A method and system for medical document report error correction early warning

Abstract

The application provides a method and system for medical document report error correction early warning, which directly captures the content of the text input area from the visual interface through the design of screen pixel area recognition positioning and OCR recognition, expands the applicable scene of the scheme, dynamically monitors the change of the screen pixel area, updates the recognition range in real time, provides data support for subsequent real-time checking, adopts light-weight OCR technology, reduces the influence of invalid recognition data on the subsequent process, uses a medical special semantic model to perform semantic analysis and extraction on the text content, obtains medical entities and entity relationships, improves the accuracy of medical entity recognition, adapts to the professional term scene, obtains structured data through the entity relationship, can be directly called by the subsequent checking module, improves the processing efficiency of the checking process, generates to-be-confirmed information for early warning and reminding through the position and error type of the error information, accurately locates the error, and improves the error correction efficiency.

Description

Technical Field

[0001] This invention relates to the field of medical report technology, and in particular to a method and system for error correction and early warning in medical document reports. Background Technology

[0002] Medical records are crucial documents in clinical diagnosis and treatment, including progress notes, discharge summaries, diagnostic certificates, and interpretations of laboratory and examination reports. Their quality directly impacts medical safety, the protection of patients' and doctors' rights, and the handling of medical disputes. With the widespread application of computer technology in the medical field, medical record writing has shifted entirely from traditional handwritten methods to electronic writing. While this has improved writing efficiency and facilitated archiving, it has also brought new problems and challenges.

[0003] The following technical defects are commonly found in the current electronic medical record writing process: First, frequent typos and omissions occur. In electronic input mode, operators are prone to typos such as entering too many characters, miswriting units, and misspelling drug names due to factors such as excessive input speed and slipping fingers. They are also prone to overlooking key diagnostic information. Such errors directly affect the accuracy of medical records and may even mislead subsequent treatment decisions. Second, inconsistencies exist between diagnostic conclusions and clinical descriptions. Specifically, the diagnostic opinions do not match the imaging descriptions and physical signs descriptions, undermining the logic and rigor of the medical records. Third, the verification of normal ranges for laboratory test values ​​is lacking. In clinical work, especially late at night when staff are fatigued, the comparison between test values ​​and corresponding normal ranges is easily overlooked when writing interpretations of laboratory test reports. This results in abnormal values ​​not being identified and marked in a timely manner, affecting the timeliness and accuracy of diagnostic judgments. Fourth, there are compatibility and update lag issues with existing medical software systems. In the era of traditional handwritten medical documents, the probability of such errors was relatively low due to the slow writing speed and the ability of operators to manually verify them in real time. However, in the electronic writing mode, the workload of medical staff has increased significantly, and operators are under high pressure for extended periods, leading to increased fatigue and a sharp rise in the error rate. To avoid these problems, existing medical institutions generally adopt a two- or even three-level manual verification model, where senior physicians review and correct medical documents written by junior physicians after the fact. However, this model has inherent technical defects: on the one hand, manual review relies on the clinical experience of the reviewers, and the experience of physicians in different departments and at different levels varies greatly, resulting in inconsistent review standards and making it difficult to guarantee the comprehensiveness and accuracy of error identification; on the other hand, the review process occurs after the document is written, which is an ex-post correction and cannot prevent errors at the source. Some erroneous information that has already been entered may have been used in diagnosis and treatment decisions, still posing a potential medical risk. Moreover, ex-post review requires a significant additional manpower cost, reducing the overall efficiency of medical work. Summary of the Invention

[0004] This invention provides a method and system for error correction and early warning in medical document reports, in order to solve the problems mentioned in the background art.

[0005] A method for error correction and early warning in medical document reports includes: S1: Identify and locate the screen pixel area of ​​the medical document report to obtain the text input area, and recognize the text content of the text input area based on lightweight OCR; S2: Use a medical-specific semantic model to perform semantic analysis and extraction on the text content to obtain medical entities and entity relationships; S3: Based on a medical knowledge base, perform multi-dimensional verification of medical entities and entity relationships to obtain error information; S4: Based on the location and type of the error message, generate a notification message to be confirmed.

