Physical examination table full quantity intelligent auditing method based on natural language processing

Abstract

The application discloses a physical examination table full quantity intelligent auditing method based on natural language processing, and particularly relates to the technical field of natural language processing. First, the heterogeneous data of a physical examination table of a subject is acquired, and preprocessing is performed to obtain a standardized heterogeneous data set. A first structured fact list is generated based on the first data source after standardization, and a second structured fact list is obtained by reconstruction based on the second data source after standardization. Then, the first structured fact list and the second structured fact list are associated based on the unique identification of the physical examination table of the subject, and multi-dimensional logical consistency comparison is performed. An intelligent auditing result is obtained according to the comparison result. Finally, an intelligent auditing report is generated based on the intelligent auditing result, and human-computer interaction is performed. With the powerful computing power of the AI model, the application realizes full quantity automatic auditing of the physical examination table, does not need manual sampling, improves the auditing efficiency, reduces the misjudgment rate, and is suitable for the auditing scene of large-scale physical examination data.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, specifically to a method for full intelligent review of medical examination forms based on natural language processing. Background Technology

[0002] In the physical examination work of basic public health services, the standardized review of the health evaluation on the physical examination form is a key link to ensure the quality of physical examination data and implement the assessment requirements of public health services. During the physical examination, doctors are required to describe any abnormal indicators found in a standardized manner in the "health evaluation" column. During the supervision and management process, the standardization of the completion of the health evaluation content will be the focus of verification. Currently, the existing methods for reviewing health evaluations on physical examination forms mainly include: automatic review based on keyword recognition, which determines whether the review is qualified by matching keywords corresponding to preset abnormal indicators with keywords in the health evaluation content; manual sampling review, which requires professional reviewers to conduct sampling inspections of physical examination forms and manually judge the standardization of the health evaluation content; and automated text review technology, whose paradigm is usually to perform grammatical, semantic, or consistency checks on a single text stream or with a single external reference, thereby realizing the review of health evaluations on physical examination forms.

[0003] Existing methods for reviewing physical examination forms still have some shortcomings. For example, keyword-based rule matching in keyword review cannot handle personalized expressions and typos, nor can it determine whether the text matches specific numerical values, resulting in a high false positive rate and poor review quality. Manual sampling is costly and inefficient, and the sampling results cannot represent the overall quality of all data, leaving quality blind spots and failing to meet the management requirements of full-volume quality control. Therefore, a sophisticated intelligent review solution is needed to efficiently, accurately, and comprehensively review massive amounts of unstructured, diverse health evaluation texts. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method for full intelligent review of medical examination forms based on natural language processing, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for full intelligent review of medical examination forms based on natural language processing, comprising: S1: Obtain heterogeneous data from the examinee's physical examination form, including the first data source and the second data source, and preprocess the two types of data respectively to obtain the preprocessed standardized heterogeneous dataset. S2: Based on the standardized heterogeneous dataset, the numerical physical examination indicators in the standardized first data source are analyzed one by one and abnormal indicators are identified. The first structured fact list is generated based on the identification results. S3: Based on the standardized heterogeneous dataset, the standardized second data source is input into the natural language processing model that has been fine-tuned by the physical examination text. The model performs two core tasks: entity and relation extraction, standardization and structured output, and reconstructs the second structured claim list. S4: Based on the unique identifier of the examinee's medical examination form, the first structured fact list and the second structured claim list are associated, and a multi-dimensional logical consistency comparison is performed. The intelligent review result is obtained based on the comparison result. S5: Based on the intelligent audit results, a personalized intelligent audit report is generated for each examinee, and human-computer interaction is conducted with the supervision and management terminal through a visual interactive interface.

[0006] Preferably, the first data source includes numerical physical examination indicators and non-numerical physical examination indicator conclusion texts, and the second data source is unstructured free text health evaluation.

[0007] Preferably, the first structured fact list: based on a standardized heterogeneous dataset, the standardized first data source is parsed one by one, and the measured values ​​of each physical examination indicator are compared with the preset normal value range corresponding to the indicator, automatically identifying all abnormal indicators that exceed the normal value range; for each abnormal indicator in the numerical physical examination data, the indicator name and measured value are extracted, and the abnormal status is marked according to the value exceeding the range; for the conclusion text of non-numerical physical examination indicators, the standard text word vector B of the corresponding indicator in the preset physical examination qualitative standard comparison library is called, and the string is precisely matched M. str (A,B) and cosine similarity semantic comparison Sim cos Given the combination of (A, B), determine whether the qualitative descriptive word vector A in the physical examination conclusion text is an abnormal qualitative description. If M str If (A,B)=0, then the combination comparison judgment formula is F_Sim(A,B)=Sim cos (A,B) and satisfying F_Sim(A,B)≥s the semantic matching threshold, if M str If (A,B)=1, then the combination comparison judgment formula F_Sim(A,B)=1, indicating that the non-numerical indicator is positive. Output the abnormal indicator name and corresponding conclusion text, generating a structured first structured fact list. obj It includes the examinee's unique identifier, the number of items, the indicator name corresponding to each item, the measurement value, and any abnormal status.

[0008] Preferably, S3 includes: S3.1: processing the standardized second data source T using natural language processing technology. pre By analyzing each sentence according to its punctuation, we obtain m sentences S. j Parsing result Seg(T) pre ), Seg(T pre )={S1,S2,...,S m Then, the sub-word segmentation function SubSeg(S) is used. j )={sub1,sub2,...,sub K}, for each sentence S j Divide into K words (sub) k Then SubSeg(S) j ) and the pre-set core terminology dictionary for physical examinations D med The input is fed into the core function Recog(SubSeg(S) for physical examination entity recognition. j ),D med In the output recognition result set Recog={E ) ind E sta E num E unit ,F num}, E ind E sta E num E unit and F num These are, respectively, the set of physical examination indicator entities, the set of state words corresponding to each indicator entity in the physical examination indicator entity set, the set of numerical values ​​that correspond one-to-one with each indicator entity, the set of units that correspond one-to-one with each indicator entity, and the non-numerical indicator labeling function. If it is a non-numerical indicator F... num =1, using a pre-trained text parsing algorithm, the corresponding set of qualitative positive descriptions, Recog, is identified. num If it is a numerical indicator F num =0.

