Writing content wrongly written character detection reminding method and system based on report

Abstract

The invention provides a report writing content-based wrongly written character detection reminding method, which comprises the following steps of: obtaining a preliminary report statement needing to be detected, if characters and / or words in the preliminary report statement are characters and / or words in a blacklist medical wrongly written character database, judging that the characters and / or words are wrongly written characters and defining as first wrongly written characters, and at the moment, judging that the characters and / or words are wrongly written characters and defining as second wrongly written characters; extracting correct content corresponding to the first wrongly written character from the blacklist medical wrongly written character database, and replacing the first wrongly written character with the correct content; based on a white list NLP word library, performing white list word segmentation on the modified preliminary report statement, and if a single character which cannot be divided appears, determining that the single character is a wrongly written character, and defining the single character as a second wrongly written character; and extracting the word with the highest probability from the white list NLP word bank according to the pinyin as the modification content, and returning the modification content to the report. The invention further discloses a system for detecting and reminding wrongly written characters based on the written content of the report. According to the method, the wrong content of the report text is recognized, a doctor is reminded to correct wrongly written characters, and the problems of false report and missing report of the wrongly written characters in the text report are avoided.

Description

Technical Field

[0001] The present invention relates to the field of medical information, and more specifically, to a method and system for detecting and reminding typos based on report writing content. Background Art

[0003] The use of text error correction technology can improve the quality of reports and the efficiency of doctors' work. Commonly used error correction technologies include probabilistic calculation models and neural network models based on deep learning. They can identify common semantic errors relatively well, but they are insufficient when facing the task of detecting typos in medical report texts: First, because medical vocabulary is professional and complex, if only the model trained with non-medical general corpus is used, it is difficult to perform recognition and detection work, and it is easy to produce false positives or omissions; second, the professionalism of medical vocabulary itself leads to the difficulty of material sorting and annotation; third, the preciousness of medical text materials themselves makes it difficult to provide sufficient materials for neural network model training, and the model trained with neural network cannot be improved in subsequent use and feedback; fourth, the medical text obtained by conventional methods is usually the correct text after modification, and the precious information such as the erroneous content and modification method contained therein has been lost. Summary of the invention

[0004] In view of this, the main purpose of the present invention is to provide a method and system for detecting and reminding typos based on report writing content. The method can compare the modification traces of historical preliminary reports and audit reports, analyze the original typos, misspellings, incorrect punctuation marks in the report and the correct content after modification, simplify parameter sorting and labeling work, establish a high-quality blacklist medical typo database, and then combine conventional typo detection technology to supplement the detection, realize the recognition and detection of erroneous content in report texts, remind the reporting doctor to correct typos, and solve the problems of false reporting and missed reporting of typos in text reports existing in the prior art.

[0005] To achieve the above object, the technical solution of the present invention is achieved as follows:

[0006] On the one hand, the present invention provides a method for detecting and reminding typos based on report writing content, comprising: creating a blacklist medical typo database based on preliminary report content, audit report content and report modification traces; performing word segmentation processing on the audit report content to generate a whitelist NLP vocabulary; obtaining a preliminary report sentence to be detected, and comparing it with the blacklist medical typo database; if the characters and / or words in the preliminary report sentence are characters and / or words in the blacklist medical typo database, it is determined that the characters and / or words are typos, which are defined as the first typo, and at this time, the characters and / or words are compared with the blacklist medical typo database. The correct content corresponding to the first typo is extracted from the blacklist medical typo database, and the first typo is replaced with the correct content, and a modified preliminary report sentence is output; based on the whitelist NLP vocabulary, the modified preliminary report sentence is subjected to whitelist word segmentation, and if a single word that cannot be divided appears, the single word is determined to be a typo and defined as a second typo; the word with the highest probability is extracted from the whitelist NLP vocabulary according to pinyin as the modified content of the first typo and the second typo, and returned to the report writing interface for reference by the reporting doctor.

[0007] Preferably, the method also includes: verifying whether the single word is a typo through calculation using a preset model, and if the probability of the single word appearing reaches a preset threshold, finally determining that the single word is a typo, that is, the second typo; wherein the preset model is a probabilistic language model trained based on the content of the audit report.

[0008] Preferably, the steps of creating a blacklist medical typo database include: obtaining all the preliminary report contents, the audit report contents and report modification traces within a preset period of a single disease, as well as patient-related clinical information; according to a single disease, locating the modified content and extracting all typos based on the preliminary report content, the audit report content and the report modification traces; classifying and labeling the report modification traces, the location of the modified content, all typos, and the modified content to generate the blacklist medical typo database.

[0009] Preferably, the method also includes: identifying semantic errors, the steps of which are: obtaining preliminary report statements that need to be tested; obtaining knowledge base decision rules; based on clinical background and medical knowledge, using regular expressions to identify semantic errors, and returning to the report writing interface for reference by the reporting doctor.

