Method for extracting standard terms for human phenotypes
By using an entropy-based method for extracting standard human phenotypic terms, the problem of extracting these terms from unstructured electronic medical records has been solved. This method achieves highly accurate automated disease feature recognition, shortens disease diagnosis time, and reduces reliance on the subjective opinions of clinicians.
Patent Information
- Application Number
- CN201911287912.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2019-12-15
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2039-12-15
AI Technical Summary
Existing technologies struggle to efficiently and automatically extract standard human phenotypic terms from unstructured electronic medical records, resulting in lengthy disease diagnosis times that are limited by the subjective influence of clinicians and hindering the rapid identification of the relationship between a patient's disease and genetic variations.
An entropy-based method for extracting standard human phenotype terms is adopted. By using custom rules to segment disease feature strings, perform string matching, block partitioning, and entropy calculation, and combine the F-metric value to optimize similarity calculation, the most similar records to the standard human phenotype terminology database are extracted.
It significantly improved the accuracy of extracting standard human phenotypic terms, reduced disease diagnosis time, improved the accuracy of automated identification, and reduced the subjective impact on clinicians.
Smart Images

Figure CN112992303B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of bioinformatics, and in particular relates to a method for extracting standard terms of human phenotypes in medical electronic medical records. Background Art
[0002] The statements in this section merely provide background art related to the present disclosure and do not necessarily constitute prior art.
[0003] Human phenotype is an important entity information in clinical data and medical literature data, and is the main basis for doctors' diagnosis and treatment. However, phenotypic information in electronic medical records is often unstructured data and exists in the form of natural language, which brings difficulties to the subsequent application of medical record data. Therefore, for the application of medical record data, text analysis and information extraction are required according to specific circumstances.
[0004] The incidence and mortality of human genetic diseases have been increasing year by year. Human genetic diseases have become an important factor threatening human health. Severe genetic diseases affect more than 7 million newborns worldwide each year, and these diseases often progress quickly. Therefore, it is urgent to diagnose these diseases, and the time for disease diagnosis should be shortened as much as possible. However, manual diagnosis and evaluation of medical records will waste a lot of time and energy. Generally speaking, it takes about a week to identify the patient's pathogenic gene, and it may also be affected by the clinician's subjective influence. Limited by the medical level of clinical medical experts, it is impossible to diagnose diseases with unclear correlation. If the patient's disease can be automatically identified, or the correspondence between the patient's disease and his own gene mutation can be automatically found to assist the doctor in diagnosis, it will greatly reduce the time for the doctor to make a diagnosis and win precious treatment time for the patient. However, the description of the patient's disease phenotype in electronic medical records, medical literature and other information is often unstructured data and exists in natural language, which seriously hinders the use of automatic methods to identify diseases or the correspondence between gene mutations and diseases. Therefore, the automatic extraction of standard terms for human phenotypes in clinical data is very important.
[0005] Among traditional methods, rule-based methods rely on existing medical vocabulary, such as the use of regular expressions. However, rules and templates based on manual summaries often fail to cover various language phenomena, have poor generalization capabilities, and are not portable across domains. It is difficult for researchers to design complete rules to cover various situations. Dictionary-based methods are widely used due to their simplicity. By calculating semantic similarity, target concepts can be annotated with existing words in the dictionary. However, this method is overly dependent on the scale and quality of external data, and the process of collecting corpus consumes a lot of human resources.
[0006] At present, most methods for automatically extracting standard terms for human phenotypes match natural language descriptions with standard term libraries, and find the corresponding HPO standard vocabulary based on the calculated similarity. Human Phenotype Ontology (HPO) is the human phenotype ontology. HPO currently contains more than 13,000 terms and more than 156,000 genetic disease annotations, providing standardized vocabulary for phenotypic abnormalities encountered in human diseases. Each term in HPO describes a phenotypic abnormality, for example: HP:0001250 is the ID for Seizures. Summary of the invention
[0007] In view of the problems existing in the prior art, the present disclosure provides a method for extracting standard terms of human phenotypes from clinical electronic medical records.
