Multi-source clinical data quality inspection method and device based on large model and electronic equipment

Abstract

The invention discloses a multi-source clinical data quality inspection method and device based on a large model and electronic equipment, and belongs to the technical field of data processing.The multi-source clinical data quality inspection method based on the large model comprises the steps that multi-source clinical data are obtained, and the multi-source clinical data comprise electronic medical records, laboratory inspection reports, medical image reports and operation records; performing textualization processing on the multi-source clinical data to obtain a format text; determining cue words according to the format text, wherein the cue words are correspondingly matched with the quality inspection rules; utilizing cue words to guide the trained deep learning model to extract keywords in the format text; and carrying out logic judgment on the keywords by utilizing a quality inspection rule to obtain a quality inspection result. According to the method, the problem that the rule engine is insufficient in fixity and adaptability is solved, and the accuracy and flexibility of multi-source clinical data quality inspection are remarkably improved.

Description

Technical Field

[0001] The present application belongs to the field of data processing technology, and specifically relates to a multi-source clinical data quality inspection method, device and electronic equipment based on a large model. Background Art

[0002] In modern medical research and clinical trials, data quality control is the core link to ensure the reliability and validity of research results.

[0003] With the development of medical information technology, the data sources in clinical research have become increasingly diverse, including not only traditional electronic medical records (EMR) and laboratory test data, but also various forms of clinical data such as gene sequencing, medical imaging, and patient self-report. The quality of these data directly affects the conclusions and decisions of clinical research. Therefore, how to efficiently and accurately perform data quality inspection and verification has become an urgent problem to be solved in clinical research. Summary of the invention

[0004] The purpose of this application is to provide a multi-source clinical data quality inspection method, device and electronic equipment based on a large model to efficiently and accurately inspect various forms of clinical data.

[0005] According to a first aspect of an embodiment of the present application, a multi-source clinical data quality inspection method based on a large model is provided, and the method may include: Obtain multi-source clinical data, including electronic medical records, laboratory test reports, medical imaging reports and surgical records; Perform text processing on multi-source clinical data to obtain formatted text; Determine prompt words based on format text, and match prompt words with quality inspection rules; Use prompt words to guide the trained deep learning model to extract keywords from formatted text; Use quality inspection rules to perform logical judgment on keywords and obtain quality inspection results.

[0006] In some optional embodiments of the present application, obtaining multi-source clinical data includes: Import multi-source clinical data from hospital information systems and laboratory information systems to ensure the diversity and integrity of data sources.

[0007] In some optional embodiments of the present application, multi-source clinical data is processed into text format to obtain a formatted text, including: Performing optical character recognition processing and standardization processing on the image data to obtain a first text; Perform text parsing on the PDF file and the DOCX file to obtain a second text; Performing standardization on the TXT file to obtain a third text; The first text, the second text and the third text are processed in a unified format to obtain a formatted text.

[0008] In some optional embodiments of the present application, the first text, the second text, and the third text are subjected to unified format processing to obtain formatted texts, including: Segmenting the first text, the second text, and the third text to obtain a first text segment set, a second text segment set, and a third text segment set; The first text segment set, the second text segment set and the third text segment set are uniformly reconstructed to obtain a structurally unified text set; The text sets with unified structure are merged to obtain formatted text.

[0009] In some optional embodiments of the present application, using prompt words to guide the trained deep learning model to extract keywords in the formatted text includes: The prompt word and the format text are combined into a structured text, wherein the prompt word and the format text are in different fields in the structured text; Input structured text into the trained deep learning model to obtain keywords.

[0010] In some optional embodiments of the present application, the trained deep learning model is obtained through LoRA fine-tuning training.

[0011] In some optional embodiments of the present application, after performing logical judgment on keywords using quality inspection rules, the multi-source clinical data quality inspection method based on a large model further includes: Use the big model to perform semantic analysis on keywords and obtain semantic understanding results; The final verification result is obtained by combining the logical judgment results and semantic understanding results.

[0012] According to a second aspect of an embodiment of the present application, a multi-source clinical data quality inspection device based on a large model is provided, and the device may include: The acquisition module is used to acquire multi-source clinical data, including electronic medical records, laboratory test reports, medical imaging reports and surgical records; The text processing module is used to process multi-source clinical data into text format to obtain formatted text; A determination module is used to determine prompt words according to the format text, and the prompt words are matched with the quality inspection rules; An extraction module is used to use prompt words to guide the trained deep learning model to extract keywords from multi-source clinical data; The quality inspection module is used to perform logical judgment on keywords using quality inspection rules to obtain quality inspection results.

[0013] According to a third aspect of an embodiment of the present application, an electronic device is provided, which may include: processor; a memory for storing processor-executable instructions; The processor is configured to execute instructions to implement the multi-source clinical data quality inspection method based on a large model as shown in any one of the embodiments of the first aspect.

[0014] According to the fourth aspect of the embodiments of the present application, a storage medium is provided. When the instructions in the storage medium are executed by a processor of an information processing device or a server, the information processing device or the server implements a multi-source clinical data quality inspection method based on a large model as shown in any one of the embodiments of the first aspect.

[0015] The above technical solution of the present application has the following beneficial technical effects: The method of the embodiment of the present application combines the deep semantic understanding ability of the deep learning model with the logical verification ability of the traditional rule engine, realizing intelligent quality inspection and dynamic verification of multi-source clinical data. The deep learning model pre-judges the newly entered multi-source clinical data based on the prompt words, and combines with the rule engine to dynamically adjust the quality inspection rules to ensure that it can efficiently adapt to and verify complex and diverse clinical data. It solves the problem of insufficient fixity and adaptability of the rule engine, and significantly improves the accuracy and flexibility of multi-source clinical data quality inspection. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] FIG1 is a schematic flow chart of a multi-source clinical data quality inspection method based on a large model in an exemplary embodiment of the present application; FIG2 is a schematic diagram of the structure of a multi-source clinical data quality inspection device based on a large model in an exemplary embodiment of the present application; FIG3 is a schematic diagram of the structure of an electronic device in an exemplary embodiment of the present application; FIG. 4 is a schematic diagram of the hardware structure of an electronic device in an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is further described in detail below in conjunction with specific implementations and with reference to the accompanying drawings. It should be understood that these descriptions are only exemplary and are not intended to limit the scope of the present application. In addition, in the following description, the description of known structures and technologies is omitted to avoid unnecessary confusion of the concepts of the present application.

[0018] The accompanying drawings show schematic diagrams of layer structures according to embodiments of the present application. These figures are not drawn to scale, and some details are magnified and some details may be omitted for the purpose of clarity. The shapes of various regions and layers shown in the figures and the relative sizes and positional relationships therebetween are only exemplary, and may deviate in practice due to manufacturing tolerances or technical limitations, and those skilled in the art may additionally design regions / layers with different shapes, sizes, and relative positions according to actual needs.