[0010] Preferably, the obtaining of the preliminary report sentence to be detected and comparing it with the blacklist medical typo database comprises: traversing the preliminary report sentence using a directed acyclic word graph model, searching the blacklist medical typo database for all rules that satisfy the preset minimum match length and maximum match length based on the preset minimum match length and maximum match length; sorting all the searched rules according to weights; and matching the preliminary report sentence with the sorted rules one by one to determine whether there are typos in the preliminary report sentence.

[0011] Preferably, the acquisition of knowledge base decision rules is based on clinical background and medical knowledge, and uses regular expressions to identify semantic errors, including: using a directed acyclic word graph model to traverse the preliminary report statement, based on a preset minimum match length and a maximum match length, searching from the knowledge base decision rules for all rules that meet the preset minimum match length and the maximum match length; sorting all searched rules according to weights; and matching the preliminary report statement with the sorted rules one by one to determine whether there are semantic errors in the preliminary report statement.

[0012] Preferably, the method also includes: optimizing error recognition, including: recording whether the reporting doctor recognizes and uses the identified error content and recommended modification content; based on the identified error content and recommended modification content used by the reporting doctor, counting the usage frequency of the blacklist medical typo database and the knowledge base decision rules to change the weights of the corresponding rules and optimize error recognition.

[0013] On the other hand, the present invention also provides a typo detection and reminder system based on report writing content, the system comprising: a blacklist database module, a white-name word library module, a first typo recognition module, a second typo recognition module and a modified content extraction module, wherein the blacklist data module is connected to the first typo recognition module, and is used to create a blacklist medical typo database based on preliminary report content, audit report content and report modification traces; the white-name word library module is respectively connected to the second typo recognition module and the modified content extraction module, and is used to perform word segmentation processing on the audit report content to generate a whitelist NLP word library; the first typo recognition module is respectively connected to the blacklist data module, the second typo recognition module and the modified content extraction module, and is used to obtain a preliminary report sentence to be detected, and compare it with the blacklist medical typo database. If the characters and / or words in the preliminary report sentence are in the blacklist medical typo database, If a character and / or word is found in the blacklist medical typo database, it is determined that the character and / or word is a typo, which is defined as the first typo. At this time, the correct content corresponding to the first typo is extracted from the blacklist medical typo database, and the first typo is replaced by the correct content, and the modified preliminary report sentence is output; the second typo recognition module is respectively connected to the first typo recognition module, the white-name word library module, and the modified content extraction module, and is used to perform white-list word segmentation on the modified preliminary report sentence based on the white-list NLP word library. If a single character that cannot be segmented appears, the single character is determined to be a typo, which is defined as the second typo; the modified content extraction module is respectively connected to the first typo recognition module, the second typo recognition module, and the white-name word library module, and is used to extract the word with the highest probability from the white-list NLP word library according to pinyin as the modified content of the first typo and the second typo, and return it to the report writing interface for reference by the reporting doctor.

[0014] Preferably, the system also includes a verification module, which is respectively connected to the second typo recognition module and the modified content extraction module, and is used to verify whether the single word is a typo through a preset model calculation. If the probability of the single word occurring reaches a preset threshold, the single word is finally determined to be a typo, that is, the second typo; wherein the preset model is a probabilistic language model trained based on the content of the audit report.

[0015] Preferably, the system also includes a knowledge base decision rule module and a semantic error recognition module, wherein the knowledge base decision rule module is respectively connected to the semantic error recognition module and the modified content extraction module, and is used to store the knowledge base decision rules; the semantic error recognition module is respectively connected to the knowledge base decision rule module and the modified content extraction module, and is used to obtain the preliminary report statements that need to be tested, obtain the knowledge base decision rules, and use regular expressions to identify semantic errors based on clinical background and medical knowledge, and return them to the report writing interface for reference by the reporting doctor.

[0016] Technical effects of the present invention:

[0017] 1. The method of the present invention makes full use of the content and modification traces of the preliminary report and the audit report to construct a high-quality blacklist medical typo database, combined with conventional typo recognition technology, to solve the problem that in the field of medical text error recognition, non-medical corpus models are difficult to handle, while medical corpus is difficult to organize and annotate, thereby improving the quality of imaging reports and the work efficiency of reporting doctors, and reducing the burden on reporting doctors;

[0018] 2. The method of the present invention verifies whether the single word is a typo by calculating through a preset model. If the probability of the single word reaching a preset threshold, the single word is finally determined to be a typo. It can also be used to supplement the determination of whether the single word not in the whitelist NLP vocabulary is a typo, thereby preventing the possibility of false alarms.