[0009] Preferably, the implementation of S3 further includes: S3.2: constructing a suspected error identification function De based on character shape similarity and pinyin similarity. err (E ind D med D abb ), D abb Output a set of suspected typos Err based on a pre-defined dictionary of standard abbreviations for physical examination terminology. cha and the collection of suspected non-standard abbreviations Err abb Err cha It needs to satisfy either a character shape similarity ≥ the corresponding threshold th1 or a pinyin similarity ≥ the corresponding threshold th2 and be consistent with D. med Middle and Dmed The terminology used in the Chinese standard is inconsistent and does not meet the criteria for inconsistency in the combined comparison judgment formula F_Sim. Err abb Must satisfy not in D abb And it meets the requirements of similarity in shape or pinyin with the standard abbreviation, M ma This is a natural language processing model fine-tuned based on physical examination text; it uses glyph similarity, pinyin similarity, and S... j Based on the contextual semantics of physical examination, a collaborative error correction function Co is constructed. syn (Err cha Err abb D med D abb ,S j M ma The collaborative error correction function refers to each sentence S j The contextual semantics of the physical examination generate candidate correction results for each suspected error, and then input M. ma Perform semantic feedback verification to obtain the error correction result set Co. res Finally, the results of physical examination entity recognition (Recog) and the non-numerical qualitative positive description set (Recog) are used. num and Co res Construct the entity and relation integration function In(Recog,Recog) num Co res )={E ind corr E sta E num E unit D pos corr}, E ind corr and D pos corr The corrected set of indicator entities and the set of non-numerical qualitative positive descriptions are respectively used to obtain the physical examination entity extraction results, including the correspondence between indicator entity-state word-value-unit and non-numerical indicator entity-qualitative positive description.

[0010] Preferably, the implementation of S3 further includes: S3.3: judging the formula F_Sim(e) by combining and comparing. ind corr ,e ind std The threshold corresponding to )≥ is used as the set of corrected indicator entities E. ind corr entity element e in ind corr Terminology database of pre-set standard physical examination indicators D ind std entity element e inind std The mapping logic yields the standardized entity E. ind std Then, based on the set of state words E corresponding to each indicator entity... sta 1. Pre-set standardized abnormal state label library D sta std Non-numerical index labeling function F num Construct a standardized mapping logic for state words, if F num =1, the set of qualitative positive descriptions D pos corr State words are mapped to D sta std Standard qualitative description D in pos std If F num =0, set of state words E sta One-to-one mapping to D sta std The corresponding standardized abnormal state set E in sta std .

[0011] Preferably, the implementation of S3 further includes: S3.4: Next, perform numerical standardization processing, if the non-numerical index labeling function F num =1, then the claimed numeric field will be marked as "non-numeric" and the unit field will be marked as Ø. If F num =0, set the number E num and the unit set E unit Each element in the set is standardized in terms of both value and unit. If the set of values ​​E num If the element in the formula is Ø, then after standardization it remains Ø, resulting in the standardized set of claimed values ​​E. num std and the corresponding unit set E unit td Next, establish a one-to-one correspondence between standard indicator entities, standardized claim states, standardized claim values, and standardized units, retaining only the relationship with D. ind std The effective information associated with the standard indicators is used to obtain the effective data set Ass after association filtering. cla Finally, using the unique identifier of the examinee, Ass cla The valid data in the data is organized according to a structured format to generate a second structured claim list. cla This includes the subject's unique identifier, indicator name, claimed status, and claimed value.

[0012] Preferably, the intelligent audit result is as follows: S4.1: The first structured fact list List is generated using the examinee's unique identifier ID.obj With the second structured claim list cla ID Perform the association to obtain the dual-list association set (List) corresponding to this ID. obj ID List cla ID ); S4.2: Based on a dual-list associative set (List obj ID List cla ID If List obj ID The set of all indicator names E obj Not a List cla ID The set of all indicator names E ind std If a subset is found, the indicator name is deemed invalid, and the missing indicator name is marked; if the claimed value x of the indicator in the first structured fact list is... sub =Ø, then the numerical value is missing. If the measured value x of the indicator in the first structured fact list is Ø, then the numerical value is missing. no With x sub The absolute difference Δx ≤ the corresponding set error range Δx std If the absolute difference Δx > Δx, then the value is considered acceptable. std Other values ​​are marked as abnormal; the judgment is made by combining and comparing the formula F_Sim(Tag) obj Tag sub The threshold corresponding to ) ≥ is used as the anomalous state tag in the first structured fact list. obj With the anomalous state tag in the second structured claim list sub The description matching logic is as follows: if the condition is met, the description is considered acceptable; otherwise, it is rejected. ind std ⊆(E obj ∪E nor E obj and E nor If the indicators in the subject's first structured fact list are the set of indicators and the set of normal indicators, then the redundancy review is passed; otherwise, it is marked as a redundant evaluation. Finally, if the review results of the four dimensions of indicator name, value, abnormal status, and redundancy are all passed, it means that the health evaluation standard corresponding to the ID is met. If any dimension fails, it is considered non-standard. Based on the subject's unique identifier, the names of the abnormal indicators that passed and failed are summarized along with the problem descriptions. The problem descriptions of the passed ones are "correct evaluation," and the problem descriptions of the failed ones are "marked content," thus obtaining the intelligent review result.