[0006] Preferably, in step S1, the screen pixel area of ​​the medical document report is identified and located to obtain the text input area, including: Global identification of screen pixel areas is performed to obtain full-screen pixel features. Based on the preset pixel features of medical document reports, the full-screen pixel features are filtered to obtain the report area. The report area is divided into first pixel blocks according to the size of the first pixel. The gray-level variance and edge density of the first pixel block are calculated. Based on the gray-level variance and edge density, the medical text region feature value of the first pixel block is calculated. The first pixel block with a medical text region feature value greater than a preset threshold is selected as a candidate pixel block. The candidate pixel blocks are divided according to the second pixel size to obtain the second pixel blocks, and the pixel gradient of each pixel in the second pixel blocks is obtained. Pixels with pixel gradients greater than the preset average value are selected from the second pixel blocks, and adjacent pixels are merged to obtain the core region. The core region is traversed according to the third pixel size to identify boundary features. The boundary features are then divided to obtain hard boundaries, soft boundaries, and cursor anchoring features. Determine whether there are interference areas in the core region based on hard boundaries, soft boundaries, and cursor anchoring features; If so, obtain background features, text features, and boundary features from the core region, compare the background features, text features, and boundary features with the pre-established input region standard feature library, select the core region with a similarity greater than the preset similarity as the target region, and obtain the pixel dynamic change features of the target region. Compare the pixel dynamic change features with the input region change features and the interference region change features respectively, retain the target region consistent with the input region change features as the text input region, and remove the target region consistent with the interference region change features. Otherwise, the core area will be used as the text input area.

[0007] Preferably, the pixel gradient of each pixel in the second pixel block is obtained, and pixels with pixel gradients greater than a preset mean are selected from the second pixel block, including: The horizontal and vertical gradients of each point in the second pixel block are calculated, and the magnitudes of the horizontal and vertical gradients are taken as the pixel gradients. From the pixel training results of historical medical documents, an initial preset mean is obtained. The initial preset mean is then adjusted downward by a preset adjustment ratio for the second pixel block that detects cursor blinking, and upward by a preset adjustment ratio for the second pixel block that detects static labels, thus obtaining a dynamic preset mean. Select pixels whose pixel gradient is greater than the dynamic preset mean from the second pixel block.

[0008] Preferably, in step S1, recognizing the text content of the text input area based on lightweight OCR includes: Lightweight preprocessing is performed on the text input area to obtain the target input area; This technology uses lightweight OCR to recognize text content in the target input area.

[0009] Preferably, obtaining the text input area also includes: The text input area is initially identified to obtain basic patient information; The patient's basic information is matched with the trust permission mechanism between the early warning system and the hospital information system to determine whether the patient's basic information meets the trust requirements. If so, agree to perform text content recognition on the text input area; Otherwise, text content recognition in the text input area will be prohibited, and a warning will be issued.

[0010] Preferably, in step S2, a medical-specific semantic model is used to perform semantic analysis and extraction on the text content to obtain medical entities and entity relationships, including: The text content is split based on a medical thesaurus to obtain basic word segments, and the basic word segments are normalized to obtain standard word segments. The standard word segmentation is matched with the entity of the ontology layer of the medical knowledge graph. A preset number of candidate entities are matched for each standard word segmentation. The confidence of the standard word segmentation and the candidate entities is established based on the fuzzy matching degree, and a triplet of word segmentation-entity-confidence is established. Candidate entities with confidence scores greater than a preset confidence score are extracted from triples as initial basic entities. Based on the character length and combination rules of entities in the medical knowledge graph, the boundaries of the basic entities are corrected to obtain the target basic entities. Based on the entity features in the target basic entity, the corresponding department is determined, and a subset of the knowledge graph of the corresponding department is retrieved from the medical knowledge graph. Based on the subset of the knowledge graph, the target basic entity is supplemented with specialized terms to obtain the medical entity. Candidate relationship paths that are associated with the medical entity are obtained from the medical knowledge graph. The semantic association degree between entities in the candidate relationship paths is determined based on a medical-specific semantic model. The target relationship path with a semantic association degree greater than a preset association degree is selected. Entity relationships are extracted from the target relationship path to obtain initial relationships. The initial relationships are then calibrated based on the patient's physiological characteristics to obtain entity relationships.

[0011] Preferably, in step S3, based on a medical knowledge base, multi-dimensional verification is performed on medical entities and entity relationships to obtain error information, including: Numerical entity relationship features, semantic entity relationship features, and spelling features are obtained from the medical entities and entity relationships. Numerical knowledge verification information, semantic knowledge verification information, and spelling verification information are retrieved from the medical knowledge base. The numerical entity relationship features, semantic entity relationship features, and spelling features are verified respectively, and error information is obtained from the verification results.

[0012] Preferably, in step S4, generating a confirmation message for early warning based on the location and type of the error message includes: The warning level is determined based on the error type, and the information to be confirmed is determined based on the location and error message. The information to be confirmed is used to generate a warning push message according to the warning level and displayed on the screen interface.

[0013] Preferred options also include: Obtain the doctor's operation information regarding the warning push information, including whether to ignore or confirm modification; The operational information was used as instance data to optimize the medical knowledge base.

[0014] A system for error correction and early warning in medical document reports includes: The region content recognition module is used to identify and locate the screen pixel area of ​​medical document reports to obtain the text input area, and recognize the text content of the text input area based on lightweight OCR. The semantic analysis module is used to perform semantic analysis and extraction on text content using a medical-specific semantic model to obtain medical entities and entity relationships. The multi-dimensional verification module is used to perform multi-dimensional verification of medical entities and entity relationships based on a medical knowledge base, and to obtain error information. The early warning module is used to generate early warning information based on the location and type of the error message.