[0019] 3. The method of the present invention incorporates an auxiliary recognition method of medical knowledge base rules and uses a directed acyclic word graph model to improve recognition efficiency and enhance recognition quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0021] Figure 1 A flow chart of a method for detecting and reminding typos based on report writing content according to Embodiment 1 of the present invention is shown;

[0022] Figure 2 A schematic diagram of the interface of the preliminary report, the review report and the modification traces in the method for detecting and reminding typos in report writing content according to the first embodiment of the present invention is shown;

[0023] Figure 3 A schematic diagram of the structure of a system for detecting and reminding typos based on report writing content according to Embodiment 2 of the present invention is shown;

[0024] Figure 4A schematic diagram of the structure of a system for detecting and reminding typos based on report writing content according to Embodiment 3 of the present invention is shown;

[0025] Figure 5 A schematic diagram of the structure of a system for detecting and reminding typos based on report writing content according to a fourth embodiment of the present invention is shown;

[0026] Figure 6 A schematic diagram of the interface of the preliminary report, the review report and the modification traces in the system for detecting and reminding typos in report writing content according to the fourth embodiment of the present invention is shown. DETAILED DESCRIPTION

[0027] The present invention will be described in detail below with reference to the accompanying drawings and in combination with embodiments.

[0028] Embodiment 1

[0029] Figure 1 FIG. 1 shows a flow chart of a method for detecting and reminding typos based on report writing content according to Embodiment 1 of the present invention; Figure 1 As shown, the method comprises the following steps:

[0030] Overall description: First, typos are identified based on the blacklist medical typo database, and then the correct content is extracted from the blacklist medical typo database for correction. Then, the modified complete text report is segmented according to the whitelist NLP vocabulary, and finally the N-Gram model is used to supplement the detection of the probability that the remaining single words are typos.

[0031] Step S101, creating a blacklist medical typo database based on the preliminary report content, the audit report content and the report modification traces;

[0032] The steps of creating a blacklist medical typo database include:

[0033] Acquisition of report content: Acquisition of all the preliminary report contents, audit report contents and report modification traces within the preset period of a single disease, as well as relevant clinical information of the patient;

[0034] Since imaging reports are templates set according to disease types, when creating a blacklist medical typo database, they are also classified according to single disease types, such as lung nodule imaging reports, prostate cancer imaging reports, gastric cancer imaging reports, etc.

[0035] Corpus collation: According to the single disease type, based on the content of the preliminary report, the content of the audit report and the report modification traces, locate the modified content and extract all typos;

[0036] Typos marking: Classify and mark the report modification traces, the location of the modified content, all typos, and the modified content to generate the blacklist medical typos database.

[0037] Step S102, performing word segmentation processing on the audit report content to generate a whitelist NLP word library;

[0038] Step S103, obtaining a preliminary report sentence to be detected, and comparing it with the blacklist medical typo database. If a character and / or word in the preliminary report sentence is a character and / or word in the blacklist medical typo database, it is determined that the character and / or word is a typo, which is defined as a first typo. At this time, the correct content corresponding to the first typo is extracted from the blacklist medical typo database, and the first typo is replaced with the correct content, and the modified preliminary report sentence is output;

[0039] Step S104, based on the whitelist NLP word library, the modified preliminary report sentence is subjected to whitelist word segmentation, and if a single word that cannot be segmented appears, the single word is determined to be a wrong character, which is defined as a second wrong character;

[0040] Step S105, extracting the word with the highest probability from the whitelist NLP vocabulary according to the pinyin as the modification content of the first typo and the second typo, and returning it to the report writing interface for reference by the reporting doctor.

[0041] In the report writing interface, the reporting doctor is reminded of possible errors, the location of the errors, the reasons for the errors, and recommended modifications. The reporting doctor can accept the modification suggestion or reject it, and the system will count the doctor's selection results.

[0042] Types of typo reminders:

[0043] Spelling errors: adjacent pinyin, homophones

[0044] Numerical unit detection

[0045] Incorrect medical content, outside the defined categories, such as incorrect TNM or pirads staging.

[0046] Content beyond the current anatomical site.

[0047] The method further includes: verifying whether the single word is a typo by calculating through a preset model; if the probability of occurrence of the single word reaches a preset threshold, finally determining that the single word is a typo, i.e., the second typo; wherein the preset model is a probabilistic language model trained based on the content of the audit report.

[0048] Among them, the probabilistic language model is an N-Gram model. After training this model with the text content of the audit report, it can be used to calculate the probability of occurrence of each word in the input sentence. The lower the probability, the higher the possibility of error. The technical solution of the present invention is mainly based on black and white list word segmentation recognition, and the N-Gram model is used as a supplementary recognition to make the quality of typos recognition higher. The N-Gram model is mainly used to supplement the judgment of whether a single word that is not in the white list NLP vocabulary is a typo.

[0049] The design of this application is characterized by being able to simultaneously obtain the content of the medical text report before modification, the modification traces, and the content after modification, so as to obtain the real error situation. The conventional error dictionary construction usually uses the method of analyzing low-frequency words, and the conventional model training method is to randomly replace the wrong characters, and these methods cannot reflect the real error characteristics.

[0050] The method further includes: identifying semantic errors, the steps of which are:

[0051] Get the preliminary report statement that needs to be tested;

[0052] Obtaining knowledge base decision rules;

[0053] Based on clinical background and medical knowledge, regular expressions are used to identify semantic errors and return them to the report writing interface for reference by the reporting doctor.