[0019] Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present application.

[0020] In the description of the present application, it should be noted that the terms "first", "second" and "third" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0021] In addition, the technical features involved in the different embodiments of the present application described below can be combined with each other as long as they do not conflict with each other.

[0022] In the following, in combination with the accompanying drawings, the multi-source clinical data quality inspection method, device and electronic device based on a large model provided in the embodiments of the present application are described in detail through specific embodiments and their application scenarios.

[0023] like Figure 1 As shown, in a first aspect of an embodiment of the present application, a multi-source clinical data quality inspection method based on a large model is provided, and the method may include: S110: Acquire multi-source clinical data, including electronic medical records, laboratory test reports, medical imaging reports and surgical records; S120: Text processing is performed on multi-source clinical data to obtain formatted text; S130: Determine a prompt word according to the format text, and match the prompt word with the quality inspection rule; S140: Using prompt words to guide the trained deep learning model to extract keywords from the formatted text; S150: Perform logical judgment on the keywords using quality inspection rules to obtain quality inspection results.

[0024] The method of this embodiment combines the deep semantic understanding ability of the deep learning model with the logical verification ability of the traditional rule engine to realize intelligent quality inspection and dynamic verification of multi-source clinical data. The deep learning model predicts the newly entered multi-source clinical data based on the prompt words, and combines with the rule engine to dynamically adjust the quality inspection rules to ensure that it can efficiently adapt to and verify complex and diverse clinical data. It solves the problem of insufficient fixity and adaptability of the rule engine, and significantly improves the accuracy and flexibility of multi-source clinical data quality inspection.

[0025] The deep learning model in this embodiment can be the Baichuan big model, which combines the deep semantic understanding ability of the Baichuan big model with the logical verification ability of the traditional rule engine to realize intelligent quality inspection and dynamic verification of multi-source clinical data. Pre-judge the new text data according to the prompt words, and dynamically adjust the quality inspection rules in combination with the rule engine to ensure efficient adaptation and verification of complex and diverse clinical data. Keywords may include: disease names in electronic medical records, key values ​​in laboratory reports, and diagnostic information in image descriptions. Quality inspection rules describe different types of quality inspection tasks (such as extracting patient names, determining report categories, etc.), which are used to compare the extracted keywords with the records in the system database, and the comparison results are quality inspection results.

[0026] In some embodiments, obtaining multi-source clinical data includes: Import multi-source clinical data from hospital information systems and laboratory information systems to ensure the diversity and integrity of data sources.

[0027] In this embodiment, the multi-source clinical data may include electronic medical records, laboratory test reports, medical imaging reports, surgical records, etc. The original format of these data may be pictures (such as scanned medical records and imaging reports), PDF documents, DOCX files or plain text (TXT).

[0028] In some embodiments, text processing is performed on multi-source clinical data to obtain formatted text, including: Performing optical character recognition processing and standardization processing on the image data to obtain a first text; Perform text parsing on the PDF file and the DOCX file to obtain a second text; Performing standardization on the TXT file to obtain a third text; The first text, the second text and the third text are processed in a unified format to obtain a formatted text.

[0029] In this embodiment, for scanned images or imaging reports, optical character recognition (OCR) technology (PaddleOCR) can be used to extract text content in the image. OCR processing is not just a simple text extraction, but also includes a preliminary cleaning step to remove noise, correct recognition errors, and ensure high accuracy of the text. Combined with a custom preprocessing step, OCR technology effectively processes noisy medical imaging texts to ensure the accuracy and reliability of the extracted content.

[0030] You can use PyMuPDF and python-docx tools to parse PDF and DOCX files and extract the document content as structured text data. During the parsing process, the structural information such as paragraphs, titles, and tables in the document is retained to ensure that the integrity of the document content is not damaged when converted to text. This processing of retaining the document structure enhances the subsequent model's ability to understand the text semantics, especially when processing complex medical reports.

[0031] For plain text files, the content can be directly read and standardized. The standardization process includes removing extra spaces and format characters, unifying the text format, and ensuring that all data has a consistent structure before entering the Baichuan large model. The standardization process effectively eliminates the analysis bias that may be caused by inconsistent text formats and ensures the reliability and consistency of processing.

[0032] In some embodiments, the first text, the second text, and the third text are subjected to unified format processing to obtain formatted texts, including: Segmenting the first text, the second text, and the third text to obtain a first text segment set, a second text segment set, and a third text segment set; The first text segment set, the second text segment set and the third text segment set are uniformly reconstructed to obtain a structurally unified text set; The text sets with unified structure are merged to obtain formatted text.

[0033] This embodiment can organically combine OCR, PDF parsing, DOCX parsing and TXT processing, thereby improving the accuracy and consistency of multi-source document processing. When processing documents in multiple formats, intelligent segmentation can be performed based on the document content to ensure that the best effect can be achieved in the appropriate part.

[0034] Exemplarily, density-based text analysis technology is used to analyze the layout and content features of the document. This includes detecting the text density, paragraph structure, table location, etc. in different areas of the document. When processing scanned images, edge detection algorithms are used to identify text blocks and non-text areas, and the complexity of the text areas is determined based on this information. Based on the content feature analysis, the document is segmented into multiple logical blocks. Each logical block is mapped to the most appropriate parsing module based on its features. Blocks containing complex tables are assigned to the PDF parsing module, while images and complex typesetting areas are processed by the OCR module. This adaptive segmentation and mapping mechanism ensures that the document content can be parsed in the best way.

[0035] During the parsing process, the document structure needs to be extracted from multiple sources and reconstructed into a unified representation to ensure consistency across document formats.

[0036] Exemplarily, a rule-based structure extraction method is used to extract structural information from OCR, PDF parsing and DOCX parsing modules. This includes paragraph hierarchy, table layout, title format, etc. Specifically, the position information extracted by the OCR module is processed using position encoding technology to ensure that the spatial relationship of the text block is correctly captured. The PDF and DOCX parsing modules use the DOM tree structure parsing method to extract the hierarchical structure of the document. After extracting the structural information, the output of each module is integrated into a unified document structure representation through a template-driven structure reconstruction algorithm. Align the text blocks extracted by OCR with the paragraph information of PDF parsing to ensure accurate correspondence between the text content and the paragraph boundaries. The position of the text block can be automatically adjusted by the paragraph alignment algorithm to ensure that all text content is correctly placed. The table data is reconstructed by the table merging algorithm to ensure the integrity and accuracy of the table structure.

[0037] To ensure that the final text data is consistent with the original document in content and structure, the output results from different sources need to be integrated and fully verified.

