An ai large model-based medical data cascade screening and form configuration method

Abstract

The present application relates to the technical field of medical data processing, and in particular to a medical data cascade screening and form configuration method based on an AI large model. The method comprises intent recognition, decomposing the natural language instruction input by the user into a rule recognition task and an AI recognition task through an Agent planner; form configuration, generating form configuration parameters matched therewith; multi-dimensional data processing, obtaining medical data from a medical database and performing data preprocessing; data screening, constructing a cascade data screening process based on the form configuration parameters, and performing rule screening and AI recognition screening on the preprocessed data. The present application uses the Agent planner to generate and match the form configuration parameters from the natural language instruction input by the user, extracts data from the medical database according to the parameters, and performs rule screening and AI recognition screening on the data, which helps to greatly improve the form configuration efficiency and ease of use.

Description

Technical Field

[0001] This invention relates to the field of medical data processing technology, and in particular to a method for cascading screening and form configuration of medical data based on a large AI model. Background Technology

[0002] Currently, data filtering and form configuration in the medical field mainly rely on two types of technical solutions:

[0003] 1. Traditional rule-driven solution: This method involves manually pre-setting fixed SQL statements, field matching rules, and logical operators to complete the filtering based on structured medical data (such as gender, age, and diagnosis codes in electronic medical records). Forms require medical staff or technicians to manually configure field relationships and logical operation rules according to specific needs. This solution requires operators to possess both medical expertise and technical configuration skills, and the configuration process is cumbersome and prone to errors.

[0004] 2. Pure NLP-driven solution: This approach uses natural language processing technology to perform semantic analysis on medical data. However, the model often functions as an isolated text understanding tool, lacking systematic integration with specific form configurations, data structure mapping, and automated task execution. For complex screening logic that requires a mix of quantitative rules and qualitative descriptions, it lacks effective task decomposition and routing mechanisms. Furthermore, it generally neglects automated verification of the compliance and logical consistency of configuration results, and lacks a system self-optimization loop based on actual usage feedback.

[0005] Therefore, existing technologies generally face key shortcomings when addressing the needs for intelligent form configuration and data filtering in the medical field, such as low automation, weak cross-modal data processing capabilities, insufficient system flexibility and adaptability, lack of continuous evolution capabilities, and data security and privacy compliance risks. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention discloses a method for cascading screening and form configuration of medical data based on a large AI model, characterized by the following steps:

[0007] Intent recognition: The Agent planner breaks down the natural language commands input by the user into rule recognition tasks and AI recognition tasks that are related to each other but are configured independently.

[0008] Form configuration: Generate form configuration parameters that match the task type;

[0009] Multi-dimensional data processing: Based on the form configuration parameters, medical data is retrieved from the medical database and data preprocessing is performed;

[0010] Data filtering: A cascading data filtering process is built based on form configuration parameters, and rule-based filtering and AI recognition filtering are performed on the preprocessed data to output the filtering results.

[0011] The data preprocessing includes multimodal data conversion and privacy desensitization processing;

[0012] The Agent planner specifically refers to the large-scale AI Agent planner, whose main function is to transform the complex and ambiguous goals given by the user into a series of structured, executable, and logically ordered sub-tasks, thereby guiding the entire AI system to achieve the goals efficiently.

[0013] Multimodal data conversion refers to the process of processing and fusing data (such as text, images, audio, and video) from different sources and in different formats, and converting them into a unified, structured form or one that can be deeply understood by machines.

[0014] This solution utilizes an agent planner to break down user-input natural language commands into rule-based recognition tasks and AI-based recognition tasks, generating matching form configuration parameters based on the task type. Data is then extracted from a medical database using these parameters, and rule-based and AI-based filtering are applied to the data. This significantly improves the efficiency and usability of form configuration, automating the conversion of medical natural language requirements into standardized forms through agent task planning. Furthermore, it performs multimodal data conversion on user-input natural language commands using different rule-based and AI-based recognition logics.

[0015] Specifically, the agent planner performs structural priors on the natural language input by the user, including:

[0016] Structural anchor point recognition: The planner will prioritize recognizing explicit structural markers (such as headings, chapter numbers, line breaks, and Markdown tags) in the input natural language.

[0017] Coordinate-based reading: The class locates the relevant chapter coordinates by scanning, and then calls a deep reading tool to read the continuous paragraphs within that range.

[0018] This solution ensures that even if the natural language input in the medical field is very long, as long as it belongs to the same chapter or paragraph structure, the Agent planner will treat it as a whole and avoid losing context due to blind truncation.