[0054] For example, it can be set that the length unit "m (meter)" that appears alone is wrong, because this length unit usually does not appear in the report. It is recommended to change it to "cm".

[0055] Among them, the obtaining of the preliminary report sentence to be detected and comparing it with the blacklist medical typo database includes: using a directed acyclic word graph model to traverse the preliminary report sentence, based on a preset minimum matching length and a maximum matching length, searching the blacklist medical typo database for all rules that meet the preset minimum matching length and the maximum matching length; sorting all the searched rules according to the weights; matching the preliminary report sentence with the sorted rules one by one to determine whether there are typos in the preliminary report sentence.

[0056] Among them, the acquisition of knowledge base decision rules is based on clinical background and medical knowledge, and uses regular expressions to identify semantic errors, including: using a directed acyclic word graph model to traverse the preliminary report statement, based on a preset minimum matching length and a maximum matching length, searching from the knowledge base decision rules for all rules that meet the preset minimum matching length and the maximum matching length; sorting all searched rules according to weights; matching the preliminary report statement with the sorted rules one by one to determine whether there are semantic errors in the preliminary report statement.

[0057] The blacklist medical typos database and knowledge base decision rules contain prefixes, matching word count requirements, and weight-related information. A directed acyclic word graph model is used to quickly match rules and improve the efficiency of identifying errors. A directed acyclic word graph (DAWG) is a data structure that can quickly search for words and can quickly search for targets containing specified content. When checking whether a certain report text has typos, when traversing the text, you can quickly search for possible rules from all rules and sort the rules by weight, and then match the rules one by one to improve efficiency. When the text length does not meet the rule requirements, the rule will not be matched.

[0058] For example, to check for unpaired parentheses (only left parentheses, no right parentheses), the rule can be set to use the regular expression:

[0059] ([^,,.。;;!!??)]+$ and the minimum matching length is 50 and the maximum matching length is 60.

[0060] The meaning of this regular expression is that it starts with a left bracket and there is no right bracket until the end of the text. With the constraint of a minimum matching length of 50, the match will be sent to the end of the traversed sentence or paragraph, otherwise the early match may get incorrect results (because there is still a chance that the right bracket will appear later). The maximum matching length of 60 can prevent the endless matching from reducing efficiency in some unexpected situations when the length of the traversed short paragraph of text continues to grow but the rule has not been met. In fact, when the length of this short paragraph of text grows to the point where there is no longer any rule that may be matched, the traversal will stop early.

[0061] Finally, we get the content, location, cause and recommended modification of the typo.

[0062] Among them, the method also includes: optimizing error recognition, including: recording whether the reporting doctor recognizes and uses the identified error content and recommended modification content; based on the identified error content and recommended modification content used by the reporting doctor, counting the frequency of use of the blacklist medical typo database and the knowledge base decision rules to change the weights of the corresponding rules and optimize error recognition.

[0063] The following uses an example to illustrate the specific implementation process of this embodiment:

[0064] Figure 2 Shows a schematic diagram of the preliminary report, review report, and modification trace interface in the method for detecting and reminding of typos in the content of a report writing according to Embodiment 1 of the present invention, as Figure 2 shown

[0065] 1. Part of the preliminary report content

[0066] 1. The esophagus is significantly dilated, and food residue is retained inside. Please combine with gastrointestinal imaging.

[0067] Part of the review report content

[0068] 1. The esophagus is significantly dilated, and food residue is retained inside. Please combine with gastrointestinal imaging.

[0069] Report trace: retained -> retention

[0070] 2. Organize the corpus, segment the words and add them to the whitelist NLP word library; train the N-Gram model

[0071] 3. Confirm "retained -> retention", mark it as a typo, and add it to the blacklist medical typo database, where the prefix is set to "retained", and both the maximum matching length and the minimum matching length are set to 2. (The above is the process of creating the blacklist medical typo database)

[0072] 4. Create decision rules for other knowledge bases;

[0073] 5. Establish a directed acyclic word graph (DAWG) for the prefixes of the rules and the blacklist medical typo database for searching;

[0074] 6. Given the preliminary report statement to be detected, perform typo analysis, for example, "There is serious food residue retention inside, XXX, and PIRADS level 7 can be considered";

[0075] When traversing to the position of the word "retained", continue to traverse "retention", "retention serious", and "retention seriously" in sequence. When traversing to "retention", a rule with a prefix match of "retention" is searched from the directed acyclic word graph, so one typo is judged.

[0076] When matching to PIRADS level 7, the rule for checking the PIRADS content in the decision rule of the knowledge base is triggered by the prefix "PIRADS", and it is found that PIRADS level 7 is not the correct PIRADS classification, so another error is judged.

[0077] Use the whitelist NLP word library to segment words and combine with the N-Gram model for recognition, and no other typos are detected;

[0078] 7. The recognition results are returned to the report writing interface. The doctor adopts the suggestion to change "storage" to "retention", and the corresponding rule weight is increased.

