteristics of the electronic medical record text and the three classification labels marked in step 2.1 need to be input during training. The training of the classifier needs to use two features of electronic medical record text: (1) TF-IDF value; (2) semantic information.

(1) TF-IDF is a statistical method to evaluate the importance of a word to one of the documents in a corpus. The term frequency (TF) refers to the frequency with which a given word appears in the document. The inverse document frequency (IDF) is obtained by dividing the total number of documents by the number of documents containing the term, and then taking the logarithm to the base 10 of the obtained quotient. The TF-IDF value is the product of these two values, and the formula is

where _nω represents the number of occurrences of the word _ω in the file, N represents the total number of words in the file, Nd represents the total number of files in the corpus, and _Nω represents the number of files in the corpus that contain the word ω.

(2) Semantic information refers to three parts of information: whether the diagnosis result is included (value 0/1), whether the inspection has been completed (value 0/1), and the length of query text (value 0-200).

Feed the trained three-classifier to step 2.3.

2.3. Input the electronic medical record text into the trained three-classifier, and provide the classification result (ie, query intent) to step 5.

Step 3: Get all subqueries for EMR text and perform initial pre-screening on them

In theory, a query containing n query words can get exponential sub-queries. For example, a query containing 3 query words "fever cough headache" can be split into sub-queries including "fever", "cough", "headache" ", "fever cough", "fever headache", "cough headache", "fever cough headache". It is impractical to exhaustively list all possible subqueries for sorting, so the subqueries need to be pre-filtered first. Subquery prefiltering consists of the following steps.

3.1. Select sub-queries whose length is between 3 and 10 in all sub-queries of the electronic medical record text. Subquery length refers to the number of words in the query. Research shows that the query length for the best retrieval effect is between 3 and 6. Considering the long text of the electronic medical record targeted by the present invention, the maximum length threshold is defined as 10. The results are provided to step 3.2.

3.2. Calculate the average mutual information of each subquery obtained in 3.1, and select 30 subqueries with the highest mutual information. The formula for calculating the average mutual information of a subquery is as follows:

where n(x,y) represents the frequency of word x and word y appearing simultaneously in documents with a window size of 25 in the entire corpus, and n(x) and n(y) represent word x and word y in the corpus, respectively frequency of occurrence. N _c represents the number of words in the entire corpus. Calculate the mutual information of any two words in a subquery, and take their weighted average as the average mutual information of the subquery.

Step 3 finally obtains 30 pre-filtered subqueries, and this result is provided to step 4.

Step 4: Train the query quality prediction model and select the optimal sub-query from the pre-screened sub-queries

4.1. Labeling query quality scores for subqueries

A round of retrieval is performed for each pre-screened subquery obtained in step 3, and the retrieved target document set comes from the pre-processed medical document text set in step 1. The search engine uses Indri5.11 from the Lemur open source project. The retrieval results are compared with the evaluation criteria provided by the TREC conference, and the average accuracy score of retrieval is calculated and marked as the query quality score of the sub-query. Subqueries annotated with query quality scores are provided to step 4.2 as the result of this step.

4.2. Train the query quality prediction model

The training of the query quality prediction model uses the existing SVMRank algorithm, and the indicators that can characterize the quality of the sub-query and the query quality score marked in step 4.1 need to be input during training.

Model training needs to use the following indicators, which are calculated for each subquery in the training set: (1) Inverse document frequency correlation indicator; (2) Simplified query clarity indicator; (3) Corpus/query similarity feature indicator; (4) Query availability Scalability metrics.

Define the meaning of the symbols used in this step before introducing these indicators individually. For a query Q, assuming that it contains query words ω ₁ ,...ω _n , n(ω _i ) in corpus C represents the frequency of query word ω _i in the corpus, and n(ω _i ,ω _j ) represents the query word in the corpus The frequency that ω _i , ω _j (i≠j) simultaneously appear in a window of length 25 words, N _c represents the total number of words contained in the corpus, N _ω represents the number of documents in which the query word ω appeared, and N _d represents the corpus The number of all files in . P _c (ω) represents the probability of occurrence of the query word ω in the corpus, P(ω|Q) represents the probability of the occurrence of ω in the query sentence Q, and S _ω represents the synonym set of the word ω.

(1) The calculation formula of the inverse document frequency correlation index is:

where _Nω is the number of documents containing the word _ω , and Nd is the total number of documents in the corpus. For each sub-query, the sum, maximum value, standard deviation, arithmetic mean, geometric mean and harmonic mean of each query term IDF value are calculated together as query quality indicators.

(2) The calculation formula of the simplified query clarity index is:

where P _ml (ω|Q) is the frequency of the word ω in the query Q, and P _c (ω) is the frequency of the word ω in the corpus.