[0015] Compared with the prior art, the present invention has achieved the following beneficial effects: By employing screen pixel region recognition and OCR design, the system directly extracts text input content from the visual interface, addressing the compatibility issues of text acquisition in heterogeneous medical systems and expanding the applicability of the solution. Through dynamic monitoring of screen pixel region changes and real-time updates to the recognition range, it captures text content during dynamic writing, providing data support for subsequent real-time verification. This overcomes the limitation of not being able to correct errors in real time. The use of lightweight OCR technology reduces the impact of invalid recognition data on subsequent processes, balancing real-time performance and recognition accuracy. Furthermore, by utilizing a medical-specific semantic model to perform semantic analysis and extraction of text content, medical entities and entity relationships are obtained. To improve the accuracy of medical entity recognition and adapt to professional terminology scenarios, this system acquires structured data by obtaining entity relationships. This data can be directly called by subsequent verification modules, avoiding the complexity of unstructured text parsing and improving the processing efficiency of the verification process. Based on a medical knowledge base, it performs multi-dimensional verification of medical entities and entity relationships to obtain error information, achieving full-type error coverage and solving the problem of insufficient traditional single-dimensional verification. Based on the location and type of error information, it generates confirmation information for early warning, accurately locating errors, improving error correction efficiency, clarifying error types and basis, and assisting physicians in judgment.

[0016] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in this application.

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a method for error correction and early warning in medical document reports according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating the early warning notification process in an embodiment of the present invention; Figure 3 This is a structural diagram of a system for error correction and early warning in medical document reports according to an embodiment of the present invention. Detailed Implementation

[0019] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0020] Example 1: This embodiment of the invention provides a method for error correction and early warning in medical document reports, such as... Figure 1 As shown, it includes: S1: Identify and locate the screen pixel area of ​​the medical document report to obtain the text input area, and recognize the text content of the text input area based on lightweight OCR; S2: Use a medical-specific semantic model to perform semantic analysis and extraction on the text content to obtain medical entities and entity relationships; S3: Based on a medical knowledge base, perform multi-dimensional verification of medical entities and entity relationships to obtain error information; S4: Based on the location and type of the error message, generate a notification message to be confirmed.

[0021] In this embodiment, medical entities include diseases, symptoms, monitoring, medication, numerical values, etc., and entity relationships include causal relationships, sequential relationships, diagnostic criteria, etc.

[0022] In this embodiment, the medical knowledge base is regularly synchronized with the latest standards of the Pharmacopoeia Commission and the Society of Laboratory Medicine to ensure that the verification rules are consistent with clinical norms, and is established based on massive medical big data.

[0023] In this embodiment, the identification and positioning of screen pixel areas of medical document reports is achieved based on a novel non-invasive human-computer interaction method.

[0024] In this embodiment, all information regarding the patient's medical treatment in the text content was obtained with the patient's consent.

[0025] In this embodiment, for the information to be confirmed, the system receives a confirmation modification instruction or an ignore instruction from the doctor, and records the reason for the ignore instruction.

[0026] In this embodiment, the medical-specific semantic model is pre-trained based on an AI model.

[0027] The beneficial effects of the above design scheme are as follows: By using screen pixel area recognition and OCR recognition, the text input area content can be directly captured from the visual interface, solving the compatibility pain point of text acquisition in heterogeneous medical systems and expanding the applicable scenarios of the solution. By dynamically monitoring changes in screen pixel areas and updating the recognition range in real time, it can capture text content during dynamic writing, providing data support for subsequent real-time verification and overcoming the limitation of not being able to correct errors in real time. The use of lightweight OCR technology reduces the impact of invalid recognition data on subsequent processes, balancing real-time performance and recognition accuracy. By using a medical-specific semantic model to perform semantic analysis and extraction of text content, medical entities and entity relationships are obtained, improving the accuracy of medical entity recognition and adapting to professional terminology scenarios. By obtaining entity relationships, structured data acquisition is achieved, which can be directly called by subsequent verification modules, avoiding the complexity of unstructured text parsing and improving the processing efficiency of the verification process. By using a medical knowledge base to perform multi-dimensional verification of medical entities and entity relationships, error information is obtained, achieving full-type error coverage and solving the problem of insufficient traditional single-dimensional verification. By generating confirmation information based on the location and type of error information, early warning reminders are provided. Accurately pinpointing errors improves error correction efficiency, clarifies error types and basis, and assists physicians in making judgments.