[0038] Exemplarily, a confidence-weighted fusion method can be used to fuse text data processed by OCR, PDF parsing, DOCX parsing, and TXT. A confidence score is assigned to each parsing result, which is based on preset rules and parsing accuracy assessment. Parsing results with higher confidence levels will be given priority, and the text content will be integrated using a weighted average fusion algorithm. If there are differences between the OCR and PDF parsing results, the content and order of the text output can be dynamically adjusted according to their respective confidence levels. The fused text data needs to undergo a global consistency check to ensure consistency in structure and content. Similarity matching can be used to compare text content, paragraph structure, and table layout to ensure consistency in output from each module. Check whether the text extracted by OCR is aligned with the paragraph structure of PDF parsing, and use semantic consistency detection to verify whether the text content is consistent between different sources. If inconsistencies are found, mark these areas and make necessary adjustments.

[0039] After data fusion and verification, a standardized output can be generated to ensure that it can be used in various practical applications.

[0040] Exemplarily, a structured text output can be generated, including all extracted text content, paragraph structure, and table layout information. Output formats such as JSON, XML, etc. ensure data compatibility and scalability. Formatting validation tools can be applied to ensure that the output data meets the specified format requirements and can be directly used in a variety of applications. API interfaces and data conversion tools can be provided to seamlessly integrate the processed data into electronic medical record systems, data analysis platforms, or other clinical applications. Such interfaces and tools ensure the wide applicability of the processing results and can support a variety of subsequent analysis and processing tasks.

[0041] In some embodiments, using prompt words to guide the trained deep learning model to extract keywords from the formatted text includes: The prompt word and the format text are combined into a structured text, wherein the prompt word and the format text are in different fields in the structured text; Input structured text into the trained deep learning model to obtain keywords.

[0042] The trained deep learning model in this embodiment is obtained by training through the following steps: 1. Data preprocessing and standardization In order to ensure that different types of clinical data (such as electronic medical records, laboratory reports, image descriptions, etc.) can be effectively understood by the big model, these data must first be preprocessed and standardized to unify the format and eliminate noise. This can be achieved through text cleaning and noise removal.

[0043] Regular expression cleaning: Through Regular Expressions, noise characters in the text can be removed, such as unnecessary special symbols, repeated blank lines, etc.

[0044] The specific implementation includes: Special symbol filtering: Using regular expressions, such as re.sub(r'[^\w\s]', '', text), to remove non-alphanumeric special symbols in the text, such as @, #, $, etc. These symbols are usually not beneficial to semantic analysis and may interfere with the model's understanding.

[0045] Whitespace cleaning: Standardize redundant spaces, line breaks, etc. in the text to a single space or line break through re.sub(r'\s+', ' ', text), ensuring the simplicity and consistency of the text structure.

[0046] Term standardization: Glossary application: A predefined medical glossary can be used to standardize terms and abbreviations in the text to ensure term consistency. For example, standardize "glucose" to "blood sugar". The implementation methods include: Lexical replacement: Use a Python dictionary for term replacement, such as text = text.replace("glucose", "blood sugar"), to ensure that terms in the entire text are consistent.

[0047] Structured processing: The spaCy natural language processing tool can be used to segment and split sentences in the document. For example, use spaCy for paragraph recognition and segmentation parsing. Through parsing, the text is split into logical blocks, such as patient information, laboratory results, diagnosis conclusions, etc.

[0048] Field tagging: Each text block can be field-tagged through NLP techniques. For example, using the entity recognition function of spaCy, tag "patient name" as "NAME" and "laboratory results" as "LAB_RESULT" to ensure that these key fields can be accurately recognized and processed in subsequent processing.

[0049] 2. Prompt generation and mapping In the quality inspection task, prompts are used to clarify the task objectives of the large model. After the prompts are generated, they are combined with the preprocessed and structured data to form the input of the model, thus helping the model focus on specific quality inspection tasks.

[0050] A built-in prompt word template library can include prompt word templates for common quality inspection tasks, such as "extracting key indicators from laboratory reports" or "comparing diagnostic information with prescribed medication consistency." These templates are designed by experts in the medical field to ensure that they can cover a wide range of quality inspection needs.

[0051] Dynamically select the appropriate prompt word template based on the type and content of the input data. For example, for laboratory report data, the prompt word "extract blood glucose level and compare with system records" will be selected and combined with the corresponding data block to ensure that the model understands the specific goal of the task.

[0052] The prompt word is combined with the structured data to form the input data pair of the model. For example, the prompt word "extract blood glucose level" is combined with the preprocessed data block "patient blood glucose: 8.5 mmol / L" to generate the data pair { "prompt": "extract blood glucose level", "text": "patient blood glucose: 8.5 mmol / L"}.

[0053] Use JSON format to combine prompt words and structured data to ensure the structure and consistency of input data. For example, {"prompt": "Extract blood glucose level", "content": "Patient blood glucose: 8.5 mmol / L"}. This structured input helps large models clarify task objectives and process data effectively.

[0054] The prompt words and input data are organized in a specific format (such as JSON). In this structured format, the prompt words and input data are placed in different fields so that the model can clearly identify the purpose of each part when processing.

[0055] Example format (JSON): { "prompt": "Extract the blood glucose level from the lab report and compare it with the blood glucose value recorded in the system", "data": "Patient blood sugar: 8.5 mmol / L, normal range: 4.0 - 6.0 mmol / L" } In this example: The "prompt" field contains the prompt word for the QA task.

[0056] The "data" field contains the standardized input data, i.e. the text content of the laboratory report.

[0057] This formatted structure allows the model to directly identify and distinguish prompt words and input data through field names. When processing, the model will first parse the prompt words, and then focus on the data part according to the guidance of the prompt words to perform corresponding task processing.

[0058] The output of the model varies depending on the type of task. For entity extraction tasks, the output includes the identified entity (such as blood sugar value), its location in the text, and the comparison results with the existing records in the system. The model outputs "8.5mmol / L, located at characters 45-47 of the text, which is inconsistent with the 7.2 mmol / L recorded in the system, with a difference of 1.3 mmol / L" to help quality inspectors quickly locate and handle inconsistent data.

[0059] In some embodiments, the trained deep learning model is obtained through LoRA fine-tuning training.

[0060] This embodiment optimizes the model through LoRA fine-tuning so that it can better understand and process disease names in electronic medical records, key values ​​in laboratory reports, diagnostic information in image descriptions, etc.

[0061] Specifically, it may include: Introduction of low-rank matrices: During the fine-tuning process, LoRA technology replaces the weight updates of certain layers in the model by introducing low-rank matrices. LoRA does not change the structure or hierarchy of the model, but reduces the number of parameters that need to be fine-tuned through parameterization. For a specific weight matrix W (the weight of the self-attention layer or the feedforward layer), LoRA approximates it as two low-rank matrices A and B, thereby reducing the number of parameters that need to be updated: W'=W+α×(A×B) Among them, W is the frozen original weight, A and B are low-rank matrices, and α is a scaling factor used to control the impact of the fine-tuning part.

[0062] Fine-tuning process: For example, based on the Baichuan model, the LoRA technology is introduced to perform fine-tuning on the low-rank matrices A and B.