[0008] Usually, for unstructured electronic medical records, natural language processing algorithms are needed to perform preprocessing such as word segmentation, word form restoration, and stop word removal. In the preprocessed electronic medical records, disease feature character strings are then identified and converted into standard human phenotype terms.
[0009] An embodiment of the present invention provides a method for extracting human phenotype standard terms based on entropy, comprising the following steps:
[0010] (1) Analyze and extract electronic medical record data to obtain disease feature strings;
[0011] (2) string matching is performed between the disease feature string and each record in the human phenotype standard terminology library, the matching relationship between the disease feature string and each record in the human phenotype standard terminology library is marked, and an alignment file is generated;
[0012] (3) performing block division on the generated alignment file to obtain a block division file;
[0013] (4) Entropy calculation is performed based on the block partition file to obtain the similarity between the disease feature string and each record in the human phenotype standard terminology library, and the corresponding human phenotype standard terminology is extracted based on the highest similarity.
[0014] In step (1), the electronic medical record data is analyzed and extracted, and the electronic medical record data is segmented according to the custom rules: punctuation marks, negative words and conjunctions (including and, but, etc.). First, the first-level segmentation is performed according to the first-class punctuation marks and negative words, and then in the first-level segmentation results, if negative words (such as not) appear, they are directly screened out, and then on this basis, the second-level segmentation is performed according to the second-class punctuation marks and conjunctions, so as to obtain the disease feature string; wherein, the first-class punctuation marks include period, exclamation mark and semicolon, and the second-class punctuation marks include comma and colon.
[0015] In the present disclosure, the human phenotype standard terminology library includes: the names of human phenotype standard terms and their definitions, as well as synonyms of human phenotype standard terms. Each standard term and its corresponding definition and synonym in the human phenotype standard terminology library is divided into a group, and each item in the group is a record. Some examples are as follows: Figure 2 As shown. The extracted disease feature character string is aligned with the record in the human phenotype standard terminology library, and after the alignment, a block division operation is performed to obtain a block division file, and then entropy calculation is performed based on the block division file to obtain the similarity between the disease feature character string and the record in the human phenotype standard terminology library, and finally the corresponding human phenotype standard terminology is extracted based on the highest similarity. The method of the present invention gets rid of the limitation of N-gram, and can obtain a character string with the maximum matching length. It has been verified by experiments that the accuracy of extraction can be significantly improved. Compared with the ClinPhen method, the present invention has a significant improvement in accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] Figure 1 This is a flowchart of a method for extracting standard terms for human phenotypes in an embodiment of the present disclosure;
[0017] Figure 2 This is a partial schematic diagram of the human phenotype standard terminology library;
[0018] Figure 3 It is a schematic diagram of an alignment method in an embodiment of the present disclosure;
[0019] Figure 4 Schematic diagram of a block partitioning method in an embodiment of the present disclosure. DETAILED DESCRIPTION
[0020] In order to make the present invention more clear, the present invention will be described in detail below with reference to the accompanying drawings and specific examples.
[0021] It should be noted that the human phenotype standard term library used in the following embodiments is in English and is verified with English electronic medical records, but the method disclosed herein is also applicable to the Chinese version of the human phenotype standard term library and medical records.
[0022] The embodiments of the present disclosure provide a method for extracting human phenotype standard terms based on entropy to improve the accuracy of extracting human phenotype standard terms.
[0023] Shannon defined information entropy as the probability of a discrete random event. Assuming X is a discrete random variable with a value space of R, its probability distribution is:
[0024] p(x)=P(X=x),x∈R (1.1)
[0025] The entropy H(X) of X is defined as:
[0026]
[0027] In formula (1.2), the logarithm is based on 2, and the unit of entropy defined by the formula is binary bit, i.e., bit. log2p(x) is usually abbreviated as logp(x).
