Method for assisting in writing electronic medical record through artificial intelligence

Abstract

The invention discloses an artificial intelligence assisted electronic medical record writing method, which relates to the technical field of data processing, and comprises the following steps of: performing character extraction and specific region identification on unstructured image data, generating a readable text in combination with a medical knowledge base, and correcting the readable text by using a preset rule in cooperation with a pre-trained language model; whether the readable text is abnormal or not is judged through a medical knowledge base, a corresponding marking mode is selected according to the recognized abnormal degree, and the readable text and the structured data are fused to obtain structured abnormal data: the structured abnormal data are converted into a generated text and then corrected, blocking is conducted according to the structure of a medical record template, and the generated text is marked; filling the priority abstract into a corresponding position to generate a medical record text, and cleaning repeated contents in the priority abstract generated by the medical record text; contradictory points and missing points in the text are detected, the logic preciseness of medical record content is improved, and potential problems are found and marked by comparing with logic rules.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and specifically to a method for writing electronic medical records assisted by artificial intelligence. Background Art

[0002] With the rapid development of informatization in the medical industry, a large amount of patient data is stored in electronic form, including electronic medical records of scanned copies, laboratory test reports, imaging reports, etc. However, these medical data are in various forms and often exist in unstructured or semi-structured forms, such as handwritten test reports, scanned copies, imaging images, etc., and cannot be directly understood and processed by the system. Medical institutions need to convert these data into a structured and readable text form for further analysis, archiving, or for use in assisting diagnosis. At the same time, due to the professionalism and complexity of medical data, not only high-precision text extraction is required during the conversion process, but also it is necessary to ensure that the generated text conforms to medical logic and clinical norms.

[0003] In addition, combining context information such as the patient's medical history and examination results to achieve efficient identification and marking of abnormal data, and generating standardized electronic medical records is of great significance for improving the work efficiency and accuracy of the medical industry.

[0004] In a Chinese invention patent with the application publication number CN112397170A, an electronic medical record generation method is disclosed. The method is as follows: a. Registration and login; b. Front-end interface selection, selecting a medical record query module, an interrogation module, and a drug contraindication entry and query module through the front-end interface; the medical record query module and the interrogation module are respectively linked to a label selection template and a label entry template; c. Interrogation, the label entry template is linked to a consultation module, and the consultation module includes an interrogation template and an auxiliary interrogation template; both the interrogation template and the auxiliary interrogation template are linked and stored in a backend cloud storage platform; d. The medical record query module selects the patient's medical records under a certain label through the label selection template, and medical staff can perform intelligent medical record queries according to needs and can achieve triage at the front end. Its medical record data includes original voice, voice-to-text, and text data, and can fully display the patient's diagnosis and treatment process and the subsequent treatment process.

[0005] Currently, the medical data processing faces the following core technical problems: how to accurately and efficiently extract text information from unstructured images such as handwritten test reports and scanned medical records, and generate a standardized text that conforms to medical semantics and logic. Problems such as semantic inconsistency, logical conflict, and ambiguous description during text generation and anomaly detection are likely to mislead the judgment of medical workers, further increasing the workload of manual review and correction, and it is difficult to automatically complete the whole process from data extraction to structured processing and standardized text generation.

[0006] Therefore, the present invention provides a method for writing electronic medical records assisted by artificial intelligence. Summary of the Invention

[0007] (1) Technical problems to be solved

[0008] In view of the deficiencies of the prior art, the present invention provides a method for writing electronic medical records assisted by artificial intelligence. After combining with a medical knowledge base, readable text is generated, and the readable text is corrected using preset rules in cooperation with a pre-trained language model; the medical knowledge base is used to judge whether there are abnormalities in the readable text, and corresponding marking methods are selected according to the identified degree of abnormality, and the readable text and structured data are fused to obtain structured abnormal data: the structured abnormal data is converted into generated text and then corrected, divided into blocks according to the structure of the medical record template, and filled into the corresponding positions to generate medical record text, and the repeated content in the priority summary generated from the medical record text is cleaned up; the contradictions and omissions in the text are detected to improve the logical rigor of the medical record content, compared with the logical rules, and potential problems are found and marked; thus, the technical problems recorded in the background art are solved.