[0063] During fine-tuning, the entire weight matrix W is not updated, but only the low-rank matrices A and B are trained. This approach can significantly reduce the consumption of computing resources while optimizing for quality inspection tasks. For example, when processing electronic medical records, only the weights related to a specific task need to be adjusted, allowing the model to better extract and compare disease information.

[0064] Through LoRA technology, the graphics memory and computing resources required during the fine-tuning process can be significantly reduced, which makes it possible to efficiently complete the fine-tuning of large-scale models in resource-constrained environments, such as on internal hospital servers.

[0065] The fine-tuned model is used to process large amounts of medical text data (such as laboratory reports, imaging diagnoses, etc.). Through optimized parameters, the model can complete quality inspection tasks more efficiently and quickly adapt to new medical data sources or task requirements.

[0066] Through this fine-tuning approach, LoRA technology can efficiently enhance the performance of large models in specific medical tasks while avoiding the high computational cost required for full model fine-tuning.

[0067] Processing multi-source medical data usually requires processing a large amount of text and image data, which is huge and requires high model computing resources. In an environment with limited hardware resources, using gradient accumulation and mixed precision training can effectively reduce video memory usage and improve computing efficiency.

[0068] Specific implementation steps: Small batch data processing: Due to the large amount and complex structure of medical data, thousands of electronic medical records or laboratory reports may not be able to load large batches of data for training at one time. Therefore, small batches (mini-batch) are used to process data.

[0069] Implementation of gradient accumulation: During each mini-batch training process, the gradient of the current mini-batch is calculated, but the model weights are not updated immediately. Instead, these gradients are accumulated into a cumulative gradient variable. When processing each medical record text, the gradients of the disease status extracted from the text are calculated one by one and accumulated.

[0070] When the accumulated gradients reach the preset large batch size, these accumulated gradients are used to perform a backpropagation and weight update, thereby simulating the training effect of a large batch and ensuring that the model learns under stable conditions.

[0071] In this way, efficient training can be performed under limited video memory conditions, avoiding training interruptions caused by insufficient video memory, while ensuring the stability and effectiveness of the model in training large amounts of medical data.

[0072] Application of mixed precision training in medical text and image data processing: Half-precision processing: During forward propagation, half-precision floating point numbers (FP16) are used for calculations. At this time, when the model processes input medical data (such as imaging report text or laboratory data), it uses FP16 format calculations to reduce video memory usage and speed up processing. For example, input text content such as "Patient blood sugar level: 8.5 mmol / L" is processed under FP16.

[0073] In key operations such as gradient calculation and model weight update, single-precision floating point numbers (FP32) are used for calculation to ensure stability and accuracy of calculation. This is especially important when processing medical data that requires high precision (such as verification of laboratory test values).

[0074] Use the mixed precision training tool torch.cuda.amp to manage and execute mixed precision operations, ensuring that gradients maintain accuracy during scaling and updates.

[0075] Mixed precision training can significantly reduce resource usage and increase training speed when processing complex medical text and image data, thereby completing large-scale model fine-tuning in an environment with limited hardware resources. For example, massive patient data in hospital information systems can be trained in a shorter time in this way, improving the real-time response capability of quality inspection systems.

[0076] In some embodiments, after using the quality inspection rules to perform logical judgment on the keywords, the multi-source clinical data quality inspection method based on the big model further includes: Use the big model to perform semantic analysis on keywords and obtain semantic understanding results; The final verification result is obtained by combining the logical judgment results and semantic understanding results.

[0077] In this embodiment, all quality inspection rules and their corresponding prompt words are predefined and stored in the database. These rules describe different types of quality inspection tasks (such as extracting patient names, determining report categories, etc.), and each rule is associated with one or more prompt words, which are used to guide the Baichuan large model to perform specific tasks. Each quality inspection rule is associated with its prompt word through a unique identifier (rule ID). This association method ensures that when receiving specific uploaded content, the corresponding quality inspection rules and prompt words can be quickly retrieved. The uploaded content type (such as medical record text, laboratory report, image description, etc.) can be monitored in real time. According to the type of uploaded content, it is determined whether quality inspection is required. If necessary, match the applicable quality inspection rules. After obtaining the quality inspection results, the logical relationship between the extracted diagnostic information and the medication record is also verified to verify its rationality and obtain the verification result.

[0078] For example, when it detects that the uploaded file is a medical record text, it will check whether there are corresponding quality inspection rules for this type of content.

[0079] Rule triggering mechanism: Once the type of uploaded content is determined, the quality inspection rules in the database are retrieved according to the predefined trigger conditions. For example, if the uploaded content is a medical record text, query all quality inspection rules for medical record text, such as "extract the patient name in the medical record text and compare it with the system record." By matching the trigger conditions of the rules, find the corresponding prompt words. Extract the prompt words related to the current uploaded content from the database, and pass these prompt words as part of the input to the Baichuan large model. For example, for the uploaded medical record text, generate the prompt word "extract the patient name in the medical record text and compare it with the patient name recorded in the system."

[0080] Data input format: The generated prompt words are combined with the actual uploaded content (text data) to form the input data pair of the Baichuan model. The format is as follows: { "prompt": "Extract the patient name from the medical record text and compare it with the patient name recorded in the system", "data": "Patient name: Zhang San, gender: male, hospital number: 12345" } In this format, the prompt field carries the prompt word, which clearly tells the model the specific goal of the current task; the data field contains the actual text data.

[0081] Execution and result processing of Baichuan large model: After receiving the input data, the Baichuan large model performs the corresponding quality inspection tasks according to the prompt words. For example, the model recognizes the prompt word "extract patient name", then extracts the entity "Zhang San" from the data and compares it with the records in the system. After the model processing is completed, further processing is performed based on the results of the model output. For example: if the model determines that the patient's name is consistent with the system record, it will mark the quality inspection as passed. If it is inconsistent, an error report is generated to prompt the user or automatically initiate the data correction process.

[0082] The quality inspection method of this embodiment not only performs tasks based on predefined prompt words, but also combines the logic verification function of the rule engine to dynamically adjust the quality inspection rules when processing data in real time. For example, the rule priority can be adjusted according to the characteristics of the input data (such as data complexity, content type), or new verification rules can be introduced when anomalies are detected. This dynamic adjustment mechanism can flexibly respond to complex and changing data scenarios and ensure the adaptability and accuracy of the quality inspection process.

[0083] Each QA rule has a default priority, usually represented by a number or weight (1 to 10, with higher numbers indicating higher priorities). Priorities are set based on the importance of the rule, the type of data the rule processes, and the degree of impact the rule has on the QA results. For example, a rule that verifies patient identity information may be given a higher priority, while a format check may be given a lower priority.

