KR101607672B1 - Apparatus and method for permutation based pattern discovery technique in unstructured clinical documents - Google Patents

Abstract

The apparatus for searching for a substitution-based pattern of an unstructured clinical document according to the present invention includes a natural language processing unit for normalizing an unstructured medical document received through a Natural Language Processing (NLP) process to generate normalized text, An object name recognizing unit for identifying the Named Entity Recognition information by matching the text and the received domain model, a structured information from the object name recognition information using a permutation-based pattern discovery approach, And a template creating unit for creating a template defined based on the structured information.

Description

[0001] APPARATUS AND METHOD FOR PERMUTATION BASED PATTERN DISCOVERY TECHNIQUE IN UNSTRUCTURED CLINICAL DOCUMENTS [0002]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to natural language processing, and more particularly, to a technique for extracting information from an unstructured document by utilizing a natural language processing technique.

Most of the available medical information used in the biomedical field is recorded in free-text format. Especially, narrative text written in free form is difficult to have a unified form because of abbreviation use and spelling. Documents written in a free format are easy to create for healthcare providers, but they are difficult to use in clinical research, statistical analysis, Decision Support and other end-dynamics studies.

Conventional information extraction techniques used in the medical field extract the medication names and signatures from the clinical narratives described in the clinical data such as discharge summaries and surgical pathology reports. However, as described above, many clinical data do not have a predetermined format or template, and in many cases, they are described as descriptive texts such as a research note or a clinical note. Therefore, in the conventional information extraction technology, it is desirable to extract necessary information from the descriptive text having no predetermined template, and it takes a considerable time to directly retrieve necessary information.

Korean Patent Laid-Open No. 10-2011-0110683 discloses a term system authoring support system based on reference terms. The patent includes the construction of a structured term system from medical terminology used in a medical institution. However, the above-mentioned patent only constructs a structured term system through the mapping relation with the reference term system, and does not effectively extract the structured information from the descriptive record described in the unclear form or template.

Korean Patent Publication No. 10-2011-0110683

A problem to be solved by the present invention is to provide a substitution-based pattern search apparatus and a search method capable of structuring a large number of unstructured medical documents written in a description format or in a free format.

The apparatus for searching for a substitution-based pattern of an unstructured clinical document according to the present invention includes a natural language processing unit for normalizing an unstructured medical document received through a Natural Language Processing (NLP) process to generate normalized text, An object name recognizing unit for identifying the Named Entity Recognition information by matching the text and the received domain model, a structured information from the object name recognition information using a permutation-based pattern discovery approach, And a template creating unit for creating a template defined based on the structured information. A domain model unit for storing a domain model in which an entry for all required entities is built, which is composed of domain lexicons, and a domain model unit for storing structured information in consideration of an explicit relationship between different entities And a structured text storing unit for storing the structured text.

The natural language processing unit classifies meaningful words in unstructured medical documents, generates meaningful words by dividing them into individual words based on the selected special characters, generates a tokenized word, transforms the tokenized word into spelling, Adjusts the reduction to produce normalized text.

The pattern recognition unit identifies the entity name recognition information by a non-structured value and a non-numeric key, and recognizes the pattern of the unstructured medical document by replacing the distinguished key and the value. The pattern recognition unit matches the extracted array of the key and the character string of the value in the object name recognition information.

The replacement-based pattern search method of an unstructured clinical document according to the present invention normalizes the received unstructured medical document through a Natural Language Processing (NLP) process to generate normalized text. Then, the normalized text and the received domain model are matched to identify Named Entity Recognition information. Next, structured information is generated from the entity name recognition information using a permutation-based pattern discovery approach, and a prescribed template is created based on the structured information.

The step of generating the normalized text classifies the meaningful words in the unstructured medical document, generates a tokenized word by dividing the meaningful word into individual words based on the selected special character, Generates normalized text by adjusting spelling variations, lowercase alphabetic characters, and abbreviations. The step of generating the structured information may further include dividing the entity name recognition information into a set value and a non-numeric key, and replacing the separated key and the value, Lt; / RTI >

By extracting structured information from unstructured medical documents through the substitution-based pattern searching apparatus and the searching method of the unstructured clinical document according to the present invention, it is possible for the clinician or the researcher to search and classify information more quickly, Sharing can be facilitated. In addition, patient cases of extracted structured information can be reflected in the knowledge base of the clinical decision system to effectively support clinical decision making.