[0009] (2) Technical solutions

[0010] To achieve the above objectives, the present invention is realized through the following technical solutions: A method for writing electronic medical records assisted by artificial intelligence, including extracting text and identifying specific areas from unstructured image data, generating readable text after combining with a medical knowledge base, and correcting the readable text using preset rules in cooperation with a pre-trained language model;

[0011] The medical knowledge base is used to judge whether there are abnormalities in the readable text, and corresponding marking methods are selected according to the identified degree of abnormality, and the readable text and structured data are fused to obtain structured abnormal data:

[0012] The structured abnormal data is converted into generated text and then corrected, the corrected generated text is divided into blocks according to the structure of the medical record template, and filled into the corresponding positions to generate medical record text, and the repeated content in the priority summary generated from the medical record text is cleaned up;

[0013] The generated text is logically detected to obtain the abnormal values therein, and the abnormality degree Aop is generated according to its abnormal state. When the obtained abnormality degree Aop exceeds the expectation, the medical record text is associated and detected, and a standardized electronic medical record is generated when there is no error; among them, the marked logical conflict points or abnormal values and the medical record background information are used as inputs, and the chief complaint or the description of the current medical history is regenerated according to the logical conflict points; the adjusted medical record text is checked again using the medical logic rule engine, and if there is no error, a standardized electronic medical record is generated.

[0014] Furthermore, after preprocessing the handwritten test report or scanned electronic medical record, use OCR to identify the text and correct errors in the string through a convolutional neural network and a sequence-to-sequence model;

[0015] Use the pre-trained YOLOv5 detection algorithm to identify the row and column regions of the table, determine the boundaries of the cells, extract the text in each cell using OCR, and record the key-value pairs corresponding to the rows and columns. Use the segmentation model obtained by training with the U-Net algorithm to extract specific regions from the medical images, and extract the features of the lesions through a deep learning network.

[0016] Furthermore, optimize through rule checking and context semantic correction models to remove character recognition errors or semantic inconsistencies in the readable text; check the extracted content by setting rules in the medical knowledge base, traverse the rule set of the knowledge base, and match the entities and their context descriptions in the extraction results one by one.

[0017] Furthermore, use the medical knowledge base to judge whether there are abnormalities in the readable text; use the knowledge base rules to judge whether the index range is abnormal and output the abnormal values. For complex indicators not covered by the knowledge base, use the trained supervised learning model to detect the complex patterns of abnormal values and obtain the structured data after detection;

[0018] Conduct correlation analysis on the detected abnormal values and the medical knowledge base to provide interpretable abnormal reasons or suggestions. If the abnormalities detected by the model conflict with the medical knowledge base, mark them.

[0019] Furthermore, use the pre-trained multi-modal neural network to process the readable text and structured data, map the two to the same feature space and then fuse them to obtain structured abnormal data:

[0020] Collect and record the abnormal factor data of the identified abnormal values. Use the trained priority evaluation model to evaluate with the abnormal factor data of the abnormal values as the input, and obtain the marked priority index MPI. If the obtained marked priority index MPI exceeds the preset marking threshold, directly point out the abnormal indicators, abnormal types and their deviation degrees using explicit marking. If the opposite is true, use implicit marking to prompt the associated problems of upstream and downstream indicators.

[0021] Furthermore, after converting the structured abnormal data into an acceptable input format, use the pre-trained language model to convert the structured abnormal data into a natural language description to obtain the generated text, and correct the generated text using regular expressions;

[0022] Load and preprocess the medical knowledge graph, automatically match relevant entities and their upstream and downstream relationships in the knowledge graph for each piece of abnormal data, use a semantic-based matching algorithm for mapping, utilize the established mapping to associate the semantics of the abnormal data with the patient's chief complaint, medical history, and examination results, and supplement the generated preliminary natural language description with knowledge graph relationships.

[0023] Furthermore, construct a semantic model of the medical record content, use natural language processing technology to match the generated text sentence by sentence with the extracted medical record content, and utilize semantic similarity calculation to identify logical contradictions existing in the generated text.

[0024] Determine the standard template structure of the medical record, divide the generated text into blocks according to the structure of the medical record template, fill them into the corresponding positions, and automatically match keywords. After classifying the descriptions in the generated text that conform to a certain piece of content, obtain the medical record text; construct a medical terminology library, and use regular matching to uniformly correct the terms in the filled content.