[0028] Example 2: Based on Example 1, this embodiment of the invention provides a method for error correction and early warning in medical document reports. In step S1, the screen pixel area of ​​the medical document report is identified and located to obtain the text input area, including: Global identification of screen pixel areas is performed to obtain full-screen pixel features. Based on the preset pixel features of medical document reports, the full-screen pixel features are filtered to obtain the report area. The report area is divided into first pixel blocks according to the size of the first pixel. The gray-level variance and edge density of the first pixel block are calculated. Based on the gray-level variance and edge density, the medical text region feature value of the first pixel block is calculated. The first pixel block with a medical text region feature value greater than a preset threshold is selected as a candidate pixel block. The candidate pixel blocks are divided according to the second pixel size to obtain the second pixel blocks, and the pixel gradient of each pixel in the second pixel blocks is obtained. Pixels with pixel gradients greater than the preset average value are selected from the second pixel blocks, and adjacent pixels are merged to obtain the core region. The core region is traversed according to the third pixel size to identify boundary features. The boundary features are then divided to obtain hard boundaries, soft boundaries, and cursor anchoring features. Determine whether there are interference areas in the core region based on hard boundaries, soft boundaries, and cursor anchoring features; If so, obtain background features, text features, and boundary features from the core region, compare the background features, text features, and boundary features with the pre-established input region standard feature library, select the core region with a similarity greater than the preset similarity as the target region, and obtain the pixel dynamic change features of the target region. Compare the pixel dynamic change features with the input region change features and the interference region change features respectively, retain the target region consistent with the input region change features as the text input region, and remove the target region consistent with the interference region change features. Otherwise, the core area will be used as the text input area.

[0029] In this embodiment, the full-screen pixel features are filtered based on the preset pixel features of the medical document report. For example, the complete pixel frame of the currently active medical window is captured at a frame rate of 30fps, and an RGB pixel matrix is ​​output. The background of the medical document input area is mostly white / light gray. The color histogram of the pixel frame is calculated, and areas with a background color ratio of <80% are filtered out to quickly lock the candidate writing area and reduce the amount of subsequent calculations.

[0030] In this embodiment, the size of the first pixel, the size of the second pixel, and the size of the third pixel can be, for example, 8*8, 2*2, and 1*1.

[0031] In this embodiment, the medical text region feature value of the first pixel block is calculated based on grayscale variance and edge density as grayscale variance * 0.6 + edge density * 0.4 = medical text region feature value.

[0032] In this embodiment, the hard boundary is the border pixels of the input box, the soft boundary is the frameless input area, and the cursor anchoring feature is the blinking cursor.

[0033] In this embodiment, the presence of interference regions in the core area is determined based on the boundary features of preset interference regions, using hard boundaries, soft boundaries, and cursor anchoring features.

[0034] In this embodiment, the interfering area is, for example, a pop-up window, a prompt box, or an advertisement.

[0035] In this embodiment, background features include background color, background uniformity, etc., text features include text line spacing, character size, etc., and boundary features include the presence of an input box, border width, etc.

[0036] In this embodiment, the input area change characteristics are: characters are added / deleted one by one (pixel changes are linear and distributed line by line, with only 1-5 pixel blocks changing for each character input), and the interference area change characteristics are pop-up windows / prompt boxes (shown / hidden at once, pixel changes are block-shaped, with more than 100 pixel blocks changing in a single instance).

[0037] The beneficial effects of the above design scheme are as follows: By globally recognizing the screen pixel area, the full-screen pixel features are obtained. Based on the preset pixel features of the medical document report, the full-screen pixel features are filtered to obtain the report area, quickly locking the core range of the medical document, excluding irrelevant areas of the entire screen, reducing subsequent computation, and improving positioning efficiency. By dividing the report area according to the size of the first pixel, the first pixel block is obtained. The gray-level variance and edge density of the first pixel block are calculated, and the medical text region feature value of the first pixel block is calculated based on the gray-level variance and edge density. The medical text region feature value is then selected. The first pixel block larger than a preset threshold is used as a candidate pixel block to coarsely screen high-probability text blocks and filter out blank non-text areas, laying the foundation for accurate positioning. The candidate pixel blocks are then divided according to the second pixel size to obtain the second pixel block, and the pixel gradient of each pixel in the second pixel block is obtained. Pixels with pixel gradients greater than a preset average value are selected from the second pixel block, and adjacent pixels are merged to obtain the core region. Focusing on pixel-level gradient features, merging adjacent effective pixels, and eliminating isolated noise points, the core text area is accurately delineated, improving positioning accuracy. The core region is further divided according to the third pixel size... The algorithm iterates through the text to identify boundary features, which are then divided into hard boundaries, soft boundaries, and cursor anchoring features. This adapts to different input box designs, covering diverse medical document interfaces and improving positioning compatibility. Based on hard boundaries, soft boundaries, and cursor anchoring features, it determines whether there are interfering regions in the core area. If so, it extracts background, text, and boundary features from the core area and compares them with a pre-established standard feature library for input areas. The core area with a similarity greater than a preset similarity is selected as the target area. The algorithm then obtains the pixel dynamic change features of the target area and compares them with the input area change features and the interfering area change features. Target areas consistent with the input area change features are retained as text input areas, while those consistent with the interfering area change features are removed. Otherwise, the core area is used as the text input area. Dual verification (static feature matching and dynamic change mode) accurately removes pop-ups, prompts, and other interference, ensuring the purity of the text input area and avoiding misjudgments. When there is no interference, the core area is directly output; when there is interference, it is filtered before output, balancing efficiency and adaptability to complex clinical interface scenarios.