1 is a block diagram illustrating an apparatus for searching for a substitution-based pattern in an unstructured clinical document according to an embodiment of the present invention.
FIG. 2 is a block diagram illustrating a substitution-based pattern retrieval method of an unstructured clinical document according to an embodiment of the present invention. Referring to FIG.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. The terms and words used in the present specification are selected in consideration of the functions in the embodiments, and the meaning of the terms may vary depending on the intention or custom of the invention. Therefore, the terms used in the following embodiments are defined according to their definitions when they are specifically defined in this specification, and unless otherwise specified, they should be construed in a sense generally recognized by those skilled in the art.

In the medical field, a large number of medical data is generated in the form of medical documents written in free-form narrative text, such as notes / reports. Such medical data often require additional analysis by clinicians, medical researchers or analysts. Since the data in general medical documents are not structured, the clinician needs to input the medical data directly into the structured template. However, this can be time consuming and can lead to various problems in the daily practice of medical practice. The apparatus for searching a substitution-based pattern of an unstructured clinical document according to the present invention includes at least one of Domain Modeling, Named Entity Recognition, a property value sorting according to a permutation-based pattern recognition approach, (Template Filling) and Natural Language Processing (NLP) techniques to extract structured information from unstructured medical documents. In this way, we overcome the limitations of the conventional rule-based system.

1 is a block diagram illustrating an apparatus for searching for a substitution-based pattern in an unstructured clinical document according to an embodiment of the present invention.

1, the substitution-based pattern searching apparatus 100 of an unstructured clinical document according to the present invention includes a natural language processing unit 110, an entity name recognizing unit 120, a domain modeling unit 130, a pattern recognizing unit 140 A structured text storage unit 150, and a template creation unit 160. [

The natural language processing unit 110 normalizes an unstructured medical document that is not structured through a Natural Language Processing (NLP) process. Many medical documents often do not have a defined format or template. When a clinician or researcher writes a research note or a clinical note, it can not be forced to write it in a specific style. Therefore, in many cases, such as a research note or a clinical note, it is described as a descriptive text in a free form by the author. Therefore, in the conventional information extraction technology, it is desirable to extract necessary information from the descriptive text having no predetermined template, and it takes a considerable time to directly retrieve necessary information. In order to solve such a problem, the natural language processing unit 110 first utilizes a natural language processing technique to recognize the unstructured medical document of such a free form. The natural language processing technique includes a clause / sentence boundary identification and a tonenization process. The natural language processing unit 110 can perform this through a general natural language processing (NLP).

The natural language processing unit 110 classifies Meaningful Words in the unstructured medical document through clause / sentence boundary identification and tokenizing process, and assigns meaningful words to individual words And generates a tokenized word by dividing it. A meaningful word is a set of words excluding Stop Words and symbols (parentheses, special characters, etc.). Then, the natural language processing unit 110 normalizes the tokenized words by adjusting spell variation, alphabetic lower case, and abbreviation. For example, the 'Lymph Nodes Level 1' described in the unstructured medical document is normalized to 'lymph node level 1' using lower case alphabetic characterization and multiple word shortening. The natural language processing unit 110 transfers the normalized text to the entity name recognizing unit 120.

The entity name recognizing unit 120 identifies the Named Entity Recognition information by matching the normalized text received from the natural language processing unit 110 and the received Domain Model from the domain model unit 130 . Named Entity is a concept that is required when recognizing input text. The domain model received from the domain modeling unit 130 is domain lexicons, and an entry for all required entities is embedded. In the process of recognizing entity names, Lexicon words can be precisely matched or partially matched. In exact matching, the matched words are extracted for further processing during partial matching. Then, the partially matched words during the partial matching are extracted and held until the next word is extracted. The recognition of the entity name is concatenated with the next word extracted, and it is confirmed whether or not it is an exact match. The entity name recognition information recognized by the entity name recognition unit 120 includes a key and a value. The entity name recognition unit 120 transmits the entity name recognition information generated through the entity name recognition process to the pattern recognition unit 140.