(3) The calculation formula of the corpus/query similarity feature index is:

Like the inverse document frequency-related metrics, we calculate the sum, maximum, standard deviation, arithmetic mean, geometric mean, and harmonic mean of each query term SCQ value together as a query quality index.

(4) Query scalability indicators

The present invention proposes an index reflecting query expansion performance for the first time - query expandability index. The calculation formula is:

Among them, Sω is the synonym set of the query word ω, and P(α|Q) refers to the occurrence probability of the query word α in the query model. Queries with higher query scalability can be considered to have higher query quality because more relevant documents can be retrieved after query expansion on them.

Feed the trained query quality prediction model to step 4.3.

4.3. For each pre-screened subquery obtained in step 3, calculate the 4 indicators that characterize the quality of the subquery in step 4.2, and input them into the query quality prediction model trained in step 4.2 to obtain the query quality of the subquery. Score. The sub-query with the highest query quality score among the 30 sub-queries is selected as the optimal sub-query, and the result is provided to step 5.

Step 5: Combine the query intent and the optimal subquery to get the final reconstructed query

The query intent obtained in step 2 and the optimal sub-query obtained in step 4 are combined as the final result of the reconstructed query.

Claims (1)

1. A query reconstruction method for electronic medical record description is characterized by comprising

Step 1, preprocessing an electronic medical record text and a medical literature text in a data set;

step 2, training an SVM classifier to predict the query intention of the electronic medical record text;

step 3, acquiring all sub-queries of the electronic medical record text and performing preliminary pre-screening on the sub-queries;

step 4, training a query quality prediction model, and selecting the optimal sub-query from the sub-queries output in the pre-screening in the step 3;

step 5, combining the query intention obtained in the step 2 and the optimal sub-query output in the step 4 to obtain a final reconstruction query;

wherein

Step 1: preprocessing electronic medical record text and medical literature text in a data set

1.1, extracting plain text

This step is not required because the electronic medical record text itself is plain text. The medical literature text is a webpage file stored in an XML format, useless CSS and JS codes are required to be removed, and required plain text data including titles, abstracts, keywords and body parts of the literature are taken out according to an XML tag, so that the preprocessed literature text has a uniform format.

Providing the plain text of the electronic medical record and the extracted plain text of the medical literature to the step 1.2;

1.2, removing stop words

And removing stop words in the plain text by utilizing the preprocessed vocabulary, wherein the stop words comprise words without semantic information and words with high use frequency.

The result after the stop word is removed is provided to step 1.3;

1.3 restoring part of speech

Different parts of speech are integrated and restored to root words, words with the same meaning in English have different tense changes, and the words are restored in terms of speech.

After the part of speech is restored, the text preprocessing work of the step 1 is completed, the preprocessed electronic medical record text is provided for the step 2 and the step 3, and the preprocessed medical literature text is provided for the step 4.

Step 2: method for predicting query intention of electronic medical record text by training SVM (support vector machine) three classifiers

And 2, training an SVM classifier to judge the query intention by using the preprocessed electronic medical record text obtained in the step 1 as a training set, and specifically comprising the following steps.

2.1, labeling three classification labels for each electronic medical record text in the training set: if the text content of the electronic medical record belongs to Diagnosis (Diagnosis), marking as 1; if the text content of the electronic medical record belongs to a Treatment scheme (Treatment), the text content is marked as 2; if the text content of the electronic medical record belongs to the diagnosis and detection means (Test), the label is 3. The annotated results are provided to step 2.2.

2.2, training a three-classifier.

The training of the three classifiers uses the existing SVM algorithm, and the features of the electronic medical record text and the three classification labels labeled in the step 2.1 are required to be input during the training. The training of the classifier requires the use of two features of the electronic medical record text: (1) a TF-IDF value; (2) and (4) semantic information.

(1) TF-IDF is a statistical method to assess how important a word is to one of the documents in a corpus. Where Term Frequency (TF) refers to the frequency with which a given term appears in the document. The Inverse Document Frequency (IDF) is obtained by dividing the total document number by the document number containing the word, and taking the obtained quotient to be a logarithm with the base 10. The TF-IDF value is the product of these two values and has the formula

Wherein n is_ωRepresenting the number of occurrences of the word omega in the document, N representing the total number of words in the document, N_dRepresenting the total number of documents in the corpus, N_ωRepresenting the number of files in the corpus that contain the word ω.

(2) The semantic information refers to three parts of information: whether diagnostic results are included (value 0/1), whether the check is complete (value 0/1), query text length (value 0-200).

The trained three classifiers are provided to step 2.3.

And 2.3, inputting the electronic medical record text into the trained three classifiers, and providing classification results (namely query intentions) to the step 5.