[0038] Example 3: Based on Example 2, this embodiment of the invention provides a method for error correction and early warning in medical document reports, which involves obtaining the pixel gradient of each pixel in a second pixel block and selecting pixels from the second pixel block whose pixel gradient is greater than a preset average value, including: The horizontal and vertical gradients of each point in the second pixel block are calculated, and the magnitudes of the horizontal and vertical gradients are taken as the pixel gradients. From the pixel training results of historical medical documents, an initial preset mean is obtained. The initial preset mean is then adjusted downward by a preset adjustment ratio for the second pixel block that detects cursor blinking, and upward by a preset adjustment ratio for the second pixel block that detects static labels, thus obtaining a dynamic preset mean. Select pixels whose pixel gradient is greater than the dynamic preset mean from the second pixel block.

[0039] In this embodiment, the preset downward adjustment ratio is, for example, 0.8, and the preset upward adjustment ratio is, for example, 1.2.

[0040] The beneficial effects of the above design scheme are as follows: by fusing horizontal and vertical gradient calculations to fully capture the edge features of medical text characters, it avoids missing effective pixels due to gradients in a single direction, thereby improving the completeness and accuracy of pixel gradient recognition. By dynamically adjusting the preset mean based on the scene, it dynamically adapts to the different regional features of the medical document interface, solving the problem of poor adaptability of fixed mean. By filtering pixels based on the dynamic mean, it accurately locks effective text pixels, eliminates interfering pixels, provides high-purity pixel data for core area extraction, and improves the accuracy of subsequent text input area positioning.

[0041] Example 4: Based on Example 1, this embodiment of the invention provides a method for error correction and early warning in medical document reports. In step S1, the text content of the text input area is recognized based on lightweight OCR, including: Lightweight preprocessing is performed on the text input area to obtain the target input area; This technology uses lightweight OCR to recognize text content in the target input area.

[0042] In this embodiment, lightweight preprocessing includes, for example, grayscale conversion (reducing channel calculation), adaptive binarization (adapting to brightness differences in medical interfaces), and noise reduction (filtering screen noise / cursor interference).

[0043] The beneficial effects of the above design scheme are: by using screen pixel area recognition and positioning and OCR recognition, the text input area content can be directly captured from the visual interface, which solves the compatibility pain point of text acquisition in heterogeneous medical systems and expands the applicable scenarios of the solution.

[0044] Example 5: Based on Example 1, this embodiment of the invention provides a method for error correction and early warning in medical document reports, which obtains a text input area and further includes: The text input area is initially identified to obtain basic patient information; The patient's basic information is matched with the trust permission mechanism between the early warning system and the hospital information system to determine whether the patient's basic information meets the trust requirements. If so, agree to perform text content recognition on the text input area; Otherwise, text content recognition in the text input area will be prohibited, and a warning will be issued.

[0045] The beneficial effects of the above design scheme are: by matching patient basic information with cross-system trust and permission mechanisms, text recognition is only authorized in compliant scenarios, eliminating unauthorized data access and reducing the risk of data leakage and abuse.

[0046] Example 6: Based on Example 1, this embodiment of the invention provides a method for error correction and early warning in medical document reports. In step S2, a medical-specific semantic model is used to perform semantic analysis and extraction on the text content to obtain medical entities and entity relationships, including: The text content is split based on a medical thesaurus to obtain basic word segments, and the basic word segments are normalized to obtain standard word segments. The standard word segmentation is matched with the entity of the ontology layer of the medical knowledge graph. A preset number of candidate entities are matched for each standard word segmentation. The confidence of the standard word segmentation and the candidate entities is established based on the fuzzy matching degree, and a triplet of word segmentation-entity-confidence is established. Candidate entities with confidence scores greater than a preset confidence score are extracted from triples as initial basic entities. Based on the character length and combination rules of entities in the medical knowledge graph, the boundaries of the basic entities are corrected to obtain the target basic entities. Based on the entity features in the target basic entity, the corresponding department is determined, and a subset of the knowledge graph of the corresponding department is retrieved from the medical knowledge graph. Based on the subset of the knowledge graph, the target basic entity is supplemented with specialized terms to obtain the medical entity. Candidate relationship paths that are associated with the medical entity are obtained from the medical knowledge graph. The semantic association degree between entities in the candidate relationship paths is determined based on a medical-specific semantic model. The target relationship path with a semantic association degree greater than a preset association degree is selected. Entity relationships are extracted from the target relationship path to obtain initial relationships. The initial relationships are then calibrated based on the patient's physiological characteristics to obtain entity relationships.