The domain modeling unit 130 transfers the domain model to the object name analyzing unit 120. The domain model is designed to learn the context and boundary of each concept. The domain model helps to understand the concepts of concept, standardization and scope. Lexicon is required to find variations of the vocabulary using all concepts and specific settings. In such a domain model, the entity name recognizing unit 120 enables the entity name recognition of the text that has been regularized through the matching process.

The pattern recognition unit 140 identifies the pattern from the object name recognition information using the permutation-based pattern discovery approach in the received object name recognition information. The required template, which is a required form, needs to be filled with relevant extracted information. The complex part of the template type is the correct identification of the correlated information that occurs in the sentence, and more specifically the assignment of the exact value of the attribute. In text, some concepts are represented by values, not numbers, and some by numeric values. The pattern recognition unit 140 considers only the numerical value. Sequences of other concepts play an important role in the accurate assignment of attribute values. In some cases, a value follows its attribute, while in other cases an attribute follows its value. Occasionally, the values of all attributes appear first immediately after, and sometimes all values appear immediately after their attribute. First, all permutations are generated and grouped together to represent a single pattern with common properties. In other words, the general unstructured medical document can have the value of each attribute and the corresponding attribute sequentially as shown in Table 1. For example, the attribute, the lymphatic chain, and the attribute value of the lymph node, 5, may appear sequentially. In order to recognize such a pattern, the pattern recognition unit 140 performs pattern recognition by distinguishing a non-numeric value from a non-numeric value.

The input / output data of the pattern recognition unit 140 Input Named Entity Recognition (Key and Value)

Output Pattern Recognition V1 K1 V2 K2 V3 K3 V4 K4 V5 K5 (Permutation)
value (key)

The entity recognized by the entity name recognizing unit 120 has two parts including a key and a value. Looking at Table 1, concepts recognized from the text include keys that are numeric and not numeric. The problem is how to identify the values belonging to a key (like a clinician) that can create an object from various orders, such as a key according to a value (or a value according to a key). For this, the pattern recognition unit 140 recognizes the last pattern through the Permutation based Pattern Discovery Approach. The permutation-based pattern approach matches an extracted array (Arrange) and a valued string (String) of the key with the possible permutations generated by Equation (1).

In Equation (1), N represents the size of the set of keys and values obtained from substitution. And, r represents each replacement size. A substitution represents an array of keys and values generated from the text. This process identifies the last pattern in the recognized object name. Then, the pattern recognition unit 140 generates structured information based on the recognized pattern. That is, the pattern recognition unit 140 can recognize the structure and the contents (and the flow) of the unstructured medical document by recognizing the pattern by replacing the key and the valleys classified by the attribute instead of the number and the number. Through this process, the pattern recognition unit 140 generates structured information. The pattern recognition unit 140 transfers the generated structured information to the structured text storage unit 150.

The structured text storage unit 150 stores the structured information received from the pattern recognition unit 140 in a structured manner. Structured information maintains an explicit relationship between different entities. The structured information stored in the structured text storage unit 150 in a structured manner can then be utilized for various analyzes. Clinical and epidemiological studies are conducted according to structured information. This kind of analysis is very difficult to carry out through unstructured documents because information is distributed in an informal style. The stored structured information can be considered as Patient Cases. Multiple statistics can be retrieved from a set of patient cases. For example, the survival rates and risk factors found in patients associated with age and sex can be determined. Clinical researchers can link information domains to find more important epidemiological outcomes.

The template creating unit 160 creates a template defined by using the structured information stored in the structured text storage unit 150. Since the prescribed templates created through the structured information have the same format, it is easy to share information, and the necessary information can be automatically searched.

The structured information stored in the structured storage unit 150 may be applied to a clinical decision support system. Clinical decision support system is a system to support clinicians in important clinical decision making process. The clinical decision support system utilizes a knowledge base. The knowledge base consists of knowledge rules. Knowledge rules can be generated from patient cases stored as structured information. Clinical decision support systems can infer about rules from knowledge bases generated from patient cases extracted from unstructured text.