And step 3: all sub-queries of the electronic medical record text are obtained and subjected to preliminary pre-screening

Theoretically, a query containing n query terms may obtain sub-queries of a number level, for example, the sub-query into which the query statement "fe ver core header" containing 3 query terms may be split includes "fe ver," core, "" header, "" fe ver core, "" fe ver header, "" cog header, "" fe ver core. It is impractical to exhaust all possible sub-queries for ranking, so the sub-queries need to be pre-screened first. Sub-query pre-screening comprises the following steps.

And 3.1, selecting the subqueries with the length of 3-10 from all the subqueries of the electronic medical record text. The sub-query length refers to the number of words in the query. Research shows that the query length for optimizing the retrieval effect is between 3 and 6, and the maximum length threshold is defined to be 10 in consideration of the electronic medical record long text targeted by the invention. The result is provided to step 3.2.

And 3.2, calculating the average mutual information quantity of each sub-query obtained in the step 3.1, and selecting the 30 sub-queries with the highest mutual information quantity. The average mutual information quantity calculation formula of the sub-queries is as follows:

where n (x, y) represents the frequency with which the word x and the word y appear simultaneously in a document with a window size of 25 throughout the corpus, and n (x), n (y) represent the frequency with which the word x and the word y appear in the corpus, respectively. N is a radical of_cRepresenting the number of words in the entire corpus. And calculating mutual information quantity of any two words in one sub-query, and taking the weighted average of the mutual information quantity as the average mutual information quantity of the sub-queries.

And step 3, finally obtaining 30 sub-queries after pre-screening, and providing the result to step 4.

And 4, step 4: training a query quality prediction model, and selecting the optimal sub-query from the pre-screened sub-queries

4.1 labeling sub-queries with query quality scores

And (3) performing one round of retrieval on each pre-screened sub-query obtained in the step (3), wherein the retrieved target document set is from the medical document text set preprocessed in the step (1). The search engine used Indri5.11 in the Lemur open source project. And comparing the retrieval result with the evaluation standard provided by the TREC conference, calculating to obtain the average accuracy score of the retrieval, and marking the average accuracy score as the query quality score of the sub-query. The sub-queries with labeled query quality scores are provided to step 4.2 as the result of this step.

4.2 training query quality prediction model

The existing SVMRank algorithm is used for training the query quality prediction model, and indexes capable of representing the sub-query quality and the query quality scores marked in the step 4.1 need to be input during training.

Model training requires the use of the following indices, calculated for each sub-query in the training set: (1) an inverse document frequency correlation index; (2) simplifying the query definition index; (3) corpus/query similarity feature index; (4) and querying the expandability index.

The symbolic meaning used in this step is defined before introducing these indices separately. For a query Q, assume that it contains the query term ω₁,…ω_nN (ω) in corpus C_i) Representing a query term omega_iFrequency of occurrence in corpus, n (ω)_i,ω_j) Representing query words omega in a corpus_i,ω_j(i ≠ j) frequency of simultaneous occurrence in a window of 25 words in length, N_cRepresenting the total number of words contained in the corpus, N_ωNumber of documents in which the query word ω appears, N_dRepresenting the number of all documents in the corpus. P_c(ω) represents the probability of the occurrence of the query word ω in the corpus, P (ω | Q) represents the probability of the occurrence of ω in the query sentence Q, S_ωA set of synonyms representing the word ω.

(1) The inverse document frequency correlation index calculation formula is as follows:

wherein N is_ωNumber of documents containing word ω, N_dThe total number of documents in the corpus. For each sub-query, the sum, maximum, standard deviation, arithmetic mean, geometric mean, and harmonic mean of each query term IDF value are calculated together as query quality indicators.

(2) The simplified query definition index calculation formula is as follows:

wherein P is_ml(ω | Q) is the frequency of occurrence of the word ω in the query Q, P_c(ω) the frequency with which the word ω appears in the corpus.

(3) The corpus/query similarity characteristic index calculation formula is as follows:

like the inverse document frequency correlation index, the sum, the maximum value, the standard deviation, the arithmetic mean, the geometric mean and the harmonic mean of the SCQ value of each query term are calculated together as the query quality index.

(4) Query extensibility index

The invention firstly provides an index reflecting the query expansion performance, namely a query expansion index. The calculation formula is as follows:

where S ω is a synonym set of query terms ω and P (α | Q) refers to the probability of occurrence of query terms α in the query model.

And (4) providing the trained query quality prediction model to a step 4.3.

4.3, calculating 4 indexes representing the sub-query quality in the step 4.2 for each pre-screened sub-query obtained in the step 3, and inputting the indexes into the query quality prediction model obtained by training in the step 4.2 to obtain the query quality score of the sub-query. And selecting the sub-query with the highest query quality score in the 30 sub-queries as the optimal sub-query, and providing the result to the step 5.

And 5: obtaining final reconstruction query by combining query intention and optimal sub-query

And combining the query intention obtained in the step 2 and the optimal sub-query obtained in the step 4 to obtain a reconstructed query serving as a final result.

Images