[0047] In this embodiment, the medical glossary contains 80,000+ specialized terms and 30,000+ abbreviations / colloquial expressions.

[0048] In this embodiment, normalization is performed by mapping abbreviations / colloquial expressions to standard terms.

[0049] In this embodiment, for example, creatinine is linked to creatinine-test items-renal function indicators.

[0050] In this embodiment, the boundary of the basic entity is modified, for example, cefoperazone-sulbactam sodium 2g is split into the drug cefoperazone-sulbactam sodium and the dosage 2g.

[0051] In this embodiment, the target basic entity is supplemented with specialized terms based on a subset of the knowledge graph, such as myocardial infarction type: ST segment elevation type, coronary artery: left anterior descending branch.

[0052] In this embodiment, the candidate relationship path is, for example, pneumonia-fever-cefoperazone-sulbactam sodium.

[0053] The beneficial effects of the above design scheme are as follows: Basic word segments are obtained by splitting text content based on a medical lexicon. These basic word segments are then normalized to obtain standard word segments, eliminating semantic ambiguity in medical texts, improving segmentation accuracy, reducing the difficulty of subsequent entity matching, and improving overall process efficiency. By performing fuzzy matching between the standard word segments and entities at the ontology layer of the medical knowledge graph, a preset number of candidate entities are matched for each standard word segment. A confidence level between the standard word segments and candidate entities is established based on the fuzzy matching degree, creating a segmentation-entity-confidence triplet. This improves the inclusiveness of entity matching, covers complex matching scenarios, provides a quantitative basis for subsequent entity selection, and ensures extraction accuracy. Candidate entities with confidence levels greater than a preset confidence level are extracted from the triplet as initial basic entities. Furthermore, based on the character length and combination rules of entities in the medical knowledge graph, the basic entities are boundary-corrected. The process involves several steps: First, the target basic entity is obtained. Based on the entity features of the target basic entity, the corresponding department is determined. A subset of the knowledge graph for the corresponding department is retrieved from the medical knowledge graph. This subset is used to supplement the target basic entity with specialized terms, resulting in medical entities. This process accurately identifies core entities, filters out invalid interference, corrects entity boundary errors, resolves entity mixing issues, and achieves specialized entity supplementation, covering subdivided entity needs, avoiding irrelevant entity interference, and improving extraction targeting. Candidate relationship paths related to the medical entities are obtained from the medical knowledge graph. The semantic correlation between entities in the candidate relationship paths is determined based on a medical-specific semantic model. Target relationship paths with semantic correlation greater than a preset correlation are selected. This ensures the clinical compliance of entity relationships, avoids logical contradictions, accurately filters effective relationships, and eliminates weak correlation interference. Entity relationships are extracted from the target relationship paths to obtain initial relationships. These initial relationships are calibrated based on the patient's physiological characteristics to obtain entity relationships, adapting to individual differences, improving the personalization and accuracy of relationship extraction, correcting potential relationship deviations, and reducing clinical application risks. Ultimately, this process achieves high-precision and personalized extraction of medical entities and entity relationships.

[0054] Example 7: Based on Example 1, this embodiment of the invention provides a method for error correction and early warning in medical document reports. In step S3, based on a medical knowledge base, multi-dimensional verification is performed on medical entities and entity relationships to obtain error information, including: Numerical entity relationship features, semantic entity relationship features, and spelling features are obtained from the medical entities and entity relationships. Numerical knowledge verification information, semantic knowledge verification information, and spelling verification information are retrieved from the medical knowledge base. The numerical entity relationship features, semantic entity relationship features, and spelling features are verified respectively, and error information is obtained from the verification results.

[0055] The beneficial effects of the above design scheme are: by performing multi-dimensional verification of medical entities and entity relationships based on a medical knowledge base, error information is obtained, achieving full-type error coverage and solving the problem of insufficient traditional single-dimensional verification.

[0056] Example 8: Based on Example 1, this embodiment of the invention provides a method for error correction and early warning in medical document reports, such as... Figure 2 As shown, in step S4, based on the location and type of the error information, a confirmation message is generated for early warning, including: The warning level is determined based on the error type, and the information to be confirmed is determined based on the location and error message. The information to be confirmed is used to generate a warning push message according to the warning level and displayed on the screen interface.

[0057] The beneficial effects of the above design scheme are: by generating confirmation information based on the location and type of error information, early warning reminders are provided, errors are accurately located, error correction efficiency is improved, error types and basis are clarified, and physicians are assisted in making judgments.

[0058] Example 9: Based on Example 8, this embodiment of the invention provides a method for error correction and early warning in medical document reports, further comprising: Obtain the doctor's operation information regarding the warning push information, including whether to ignore or confirm modification; The operational information was used as instance data to optimize the medical knowledge base.