[0025] Furthermore, use a rule engine or NLP technology to extract all abnormal data involved in the medical record text, assign weights to each piece of abnormal data based on a medical knowledge base or the analytic hierarchy process: sort the abnormal data from high to low according to the calculated comprehensive weight and generate a priority summary.

[0026] Furthermore, construct medical logic rules, use the generated text as input, match it item by item with the medical logic rules, and mark it after detecting logical conflicts or omissions.

[0027] Construct a normal value reference range library, extract all laboratory examination and physical sign data in the generated text, determine whether they exceed the normal range. If they exceed, mark them as abnormal values, and obtain the corresponding abnormal status data. Generate an abnormality degree Aop from the abnormal status data. If the abnormality degree Aop exceeds the expectation, check whether the abnormal value is consistent with other content in the medical record text: if not, mark it as an unreasonable abnormal value.

[0028] Furthermore, the method for generating the abnormality degree Aop from the abnormal status data is as follows:

[0029]

[0030] In the formula: V(t) is the monitoring index, Δ(t) is the normalized deviation degree, k > 0, D(t) is the dynamic change rate, tanh(D(t)) is the smoothing penalty function of the dynamic change, and C(t) is the clinical severity grading.

[0031] V(f) is the frequency component of the measured value, λ is the weight of high-frequency abnormalities, f 0 is the frequency threshold, w Δ 、w C and w Dis the weight, where, w Δ + w C + w D = 1.

[0032] (III) Beneficial Effects

[0033] The present invention provides a method for writing an electronic medical record assisted by artificial intelligence, having the following beneficial effects:

[0034] 1. Rapidly filter out obvious normal values. For complex indicators not covered by the knowledge base, by training a supervised learning model, abnormal complex patterns can be effectively identified, improving the universality of the algorithm. For clearly abnormal values, they can be directly marked and classified, providing preliminary classification information and significantly shortening the judgment time.

[0035] 2. Automatically mark abnormal conflict points when the detection result does not match the data in the knowledge base, reminding for manual review and enhancing the credibility of the final result. The multi-modal neural network combines text and structured data, ensuring that the extracted content is consistent with the patient's medical history and clinical information; by calculating the severity, frequency, and impact of abnormal values to assign priorities, quickly focus on the most urgent or important abnormal data, improving the processing efficiency, and providing hints for minor abnormalities through implicit marking to avoid interference caused by a large amount of unimportant information.

[0036] 3. Use regular expressions to correct terms and modify quantitative descriptions of the generated text, ensuring the professionalism and consistency of medical descriptions, and by combining with a medical knowledge graph, supplement the disease causes and examination suggestions related to abnormal values to form more complete medical record information.

[0037] 4. Automatically generate a priority summary based on weight sorting, highlighting key abnormalities, which can quickly focus on the most important abnormal data. By using a deduplication algorithm to clean keywords and duplicate content, the conciseness and effectiveness of the generated summary can be optimized, and abnormal data with high weights are preferentially displayed and detailed relevant suggestions are provided.

[0038] 5. Detect contradictions and omissions in the text through medical logic rules, ensuring that the generated content meets the specifications and improving the logical rigor of the medical record content. By comparing with the logic rules, potential problems can be found and marked, reducing omissions in manual review.

[0039] 6. Generate an abnormality degree Aop based on factors such as the deviation amplitude from the normal value and severity grading, providing a quantitative basis for further processing of abnormalities. Check whether the abnormal values are consistent with other content, and unreasonable abnormal values can be promptly discovered and marked, reducing the risk of misguidance. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 is a schematic flow diagram of the method for writing an electronic medical record according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0041] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0042] Please refer to Figure 1 , the present invention provides a method for artificial intelligence-assisted writing of electronic medical records, including,

[0043] Step 1: Extract text and identify specific regions from unstructured image data, generate readable text after combining with a medical knowledge base, and correct the readable text using pre-set rules in cooperation with a pre-trained language model;

[0044] The said Step 1 includes the following contents:

[0045] Step 101: For unstructured image data, including handwritten test reports or scanned copies of electronic medical records, convert the pictures into text content through OCR technology. Among them,