[0013] The technical effects and advantages of this invention are as follows: 1. The NLP model of the present invention has been fine-tuned with personalized corpus and integrated with a collaborative error correction module. It can effectively identify and correct colloquial expressions, common typos, and non-standard abbreviations. Its error tolerance is significantly better than that of existing general NLP models. It solves the problem of misjudgment caused by the inability of existing technologies to cover personalized expressions, and improves the accuracy of indicator extraction and the accuracy of typo / colloquial expression identification and correction. 2. This invention leverages the powerful computing capabilities of AI models to achieve fully automated review of all physical examination forms without the need for manual sampling. This avoids the risks of unrepresentative samples, missed reviews, and incorrect reviews during sampling inspections, while also significantly reducing the human and material resources required for manual review, improving review efficiency, and lowering the error rate. It is suitable for review scenarios involving large-scale physical examination data. 3. This invention addresses the heterogeneity between structured indicator data and unstructured health evaluation text in physical examination forms by transforming unstructured text into a structured list that can be accurately compared with objective indicators. It constructs a multi-dimensional logical consistency review mechanism, which solves the technical defects of existing technologies that cannot achieve in-depth comparison of heterogeneous physical examination data and can only perform simple keyword matching, thus greatly improving the accuracy of review. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the overall process of the present invention.

[0015] Figure 2 This is a schematic diagram of the method flow of the present invention.

[0016] Figure 3 This is a schematic diagram of the second structured claim list reconstruction process of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 As shown, the present invention provides a full-scale intelligent review system for physical examination forms based on natural language processing, including a physical examination form data acquisition module, an objective positive physical examination list generation module, a subjective evaluation physical examination list reconstruction module, a full-scale intelligent review module for physical examination forms, and an intelligent review interaction module for physical examination forms.

[0019] The physical examination form data acquisition module is connected to the objective positive physical examination list generation module and the subjective evaluation physical examination list reconstruction module, respectively. The full-volume intelligent review module of the physical examination form is connected to the objective positive physical examination list generation module, the subjective evaluation physical examination list reconstruction module, and the intelligent review interaction module of the physical examination form, respectively.

[0020] Physical examination form data acquisition module: acquires heterogeneous data from the examinee's physical examination form, including a first data source and a second data source, and preprocesses the two types of data respectively to obtain a preprocessed standardized heterogeneous dataset; Objective positive physical examination list generation module: Based on standardized heterogeneous dataset, it parses the first data source after standardization one by one and identifies abnormal indicators, and generates the first structured fact list based on the identification results; Subjective evaluation medical examination checklist reconstruction module: Based on a standardized heterogeneous dataset, the standardized second data source is input into a natural language processing model that has been fine-tuned by the medical examination text. The model performs two core tasks: entity and relation extraction, standardization and structured output, and reconstructs a second structured claim list. The full-scale intelligent review module for physical examination forms: Based on the unique identifier of the examinee's physical examination form, it associates the first structured fact list with the second structured claim list and performs multi-dimensional logical consistency comparison, and obtains the intelligent review result based on the comparison result; The intelligent review and interaction module for physical examination forms generates a personalized intelligent review report for each examinee based on the intelligent review results, and allows for human-computer interaction with the supervision and management end through a visual interactive interface; Please see Figure 2 As shown, the full-scale intelligent review method for medical examination forms based on natural language processing includes: S1: Acquiring heterogeneous data from the examinee's medical examination forms, including a first data source and a second data source, and preprocessing the two types of data to obtain a preprocessed standardized heterogeneous dataset; S2: Based on the standardized heterogeneous dataset, analyzing each numerical medical examination indicator in the standardized first data source and identifying abnormal indicators, generating a first structured fact list based on the identification results; S3: Based on the standardized heterogeneous dataset, inputting the standardized second data source into a natural language processing model that has been fine-tuned for the medical examination text, and performing two core tasks—entity and relation extraction, standardization, and structured output—through this model to reconstruct a second structured claim list; S4: Based on the unique identifier of the examinee's medical examination form, associating the first structured fact list with the second structured claim list, and performing a multi-dimensional logical consistency comparison, obtaining the intelligent review result based on the comparison result; S5: Based on the intelligent review result, generating a personalized intelligent review report for each examinee, and enabling human-computer interaction with the supervisory management end through a visual interactive interface.

[0021] S1: Obtain heterogeneous data from the examinee's physical examination forms, including a first data source and a second data source, and preprocess the two types of data respectively to obtain a preprocessed standardized heterogeneous dataset, including: S1.1: The first data source includes numerical physical examination indicators and non-numerical physical examination indicator conclusion text (e.g., the electrocardiogram shows complete right bundle branch block). The first data source must at least include the examinee's unique identifier, the name of each physical examination indicator, the indicator measurement value, the normal reference range of the indicator, and the conclusion of the non-numerical physical examination indicator (the normal reference range field is preset to "negative / normal"); the second data source is unstructured free text health evaluation (i.e., the health evaluation content at the end of the physical examination form manually edited by medical staff). This embodiment requires specific explanation of the first data source, such as: structured fields including the examinee's unique identifier (ID number, examinee number, etc.), name, age, gender, and the names, measurement values, and normal reference ranges of 15 physical examination indicators, such as waist circumference, BMI, fasting blood glucose, systolic blood pressure, and diastolic blood pressure (e.g., waist circumference: male <90cm, female <85cm; BMI: 18.5-23.9kg / m²; fasting blood glucose: 3.9-6.1mmol / L); and the second data source (unstructured free text health evaluation), such as: health evaluations manually edited by medical staff, such as "waist circumference 92cm, slightly thick; BMI 24.5, slightly high; blood glucose normal, dietary control recommended," "waist is slightly thick, wider than standard, BMI slightly high, no other abnormalities," and "fasting blood glucose 6.8, significantly exceeding the standard, re-examination required."