[0084] When new input data is received, a preliminary analysis of the complexity and content type of the data is performed. The specific analysis includes: Data complexity: Assess whether the data involves multiple types of content. For example, if the data includes multiple types such as medical history text, imaging reports, and laboratory test results, increase the priority of related quality inspection rules based on the complexity of these data. Complex medical history text will increase the priority of rules related to key information extraction, while imaging reports may increase the priority of rules related to description consistency checks.

[0085] Content type: Identify the specific type of input data, such as medical history text, laboratory report, imaging report, etc. Based on these data types, adjust the priority of related rules accordingly.

[0086] For example, since medical history texts usually contain a lot of free text and key medical terms, rules related to key entity extraction and logical consistency checking may be prioritized. Imaging reports usually need to be compared with medical records and laboratory results, which will increase the priority of rules related to description consistency and diagnostic accuracy. For laboratory reports, rules related to numerical verification may be prioritized to ensure the accuracy of key indicators.

[0087] The priority of each rule is dynamically adjusted based on the analysis results of the input data. For example, if an abnormal value is detected in a laboratory report (such as a blood sugar level that does not meet the normal range), the priority of the rules related to abnormal value processing may be increased by 10% to ensure that these issues are handled first.

[0088] In some cases, the priority of the rules is adjusted based on specific conditions. For example, when the data contains the keywords "urgent" or "abnormal", the priority of the relevant rules will be raised to the highest level to ensure that these critical issues can be handled in a timely manner.

[0089] Under the guidance of quality inspection rules, the Baichuan large model extracts key entities from the input text. For example, the patient's name, diagnosis information, medication records, laboratory data, etc. can be accurately extracted from the medical record text according to the prompt "Extract patient name". The extraction process is based on the deep semantic knowledge learned by the model during fine-tuning, ensuring that key data can be accurately extracted in complex medical scenarios.

[0090] The extracted key entities will be compared with the records in the system database. Not only will these entities be checked to see if they are consistent with the database records, but a deeper consistency check will also be performed, such as comparing laboratory data across time points to ensure data continuity and consistency. This multi-level comparison and inspection can effectively discover and correct anomalies in the data, ensuring data integrity and accuracy.

[0091] Verify the logical relationship between the extracted diagnostic information and medication records to verify their rationality. For example, check whether the diagnosed disease matches the medication regimen to ensure the logical consistency of clinical decision-making. Logical verification not only relies on the logical analysis capabilities of the rule engine, but also combines the semantic understanding of the Baichuan big model to ensure the accuracy and effectiveness of the verification. It should be noted that the key entity in this embodiment is a keyword.

[0092] First, key diagnostic information is extracted from the input medical text (such as medical records, imaging reports, laboratory reports, etc.). This information includes disease name, symptom description, diagnosis conclusion, etc.

[0093] Medication record extraction: At the same time, medication-related information is extracted from the text, including drug name, dosage, duration of use, method of use, etc.

[0094] The rule engine contains a library of predefined logical rules that are based on standard knowledge in the medical field. For example, certain diseases require specific medications, or certain drugs are contraindicated for specific diseases. The extracted diagnostic information and medication records are matched against the rules in the rule library respectively. Logical matching is performed based on these rules to check whether the diagnosed disease is consistent with the medication used. For example, if the diagnosis is "hypertension", check whether antihypertensive drugs are correctly prescribed or whether contraindicated drugs are incorrectly prescribed. Detect inconsistencies between diagnostic information and medication records. For example, if the diagnosis is "asthma", but contraindicated asthma drugs appear in the medication record, mark this inconsistency and trigger further inspection.

[0095] The Baichuan big model is used to understand the contextual relationship between diagnostic information and medication records. Through its pre-trained semantic understanding capabilities, the big model can identify more complex semantic relationships in the text, such as the association between implicit diagnostic information and medication recommendations. The big model can understand ambiguous expressions or implicit logic in the language and determine whether these expressions are medically reasonable. For example, the big model can identify whether the symptoms described in the text are semantically matched to the prescribed medication.

[0096] Combining the rule engine with large model verification: First, the rule engine will perform a preliminary logic check to handle clear rule matching and contradiction detection. Then, the big model further analyzes the complex semantic relationships in the text, especially when the rule engine cannot handle fuzzy expressions, the model can make up for this deficiency. Finally, the logical analysis results of the rule engine are combined with the semantic understanding results of the big model to make a comprehensive judgment. If the two come to the same conclusion on a certain issue (such as a problem with the rationality of medication), the priority of this issue will be increased and a detailed error report will be generated.

[0097] After receiving the medical text input by the doctor, it is preprocessed to extract the diagnosis information and medication records. The rule engine is used to perform a preliminary logical check on the extracted information to determine whether it meets the medical standards. The big model further analyzes the semantics of the text, especially to check the complex logic or implicit relationships that are difficult for the rule engine to handle. The analysis results of the rule engine and the big model are combined to generate the final verification results. If problems are found, they are marked and a report is generated for the doctor's reference.

[0098] Through this multi-level logic verification mechanism, the logical relationship between diagnosis and medication can be accurately identified in medical texts, and through the combination of the rule engine and the Baichuan big model, higher verification accuracy and effectiveness can be achieved.

[0099] During the verification process, detect anomalies in the data, such as inconsistent patient names, text classification errors, and laboratory data outside a reasonable range. Generate a detailed exception report to record each detected problem and provide corresponding error descriptions and cause analysis. These reports are not only used to prompt users, but also provide basic data support for subsequent exception processing. Feedback the exception information to the user and provide detailed correction suggestions. After receiving the feedback, the user can correct the data according to the prompts. The corrected data will re-enter the quality inspection process and be verified again to ensure the accuracy of all data. By combining user input and the rule engine, the quality inspection rules can be automatically adjusted in the subsequent quality inspection process to reduce the recurrence of similar problems and further improve the overall efficiency of the system.

[0100] By continuously monitoring the feedback information during the quality inspection process, the quality inspection rule base is dynamically optimized. For example, if a certain type of prompt word frequently causes anomalies, the relevant rules will be automatically adjusted or the prompt word will be optimized based on user feedback and quality inspection results to ensure that the quality inspection rules are more accurate and effective. The quality inspection rules can be continuously optimized iteratively based on the input of new data and changes in quality inspection results, so that the system always maintains efficient verification capabilities.

[0101] Incremental learning techniques can be used to continuously optimize the performance of the Baichuan large model based on new data and user feedback. Through this continuous learning, the model can gradually adapt to new medical data types and semantic requirements, ensuring that the quality inspection system remains efficient and accurate. By learning and analyzing abnormal data, the model's adaptability and accuracy in processing complex and changing data can be further improved.

[0102] In some embodiments, the Baichuan big model can be combined with the knowledge base rule engine, which may specifically include: Step 1: Manual rule base construction and model adaptation Construction of knowledge base: Step 1.1: Medical experts manually create a knowledge base containing various medical rules and verification logic based on different clinical research and medical standards. These rules cover a variety of scenarios, from medical records, laboratory tests, medical images to patient symptom descriptions. For example, the rule for checking blood sugar levels is: "Blood sugar levels should be between 4.0-6.0 mmol / L."