[0046] Preprocess these image data, including: noise removal: remove background noise caused by light, blur, etc. in the image through filtering technology; image enhancement: improve the clarity of the text area through means such as contrast adjustment and gray-scale transformation; image binarization: process the image into black and white for easy character segmentation and recognition;

[0047] Use OCR on the preprocessed image through a convolutional neural network (CNN) and a sequence-to-sequence (Seq2Seq) model to identify the text and correct possible errors in the string. For example, extract "Chest CT result: Lung nodule" from a CT scan form;

[0048] Use the pre-trained YOLOv5 detection algorithm to identify the row and column regions of the table, determine the boundaries of the cells, extract the text in each cell using OCR technology, and record the key-value pairs corresponding to the rows and columns;

[0049] Use a segmentation model trained by the U-Net algorithm to extract specific regions (such as lung nodules) from medical images, extract the characteristics of the lesions through a deep learning network, such as shape (spherical, flat), density (low density, high density), etc., and generate readable text in combination with a medical knowledge base. For example, "A low-density nodule with a diameter of 8 mm was found in the middle lobe of the right lung, and the edge is smooth."

[0050] Through OCR, automated processing can be achieved to quickly convert scanned documents or handwritten test reports into text information, reducing the manual input time of medical staff, and significantly improving efficiency in high-frequency operations. Image preprocessing optimizes the clarity of imperfect images, ensuring accurate text extraction in complex lighting, handwriting and other scenarios. The YOLOv5 algorithm is used to locate the boundaries of table rows and columns, avoiding the non-standard problems of traditional OCR in extracting table data, and directly outputting the structured data corresponding to the rows and columns, which is suitable for highly structured files such as test reports and medical record summaries. At the same time, the U-Net deep learning model is used to achieve precise segmentation (such as extracting lung nodule areas), and automatically generate standardized medical descriptions including lesion locations and features, providing a basis for subsequent image analysis and diagnosis.

[0051] Step 102: After the OCR extraction or text analysis is completed, the rule checking and context semantic correction model optimization are performed to ensure that all the extracted readable texts conform to the medical logic and context semantics, and remove the readable texts with character recognition errors or semantic inconsistencies (such as unit errors, symbol errors, etc.);

[0052] The normal value range and logical rules provided by the pre-built knowledge base are the basis for rule checking. The extracted content is checked by setting rules in the medical knowledge base, traversing the rule set of the knowledge base, matching the entities and their contextual descriptions in the extracted results one by one, and combining the contextual information of the extracted content with the pre-trained RoBERTa language model and rule constraints to further correct errors.

[0053] When used, combine the contents of steps 101 and 102;

[0054] Through rule checking, errors that may occur during the OCR process (such as unit errors and character confusion) can be corrected to improve data accuracy and ensure the effectiveness of extracted information. Combined with the contextual semantic model, it can discover implicit inconsistencies in the text content, such as contradictions between diagnostic results and inspection conclusions, to ensure that the generated text meets professional requirements. Automated semantic verification and knowledge base checking greatly reduce the proportion of manual review content and reduce the burden of repetitive work.

[0055] Step 2: Use the medical knowledge base to determine whether there are abnormalities in the readable text, select the corresponding marking method according to the degree of abnormality identified, and fuse the readable text and structured data to obtain structured abnormal data:

[0056] The step 2 includes the following contents:

[0057] Step 201: Use a medical knowledge base to determine whether the readable text is abnormal. The medical knowledge base provides the normal value ranges of various laboratory indicators and, in combination with the patient's specific situation (such as gender, age, medical history, etc.), realizes the automatic verification of indicators. For example, if a certain laboratory value exceeds the normal range, it is marked as abnormal;

[0058] For simple and clear indicator ranges (such as blood pressure, blood sugar, body temperature, etc.), use the knowledge base rules to determine whether it is abnormal and output the abnormal value, which can achieve efficient primary screening, filter out obviously normal samples, and reduce the model calculation cost; for complex indicators not covered by the knowledge base (such as electrocardiogram features, imaging data features) or scenarios that require joint feature judgment, train a supervised learning model through historical laboratory data, and use the trained supervised learning model to detect complex patterns of abnormal values. For example, classify the normal / abnormal situation of each laboratory indicator based on a classification model to obtain the structured data after detection;

[0059] The knowledge base provides the normal range of indicators and personalized conditions, automatically screens abnormal values in a large number of laboratory indicators, and quickly filters out obviously normal values. For complex indicators not covered by the knowledge base, by training a supervised learning model, it can effectively identify complex patterns of abnormalities, improve the universality of the algorithm, directly mark and classify clear abnormal values, and provide preliminary classification information to significantly shorten the judgment time.