[0079] The embodiment of the present invention makes full use of the content and modification traces of the preliminary report and the audit report to construct a high-quality blacklist medical typo database, combined with conventional typo recognition technology to solve the problem that in the field of medical text error recognition, non-medical corpus models are difficult to handle, while medical corpus is difficult to organize and annotate, thereby improving the quality of image reports and the work efficiency of reporting doctors, and reducing the burden on reporting doctors; the embodiment of the present invention verifies whether the single word is a typo through a preset model calculation, and if the probability of the single word reaching a preset threshold, the single word is finally determined to be a typo, and it can be supplemented to determine whether the single word not in the whitelist NLP vocabulary is a typo, thereby preventing the possibility of false alarms; the embodiment of the present invention adds an auxiliary recognition method of medical knowledge base rules, uses a directed acyclic word graph model to improve recognition efficiency, and improves recognition quality.

[0080] Embodiment 2

[0081] Overall description: First, typos are identified based on the blacklist medical typo database, and then the correct content is extracted from the blacklist medical typo database for correction. Then, the modified complete text report is segmented according to the whitelist NLP vocabulary, and finally the N-Gram model is used to supplement the detection of the probability that the remaining single words are typos.

[0082] Figure 3 FIG. 2 shows a schematic diagram of the structure of a system for detecting and reminding typos in report writing content according to Embodiment 2 of the present invention; Figure 3 As shown, the system includes: a blacklist database module 10, a white-name word library module 20, a first typo recognition module 30, a second typo recognition module 40 and a modified content extraction module 50, wherein:

[0083] The blacklist data module 10 is connected to the first typo identification module 30 and is used to create a blacklist medical typo database based on the preliminary report content, the audit report content and the report modification traces;

[0084] The step of creating a blacklist medical typo database includes:

[0085] Acquisition of report content: Acquisition of all the preliminary report contents, audit report contents and report modification traces within the preset period of a single disease, as well as relevant clinical information of the patient;

[0086] Since imaging reports are templates set according to disease types, when creating a blacklist medical typo database, they are also classified according to single disease types, such as lung nodule imaging reports, prostate cancer imaging reports, gastric cancer imaging reports, etc.

[0087] Corpus collation: According to the single disease type, based on the content of the preliminary report, the content of the audit report and the report modification traces, locate the modified content and extract all typos;

[0088] Typos marking: Classify and mark the report modification traces, the location of the modified content, all typos, and the modified content to generate the blacklist medical typos database.

[0089] The white-list word library module 20 is connected to the second wrong character recognition module 40 and the modified content extraction module 50 respectively, and is used to perform word segmentation processing on the audit report content to generate a white-list NLP word library;

[0090] The first wrong character recognition module 30 is connected to the blacklist data module 10, the second wrong character recognition module 40, and the modified content extraction module 50, respectively, and is used to obtain a preliminary report sentence to be detected, and compare it with the blacklist medical wrong character database. If the character and / or word in the preliminary report sentence is a character and / or word in the blacklist medical wrong character database, it is determined that the character and / or word is a wrong character, which is defined as a first wrong character. At this time, the correct content corresponding to the first wrong character is extracted from the blacklist medical wrong character database, and the first wrong character is replaced with the correct content, and the modified preliminary report sentence is output;

[0091] Among them, the obtaining of the preliminary report sentence to be detected and comparing it with the blacklist medical typo database includes: using a directed acyclic word graph model to traverse the preliminary report sentence, based on a preset minimum matching length and a maximum matching length, searching the blacklist medical typo database for all rules that meet the preset minimum matching length and the maximum matching length; sorting all the searched rules according to the weights; matching the preliminary report sentence with the sorted rules one by one to determine whether there are typos in the preliminary report sentence.

[0092] The second wrong character recognition module 40 is connected to the first wrong character recognition module 30, the white-list word library module 20, and the modified content extraction module 50, respectively, and is used to perform white-list word segmentation on the modified preliminary report sentence based on the white-list NLP word library. If a single word that cannot be segmented appears, it is determined that the single word is a wrong character, which is defined as a second wrong character;

[0093] The modification content extraction module 50 is respectively connected to the first typo recognition module 30, the second typo recognition module 40, and the white-list word library module 20, and is used to extract the word with the highest probability from the white-list NLP word library according to pinyin as the modification content of the first typo and the second typo, and return it to the report writing interface for reference by the reporting doctor.

[0094] In the report writing interface, the reporting doctor is reminded of possible errors, the location of the errors, the reasons for the errors, and recommended modifications. The reporting doctor can accept the modification suggestion or reject it, and the system will count the doctor's selection results.

[0095] Types of typo reminders:

[0096] Spelling errors: adjacent pinyin, homophones

[0097] Numerical unit detection

[0098] Incorrect medical content, outside the defined categories, such as incorrect TNM or pirads staging.

[0099] Content beyond the current anatomical site.