[0019] Optionally, the Agent planner adopts an architecture of orthogonal decoupling, dynamic routing, and closed-loop verification.

[0020] Step 1: The Agent planner transforms natural language into an intermediate medical logic primitive tree, and uses a dedicated BERT model or LLM with a small number of parameters to identify entities and their logical relationships in the text.

[0021] Step 2: Based on the parsed medical logic primitive tree, the Agent planner executes an orthogonal decoupling algorithm to split the task flow into two parallel tracks:

[0022] Track A extracts all nodes containing explicit numerical values, Boolean logic, and enumeration values.

[0023] Generate Structured Query Language or JSON Schema.

[0024] Track B extracts all nodes containing adjectives, non-standardized terms, and complex reasoning. It then generates semantic embedding vectors or natural language prompts.

[0025] Step 3: The Agent planner generates a dynamic hybrid execution graph. This graph includes rule executor nodes and AI inference nodes, and the Agent planner automatically analyzes dependencies.

[0026] Step 4: Execute logical-semantic consistency verification. After parallel execution of rule filtering and AI inference, a lightweight "consistency verification agent" is introduced to check whether the AI's output violates the constraints of the rule track.

[0027] Optionally, in the form configuration, for rule recognition tasks, a rule form containing logical operators, data tables, fields, values, and operators is generated; for AI recognition tasks, an AI recognition form containing prompt words and contextual scope is generated.

[0028] Optionally, in the form configuration, the AI-recognized form includes prompts automatically generated by a large language model based on the natural language instructions and corresponding data context ranges, wherein the data context ranges include time ranges, data services, fields, and data pruning parameters.

[0029] Optionally, in the form compliance validation, the form configuration parameters are parsed and stored through an abstract syntax tree.

[0030] The form configuration parameters use an abstract syntax tree structure to represent and store multi-level logical relationships to support the combined execution of AND, OR, and NOT logic.

[0031] Specifically, the recursive construction method of the abstract syntax tree is to recursively parse the structured configuration, generate the corresponding AST nodes, and then assemble them into a complete tree structure.

[0032] This solution ensures the accurate execution of complex logical relationships by applying an abstract syntax tree, avoiding logical errors caused by manual configuration.

[0033] Optionally, in the data filtering process, before performing AI recognition filtering, the large language model identifies paragraph anchors in the medical data according to the data clipping parameters in the form configuration, dynamically slices the unstructured text, and removes redundant information that is irrelevant to the current clipping range.

[0034] Optionally, after form configuration, form compliance verification is performed. Consistency, compliance, and logical consistency checks are applied to the generated form configuration parameters, and the verified target form configuration parameters are output. The form compliance verification includes:

[0035] Input domain compliance verification: The large language model determines whether the natural language commands input by the user belong to the medical field;

[0036] Field association validation: Determine whether the selected table and fields are associated using a preset table and field mapping dictionary;

[0037] Logical self-consistency verification: Based on the medical knowledge graph, the natural language commands input by the user are mapped to entity nodes of the medical knowledge graph, and the logic is verified by calculating the path reachability or relation constraints in the graph.

[0038] This solution comprehensively covers domain compliance, field association, and logical consistency through an intelligent form compliance verification mechanism, automatically correcting anomalies such as logical conflicts and missing fields, and effectively reducing the invalid configuration rate.

[0039] Optionally, the cascaded data filtering process is constructed as a dynamic directed acyclic graph, where each node corresponds to an atomic data processing operation, and the directed edges between nodes represent data or state dependencies. The filtering process is scheduled and executed by performing topological sorting on the dynamic directed acyclic graph.

[0040] The atomic data processing operations include at least one of the following:

[0041] OCR (Optical Character Recognition) is used to convert images into text data.

[0042] Speech-to-text processing is used to convert speech data into text data;

[0043] Document parsing tools are used to convert unstructured text into structured content;

[0044] Semantic reasoning based on a large language model is used to determine whether medical texts meet the screening criteria.

[0045] Optionally, in the AI ​​recognition and screening process, the screening process includes:

[0046] Based on predefined data services and field mapping relationships, the validity of form configuration parameters is validated.

[0047] After the legality verification is passed, the corresponding unstructured medical data is obtained;

[0048] Perform structuring transformation and data trimming on the acquired unstructured medical data;

[0049] Based on the processed unstructured medical data, a large language model determines whether the medical requirements are met and outputs the results in JSON format.