[0060] Step 202: Conduct correlation analysis on the detected abnormal values with the medical knowledge base to provide interpretable abnormal reasons or suggestions. If the abnormalities detected by the model conflict with the medical knowledge base, mark them for subsequent manual review;

[0061] Use a pre-trained multi-modal neural network (such as BERT+MLP) to process the readable text and structured data, map the two into the same feature space and then fuse them to obtain structured abnormal data:

[0062] Through the upstream and downstream logic of the knowledge base, clarify the possible reasons and impacts of each abnormal value, reduce the speculation time, automatically mark the abnormal conflict points when the detection results do not match the knowledge base data, remind for manual review, and improve the credibility of the final results. The multi-modal neural network combines text and structured data, which can ensure that the extracted content is consistent with the patient's medical history and clinical information, and helps doctors comprehensively understand the patient's status.

[0063] Step 203: Collect and record abnormal factor data such as the abnormal severity, frequency, impact, and cost of the identified abnormal values, construct a Marking Priority Index (MPI), and flexibly decide on explicit marking or implicit marking based on scoring and thresholds.

[0064] Train a machine learning algorithm with the labeled sample data to obtain a trained priority evaluation model;

[0065] Use the outlier factor data with outliers as input, evaluate it using the trained priority evaluation model to obtain the marked priority index MPI. If the obtained marked priority index MPI exceeds the preset marking threshold, directly indicate the abnormal indicators, abnormal types, and their deviation degrees using explicit marking. Conversely, use implicit marking to provide potential analysis directions by hinting at the correlation issues of upstream and downstream indicators;

[0066] When using, combine the content in steps 201 to 203:

[0067] Assign priorities by calculating the severity, frequency, and impact of outliers, quickly focus on the most urgent or important abnormal data, improve processing efficiency, provide hints for minor abnormalities through implicit marking, avoid interference caused by a large amount of unimportant information, and provide a quantitative basis through the constructed marked priority index MPI to form a reference for subsequent manual review of abnormal data.

[0068] Step 3. Convert the structured abnormal data into generated text for correction, divide the corrected generated text into blocks according to the structure of the medical record template and fill them into the corresponding positions to generate medical record text, and clean up the duplicate content in the priority summary generated from the medical record text;

[0069] The said step 3 includes the following content:

[0070] Step 301. After converting the structured abnormal data (such as laboratory test indicators, imaging descriptions) into an acceptable input format, use the pre-trained language model to convert the structured abnormal data into natural language descriptions to obtain the generated text; use regular expressions to correct the generated text, including term correction and quantitative description;

[0071] Load and preprocess the medical knowledge graph. For each abnormal data, automatically match the relevant entities and their upstream and downstream relationships in the knowledge graph, use a semantic-based matching algorithm (such as similarity calculation) or a rule engine (keyword-based matching) for mapping, and use the established mapping to associate the semantics of the abnormal data with the patient's chief complaint, medical history, and examination results;

[0072] For the preliminary natural language descriptions generated by the model, supplement them using the knowledge graph relationships. Among them, if the possible disease causes and relevant examination suggestions are not mentioned for the outliers, they are automatically supplemented; for the detected conflicting content, the model automatically adjusts the generated text; convert the structured data into coherent natural language text through the pre-trained language model to relieve the writing pressure, use regular expressions to correct the terms and quantitative description of the generated text, ensure the professionalism and consistency of the medical description, and supplement the disease causes and examination suggestions related to the outliers by combining the medical knowledge graph to form more complete medical record information.