[0100] The design of this application is characterized by being able to simultaneously obtain the content of the medical text report before modification, the modification traces, and the content after modification, so as to obtain the real error situation. The conventional error dictionary construction usually uses the method of analyzing low-frequency words, and the conventional model training method is to randomly replace the wrong characters, and these methods cannot reflect the real error characteristics.

[0101] Among them, the system also includes: optimizing error recognition, including: recording whether the reporting doctor recognizes and uses the identified error content and recommended modification content; based on the identified error content and recommended modification content used by the reporting doctor, counting the frequency of use of the blacklist medical typo database and the knowledge base decision rules to change the weights of the corresponding rules and optimize error recognition.

[0102] The embodiments of the present invention make full use of the content and modification traces of preliminary reports and audit reports to construct a high-quality blacklist medical typo database, combined with conventional typo recognition technology to solve the problem that in the field of medical text error recognition, non-medical corpus models are difficult to perform, while medical corpus is difficult to organize and annotate, thereby improving the quality of imaging reports and the work efficiency of reporting doctors, and reducing the burden on reporting doctors.

[0103] Embodiment 3

[0104] Figure 4 FIG. 4 shows a schematic diagram of the structure of a system for detecting and reminding typos in report writing content according to Embodiment 3 of the present invention; Figure 4 As shown, the system also includes a verification module 60, which is respectively connected to the second typo recognition module 40 and the modified content extraction module 50, and is used to verify whether the single word is a typo through a preset model calculation. If the probability of the single word occurring reaches a preset threshold, the single word is finally determined to be a typo, i.e., the second typo; wherein the preset model is a probabilistic language model trained based on the content of the audit report.

[0105] Among them, the probabilistic language model is an N-Gram model. After training this model with the text content of the audit report, it can be used to calculate the probability of occurrence of each word in the input sentence. The lower the probability, the higher the possibility of error. The technical solution of the present invention is mainly based on black and white list word segmentation recognition, and the N-Gram model is used as a supplementary recognition to make the quality of typos recognition higher. The N-Gram model is mainly used to supplement the judgment of whether a single word that is not in the white list NLP vocabulary is a typo.

[0106] The method of the present invention verifies whether the single word is a typo through calculation using a preset model. If the probability of occurrence of the single word reaches a preset threshold, the single word is finally determined to be a typo. Whether the single word not in the whitelist NLP vocabulary is a typo can be supplemented to determine whether it is a typo, thereby preventing the possibility of false alarms.

[0107] Embodiment 4

[0108] Figure 5 FIG. 4 shows a schematic diagram of the structure of a system for detecting and reminding typos in report writing content according to Embodiment 4 of the present invention; Figure 5 As shown, the system also includes a knowledge base decision rule module 70 and a semantic error recognition module 80, wherein:

[0109] The knowledge base decision rule module 70 is connected to the semantic error identification module 80 and the modified content extraction module 50 respectively, and is used to store the knowledge base decision rules;

[0110] The semantic error recognition module 80 is respectively connected to the knowledge base decision rule module 70 and the modified content extraction module 50, and is used to obtain the preliminary report statements that need to be tested, obtain the knowledge base decision rules, and use regular expressions based on clinical background and medical knowledge to identify semantic errors, and return them to the report writing interface for reference by the reporting doctor.

[0111] For example, it can be set that the length unit "m (meter)" that appears alone is wrong, because this length unit usually does not appear in the report. It is recommended to change it to "cm".

[0112] Among them, the acquisition of knowledge base decision rules is based on clinical background and medical knowledge, and uses regular expressions to identify semantic errors, including: using a directed acyclic word graph model to traverse the preliminary report statement, based on a preset minimum matching length and a maximum matching length, searching from the knowledge base decision rules for all rules that meet the preset minimum matching length and the maximum matching length; sorting all searched rules according to weights; matching the preliminary report statement with the sorted rules one by one to determine whether there are semantic errors in the preliminary report statement.

[0113] The blacklist medical typos database and knowledge base decision rules contain prefixes, matching word count requirements, and weight-related information. A directed acyclic word graph model is used to quickly match rules and improve the efficiency of identifying errors. A directed acyclic word graph (DAWG) is a data structure that can quickly search for words and can quickly search for targets containing specified content. When checking whether a certain report text has typos, when traversing the text, you can quickly search for possible rules from all rules and sort the rules by weight, and then match the rules one by one to improve efficiency. When the text length does not meet the rule requirements, the rule will not be matched.

[0114] For example, to check for unpaired parentheses (only left parentheses, no right parentheses), the rule can be set to use the regular expression:

[0115] ([^,,.。;;!!??)]+$ and the minimum matching length is 50 and the maximum matching length is 60.

[0116] The meaning of this regular expression is that it starts with a left bracket and there is no right bracket until the end of the text. With the constraint of a minimum matching length of 50, the match will be sent to the end of the traversed sentence or paragraph, otherwise the early match may get incorrect results (because there is still a chance that the right bracket will appear later). The maximum matching length of 60 can prevent the endless matching from reducing efficiency in some unexpected situations when the length of the traversed short paragraph of text continues to grow but the rule has not been met. In fact, when the length of this short paragraph of text grows to the point where there is no longer any rule that may be matched, the traversal will stop early.