[0022] S1.2: First, use the data deduplication function F de (D s ,ID)=D s,de Data deduplication is performed on the numerical physical examination indicators in the first data source. s This is the original structured physical examination indicator dataset, formatted as a two-dimensional matrix. Each row corresponds to one structured data set from a physical examination form, and each column corresponds to one structured field. ID is the unique identifier for the examinee, and D... s,de The deduplicated structured physical examination index dataset, formatted similarly to D s Consistent, delete redundant data submitted repeatedly, if D s,de If the measured values ​​of an indicator are missing ≥ the number of corresponding items (e.g., 3 items), it is directly judged as non-standard, and the values ​​need to be supplemented and approved before proceeding to the next step; then the numerical standardization function F is used. no (x,U std U raw )=x×(U std / U raw ), x, U std and U rawThese should include the original measured value, the standard unit of the indicator (e.g., waist circumference in cm, BMI in kg / m², etc.), and the unit of the original indicator value. A unified unit format for the indicator values ​​should be used. If the unit of the original indicator value is not specified, then U... raw =1, the measurement values ​​of non-numerical physical examination indicators are "non-numerical", and the unit is Ø. For non-numerical physical examination indicators, the conclusion text is directly retained; the text cleaning function F in the text preprocessing technology of Natural Language Processing (NLP) is used. cl (T raw ,R stop ,R sym )=T std Text cleaning is performed on the second data source, T raw For the second data source text, R stop This is a set of stop words containing meaningless redundant characters ("um", "oh", etc.), R sym For special sets of symbols ("!", "#", etc.), T std For the cleaned and standardized unstructured text (waist circumference 92cm, slightly thick, exercise recommended), the core evaluation text is retained to obtain a preprocessed standardized heterogeneous dataset; S2: Based on the standardized heterogeneous dataset, the numerical physical examination indicators in the standardized first data source are analyzed one by one, and abnormal indicators are identified. A first structured fact list is generated based on the identification results. The standardized first data source is analyzed one by one, comparing the measured values ​​of each physical examination indicator with the corresponding preset normal range, automatically identifying all abnormal indicators exceeding the normal range. For each abnormal indicator in the numerical physical examination data, the indicator name and measured value are extracted, and the abnormal state is labeled according to the value exceeding the range (e.g., "mildly high", "moderately high", "severely high", "mildly low", "moderately low", "severely low"). For the conclusion text of non-numerical physical examination indicators, the preset qualitative physical examination standard comparison library D is called. sta std (The standard text word vector B of the corresponding indicator in the built-in non-numerical indicator normal standard "negative / normal", industry domain and common abnormal qualitative description standard set) is used to precisely match the string M. str (A,B) and cosine similarity semantic comparison Sim cos The combination of (A,B) (semantic matching degree ≥ corresponding threshold (0.85) is considered a match) determines whether the qualitative description word vector A in the physical examination conclusion text is an abnormal qualitative description. If M str If (A,B)=0, then the combination comparison judgment formula is F_Sim(A,B)=Sim cos (A,B) and satisfying F_Sim(A,B)≥s the semantic matching threshold, if M strIf (A,B)=1, then the combination comparison judgment formula F_Sim(A,B)=1, indicating that the non-numerical indicator is positive. The abnormal indicator name and corresponding conclusion text are output (e.g., ECG abnormality: complete right bundle branch block, the conclusion text is "abnormal state"). A structured first structured fact list is generated. obj It includes the examinee's unique identifier, the number of items (each abnormal indicator is an item), the indicator name, measurement value, and abnormal status of each item; This embodiment requires specific explanation of how numerical physical examination indicators are identified through the abnormal indicator function F. pos Identify abnormal indicators, x no R low and R high These represent the standardized measurement value of the indicator, the lower limit and upper limit of the normal reference range for the indicator, respectively. The function outputs 0 or 1, where 1 indicates that the indicator is abnormal and 0 indicates that the indicator is normal. The abnormal state is labeled using the function F. abn Perform status labeling. η1 and η2 are the mild threshold (typically ranging from [0.01 to 0.1]) and the moderate threshold (typically ranging from [0.1 to 0.2]), respectively, and neither has units. The mild and moderate abnormal thresholds are based on the abnormal classification principles of various physical examination indicators as specified in industry standards. The distribution of the proportion of values ​​exceeding the normal range of various indicators is statistically analyzed using real abnormal physical examination indicator data from historical data (covering samples from different age groups, genders, and physical examination institutions). The quantiles of abnormal values ​​are calculated (generally, the 25th quantile corresponds to mild abnormality and the 75th quantile corresponds to moderate abnormality), such as waist circumference η1=0.05, blood glucose η1=0.08, etc. The clinical benchmark thresholds are calibrated to ensure that the thresholds are adapted to the actual data characteristics of large-scale physical examination scenarios.