[0073] Step 302: Extract key content such as the chief complaint, current medical history, past medical history, and physical examination from the structurally abnormal data, construct a semantic model of the medical record content, use natural language processing technology to match the generated text with the extracted medical record content sentence by sentence, and use semantic similarity calculation to identify possible logical contradictions in the generated text;

[0074] Determine the standard template structure of the medical record, such as: Chief complaint: Describe the patient's main symptoms and duration in one sentence; Current medical history: Describe the occurrence, development, and treatment process of symptoms in chronological order; Physical examination: Standardize the recording of physical signs found during the examination; Auxiliary examinations: Record the results of laboratory and imaging examinations; Divide the generated text into blocks according to the structure of the medical record template, fill them into the corresponding positions, and automatically match keywords. Classify the descriptions in the generated text that match a certain block of content to obtain the medical record text;

[0075] Build a medical terminology library, use regular matching to uniformly correct the terms in the filled content to ensure the consistency of medical record terms. Among them, generally use a text classification model or a similarity calculation model to detect conflicting content, compare the generated text with the medical record content, mark the conflict points, and based on the detection results of the conflict points, adjust the generated text through a pre-trained language model to ensure its semantic consistency with the medical record content;

[0076] When in use, generate structured medical records such as the chief complaint, current medical history, and physical examination according to the standard template blocks, which is convenient for quick reference and understanding. Use semantic similarity calculation to identify potential logical contradictions in the generated text, and through automatic model adjustment, the logical consistency of the medical record content can be ensured. Correct the filled content through the terminology library to eliminate problems of synonymous terms or inconsistent term formats.

[0077] Step 303: Use a rule engine or NLP technology to extract all abnormal data involved in the medical record text, assign weights to each abnormal data based on a medical knowledge base or the analytic hierarchy process: Sort the abnormal data from high to low according to the calculated comprehensive weight and generate a priority summary. Use regular expressions or text deduplication algorithms (such as Token deduplication) to clean keywords and duplicate content, and clean the duplicate or redundant content in the generated summary.

[0078] When in use, combine the content in Steps 301 to 303:

[0079] Automatically generate a priority summary based on the weight sorting, highlighting key abnormalities, which can quickly focus on the most important abnormal data. Clean keywords and duplicate content through the deduplication algorithm, which can optimize the conciseness and effectiveness of the generated summary, prioritize the display of high-weight abnormal data, and provide detailed relevant suggestions for further analysis.

[0080] Step 4: Perform logical detection on the generated text, obtain the outliers therein, generate the anomaly degree Aop based on their anomaly status, and perform correlation detection on the medical record text when the obtained anomaly degree Aop exceeds the expectation, and generate a standardized electronic medical record when there is no error;

[0081] The said Step 4 includes the following contents:

[0082] Step 401: Construct medical logic rules based on medical guidelines, clinical experience, and norms, take the generated text (including chief complaint, present illness history, physical examination, laboratory examination, etc.) as input, analyze the content of the generated text, structurally extract data such as chief complaint, symptoms, examination results, diagnosis, etc., match them item by item with the medical logic rules, and mark them after detecting logical conflicts or omissions;

[0083] When in use, detect the contradictions and omissions in the text through medical logic rules, ensure that the generated content conforms to clinical norms, improve the logical rigor of the medical record content, match data such as chief complaint, examination results, diagnosis, etc. item by item, and compare them with the logic rules, potential problems can be found and marked, which can reduce the risks of missed diagnosis and misdiagnosis and reduce the omissions in manual review.

[0084] Step 402: Construct a normal value reference range library, extract all laboratory examination and physical sign data in the generated text, judge whether they exceed the normal range, if so, mark them as outliers, and obtain the corresponding anomaly status data, including the deviation amplitude from the normal value, clinical severity grading, etc.; generate the anomaly degree Aop from the anomaly status data in the following way:

[0085]

[0086] In the formula: V(t) is the monitoring index, Δ(t) is the normalized deviation degree, used to describe the deviation of the measured value V(t) from the normal range, f Δ (Δ) is the non-linear deviation penalty function, which uses an exponential form to enhance the sensitivity to abnormal deviations, f Δ (Δ) = e k·|Δ| -1, where k > 0 is the coefficient to adjust the deviation penalty degree, D(t) is the dynamic change rate, which describes the change amplitude of the measured value over time, and is calculated by the derivative of the measured value with respect to time: tanh(D(t)) is the smoothing penalty function for dynamic changes, which limits the upper limit of the contribution value of abnormal changes. Since tanh(x) ∈ [-1, 1], C(t) is the clinical severity grading, which describes the potential harm degree of the current abnormal state to health, and the range is [1, 5]. C = 1 means slightly harmless; C = 5 means extremely dangerous;