FIG. 2 is a block diagram illustrating a substitution-based pattern retrieval method of an unstructured clinical document according to an embodiment of the present invention. Referring to FIG.

Referring to FIG. 2, a method for searching for a substitution-based pattern in an unstructured clinical document according to the present invention comprises the steps of normalizing an unstructured medical document that is not structured through a Natural Language Processing (NLP) (201). Many medical documents often do not have a defined format or template. When a clinician or researcher writes a research note or a clinical note, it can not be forced to write it in a specific style. Therefore, in many cases, such as a research note or a clinical note, it is described as a descriptive text in a free form by the author. Therefore, in the conventional information extraction technology, it is desirable to extract necessary information from the descriptive text having no predetermined template, and it takes a considerable time to directly retrieve necessary information. In order to solve such a problem, natural language processing technology is first used to recognize unstructured medical documents of such free form. The natural language processing technique includes a clause / sentence boundary identification and a tonenization process.

The natural language processing step classifies Meaningful Words in unstructured medical documents through clause / sentence boundary identification and tokenization process and divides meaningful words into individual words based on selected special characters Generates a tokenized word. A meaningful word is a set of words excluding Stop Words and symbols (parentheses, special characters, etc.). The natural language processing step is normalized by adjusting Spell Variation, Alphabetic Lower Case, and Abbreviation in the tokenized word. For example, the 'Lymph Nodes Level 1' described in the unstructured medical document is normalized to 'lymph node level 1' using lower case alphabetic characterization and multiple word shortening.

Next, the received normalized text and the Domain Model are matched to identify Named Entity Recognition information (202). Named Entity is a concept that is required when recognizing input text. The domain model received from the domain modeling unit 130 is domain lexicons, and an entry for all required entities is embedded. In the process of recognizing entity names, Lexicon words can be precisely matched or partially matched. In exact matching, the matched words are extracted for further processing during partial matching. Then, the partially matched words during the partial matching are extracted and held until the next word is extracted. The recognition of the entity name is concatenated with the next word extracted, and it is confirmed whether or not it is an exact match. The entity name recognition information recognized in the entity name identification step includes a key and a value.

The domain model is designed to learn the context and boundary of each concept. The domain model helps to understand the concepts of concept, standardization and scope. Lexicon is required to find variations of the vocabulary using all concepts and specific settings. In this domain model, it is possible to recognize the object name of the text that has been regularized through the matching process at the object name recognition step.

Next, a pattern is identified in the entity name recognition information using a permutation-based pattern discovery approach (203). The required template, which is a required form, needs to be filled with relevant extracted information. The complex part of the template type is the correct identification of the correlated information that occurs in the sentence, and more specifically the assignment of the exact value of the attribute. In text, some concepts are represented by values, not numbers, and some by numeric values. The pattern recognition unit 140 considers only the numerical value. Sequences of other concepts play an important role in the accurate assignment of attribute values. In some cases, a value follows its attribute, while in other cases an attribute follows its value. Occasionally, the values of all attributes appear first immediately after, and sometimes all values appear immediately after their attribute. First, all permutations are generated and grouped together to represent a single pattern with common properties.

The recognized object has two parts including a key and a value. Concepts recognized from the text include keys that are numeric and not numeric. The problem is how to identify the values belonging to a key (like a clinician) that can create an object from various orders, such as a key according to a value (or a value according to a key). For this, the pattern recognition unit 140 recognizes the last pattern through the Permutation based Pattern Discovery Approach. The permutation-based pattern approach matches an extracted array (Arrange) and a valued string (String) of the key with the possible permutations generated by Equation (1).

A substitution represents an array of keys and values generated from the text. This process identifies the last pattern in the recognized object name. Then, the pattern recognition unit 140 generates structured information based on the recognized pattern.

Structured information maintains an explicit relationship between different entities. The structured information stored in the structured text storage unit 150 in a structured manner can then be utilized for various analyzes. Clinical and epidemiological studies are conducted according to structured information. This kind of analysis is very difficult to carry out through unstructured documents because information is distributed in an informal style. The stored structured information can be considered as Patient Cases. Multiple statistics can be retrieved from a set of patient cases. For example, the survival rates and risk factors found in patients associated with age and sex can be determined. Clinical researchers can link information domains to find more important epidemiological outcomes.