[0023] This embodiment specifically explains the text parsing algorithm for non-numerical physical examination indicator conclusion text, based on natural language processing technology and fine-tuned from physical examination conclusion text samples. The training steps are as follows: First, a training dataset is constructed. This dataset contains historical real-world physical examination indicator conclusion text samples, covering various non-numerical indicators such as electrocardiograms, chest CT scans, and abdominal ultrasounds. The samples include normal conclusion texts ("negative / normal"), various abnormal conclusion texts (such as "complete right bundle branch block," "pulmonary nodules"), and conclusion texts containing redundant expressions (such as "ECG examination: complete right bundle branch block, regular follow-up recommended"). Simultaneously, the core qualitative description of each sample is labeled (text after removing redundancy). Second, based on the rule engine's preset basic parsing rules, including... This includes rules for redundant character removal (fixed redundant prefix / suffix removal, meaningless modifier removal, etc.) and rules for extracting core medical examination terms (pre-associating non-numerical indicator types with corresponding core terminology databases, core semantic matching, etc.). Then, the training dataset is input into the natural language processing model for training. The parameters of the rule engine are iteratively optimized and adjusted (matching thresholds for fixed redundant prefixes / suffixes, similarity thresholds ≥ 0.85 based on edit distance algorithms, etc.) to ensure the algorithm can accurately extract the core qualitative description of each sample. Simultaneously, the semantic matching similarity threshold is optimized to ensure extraction accuracy ≥ the corresponding threshold (0.9). Finally, after training, the algorithm is solidified and applied to the parsing of medical examination conclusion text in this step. It does not require real-time training during the review process and can be directly called for rapid batch processing. The a i and b i Let be the i-th component of the word vectors A and B, and n be the dimension of the word vector (text vectorization methods, such as Word2Vec, BERT, etc.).

[0024] Please see Figure 3 As shown, S3: Based on the standardized heterogeneous dataset, the standardized second data source is input into the natural language processing model that has been fine-tuned by the physical examination text. The model performs two core tasks: entity and relation extraction, standardization, and structured output, and reconstructs the second structured claim list, which includes: S3.1: Using natural language processing technology to process the standardized second data source T pre By analyzing each sentence according to its punctuation marks (such as periods, semicolons, commas, etc.), m sentences S are obtained. j Parsing result Seg(T) pre ), Seg(T pre )={S1,S2,...,S m Then, the sub-word segmentation function SubSeg(S) is used. j)={sub1,sub2,...,sub K}, for each sentence S j Divide into K words (sub) k Then SubSeg(S) j ) and the pre-set core terminology dictionary for physical examinations D med (Including all physical examination indicator names, status terms, indicator units, qualitative descriptions of physical examination conclusions, etc., obtained according to the health examination quality control standards) is input into the physical examination entity recognition core function Recog(SubSeg(S j ),D med In the output recognition result set Recog={E ) ind E sta E num E unit ,F num}, E ind E sta E num E unit and F num These are, respectively, the set of physical examination indicator entities, the set of state words corresponding to each indicator entity in the physical examination indicator entity set (including degree adverbs, non-numerical positive descriptions, etc.), the set of numerical values ​​that correspond one-to-one with each indicator entity (non-numerical indicators are temporarily labeled as non-numerical), the set of units that correspond one-to-one with each indicator entity (non-numerical indicators are temporarily labeled as Ø), and the non-numerical indicator labeling function. If it is a non-numerical indicator F... num =1, using a pre-trained text parsing algorithm, the corresponding set of qualitative positive descriptions, Recog, is identified. num If it is a numerical indicator F num =0; This embodiment requires specific explanation of the core function Recog(SubSeg(S) for physical examination entity recognition. j ),D med This is a common and mature application of NLP technology in the field of medical text processing.

[0025] S3.2: Construct a suspected error detection function De based on character shape similarity and pinyin similarity. err (E ind D med D abb ), D abb For a pre-defined dictionary of standard abbreviations for medical examination terminology (containing the correspondence between standard and non-standard abbreviations of common medical examination terms, such as "BMI" corresponding to the non-standard abbreviations "BMl, bmi", etc.), output a set of suspected typos, Err. cha and the collection of suspected non-standard abbreviations Err abb Err chaIt needs to satisfy either a character shape similarity ≥ the corresponding threshold th1 (0.8) or a pinyin similarity ≥ the corresponding threshold th2 (0.9) and be consistent with D. med The terminology used in the Chinese standard is inconsistent and does not meet the criteria for inconsistency in the combined comparison judgment formula F_Sim. Err abb Must satisfy not in D abb And it meets the requirements of similarity in shape or pinyin with the standard abbreviation, M ma This is a natural language processing model fine-tuned based on physical examination text; it uses glyph similarity, pinyin similarity, and S... j The physical examination context semantics (implemented using context prediction technology based on NLP models) are used to construct a collaborative error correction function Co. syn (Err cha Err abb D med D abb ,S j M ma The collaborative error correction function refers to using each sentence S... j The contextual semantics of the physical examination generate candidate correction results for each suspected error, and then input M. ma Perform semantic feedback verification to obtain the error correction result set Co. res The one-to-one corresponding set of error correction confidence (Coof) co Each error correction result must satisfy a confidence level ≥ the corresponding threshold th3 (0.9), Co res ={(err,corr)|err∈Err cha ∪Err abb ,corr∈D med ∪D abb}, where err and corr are elements in the suspected error set and the standard set, respectively; finally, the physical examination entity recognition result set Recog and the non-numerical qualitative positive description set Recog are used. num and Co res Construct the entity and relation integration function In(Recog,Recog) num Co res )={E ind corr E sta E num E unit D pos corr}, E ind corr and D pos corrThe corrected set of indicator entities and the set of non-numerical qualitative positive descriptions are respectively used to obtain the physical examination entity extraction results, including the correspondence between indicator entity-status word-numerical value-unit and non-numerical indicator entity-qualitative positive description. This embodiment requires specific explanation of S. j The semantic context of the physical examination: for example, the suspected error "only", combined with S j The context of "waist is a bit thick" ("thick" is a physical characteristic description) can be filtered by D. med D abb The system selects "standard terms related to vital signs" (such as "wei") and excludes irrelevant candidates (such as other homophones / similar characters of "wei" like "wei" and "wei"); it verifies the rationality of the error correction results: for example, the candidate correction result "waist circumference" is substituted into "waist circumference is a bit thick", and combined with the medical context that "waist circumference belongs to vital signs and 'thick' is a reasonable description", the main NLP model M ma The system will determine that the error correction result conforms to the logic of a medical scenario, thus giving a high confidence level.