[0028] Entropy is a variable uncertainty measure. In the method disclosed in the present invention, entropy can be used to reflect the distribution of the fragments that match the extracted disease feature string and the records in the human phenotype standard terminology library. The lower the entropy, the more concentrated the distribution of the matching fragments, that is, the more complete the matching block; conversely, the higher the entropy, the more dispersed the distribution of the matching fragments, the more dispersed the matching blocks, and the worse the fluency.
[0029] like Figure 1 As shown, this embodiment provides a method for extracting human phenotype standard terms, comprising the following steps:
[0030] (1) Analyze and extract electronic medical record data to obtain disease feature strings;
[0031] According to the custom rules: the electronic medical record data is segmented by punctuation marks, negative words (not) and conjunctions (including and and but). First, the first-level segmentation is performed according to the first type of punctuation marks (including periods, exclamation marks and semicolons). If negative words (such as not) appear in the first-level segmentation results, they are directly screened out. Then, on this basis, the second-level segmentation is performed according to the second type of punctuation marks (including commas and semicolons) and conjunctions to obtain the disease feature string.
[0032] (2) string matching is performed between the disease feature string and each record in the human phenotype standard terminology library, the matching relationship between the disease feature string and each record in the human phenotype standard terminology library is marked, and an alignment file is generated;
[0033] First, mark the matching of each record in the disease feature string and the human phenotype standard terminology library. After marking all the matching relationships, find each alignment in the set of matching relationships according to certain rules. The specific rules are as follows: Each match corresponds to a line connecting the disease feature string and the record in the human phenotype standard terminology library. For each record, select a set of matching relationships with the least cross lines as the alignment of the record with the disease feature string. The alignment method is as follows: Figure 3 The alignment between the disease feature string and each record in the human phenotype standard terminology library is one-to-one, that is, one word in the disease feature string matches at most one word in one record in the human phenotype standard terminology library, and vice versa, one word in one record in the human phenotype standard terminology library also matches at most one word in the disease feature string.
[0034] (3) performing block division on the generated alignment file to obtain a block division file;
[0035] The block division method is as follows Figure 4 As shown, after the alignment is completed, the alignment file needs to be divided into blocks. It is required that the positions of the matching words in the two strings in each block are continuous in the disease feature string, and the positions mapped to the records in the human phenotype standard terminology library are also continuous, and each block should be the maximum continuous string match, so as to obtain a block division file.
[0036] (4) Entropy calculation is performed based on the block partition file to obtain the similarity between the disease feature string and each record in the human phenotype standard terminology library, and then the human phenotype standard terminology in the group corresponding to the record with the highest similarity is extracted.
[0037] After the alignment file and block division are completed, entropy calculation is performed on the block situation. The entropy calculation includes: calculating entropy according to the block information of the block division file, normalizing the entropy, and combining the normalized entropy and F-measure to obtain the similarity between the disease feature string and each record in the human phenotype standard terminology library.
[0038] The formula for calculating entropy based on block information is as follows: where li represents the length of the i-th block, that is, the number of words; c represents the number of blocks; and L represents the number of all matching words.
[0039]
[0040] In order to limit the entropy value to between (0,1), the entropy obtained above needs to be normalized. The method adopted is to apply an exponential function with base e to the above formula. Entropy is used to represent the normalized entropy. The entropy of the block information is normalized, and the calculation formula is as follows:
[0041] entropy=e -H (1.4)
[0042] Finally, the normalized entropy and F-measure are combined to obtain the similarity between the disease feature string and each record in the human phenotype standard vocabulary. The details are as follows:
[0043] Step A: Use the F measure to evaluate the consistency between the disease feature string and each record in the human phenotype standard vocabulary. The calculation method is shown in (1.5), where β is the weight assigned to the precision rate, 0﹤β﹤1, precision is the precision rate, and recall is the recall rate.
[0044]
[0045] Step B: Introduce the F-measure into entropy, and obtain the final similarity between the disease feature string and each record in the human phenotype standard terminology library by combining the normalized entropy and F-measure, and obtain Sim, that is, the similarity. The calculation formula is shown in (1.6)
[0046] Sim=e -H ×F-score (1.6)
[0047] The human phenotype standard terms with the highest similarity corresponding to the disease feature string can be extracted through similarity analysis.