[0050] Optionally, after data filtering, there is also result output and closed-loop feedback: integrate the filtering process data into the results, generate a form including filtering rules, data sources, and configuration items, output to the user review interface, and receive user feedback for optimization and iteration of the aforementioned process steps;

[0051] This solution also provides a configuration correction mechanism, the process of which includes:

[0052] Configuration correction step one: Based on the user's confirmation, modification or rejection of the automatically generated form configuration, the user interaction behavior is mapped into discrete policy feedback signals;

[0053] Configuration correction step two: Construct a multi-dimensional reward function that includes rule method accuracy, rule configuration accuracy, data filtering accuracy, and the policy feedback signal. Different reward dimensions are applied to different policy subspaces in the Agent planner for rule recognition method, rule parameter generation, and AI recognition parameter generation.

[0054] Configuration correction step three: Through multiple rounds of iterative training, the Agent planning phase will adaptively adjust the rule threshold selection order, field mapping priority, and AI recognition prompt word generation strategy in subsequent form configuration tasks based on historical reward distribution.

[0055] Optionally, the multidimensional reward function (data accuracy) is confirmed in the following ways:

[0056] Confirmation Step 1: Select at least two general-purpose large language models, input the question, reference answer, and system answer, and evaluate the consistency of the method selection results based on the reference answer;

[0057] Step 2 Confirmation: If there are discrepancies in the evaluation results of different general-purpose models, a conflict resolution mechanism is introduced to generate stable scoring results. These scoring results are then used as input to the reward function in reinforcement learning. The scoring method is as follows: Each general-purpose model generates an answer rating (0, 2, or 4) based on the reference answer judgment system; the rating is converted to a range of [0, 1] as the judge's score, and the average of the judge's scores is taken.

[0058] This solution combines a closed-loop mechanism involving human participation with reinforcement learning algorithms. By learning from user modifications and feedback, it dynamically iterates form configuration rules. The more the system is used, the higher the configuration and filtering accuracy becomes, forming a virtuous cycle of "use-feedback-optimization". A quantitative performance evaluation system provides a clear basis for system optimization, ensuring that performance is measurable and can be improved.

[0059] The beneficial effects of this invention are:

[0060] 1. Significantly improves form configuration efficiency and ease of use. Through multimodal semantic parsing and Agent task planning, it realizes the automated transformation of medical natural language requirements into standardized forms. It supports intelligent mapping of multi-level nested logic, eliminating the need for manual writing of rule statements. Medical staff can directly complete the configuration through natural language, greatly improving configuration efficiency. At the same time, the system has pre-built medical tables and field libraries, automatically matching field relationships and reducing the operational threshold.

[0061] 2. Overcoming data modality limitations to improve screening accuracy and coverage. Through a "cascaded data screening pipeline" design, the system organically coordinates SQL-based rule-based screening (suitable for structured data) with AI-based recognition screening based on large models (suitable for unstructured text). This not only handles explicit quantitative standards but also deeply understands complex descriptions in medical record texts, thereby achieving more comprehensive and accurate screening of target cases and reducing missed and false screenings.

[0062] 3. Ensure the accuracy and compliance of configuration logic. The intelligent form compliance verification mechanism comprehensively covers dimensions such as domain compliance, field association, and logical consistency, automatically correcting anomalies such as logical conflicts and missing fields, effectively reducing the invalid configuration rate; the application of Abstract Syntax Tree (AST) ensures the accurate execution of complex logical relationships and avoids logical errors caused by manual configuration.

[0063] 4. Strengthen data security and privacy compliance. Based on de-identification algorithms, the "multi-dimensional data processing" stage can automatically and reliably encrypt and hide sensitive personal information at the source and throughout the data utilization process, ensuring that the entire data processing process complies with relevant domestic and international laws and regulations, and establishing a trustworthy data security barrier.

[0064] 5. Continuous optimization of system performance is achieved through a closed-loop mechanism involving human participation combined with reinforcement learning algorithms. The system learns from user modifications and feedback, dynamically iterates form configuration rules, and the more the system is used, the higher the configuration accuracy and filtering accuracy become, forming a virtuous cycle of "use-feedback-optimization"; a quantitative performance evaluation system provides a clear basis for system optimization, ensuring that performance is measurable and can be improved.

[0065] 6. Multi-level joint screening achieves a balance between efficiency and accuracy. Employing a cascaded screening mechanism that first identifies rules and then AI identifies data, this solution uses efficient SQL rules in the first stage to filter out most of the data that does not meet the criteria, significantly reducing the frequency of large model calls in the second stage, effectively optimizing computing costs, and balancing data screening accuracy and overall processing efficiency. Attached Figure Description

[0066] Figure 1 This is a flowchart illustrating the implementation of this application.