[0103] The knowledge base constructed contains not only fixed logical rules, but also dynamically updated medical standards. Regular expert input and verification ensure the knowledge base is up-to-date and accurate.

[0104] Knowledge base and large model input adaptation: Step 1.2: Before the large model starts reasoning, constraints and prompt words are generated according to the rules of the knowledge base. For example, when the model extracts the text of the electronic medical record, it will trigger relevant rule prompts based on the diagnosis results, such as "Is the blood sugar level within the normal range?" These prompt words are provided by the knowledge base and combined with the actual data at the input stage of the model.

[0105] The prompt words generated by the knowledge base are directly embedded into the model input, so that each model reasoning follows specific rules and constraints. This method ensures that the artificially provided knowledge base can act on the reasoning process of the large model in real time, avoiding the disconnection between semantic understanding and logical verification.

[0106] Step 1.3: Generate multi-level prompt words from the knowledge base, covering different data types and verification tasks. For example, for laboratory test data, prompt words include "extract blood sugar level and compare reference value", "extract patient diagnosis and verify consistency of medication records", etc. Each prompt word will be linked to a specific rule condition, so that the large model can not only complete semantic understanding, but also perform multi-level logical verification through the rule engine.

[0107] Through multi-level prompt words, the model can adapt to different types of data verification tasks. The rules provided by the knowledge base make the model's reasoning process more accurate and adaptable to more complex scenarios.

[0108] Step 2: Combined reasoning of large models and manual rules Regularized reasoning mode of large models: Step 2.1: During the large model reasoning process, the rule base in the knowledge base is dynamically called to perform real-time verification based on the specific input data. For example, after the large model extracts the diagnosis of "hyperglycemia" when analyzing the text, it extracts the rules related to hyperglycemia from the knowledge base and embeds them into the model reasoning for further analysis and verification to detect whether there are corresponding treatment records.

[0109] Through the reasoning mode that combines the knowledge base with the large model, the model can directly call the predefined rule library to perform real-time verification on the extracted entities and information, so that the reasoning process and logical verification are organically combined.

[0110] Knowledge base rule priority processing: Step 2.2: During the inference process, the rules in the knowledge base are processed according to their priority. For example, the verification rules for diagnostic information take precedence over format checking, and the system automatically adjusts the execution order of the rules according to the different categories of input data. The rule settings of the knowledge base are controlled by priority and triggered in real time during the model inference process.

[0111] Through priority management, the rules in the knowledge base can be flexibly applied to different data scenarios, ensuring that high-priority rules can be executed at the early stage of reasoning, thereby improving the overall efficiency and accuracy of the system.

[0112] Dynamic feedback mechanism of rule engine: Step 2.3: During the reasoning process, the rule engine will dynamically check the intermediate output of the large model. If it is found that it does not match the rules in the knowledge base, the feedback mechanism will be triggered immediately. When the model extracts a medication record that does not match the diagnosis, the rule engine will issue an alarm and notify the model in real time to adjust the reasoning path or perform secondary verification.

[0113] This dynamic feedback mechanism ensures that the model's reasoning results are always consistent with the rules in the knowledge base, which not only reduces misjudgments, but also dynamically adjusts the model's reasoning logic and enhances adaptive capabilities.

[0114] Step 3: Real-time application and dynamic adjustment of manual rule base Real-time rule updates and adjustments: Step 3.1: Allow medical experts to adjust the rules in the knowledge base in real time through the interface. For example, if the judgment criteria for a certain type of diagnosis result changes, the expert can directly update the relevant rules and take effect immediately. The big model will apply these updated rules to verify the data during the next inference.

[0115] By manually adjusting the knowledge base rules, the system can flexibly respond to changes in medical standards while ensuring that the large model can always reason based on the latest rules.

[0116] Automatic rule generation based on knowledge base: Step 3.2: The system supports automatic generation of some new rules during the model reasoning process. For example, when it is detected that certain data frequently triggers rule anomalies, new rules can be automatically generated or existing rules can be adjusted based on historical records. Experts can review this and formally incorporate it into the knowledge base to improve verification capabilities.

[0117] By combining model feedback and historical data, rules can be automatically generated or adjusted, reducing the burden of manual maintenance of the knowledge base and making the quality inspection process more intelligent.

[0118] Step 4: Organic combination of large model reasoning and manual knowledge base Knowledge base driven reasoning process optimization: Step 4.1: The rules of the knowledge base will affect the reasoning path and strategy of the large model. For example, the rules stored in the knowledge base can help the large model choose different reasoning paths and dynamically adjust the prompt word generation and reasoning mode according to the constraints in the rules. The information provided by the knowledge base enables the model to automatically choose the optimal processing method when encountering specific input data.

[0119] The rules of the knowledge base are not only used for verification, but also participate in the optimization of the reasoning strategy of the large model to ensure that the reasoning results are more accurate and meet medical standards.

[0120] Multiple iterations of manual rule verification: Step 4.2: During the reasoning process, the large model is verified through multiple iterations. The model first extracts preliminary results based on the knowledge base rules, and then returns the results to the rule engine for re-verification. If problems are found, the model will perform a second round of reasoning. The iterative reasoning process ensures the close integration of the large model and the knowledge base, achieving more accurate quality inspection.

[0121] Through iterative verification, not only can the results be verified after one inference, but multiple corrections can also be made based on the feedback from the rule engine to ensure high-quality data output.

[0122] By deeply combining the knowledge of the manual rule base with the Baichuan large model reasoning process, the unification of human knowledge and automated semantic understanding is achieved. Through the knowledge base dynamically generating prompt words, priority processing, iterative verification and feedback mechanism, professional medical knowledge can be applied in different data scenarios, and large-scale data quality inspection tasks can be efficiently processed, which has obvious technical barriers and innovation advantages.

[0123] It should be noted that the execution subject of the multi-source clinical data quality inspection method based on a big model provided in the embodiment of the present application can be a multi-source clinical data quality inspection device based on a big model, or a control module in the multi-source clinical data quality inspection device based on a big model for executing the method for multi-source clinical data quality inspection based on a big model. In the embodiment of the present application, the multi-source clinical data quality inspection device based on a big model executing the method for multi-source clinical data quality inspection based on a big model is taken as an example to illustrate the device for multi-source clinical data quality inspection based on a big model provided in the embodiment of the present application.