[0059] The beneficial effects of the above design scheme are: by obtaining the doctor's operation information on the early warning push information, the operation information is either to ignore or confirm and modify; the operation information is used as instance data to optimize the medical knowledge base, thereby realizing real-time updates of the medical knowledge base and improving the accuracy of verification.

[0060] Example 10: This embodiment of the invention provides a system for error correction and early warning in medical document reports, such as... Figure 3 As shown, it includes: The region content recognition module is used to identify and locate the screen pixel area of ​​medical document reports to obtain the text input area, and recognize the text content of the text input area based on lightweight OCR. The semantic analysis module is used to perform semantic analysis and extraction on text content using a medical-specific semantic model to obtain medical entities and entity relationships. The multi-dimensional verification module is used to perform multi-dimensional verification of medical entities and entity relationships based on a medical knowledge base, and to obtain error information. The early warning module is used to generate early warning information based on the location and type of the error message.

[0061] In this embodiment, medical entities include diseases, symptoms, monitoring, medication, numerical values, etc., and entity relationships include causal relationships, sequential relationships, diagnostic criteria, etc.

[0062] In this embodiment, the medical knowledge base is regularly synchronized with the latest standards from the Pharmacopoeia Commission and the Society of Laboratory Medicine to ensure that the verification rules are consistent with clinical practices.

[0063] In this embodiment, all information regarding the patient's medical treatment in the text content was obtained with the patient's consent.

[0064] In this embodiment, for the information to be confirmed, the system receives a confirmation modification instruction or an ignore instruction from the doctor, and records the reason for the ignore instruction.

[0065] The beneficial effects of the above design scheme are as follows: By using screen pixel area recognition and OCR recognition, the text input area content can be directly captured from the visual interface, solving the compatibility pain point of text acquisition in heterogeneous medical systems and expanding the applicable scenarios of the solution. By dynamically monitoring changes in screen pixel areas and updating the recognition range in real time, it can capture text content during dynamic writing, providing data support for subsequent real-time verification and overcoming the limitation of not being able to correct errors in real time. The use of lightweight OCR technology reduces the impact of invalid recognition data on subsequent processes, balancing real-time performance and recognition accuracy. By using a medical-specific semantic model to perform semantic analysis and extraction of text content, medical entities and entity relationships are obtained, improving the accuracy of medical entity recognition and adapting to professional terminology scenarios. By obtaining entity relationships, structured data acquisition is achieved, which can be directly called by subsequent verification modules, avoiding the complexity of unstructured text parsing and improving the processing efficiency of the verification process. By using a medical knowledge base to perform multi-dimensional verification of medical entities and entity relationships, error information is obtained, achieving full-type error coverage and solving the problem of insufficient traditional single-dimensional verification. By generating confirmation information based on the location and type of error information, early warning reminders are provided. Accurately pinpointing errors improves error correction efficiency, clarifies error types and basis, and assists physicians in making judgments.

[0066] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this application and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for error correction and early warning in medical document reports, characterized in that, include: S1: Identify and locate the screen pixel area of ​​the medical document report to obtain the text input area, and recognize the text content of the text input area based on lightweight OCR; S2: Use a medical-specific semantic model to perform semantic analysis and extraction on the text content to obtain medical entities and entity relationships; S3: Based on a medical knowledge base, perform multi-dimensional verification of medical entities and entity relationships to obtain error information; S4: Based on the location and type of the error message, generate a notification message to be confirmed.

2. The method for error correction and early warning in medical document reports according to claim 1, characterized in that, In step S1, the screen pixel area of ​​the medical document report is identified and located to obtain the text input area, including: Global identification of screen pixel areas is performed to obtain full-screen pixel features. Based on the preset pixel features of medical document reports, the full-screen pixel features are filtered to obtain the report area. The report area is divided into first pixel blocks according to the size of the first pixel. The gray-level variance and edge density of the first pixel block are calculated. Based on the gray-level variance and edge density, the medical text region feature value of the first pixel block is calculated. The first pixel block with a medical text region feature value greater than a preset threshold is selected as a candidate pixel block. The candidate pixel blocks are divided according to the second pixel size to obtain the second pixel blocks, and the pixel gradient of each pixel in the second pixel blocks is obtained. Pixels with pixel gradients greater than the preset average value are selected from the second pixel blocks, and adjacent pixels are merged to obtain the core region. The core region is traversed according to the third pixel size to identify boundary features. The boundary features are then divided to obtain hard boundaries, soft boundaries, and cursor anchoring features. Determine whether there are interference areas in the core region based on hard boundaries, soft boundaries, and cursor anchoring features; If so, obtain background features, text features, and boundary features from the core region, compare the background features, text features, and boundary features with the pre-established input region standard feature library, select the core region with a similarity greater than the preset similarity as the target region, and obtain the pixel dynamic change features of the target region. Compare the pixel dynamic change features with the input region change features and the interference region change features respectively, retain the target region consistent with the input region change features as the text input region, and remove the target region consistent with the interference region change features. Otherwise, the core area will be used as the text input area.