[0026] This embodiment requires a detailed explanation of the candidate correction result generation logic: For a single suspected error, a maximum of n1 (e.g., 2) candidate correction results are generated to avoid excessive candidates increasing the computational cost of semantic verification; candidates are selected based on a priority order of descending character similarity → descending pinyin similarity → descending frequency of medical term occurrence: First priority: Select standard terms with a character similarity ≥ th1 to the suspected error, sorted from highest to lowest similarity; Second priority: If there are fewer than n1 candidates with character similarity meeting the standard, supplement with standard terms with a pinyin similarity ≥ th2, sorted from highest to lowest similarity; Third priority: If the first two categories of candidates are still less than n1, supplement with D... med High-frequency terms of similar medical terms (such as indicator names / status terms) are used to ensure that candidates match the suspected erroneous entity types.

[0027] S3.3: Determine the formula F_Sim(e) through combination comparison. ind corr ,e ind std The threshold corresponding to )≥ is used as the set of corrected indicator entities E. ind corr entity element e in ind corr Terminology database of pre-set standard physical examination indicators D ind std entity element e in ind std The mapping logic yields the standardized entity E. ind std Then, based on the set of state words E corresponding to each indicator entity... sta1. Pre-set standardized abnormal state label library D sta std (Completely consistent with the anomaly labels in the first structured fact list of step S2) and non-numerical indicator labeling function F num Construct a standardized mapping logic for state words, if F num =1, the set of qualitative positive descriptions D pos corr State words are mapped to D sta std Standard qualitative description D in pos std If F num =0, set of state words E sta One-to-one mapping to D sta std The corresponding standardized abnormal state set E in sta std ; S3.4: Next, perform numerical standardization processing, if the non-numerical index labeling function F num =1, then the claimed numeric field will be marked as "non-numeric" and the unit field will be marked as Ø. If F num =0, set the number E num and the unit set E unit Each element in the set is standardized in terms of value and unit (e.g., standardized units, and standardized decimal places). If the set of values ​​E num If the element in the formula is Ø, then after standardization it remains Ø, resulting in the standardized set of claimed values ​​E. num std and the corresponding unit set E unit td Next, establish a one-to-one correspondence between standard indicator entities, standardized claim states, standardized claim values, and standardized units, retaining only the relationship with D. ind std The effective information associated with the standard indicators is used to ensure that each standard indicator entity corresponds to a unique state, value / unit, resulting in a valid data set (Ass) after association filtering. cla Finally, using the unique identifier of the examinee, Ass cla The valid data in the data is organized according to a structured format to generate a second structured claim list. cla It includes the examinee's unique identifier, indicator name, claimed status, and claimed value, covering both numerical and non-numerical abnormal indicators, and clarifying the abnormal indicators and subjective descriptions mentioned by medical staff in health evaluations. This embodiment specifically describes the use of a natural language processing model fine-tuned for medical text analysis to parse unstructured free text health evaluations sentence by sentence. Historical health evaluation texts are collected as training corpus, and medical experts annotate the physical examination indicators, including entities, status words, numerical values, units, and non-numerical positive descriptions. A pre-trained language model based on the Transformer architecture (such as BERT) is then used for supervised fine-tuning with the annotated corpus, enabling the model to learn entity and relation recognition in medical contexts. A pre-defined standard medical terminology mapping table is then constructed, mapping to pre-defined standard medical terms. The construction method involves extracting standard terms from medical textbooks, clinical guidelines, and historical normative data, and collecting their common variations (including colloquial expressions, abbreviations, and misspellings) to establish a "variant-standard terminology" mapping. Similarly, the mapping relation library maps the identified state words to standardized abnormal state labels according to the preset state word standardization mapping rule library, consistent with the abnormal state labeling in step 2 (e.g., "too thick → slightly high" "significantly excessive → severely high"). Simultaneously, the model integrates a suspected error identification module and a collaborative error correction module based on character shape similarity, pinyin similarity, and medical context semantics to determine the identification thresholds for misspelled words / non-standard abbreviations (e.g., character shape similarity, pinyin similarity, etc.). After fine-tuning, the following conditions must be met: entity extraction accuracy ≥ corresponding preset threshold (0.95), collaborative error correction accuracy ≥ corresponding preset threshold (0.9), and standard term mapping semantic matching degree ≥ corresponding preset threshold (consistent with the threshold for combined comparison judgment in step S2). This yields the NLP model M, which has completed basic training and physical examination text fine-tuning in advance. ma .