[0124] like Figure 2 As shown, in a second aspect of the embodiment of the present application, a multi-source clinical data quality inspection device based on a large model is provided, and the device may include: An acquisition module 210 is used to acquire multi-source clinical data, wherein the multi-source clinical data includes: electronic medical records, laboratory test reports, medical imaging reports and surgical records; A text processing module 220, used for text processing the multi-source clinical data to obtain a formatted text; A determination module 230, configured to determine a prompt word according to the format text, wherein the prompt word corresponds to and matches a quality inspection rule; An extraction module 240 is used to use the prompt words to guide the trained deep learning model to extract keywords from the multi-source clinical data; The quality inspection module 250 is used to perform logical judgment on the keywords using the quality inspection rules to obtain quality inspection results.

[0125] The multi-source clinical data quality inspection device based on the large model in the embodiment of the present application can be a device, or a component, integrated circuit, or chip in a terminal. The device can be a mobile electronic device or a non-mobile electronic device. Exemplarily, the mobile electronic device can be a mobile phone, a tablet computer, a laptop computer, a PDA, an in-vehicle electronic device, a wearable device, an ultra-mobile personal computer (UMPC), a netbook or a personal digital assistant (PDA), etc., and the non-mobile electronic device can be a server, a network attached storage (NAS), a personal computer (PC), a television (TV), a teller machine or a self-service machine, etc., which is not specifically limited in the embodiment of the present application.

[0126] The large model-based multi-source clinical data quality inspection device in this embodiment can be an intelligent quality inspection system that collaborates with local processing and the cloud.

[0127] The main function of the local device is to process simple data quality inspection rules and determine whether to pass the task to the cloud based on the complexity of the task. This design can minimize the hardware requirements of the local device while ensuring the efficiency and accuracy of data quality inspection.

[0128] A modular rule engine is designed on the local device to handle simple logic rules and data verification tasks. For example, rules such as date format check, value range check, field consistency check, etc. can be quickly completed by local hardware.

[0129] Hardware optimization: Use low-power processors and lightweight memory allocation architecture to ensure that these simple tasks can be processed quickly locally. In order to reduce energy consumption, a lightweight hardware acceleration module can be designed specifically for simple judgment operations of the rule engine, such as numerical comparison, string matching, etc. These tasks do not need to rely on high-performance GPUs, and dedicated hardware accelerators will further optimize energy efficiency and response speed.

[0130] Introducing distributed reasoning and edge caching in local devices, performing preliminary data reasoning through LoRA-optimized models, and caching intermediate quality inspection results. If the local device can completely solve the problem, there is no need to transmit data to the cloud, reducing communication delays and bandwidth consumption.

[0131] Local devices can introduce an adaptive cache system to store some intermediate results of quality inspection in the cache. Combined with the reasoning capabilities of LoRA optimization, when processing simple rule matching or repetitive quality inspection tasks, the system can reuse these cached results to reduce the need for multiple repeated calculations.

[0132] When the local device detects that the complexity of data quality inspection exceeds its processing capabilities, it can intelligently transfer the task to the cloud for in-depth analysis and processing.

[0133] A preliminary analysis is performed on the local device, and a decision is made dynamically based on the complexity of the data whether to transfer the task to the cloud. The complexity of the task is determined by the preset threshold in the rule library. Simple text comparisons or numerical checks can be processed locally, while cross-document diagnostic comparisons and integrated analysis of historical data will be marked as "advanced quality inspection tasks" in the local device and automatically forwarded to the cloud server for processing.

[0134] This process is not a simple task transfer, but rather a process of pre-processing and hierarchical analysis by local devices. The device can compress or partially pre-process the complex data to be transferred locally, making the data received by the cloud more concise and reducing the transmission and processing overhead.

[0135] Through the intelligent data synchronization module, not only complex tasks are transferred to the cloud, but also real-time synchronization is maintained between the cloud and local devices. The system can intelligently detect factors such as data transmission delay and bandwidth usage, and adjust the synchronization frequency and data transmission strategy according to the situation to avoid occupying too many resources.

[0136] Integrate efficient data compression algorithms in local devices, and compress quality inspection data before transmitting it to the cloud through real-time data compression, reducing the bandwidth and time required for transmission. The cloud can quickly decompress and process the compressed data after receiving it, ensuring the real-time nature of the quality inspection process.

[0137] Introducing an adaptive computing mode in the hardware design of local devices, the hardware operating status of the device is automatically adjusted according to the complexity of the quality inspection task. For example, when the device detects a high-load task, it automatically increases the processor's computing frequency; when processing simple tasks, the device can enter a low-power mode to keep running for a long time.

[0138] By integrating the power optimization module, the local device can dynamically adjust the processor frequency and power supply. During the continuous quality inspection process, the system can detect changes in task complexity and adjust the hardware power consumption to achieve efficient use of resources.

[0139] A multi-level cache management system is introduced to manage the quality inspection tasks processed by the device through different cache levels. The first-level cache processes temporary data (such as the intermediate results of simple rules), the second-level cache is used to store the results of ongoing quality inspection tasks, and the third-level cache stores historical data of quality inspection. This hierarchical cache architecture can avoid repeated processing and improve the execution efficiency of quality inspection tasks.

[0140] The combination of high-speed cache chip and low-power flash memory ensures fast local storage and call of data. The system can flexibly select different cache mechanisms according to the needs of different tasks.

[0141] The LoRA optimized reasoning model is directly embedded in the local hardware to handle simpler reasoning tasks. The local device uses this optimized model to process simple logical operations in structured data and rule bases, such as date verification and text consistency verification, without the need for complex hardware accelerators.

[0142] A lightweight inference engine unit is integrated into the hardware architecture, which is specifically used to process simple data quality inspection rules and reduce the device's dependence on high-performance CPUs or the cloud.

[0143] The device of this embodiment combines software and hardware to process simple quality inspection tasks on the local device and dynamically transfer complex tasks to the cloud, which optimizes the use of hardware resources and improves the overall quality inspection efficiency. Through adaptive hardware optimization, multi-level caching, dynamic task switching and real-time data compression, local devices can complete efficient quality inspection tasks with low resource usage. At the same time, the lightweight reasoning capability based on LoRA technology enables local devices to achieve intelligent and efficient working methods in data quality inspection, which is suitable for complex clinical data scenarios.

[0144] The multi-source clinical data quality inspection device based on a large model in the embodiment of the present application can be a device with an operating system. The operating system can be an Android operating system, an iOS operating system, or other possible operating systems, which are not specifically limited in the embodiment of the present application.

[0145] The multi-source clinical data quality inspection device based on a large model provided in the embodiment of the present application can achieve Figure 1 To avoid repetition, the various processes implemented by the method embodiment are not described here.

[0146] Alternatively, if Figure 3As shown, an embodiment of the present application also provides an electronic device 300, including a first processor 301, a first memory 302, and a program or instruction stored in the first memory 302 and executable on the first processor 301. When the program or instruction is executed by the first processor 301, each process of the above-mentioned large model-based multi-source clinical data quality inspection method embodiment is implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0147] It should be noted that the electronic devices in the embodiments of the present application include the mobile electronic devices and non-mobile electronic devices mentioned above.