3. The method for error correction and early warning in medical document reports according to claim 2, characterized in that, Obtain the pixel gradient of each pixel in the second pixel block, and select pixels from the second pixel block whose pixel gradient is greater than a preset mean, including: The horizontal and vertical gradients of each point in the second pixel block are calculated, and the magnitudes of the horizontal and vertical gradients are taken as the pixel gradients. From the pixel training results of historical medical documents, an initial preset mean is obtained. The initial preset mean is then adjusted downward by a preset adjustment ratio for the second pixel block that detects cursor blinking, and upward by a preset adjustment ratio for the second pixel block that detects static labels, thus obtaining a dynamic preset mean. Select pixels whose pixel gradient is greater than the dynamic preset mean from the second pixel block.

4. The method for error correction and early warning in medical document reports according to claim 1, characterized in that, In step S1, the text content of the text input area is recognized based on lightweight OCR, including: Lightweight preprocessing is performed on the text input area to obtain the target input area; This technology uses lightweight OCR to recognize text content in the target input area.

5. The method for error correction and early warning in medical document reports according to claim 1, characterized in that, The text input area also includes: The text input area is initially identified to obtain basic patient information; The patient's basic information is matched with the trust permission mechanism between the early warning system and the hospital information system to determine whether the patient's basic information meets the trust requirements. If so, agree to perform text content recognition on the text input area; Otherwise, text content recognition in the text input area will be prohibited, and a warning will be issued.

6. The method for error correction and early warning in medical document reports according to claim 1, characterized in that, In step S2, a medical-specific semantic model is used to perform semantic analysis and extraction on the text content to obtain medical entities and entity relationships, including: The text content is split based on a medical thesaurus to obtain basic word segments, and the basic word segments are normalized to obtain standard word segments. The standard word segmentation is matched with the entity of the ontology layer of the medical knowledge graph. A preset number of candidate entities are matched for each standard word segmentation. The confidence of the standard word segmentation and the candidate entities is established based on the fuzzy matching degree, and a triplet of word segmentation-entity-confidence is established. Candidate entities with confidence scores greater than a preset confidence score are extracted from triples as initial basic entities. Based on the character length and combination rules of entities in the medical knowledge graph, the boundaries of the basic entities are corrected to obtain the target basic entities. Based on the entity features in the target basic entity, the corresponding department is determined, and a subset of the knowledge graph of the corresponding department is retrieved from the medical knowledge graph. Based on the subset of the knowledge graph, the target basic entity is supplemented with specialized terms to obtain the medical entity. Candidate relationship paths that are associated with the medical entity are obtained from the medical knowledge graph. The semantic association degree between entities in the candidate relationship paths is determined based on a medical-specific semantic model. The target relationship path with a semantic association degree greater than a preset association degree is selected. Entity relationships are extracted from the target relationship path to obtain initial relationships. The initial relationships are then calibrated based on the patient's physiological characteristics to obtain entity relationships.

7. The method for error correction and early warning in medical document reports according to claim 1, characterized in that, In step S3, based on a medical knowledge base, medical entities and entity relationships are validated in multiple dimensions to obtain error information, including: Numerical entity relationship features, semantic entity relationship features, and spelling features are obtained from the medical entities and entity relationships. Numerical knowledge verification information, semantic knowledge verification information, and spelling verification information are retrieved from the medical knowledge base. The numerical entity relationship features, semantic entity relationship features, and spelling features are verified respectively, and error information is obtained from the verification results.

8. The method for error correction and early warning in medical document reports according to claim 1, characterized in that, In step S4, based on the location and type of the error information, a confirmation message is generated for early warning and reminder, including: The warning level is determined based on the error type, and the information to be confirmed is determined based on the location and error message. The information to be confirmed is used to generate a warning push message according to the warning level and displayed on the screen interface.

9. A method for error correction and early warning in medical document reports according to claim 8, characterized in that, Also includes: Obtain the doctor's operation information regarding the warning push information, including whether to ignore or confirm modification; The operational information was used as instance data to optimize the medical knowledge base.

10. A system for error correction and early warning in medical document reports, specifically used to implement the method for error correction and early warning in medical document reports as described in claim 1, characterized in that, include: The region content recognition module is used to identify and locate the screen pixel area of ​​medical document reports to obtain the text input area, and recognize the text content of the text input area based on lightweight OCR. The semantic analysis module is used to perform semantic analysis and extraction on text content using a medical-specific semantic model to obtain medical entities and entity relationships. The multi-dimensional verification module is used to perform multi-dimensional verification of medical entities and entity relationships based on a medical knowledge base, and to obtain error information. The early warning module is used to generate early warning information based on the location and type of the error message.

Images