Next, a prescribed template is created using the structured information (204). Since the prescribed templates created through the structured information have the same format, it is easy to share information, and the necessary information can be automatically searched. And structured information can be applied to clinical decision systems. Clinical decision support system is a system to support clinicians in important clinical decision making process. The clinical decision support system utilizes knowledge base. The knowledge base consists of knowledge rules. Knowledge rules can be generated from patient cases stored as structured information. Clinical decision support systems can infer about rules from knowledge bases generated from patient cases extracted from unstructured text.

The present invention including the above-described contents can be written in a computer program. And the code and code segment constituting the program can be easily deduced by a computer programmer of the field. In addition, the created program can be stored in a computer-readable recording medium or an information storage medium, and can be read and executed by a computer to implement the method of the present invention. And the recording medium includes all types of recording media readable by a computer.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is to be understood that the invention is not limited to the disclosed exemplary embodiments, but, on the contrary, It is possible.

100: Replacement-based pattern retrieval device for unstructured clinical documents
110: natural language processing unit
120: entity name recognition unit
130: Domain model part
140: pattern recognition unit
150: structured text storage unit
160: Template creation section

Claims (10)

A natural language processing unit for normalizing the received unstructured medical document through a Natural Language Processing (NLP) process to generate normalized text;
An entity name recognizing unit for identifying the Named Entity Recognition information by matching the normalized text and the received domain model;
A pattern recognition unit for generating structured information from the entity name recognition information using a permutation-based pattern discovery approach; And
A template creating unit for creating a template defined based on the structured information;
/ RTI >
Wherein the domain model is composed of domain lexicons and an entry for all required entities is embedded. The method according to claim 1,
The natural language processing unit classifies meaningful words in the unstructured medical document, generates a tokenized word by dividing the meaningful word into individual words based on the selected special character, Wherein the normalized text is generated by adjusting the lowercase letter and the abbreviation to generate the normalized text. The method according to claim 1,
Wherein the pattern recognition unit identifies the entity name recognition information by a non-structured value and a non-numeric key, and recognizes the pattern of the unstructured medical document by replacing the distinguished key and the valence Based non-structured clinical document. The method of claim 3,
Wherein the pattern recognition unit comprises:

Through the,
Matching the extracted array and value strings of the keys in the entity name recognition information, wherein N represents a size of a set of keys and values obtained from the substitution, and r represents a size of each substitution, A substitution based pattern search apparatus for a document. The method according to claim 1,
A domain model unit for storing the domain model;
Based pattern search apparatus of claim 1, further comprising: The method according to claim 1,
A structured text storage unit for storing the structured information in consideration of an explicit relationship between different entities;
Based pattern search apparatus of claim 1, further comprising: Generating normalized text by normalizing the received unstructured medical document through a Natural Language Processing (NLP) process;
Identifying the Named Entity Recognition information by matching the normalized text and the received domain model;
Generating structured information from the entity-name recognition information using a permutation-based pattern discovery approach; And
Creating a defined template based on the structured information;
/ RTI >
Wherein the domain model is composed of domain lexicons and an entry for all required entities is embedded. 8. The method of claim 7,
The step of generating the normalized text classifies the meaningful words in the unstructured medical document, generates a tokenized word by dividing the meaningful word into individual words based on the selected special character, Wherein the normalized text is generated by adjusting the spelling variant, lowercase alphabetic characterization, and abbreviation to generate the normalized text. 8. The method of claim 7,
Wherein the step of generating the structured information includes classifying the entity name recognition information into a non-structured medical document by replacing the distinguished key and the value class with a structured value and a non-numeric key, Based on the result of the comparison. 10. The method of claim 9,
Wherein the generating the structured information comprises:

Through the,
Matching the extracted array and value strings of the keys in the entity name recognition information, wherein N represents a size of a set of keys and values obtained from the substitution, and r represents a size of each substitution, A method for retrieving substitution-based patterns in documents.

Images