[0087] V(f) is the frequency component of the measured value, through Fourier transform Obtained, where λ is the weight of the high-frequency anomaly, with a value greater than 0, controlling the contribution of the frequency component to the anomaly degree, and f 0 is the frequency threshold, filtering out the high-frequency components after filtering out low-frequency fluctuations (normal physiological changes), and is used to identify abnormal fluctuations, and w Δ is the weight of the deviation amplitude; w C is the weight of the clinical severity, and w D is the weight of the dynamic change rate, where w Δ +w C +w D = 1;

[0088] If the anomaly degree Aop exceeds the expectation, check whether the abnormal value conforms to other contents in the medical record text: if not, mark it as an unreasonable abnormal value;

[0089] When in use, generate the anomaly degree Aop according to factors such as the amplitude of deviation from the normal value and the severity grading, providing a quantitative basis for the further processing of the anomaly. Check whether the abnormal value is consistent with other contents, and unreasonable abnormal values can be discovered and marked in time, reducing the risk of misguidance.

[0090] Step 403: Use the marked logical conflict points or abnormal values and the medical record background information as inputs, and regenerate the chief complaint or the description of the current medical history according to the logical conflict points; use the medical logic rule engine to check the adjusted medical record text again. If it is correct, generate a standardized electronic medical record that meets the format requirements of the usage end; use the text difference detection algorithm to compare the originally generated medical record with the automatically optimized medical record text, and mark all newly added, modified, or deleted parts.

[0091] When in use, combine the contents in steps 401 to 403:

[0092] Regenerate the text according to the logical conflict points to ensure the logical coherence between the contents such as the chief complaint and the current medical history, and can strengthen the content consistency.

[0093] Those of ordinary skill in the art can realize that, in combination with the units and algorithm steps of the examples described in the embodiments disclosed herein, they can be implemented by electronic hardware, or by a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraint conditions of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of this application.

[0094] Those skilled in the art can clearly understand that for the convenience and conciseness of description, the specific working processes of the systems, devices, and units described above can refer to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0095] In several embodiments provided by the present application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only for some logical function divisions. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces. The indirect couplings or communication connections of the devices or units can be in electrical, mechanical, or other forms.

[0096] The units described as separate components may or may not be physically separated. The components displayed as units may or may not be physical units, that is, they can be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0097] As described above, this is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for writing electronic medical records with the assistance of artificial intelligence, characterized in that: include, Extract text and identify specific areas of unstructured image data, generate readable text after combining with the medical knowledge base, and use pre-set rules and pre-trained language models to correct the readable text; Use the medical knowledge base to determine whether there are abnormalities in the readable text, select the corresponding marking method according to the degree of abnormality identified, and fuse the readable text and structured data to obtain structured abnormal data: The structured abnormal data is converted into generated text and then corrected. The corrected generated text is divided into blocks according to the structure of the medical record template and filled into corresponding positions to generate the medical record text. The duplicate content in the priority summary generated from the medical record text is cleaned up. Perform logic detection on the generated text and obtain abnormal values ​​therein, generate abnormality Aop based on its abnormal state, and perform association detection on the medical record text when the obtained abnormality Aop exceeds expectations, and generate a standardized electronic medical record when there is no error; wherein, use the marked logical conflict points or abnormal values ​​and medical record background information as input, and regenerate the chief complaint or current medical history description based on the logical conflict points; use the medical logic rule engine to check the adjusted medical record text again, and if there is no error, generate a standardized electronic medical record.

2. The method for writing electronic medical records with the aid of artificial intelligence according to claim 1, characterized in that: After preprocessing the handwritten test report or electronic medical record scan, use OCR through convolutional neural network and sequence-to-sequence model to recognize text and correct errors in the string; Use the pre-trained YOLOv5 detection algorithm to identify the row and column areas of the table, determine the boundaries of the cells, use OCR to extract the text in each cell, and record the key-value pairs corresponding to the rows and columns. Use the segmentation model trained by the U-Net algorithm to extract specific areas from medical images, and extract the characteristics of the lesions through a deep learning network.