[0117] Finally, we get the content, location, cause and recommended modification of the typo.

[0118] The specific implementation process of this embodiment is described below in the form of examples:

[0119] Figure 6 FIG. 4 shows a schematic diagram of the interface of the preliminary report, the audit report and the modification traces in the system for detecting and reminding typos in report writing according to the fourth embodiment of the present invention. Figure 6As shown

[0120] 1. Part of the preliminary report

[0121] 1. The esophagus is significantly dilated, and food residue is retained. Please combine with gastrointestinal imaging.

[0122] Part of the review report

[0123] 1. The esophagus is significantly dilated, and food residue is retained. Please combine with gastrointestinal imaging.

[0124] Report trace: retained -> retention

[0125] 2. Organize the corpus, segment the words and add them to the whitelist NLP word library; train the N-Gram model

[0126] 3. Confirm "retained" -> "retention", mark it as a typo, and add it to the blacklist medical typo database. Set the prefix as "retained", and both the maximum matching length and the minimum matching length are set to 2. (The above is the process of creating the blacklist medical typo database)

[0127] 4. Create decision rules for other knowledge bases;

[0128] 5. Build a directed acyclic word graph (DAWG) for the prefixes of the rules and the blacklist medical typo database for searching;

[0129] 6. Perform typo analysis on the given preliminary report statements to be detected, such as "There is serious food residue retention inside, XXX, and PIRADS grade 7 can be considered";

[0130] When traversing to the position of the character "retained", continue to traverse "retention", "retention serious", and "retention seriously" in sequence. When traversing to "retention", search for the "retention" rule with prefix matching from the directed acyclic word graph, and thus determine a typo.

[0131] When matching to the position of PIRADS grade 7, trigger the rule for checking the PIRADS content in the knowledge base decision rules through the prefix "PIRADS", and find that PIRADS grade 7 is not the correct PIRADS classification, so another error is determined.

[0132] Use the whitelist NLP word library for word segmentation and combine with the N-Gram model for recognition, and no other typos are detected;

[0133] 7. Return the recognition result to the report writing interface. The doctor adopted the suggestion of modifying "retained" to "retention", and the corresponding rule weight increased.

[0134] The embodiment of the present invention adds an auxiliary recognition method for medical knowledge base rules, uses a directed acyclic word graph model to improve the recognition efficiency, and improves the recognition quality.

[0135] From the above description, it can be seen that the above-mentioned embodiments of the present invention achieve the following technical effects: the embodiments of the present invention make full use of the content and modification traces of the preliminary report and the audit report to construct a high-quality blacklist medical typo database, combined with conventional typo recognition technology to solve the problem that in the field of medical text error recognition, non-medical corpus models are difficult to cope with, and medical corpus is difficult to organize and annotate, thereby improving the quality of imaging reports and the work efficiency of reporting doctors, and reducing the burden on reporting doctors; the embodiments of the present invention calculate and verify whether the single word is a typo through a preset model. If the probability of the single word reaching a preset threshold, the single word is finally determined to be a typo, and the single word not in the whitelist NLP vocabulary can be supplemented to determine whether it is a typo, thereby preventing the possibility of false alarms; the embodiments of the present invention add an auxiliary recognition method of medical knowledge base rules, and use a directed acyclic word graph model to improve recognition efficiency and recognition quality.

[0136] Obviously, those skilled in the art should understand that the above modules or steps of the present invention can be implemented by a general computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, and optionally, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. Thus, the present invention is not limited to any specific combination of hardware and software.

[0137] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for detecting and reminding typos based on report writing content, characterized in that: include: Create a blacklist medical typo database based on the preliminary report content, audit report content and report modification traces; Perform word segmentation on the content of the audit report to generate a whitelist NLP vocabulary; Obtain a preliminary report sentence to be detected, and compare it with the blacklist medical typo database. If a character and / or word in the preliminary report sentence is a character and / or word in the blacklist medical typo database, it is determined that the character and / or word is a typo, which is defined as a first typo. At this time, the correct content corresponding to the first typo is extracted from the blacklist medical typo database, and the first typo is replaced with the correct content, and the modified preliminary report sentence is output; Based on the whitelist NLP word library, the modified preliminary report sentence is subjected to whitelist word segmentation, and if a single word that cannot be segmented appears, the single word is determined to be a wrong character, which is defined as a second wrong character; The word with the highest probability is extracted from the whitelist NLP vocabulary according to the pinyin as the modification content of the first typo and the second typo, and returned to the report writing interface for reference by the reporting doctor.

2. The method for detecting and reminding typos based on report writing content according to claim 1 is characterized in that: The method also includes: verifying whether the single word is a typo by calculating through a preset model, and if the probability of the single word appearing reaches a preset threshold, finally determining that the single word is a typo, i.e., the second typo; wherein the preset model is a probabilistic language model trained based on the content of the audit report.