[0148] Figure 4 A schematic diagram of the hardware structure of an electronic device to implement an embodiment of the present application.

[0149] The hardware structure 400 of the electronic device includes but is not limited to: a radio frequency unit 401, a network module 402, an audio output unit 403, an input unit 404, a sensor 405, a display unit 406, a user input unit 407, an interface unit 408, a second memory 409, and a second processor 410 and other components.

[0150] Those skilled in the art will appreciate that the hardware structure 400 of the electronic device may also include a power source (such as a battery) for supplying power to each component, and the power source may be logically connected to the second processor 410 through a power management system, thereby implementing functions such as managing charging, discharging, and power consumption management through the power management system. Figure 4 The electronic device structure shown in the figure does not constitute a limitation on the electronic device. The electronic device may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently, which will not be described in detail here.

[0151] It should be understood that in the embodiment of the present application, the input unit 404 may include a graphics processor (GPU) 4041 and a microphone 4042, and the graphics processor 4041 processes the image data of the static picture or video obtained by the image capture device (such as a camera) in the video capture mode or the image capture mode. The display unit 406 may include a display panel 4061, and the display panel 4061 may be configured in the form of a liquid crystal display, an organic light emitting diode, etc. The user input unit 407 includes a touch panel 4071 and other input devices 4072. The touch panel 4071 is also called a touch screen. The touch panel 4071 may include two parts: a touch detection device and a touch controller. Other input devices 4072 may include but are not limited to a physical keyboard, a function key (such as a volume control button, a switch button, etc.), a trackball, a mouse, and a joystick, which will not be repeated here. The second memory 409 can be used to store software programs and various data, including but not limited to applications and operating systems. The second processor 410 can integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, the user interface, and the application program, etc., and the modem processor mainly processes wireless communication. It is understandable that the above-mentioned modem processor may not be integrated into the second processor 410 .

[0152] An embodiment of the present application also provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, each process of the above-mentioned embodiment of the multi-source clinical data quality inspection method based on a large model is implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0153] The processor is the processor in the electronic device described in the above embodiment. The readable storage medium includes a computer readable storage medium, such as a computer read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0154] An embodiment of the present application further provides a chip, which includes a processor and a communication interface, wherein the communication interface is coupled to the processor, and the processor is used to run programs or instructions to implement the various processes of the above-mentioned large model-based multi-source clinical data quality inspection method embodiment, and can achieve the same technical effect. To avoid repetition, it will not be repeated here.

[0155] It should be understood that the chip mentioned in the embodiments of the present application can also be called a system-level chip, a system chip, a chip system or a system-on-chip chip, etc.

[0156] It should be noted that, in this article, the terms "comprise", "include" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise one..." do not exclude the presence of other identical elements in the process, method, article or device including the element. In addition, it should be pointed out that the scope of the methods and devices in the embodiments of the present application is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in reverse order according to the functions involved, for example, the described method may be performed in an order different from that described, and various steps may also be added, omitted, or combined. In addition, the features described with reference to certain examples may be combined in other examples.

[0157] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a computer software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a terminal (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present application.

[0158] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present application, ordinary technicians in this field can also make many forms without departing from the purpose of the present application and the scope of protection of the claims, all of which are within the protection of the present application.

Claims

1. A multi-source clinical data quality inspection method based on a large model, characterized in that: include: Acquiring multi-source clinical data, including electronic medical records, laboratory test reports, medical imaging reports, and surgical records; Performing text processing on the multi-source clinical data to obtain a formatted text; Determine a prompt word according to the format text, wherein the prompt word corresponds to and matches a quality inspection rule; Using the prompt words to guide the trained deep learning model to extract keywords from the formatted text; The quality inspection rules are used to perform logical judgment on the keywords to obtain quality inspection results.

2. The multi-source clinical data quality inspection method based on a large model according to claim 1 is characterized in that: The obtaining of multi-source clinical data includes: Import multi-source clinical data from hospital information systems and laboratory information systems to ensure the diversity and integrity of data sources.

3. The multi-source clinical data quality inspection method based on a large model according to claim 1 is characterized in that: The text processing of the multi-source clinical data to obtain a formatted text includes: Performing optical character recognition processing and standardization processing on the image data to obtain a first text; Perform text parsing on the PDF file and the DOCX file to obtain a second text; Performing standardization on the TXT file to obtain a third text; The first text, the second text and the third text are processed in a unified format to obtain formatted texts.

4. The multi-source clinical data quality inspection method based on a large model according to claim 3 is characterized in that: The step of performing format unification processing on the first text, the second text, and the third text to obtain formatted texts includes: Segmenting the first text, the second text, and the third text to obtain a first text segment set, a second text segment set, and a third text segment set; uniformly reconstructing the first text segment set, the second text segment set and the third text segment set to obtain a structurally unified text set; The structurally unified text set is fused to obtain a formatted text.

5. The multi-source clinical data quality inspection method based on a large model according to claim 1 is characterized in that: The step of using the prompt words to guide the trained deep learning model to extract keywords from the formatted text includes: The prompt word and the format text form a structured text, wherein the prompt word and the format text are in different fields in the structured text; The structured text is input into a trained deep learning model to obtain keywords.

6. The multi-source clinical data quality inspection method based on a large model according to claim 1 is characterized in that: The trained deep learning model is obtained through LoRA fine-tuning training.

7. The multi-source clinical data quality inspection method based on a large model according to claim 1 is characterized in that: After the quality inspection rules are used to perform logical judgment on the keywords, the multi-source clinical data quality inspection method based on a large model further includes: Using the big model to perform semantic analysis on the keywords to obtain semantic understanding results; The final verification result is obtained by combining the logical judgment results and semantic understanding results.

8. A multi-source clinical data quality inspection device based on a large model, characterized in that: include: An acquisition module is used to acquire multi-source clinical data, wherein the multi-source clinical data includes: electronic medical records, laboratory test reports, medical imaging reports and surgical records; A text processing module, used for text processing the multi-source clinical data to obtain a formatted text; A determination module, used to determine a prompt word according to the format text, wherein the prompt word corresponds to and matches a quality inspection rule; An extraction module, used to use the prompt words to guide the trained deep learning model to extract keywords from the multi-source clinical data; The quality inspection module is used to perform logical judgment on the keywords using the quality inspection rules to obtain quality inspection results.

9. An electronic device, characterized in that: include: A processor, a memory, and a program or instruction stored in the memory and executable on the processor, wherein the program or instruction, when executed by the processor, implements the steps of the multi-source clinical data quality inspection method based on a large model as described in any one of claims 1 to 7.

10. A readable storage medium, characterized in that: The readable storage medium stores a program or instruction, and when the program or instruction is executed by the processor, the steps of the multi-source clinical data quality inspection method based on a large model as described in any one of claims 1 to 7 are implemented.

Images