[0028] S4: Based on the unique identifier of the examinee's medical examination form, the first structured fact list and the second structured claim list are associated, and a multi-dimensional logical consistency comparison is performed. The intelligent review result is obtained based on the comparison results, including: S4.1: Using the subject's unique identifier ID, create the first structured fact list. obj With the second structured claim list cla ID Perform the association to obtain the dual-list association set (List) corresponding to this ID. obj ID List cla ID ); S4.2: Based on a dual-list associative set (List obj ID List cla ID If List obj ID The set of all indicator names E obj Not a Listcla ID The set of all indicator names E ind std If a subset is found, the indicator name is deemed invalid, and the missing indicator name is marked; if the claimed value x of the indicator in the first structured fact list is... sub =Ø, then the numerical value is missing. If the measured value x of the indicator in the first structured fact list is Ø, then the numerical value is missing. no With x sub The absolute difference Δx ≤ the corresponding set error range Δx std (The difference between multiple tests of different indicators in historical data can be calculated. For example, the upper limit of the 95% confidence interval can be compared with the industry standard's allowable error range, and the smaller of the two can be taken.) If the absolute difference Δx > Δx, then the value is considered acceptable. std Other values ​​are marked as abnormal; the judgment is made by combining and comparing the formula F_Sim(Tag) obj Tag sub The threshold corresponding to ) ≥ is used as the anomalous state tag in the first structured fact list. obj With the anomalous state tag in the second structured claim list sub The description matching logic is as follows: if the condition is met, the description is considered acceptable; otherwise, it is rejected. ind std ⊆(E obj ∪E nor If the result is positive, it is considered to have passed the redundancy review; otherwise, it is marked as a redundancy evaluation. obj and E nor If the indicator set and normal indicator set in the subject's first structured fact list are respectively, it is determined to pass the redundancy review; otherwise, it is marked as a redundant evaluation. Finally, if the review results of the four dimensions of indicator name, value, abnormal status and redundancy are all passed, it means that the health evaluation standard corresponding to the ID is met. If any dimension fails, it is considered non-standard. Based on the subject's unique identifier, the names and descriptions of the abnormal indicators that passed and failed are summarized. The description of the problem that passed is that the evaluation is correct, and the description of the problem that failed is that the content is marked, thus obtaining the intelligent review result. S5: Based on the intelligent review results, a personalized intelligent review report is generated for each examinee, and human-computer interaction with the supervisory management end is achieved through a visual interactive interface; the review result status of all examinees is displayed through the visual interactive interface, and a verification entry point is set. When the supervisory management end clicks the verification entry point, it jumps to the details interface, which fully displays the examinee's first structured fact list, second structured claim list, and intelligent review report, facilitating manual review by users; in addition, the interface sets a review feedback entry point, and the management end can mark the review results (such as "confirmed problem" "model misjudgment"), and synchronously collect feedback data for subsequent model optimization; This embodiment requires a detailed explanation of subsequent model optimization: collecting model misjudgment samples marked during the management review process (including entity extraction misjudgment, term mapping misjudgment, and review misjudgment caused by double list comparison misjudgment), adding the misjudged samples to the model fine-tuning training set, and periodically (e.g., once a month) fine-tuning and iterating the natural language processing model, while optimizing the semantic matching threshold and reasonableness error range threshold of logical comparison, continuously improving the robustness of the model and the accuracy of review, and reducing the misjudgment rate.

[0029] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for full-scale intelligent review of medical examination forms based on natural language processing, characterized by: include: S1: Obtain heterogeneous data from the examinee's physical examination form, including the first data source and the second data source, and preprocess the two types of data respectively to obtain the preprocessed standardized heterogeneous dataset. S2: Based on the standardized heterogeneous dataset, the numerical physical examination indicators in the standardized first data source are analyzed one by one and abnormal indicators are identified. The first structured fact list is generated based on the identification results. S3: Based on the standardized heterogeneous dataset, the standardized second data source is input into the natural language processing model that has been fine-tuned by the domain text. The model performs two core tasks: entity and relation extraction, standardization and structured output, and reconstructs the second structured claim list. S4: Based on the unique identifier of the examinee's medical examination form, the first structured fact list and the second structured claim list are associated, and a multi-dimensional logical consistency comparison is performed. The intelligent review result is obtained based on the comparison result. S5: Based on the intelligent audit results, a personalized intelligent audit report is generated for each examinee, and human-computer interaction is conducted with the supervision and management terminal through a visual interactive interface.

2. The method for full intelligent review of medical examination forms based on natural language processing according to claim 1, characterized in that: The first data source includes numerical physical examination indicators and non-numerical physical examination indicator conclusion texts, while the second data source is unstructured free text health evaluations.

3. The method for full intelligent review of medical examination forms based on natural language processing according to claim 1, characterized in that: The first structured fact list: Based on a standardized heterogeneous dataset, the standardized first data source is parsed one by one. The measured values ​​of each physical examination indicator are compared with the preset normal range corresponding to that indicator, and all abnormal indicators exceeding the normal range are automatically identified. For each abnormal indicator in the numerical physical examination data, the indicator name and measured value are extracted, and the abnormal status is marked according to the value exceeding the range. For the conclusion text of non-numerical physical examination indicators, the standard text word vector B of the corresponding indicator in the preset physical examination qualitative standard comparison library is called, and the string is precisely matched to M. str (A,B) and cosine similarity semantic comparison Sim cos The combination of (A, B) determines whether the qualitative descriptive word vector A in the physical examination conclusion text is an abnormal qualitative description. If M str If (A,B)=0, then the combination comparison judgment formula is F_Sim(A,B)=Sim cos (A,B) and satisfying F_Sim(A,B)≥s the semantic matching threshold, if M str If (A,B)=1, then the combination comparison judgment formula F_Sim(A,B)=1, indicating that the non-numerical indicator is positive. Output the abnormal indicator name and corresponding conclusion text, generating a structured first structured fact list. obj It includes the examinee's unique identifier, the number of items, the indicator name corresponding to each item, the measurement value, and any abnormal status.

4. The method for full intelligent review of medical examination forms based on natural language processing according to claim 1, characterized in that: The implementation of S3 includes: S3.1: processing the standardized second data source T using natural language processing technology. pre By analyzing each sentence according to its punctuation, we obtain m sentences S. j Parsing result Seg(T) pre ), Seg(T pre )={S1,S2,...,S m Then, the sub-word segmentation function SubSeg(S) is used. j )={sub1,sub2,...,sub K }, for each sentence S j Divide into K words (sub) k Then SubSeg(S) j ) and the pre-set core terminology dictionary for physical examinations D med The input is fed into the core function Recog(SubSeg(S) for physical examination entity recognition. j ),D med In the output recognition result set Recog={E ) ind E sta E num E unit ,F num }, E ind E sta E num E unit and F num These are, respectively, the set of physical examination indicator entities, the set of state words corresponding to each indicator entity in the physical examination indicator entity set, the set of numerical values ​​that correspond one-to-one with each indicator entity, the set of units that correspond one-to-one with each indicator entity, and the non-numerical indicator labeling function. If it is a non-numerical indicator F... num =1, using a pre-trained text parsing algorithm, the corresponding set of qualitative positive descriptions, Recog, is identified. num If it is a numerical indicator F num =0.