[0048] The parameter β is used in step A to dynamically adjust the weights of precision and recall, thereby obtaining a higher F-measure value and a higher similarity Sim, making the result more accurate.
[0049] In one embodiment of the present disclosure, β is 1 / 2.
Claims
1. A method for extracting standard terms for human phenotypes, comprising the following steps: (1) Analyze and extract electronic medical record data to obtain disease feature strings; (2) string matching is performed between the disease feature string and each record in the human phenotype standard terminology library, the matching relationship between the disease feature string and each record in the human phenotype standard terminology library is marked, and an alignment file is generated; (3) performing block division on the generated alignment file to obtain a block division file; (4) performing entropy calculation based on the block partition file to obtain the similarity between the disease feature string and each record in the human phenotype standard terminology library, and extracting the corresponding human phenotype standard terminology based on the highest similarity; The step of generating the alignment file in step (2) includes: first, marking the matching situation of each record in the disease characteristic string and the human phenotype standard terminology library, and after marking all matching relationships, finding the alignment in the set of matching relationships according to certain rules, wherein the rule is: each match corresponds to a line connecting the disease characteristic string and the record in the human phenotype standard terminology library; for each record, selecting a set of matching relationships with the least number of crossing lines as the alignment of the record with the disease characteristic string; The block division method in step (3) is to divide the alignment file into blocks, requiring that the positions of the words matched in the two strings in each block are continuous in the disease feature string, and the positions mapped to the records in the human phenotype standard terminology library are also continuous, and each block should be the maximum continuous string match, thereby obtaining a block division file; The entropy calculation in step (4) includes: calculating entropy according to the block information of the block partition file, normalizing the entropy, and combining the normalized entropy and the F-measure to obtain the similarity between the disease feature string and each record in the human phenotype standard terminology library; The formula for calculating entropy based on the block information of the block-divided file is: Among them, li represents the length of the i-th block, that is, the number of words; c represents the number of blocks; L represents the number of all matching words; The calculation method of combining the normalized entropy and F-measure to obtain the similarity between the disease feature string and each record in the human phenotype standard vocabulary includes: Step A: Use the F-measure to evaluate the consistency between the disease feature string and each record in the human phenotype standard vocabulary. The calculation formula is: Among them, β is the weight assigned to the precision rate, 0﹤β﹤1, precision is the precision rate, recall is the recall rate; Step B: Introduce the F-measure into entropy, and obtain the similarity Sim between the disease feature string and each record in the human phenotype standard terminology library by combining the normalized entropy and F-measure. The calculation formula is: Sim = e -H ×F-score.
2. The method for extracting human phenotype standard terms according to claim 1, characterized in that: The analysis and extraction of the electronic medical record data in step (1) includes: firstly, performing a first-level segmentation according to the first-class punctuation marks and negative words, then directly filtering out the negative words in the first-level segmentation results if any, and then performing a second-level splitting according to the second-class punctuation marks and conjunctions on this basis, thereby obtaining a disease feature string; wherein the first-class punctuation marks include periods, exclamation marks and semicolons, and the second-class punctuation marks include commas and colons.
3. The method for extracting human phenotype standard terms according to claim 1, characterized in that: The human phenotype standard terminology library includes: the names and definitions of human phenotype standard terms, and synonyms of human phenotype standard terms.
4. The method for extracting human phenotype standard terms according to claim 1, characterized in that: β is 1 / 2.
Citation Information
Patent Citations
Automatic Web text data extraction template generating and extracting method for small samples
CN102929930A
Natural language processing method and system for Chinese disease diagnosis information
CN105095665A
Cross-language information matching method based on term extraction
CN107908712A
Data matching method and system based on ontology model
CN108710663A
Medical record data processing method, apparatus, computer device and storage medium
CN109215754A