3. The method for detecting and reminding typos based on report writing content according to claim 1 is characterized in that: The steps of creating a blacklist medical typo database include: Obtain all the preliminary report contents, review report contents and report modification traces within the preset period of a single disease, as well as relevant clinical information of the patient; According to the single disease type, based on the content of the preliminary report, the content of the audit report and the report modification traces, locate the modified content and extract all typos; The report modification traces, the location of the modified content, all typos, and the modified content are classified and annotated to generate the blacklist medical typo database.

4. The method for detecting and reminding typos based on report writing content according to claim 1, characterized in that: The method further includes: identifying semantic errors, the steps of which are: Get the preliminary report statement that needs to be tested; Obtaining knowledge base decision rules; Based on clinical background and medical knowledge, regular expressions are used to identify semantic errors and return them to the report writing interface for reference by the reporting doctor.

5. The method for detecting and reminding typos based on report writing content according to claim 3 is characterized in that: The obtaining of the preliminary report sentence to be detected and comparing it with the blacklist medical typo database includes: Using a directed acyclic word graph model to traverse the preliminary report sentence, based on a preset minimum matching length and a maximum matching length, searching the blacklist medical typos database for all rules that meet the preset minimum matching length and the preset maximum matching length; Sort all searched rules by weight; The preliminary report statement is matched against the sorted rules one by one to determine whether there are typos in the preliminary report statement.

6. The method for detecting and reminding typos based on report writing content according to claim 4, characterized in that: The acquisition of knowledge base decision rules, based on clinical background and medical knowledge, uses regular expressions to identify semantic errors, including: Using a directed acyclic word graph model to traverse the preliminary report sentence, based on a preset minimum match length and a maximum match length, searching all rules that meet the preset minimum match length and the maximum match length from the knowledge base decision rules; Sort all searched rules by weight; The preliminary report statement is matched against the sorted rules one by one to determine whether there is a semantic error in the preliminary report statement.

7. The method for detecting and reminding typos based on report writing content according to claim 5 or 6, characterized in that: The method further includes: optimizing error identification, including: Record whether the reporting physician acknowledged and used the identified errors and recommended modifications; Based on the identified errors and recommended modifications used by the reporting physician, the usage frequencies of the blacklist medical typos database and the knowledge base decision rules are counted to change the weights of the corresponding rules and optimize error identification.

8. A system for detecting and reminding typos in report writing, characterized in that: The system includes: a blacklist database module, a white-name word library module, a first typo recognition module, a second typo recognition module and a modified content extraction module, wherein: The blacklist data module is connected to the first typo recognition module and is used to create a blacklist medical typo database based on the preliminary report content, the audit report content and the report modification traces; The white-list word library module is connected to the second wrong character recognition module and the modified content extraction module respectively, and is used to perform word segmentation processing on the audit report content to generate a white-list NLP word library; The first typo recognition module is connected to the blacklist data module, the second typo recognition module, and the modified content extraction module, respectively, and is used to obtain a preliminary report sentence that needs to be detected, and compare it with the blacklist medical typo database. If a character and / or word in the preliminary report sentence is a character and / or word in the blacklist medical typo database, it is determined that the character and / or word is a typo, which is defined as a first typo. At this time, the correct content corresponding to the first typo is extracted from the blacklist medical typo database, and the first typo is replaced by the correct content, and the modified preliminary report sentence is output; The second wrong character recognition module is connected to the first wrong character recognition module, the white-list word library module, and the modified content extraction module, respectively, and is used to perform white-list word segmentation on the modified preliminary report sentence based on the white-list NLP word library. If a single word that cannot be segmented appears, the single word is determined to be a wrong character, which is defined as a second wrong character; The modification content extraction module is respectively connected to the first typo recognition module, the second typo recognition module, and the white-list word library module, and is used to extract the word with the highest probability from the white-list NLP word library according to pinyin as the modification content of the first typo and the second typo, and return it to the report writing interface for reference by the reporting doctor.

9. The system for detecting and reminding typos based on report writing content according to claim 8, characterized in that: The system also includes a verification module, which is respectively connected to the second typo recognition module and the modified content extraction module, and is used to verify whether the single word is a typo through a preset model calculation. If the probability of the single word occurring reaches a preset threshold, the single word is finally determined to be a typo, i.e., the second typo; wherein the preset model is a probabilistic language model trained based on the content of the audit report.

10. The system for detecting and reminding typos based on report writing content according to claim 8, characterized in that: The system also includes a knowledge base decision rule module and a semantic error recognition module, wherein: The knowledge base decision rule module is connected to the semantic error recognition module and the modified content extraction module respectively, and is used to store the knowledge base decision rules; The semantic error recognition module is respectively connected to the knowledge base decision rule module and the modified content extraction module, and is used to obtain the preliminary report sentences that need to be tested, obtain the knowledge base decision rules, and use regular expressions based on clinical background and medical knowledge to identify semantic errors, and return them to the report writing interface for reference by the reporting doctor.

Images