5. The method for full intelligent review of medical examination forms based on natural language processing according to claim 4, characterized in that: The implementation of S3 also includes: S3.2: Constructing a suspected error identification function De based on character shape similarity and pinyin similarity. err (E ind D med D abb ), D abb Output a set of suspected typos Err based on a pre-defined dictionary of standard abbreviations for physical examination terminology. cha and the collection of suspected non-standard abbreviations Err abb Err cha It needs to satisfy either a character shape similarity ≥ the corresponding threshold th1 or a pinyin similarity ≥ the corresponding threshold th2 and be consistent with D. med Middle and D med The terminology used in the Chinese standard is inconsistent and does not meet the criteria for inconsistency in the combined comparison judgment formula F_Sim. Err abb Must satisfy not in D abb And it meets the requirements of similarity in shape or pinyin with the standard abbreviation, M ma This is a natural language processing model fine-tuned based on physical examination text; it uses glyph similarity, pinyin similarity, and S... j Based on the contextual semantics of physical examination, a collaborative error correction function Co is constructed. syn (Err cha Err abb D med D abb ,S j M ma The collaborative error correction function refers to each sentence S j The physical examination context semantics generate candidate correction results for each suspected error, and then input M ma Perform semantic feedback verification to obtain the error correction result set Co. res Finally, the results of physical examination entity recognition (Recog) and the non-numerical qualitative positive description set (Recog) are used. num and Co res Construct the entity and relation integration function In(Recog,Recog) num Co res )={E ind corr E sta E num E unit D pos corr }, E ind corr and D pos corr The corrected set of indicator entities and the set of non-numerical qualitative positive descriptions are respectively used to obtain the physical examination entity extraction results, including the correspondence between indicator entity-state word-value-unit and non-numerical indicator entity-qualitative positive description.

6. The method for full intelligent review of medical examination forms based on natural language processing according to claim 5, characterized in that: The implementation of S3 further includes: S3.3: judging the formula F_Sim(e through combination comparison) ind corr ,e ind std The threshold corresponding to )≥ is used as the set of corrected indicator entities E. ind corr entity element e in ind corr Terminology database of pre-set standard physical examination indicators D ind std entity element e in ind std The mapping logic yields the standardized entity E. ind std Then, based on the set of state words E corresponding to each indicator entity... sta Pre-defined standardized abnormal state label library D sta std Non-numerical index labeling function F num Construct a standardized mapping logic for state words, if F num =1, the set of qualitative positive descriptions D pos corr State words are mapped to D sta std Standard qualitative description D in pos std If F num =0, set E of state words sta One-to-one mapping to D sta std The corresponding standardized abnormal state set E in sta std .

7. The method for full intelligent review of physical examination forms based on natural language processing according to claim 5, characterized in that: The implementation of S3 further includes: S3.4: Next, perform numerical standardization processing, if the non-numerical index labeling function F num =1, then the claimed numeric field will be marked as "non-numeric" and the unit field will be marked as Ø. If F num =0, set the number E num and the unit set E unit Each element in the set is standardized in terms of both value and unit. If the set of values ​​E num If the element in the formula is Ø, then after standardization it remains Ø, resulting in the standardized set of claimed values ​​E. num std and the corresponding unit set E unit td Next, establish a one-to-one correspondence between standard indicator entities, standardized claim states, standardized claim values, and standardized units, retaining only the relationship with D. ind std The effective information associated with the standard indicators is used to obtain the effective data set Ass after association filtering. cla Finally, using the unique identifier of the examinee, Ass cla The valid data in the data is organized according to a structured format to generate a second structured claim list. cla This includes the subject's unique identifier, indicator name, claimed status, and claimed value.

8. The method for full intelligent review of medical examination forms based on natural language processing according to claim 1, characterized in that: The intelligent audit result: S4.1: Using the examinee's unique identifier ID, the first structured fact list List is generated. obj With the second structured claim list cla ID Perform the association to obtain the dual-list association set (List) corresponding to this ID. obj ID List cla ID ); S4.2: Based on a dual-list associative set (List obj ID List cla ID If List obj ID The set of all indicator names E obj Not a List cla ID The set of all indicator names E ind std If a subset is found, the indicator name is deemed invalid, and the missing indicator name is marked; if the claimed value x of the indicator in the first structured fact list is... sub =Ø, then the numerical value is missing. If the measured value x of the indicator in the first structured fact list is Ø, then the numerical value is missing. no With x sub The absolute difference Δx ≤ the corresponding set error range Δx std If the absolute difference Δx > Δx, then the value is considered acceptable. std Other values ​​are marked as abnormal; the judgment is made by combining and comparing the formula F_Sim(Tag) obj Tag sub The threshold corresponding to ) ≥ is used as the anomalous state tag in the first structured fact list. obj With the anomalous state tag in the second structured claim list sub The description matching logic is as follows: if the condition is met, the description is considered acceptable; otherwise, it is rejected. ind std ⊆(E obj ∪E nor E obj and E nor If the indicators in the subject's first structured fact list are the set of indicators and the set of normal indicators, then the redundancy review is passed; otherwise, it is marked as a redundant evaluation. Finally, if the review results of the four dimensions of indicator name, value, abnormal status, and redundancy are all passed, it means that the health evaluation standard corresponding to the ID is met. If any dimension fails, it is considered non-standard. Based on the subject's unique identifier, the names of the abnormal indicators that passed and failed are summarized along with the problem descriptions. The problem descriptions of the passed ones are "correct evaluation," and the problem descriptions of the failed ones are "marked content," thus obtaining the intelligent review result.

Images