3. The method for writing electronic medical records with the assistance of artificial intelligence according to claim 2, characterized in that: Through rule checking and context semantic correction model optimization, character recognition errors or semantic inconsistencies in readable text are removed; the extracted content is checked by setting rules through the medical knowledge base, traversing the rule set of the knowledge base, and matching the entities and their context descriptions in the extracted results one by one.

4. The method of artificial intelligence-assisted electronic medical record writing according to claim 3, characterized in that: Use the medical knowledge base to determine whether there are abnormalities in the readable text, use the knowledge base rules to determine whether the indicator range is abnormal and output the abnormal value, and use the trained supervised learning model to detect the complex patterns of abnormal values ​​for complex indicators not covered by the knowledge base to obtain structured data after detection; The detected outliers are analyzed in association with the medical knowledge base, and explainable causes or suggestions for the anomalies are provided. If the anomalies detected by the model conflict with the medical knowledge base, they are marked.

5. The method for writing electronic medical records with the aid of artificial intelligence according to claim 4, characterized in that: Use a pre-trained multimodal neural network to process readable text and structured data, map them to the same feature space, and then fuse them to obtain structured abnormal data: Collect and record the abnormal factor data of the identified outliers, use the abnormal factor data of the outliers as input, use the trained priority evaluation model to evaluate, and obtain the marking priority index MPI. If the obtained marking priority index MPI exceeds the preset marking threshold, use explicit marking to directly point out the abnormal indicators, abnormal types and their degree of deviation. If the opposite is true, use implicit marking to prompt the correlation problems of upstream and downstream indicators.

6. The method of artificial intelligence-assisted electronic medical record writing according to claim 5, characterized in that: After converting the structured anomaly data into an acceptable input format, use the pre-trained language model to convert the structured anomaly data into a natural language description, obtain the generated text, and use regular expressions to correct the generated text; Load and preprocess the medical knowledge graph, automatically match the relevant entities and their upstream and downstream relationships in the knowledge graph for each abnormal data, use a semantic-based matching algorithm for mapping, use the established mapping to associate the semantics of the abnormal data with the patient's chief complaint, medical history, and examination results, and use the knowledge graph relationship to supplement the generated preliminary natural language description.

7. The method of artificial intelligence-assisted electronic medical record writing according to claim 6, characterized in that: Construct a semantic model of medical record content, use natural language processing technology to match the generated text with the extracted medical record content sentence by sentence, and use semantic similarity calculation to identify logical contradictions in the generated text; Determine the standard template structure of the medical record, divide the generated text into blocks according to the structure of the medical record template, fill in the corresponding positions, and automatically match keywords. Classify the descriptions in the generated text that match a certain block of content to obtain the medical record text; build a medical terminology library, and use regular matching to uniformly correct the terminology of the filled content.

8. The method of artificial intelligence-assisted electronic medical record writing according to claim 7, characterized in that: Use rule engines or NLP technology to extract all abnormal data involved in the medical record text, assign weights to each abnormal data based on the medical knowledge base or hierarchical analysis method, sort the abnormal data from high to low according to the calculated comprehensive weight, and generate a priority summary.

9. The method of artificial intelligence-assisted electronic medical record writing according to claim 8, characterized in that: Construct medical logic rules, take generated text as input, match it with medical logic rules item by item, detect logical conflicts or missing points and mark them; Build a normal value reference range library, extract all laboratory test and physical sign data in the generated text, determine whether it exceeds the normal range, if so, mark it as an abnormal value, and obtain the corresponding abnormal status data, generate the abnormality Aop from the abnormal status data, if the abnormality Aop exceeds expectations, check whether the abnormal value is consistent with other content in the medical record text: if not, mark it as an unreasonable abnormal value.

10. The method of artificial intelligence-assisted electronic medical record writing according to claim 9, characterized in that: The method of generating abnormality degree Aop from abnormal status data is as follows: Where: V(t) is the monitoring index, Δ(t) is the normalized deviation degree, k>0, D(t) is the dynamic change rate, tanh(D(t)) is the dynamic change smooth penalty function, and C(t) is the clinical severity grade; V(f) is the frequency component of the measured value, λ is the weight of the high-frequency anomaly, f0 is the frequency threshold, and w Δ 、w C and D is the weight, where w Δ +w c +w D =1.

Images