[0067] Figure 2 This is a schematic diagram of the unstructured data post-processing flow in the embodiments of this application. Detailed Implementation

[0068] To enable those skilled in the art to better understand the present invention, the technical solutions in the specific embodiments of the present invention will be clearly and completely described below.

[0069] Example 1

[0070] Automated configuration and automatic filtering process

[0071] This example demonstrates how to transform complex research needs into executable, automated configuration and screening processes.

[0072] Enter "Age greater than 50 years but less than 80 years, with imaging evidence of advanced recurrent or metastatic gastric and gastroesophageal junction (GEJ) adenocarcinoma within 28 days (inclusive) prior to randomization";

[0073] The Agent planner breaks down the input content: "Age greater than 50 years old but less than 80 years old" uses a structured rule recognition task, while "Advanced recurrent or metastatic gastric and gastroesophageal junction (GEJ) adenocarcinoma confirmed by imaging within 28 days (including 28 days) prior to randomization" uses an unstructured AI recognition task.

[0074] Automatically generate configuration form 1: "Age greater than 50 years old and less than 80 years old" automated form configuration:

[0075]

[0076] Automatically generated form 2: "Imaginally confirmed advanced recurrent or metastatic gastric and gastroesophageal junction (GEJ) adenocarcinoma within 28 days prior to randomization (including 28 days)" Automated configuration form

[0077]

[0078] After automatically generating the form configuration, the system returns it and performs intelligent form compliance checks, including input compliance, logical consistency, and field relevance. It automatically corrects or marks any validation exceptions and returns the results to the user.

[0079] Users conduct final review and confirmation of the configuration results and provide feedback on whether the current configuration is correct. The configuration form is saved in the form of a syntax tree.

[0080] Parse the syntax tree corresponding to configuration form 1, convert it into an SQL statement for execution, and filter out data 1 that meets the requirements;

[0081] Based on the filtered data, run the data logic corresponding to configuration form 2 and call the interface to obtain unstructured data;

[0082] Data post-processing: Determine the type of the acquired unstructured data. If it is an image or file, call a document parsing tool to process it into Markdown text. If it is speech, call a speech-to-text model to convert it into text. If it is text, proceed directly to the next step of processing. De-identify the acquired data and remove personal privacy information such as name, ID number, contact information, and address.

[0083] The form is configured with the data cropping range set to "discharge diagnosis", which extracts the discharge diagnosis content from the admission records in the inpatient medical records.

[0084] The large model is invoked to determine whether the patient meets the criteria of "advanced recurrent or metastatic gastric and gastroesophageal junction (GEJ) adenocarcinoma confirmed by imaging within 28 days (inclusive) prior to randomization". The judgment result and the basis for judgment are returned. The data format is fixed as JSON, generally as: {"flag":"This is the AI ​​judgment result, which meets / does not meet the criteria","description":"This is the reason for the judgment"}.

[0085] Example 2

[0086] Data post-processing workflow

[0087] This example demonstrates the data processing flow when the form configuration contains non-text data.

[0088] For example, "within the 6 months prior to randomization, the Fridricia-corrected QT interval (QTcF) was >450 msec for men and >470 msec for women; and the left ventricular ejection fraction was <50%":

[0089] (1) After entering "Applicable to chemotherapy or chemotherapy combined with PD-1 inhibitor therapy within 14 days prior to randomization", according to the automatic configuration scheme in Example 1, it will be configured to use AI recognition mode, and the form configuration content is as follows:

[0090]

[0091] (2) When calling the interface to obtain unstructured data according to the form configuration, the URL should be returned synchronously for image or file content;

[0092] (3) Call the document parsing tool to process the ECG PDF report into markdown format, where text content is stored in the original report reading order, and images exist in the form of links, and return the structured markdown text data;

[0093] (4) Post-process the text data, remove unnecessary page numbers, headers and footers, and encrypt sensitive information such as patient name, ID number, contact information, and home address; return the post-processed text data.

[0094] Example 3

[0095] Rule form parsing based on syntax tree (AST)

[0096] Taking "Complete blood count: absolute neutrophil count (NE#) ≥ 1.5 × 10⁹ / L" as an example, the process and operation of converting the form configuration into a syntax tree are shown:

[0097] The form configuration method is as follows:

[0098]

[0099] Define the node type and structure of the test class AST:

[0100]

[0101] Abstract Syntax Tree (AST) Recursive Construction: Recursively parse the structured configuration, generate the corresponding AST nodes, and then assemble them into a complete tree structure;

[0102] S1: Initialize the root node, creating the AST root node as the entry point for the entire syntax tree, in JSON format:

[0103] {

[0104] "type": "ROOT",

[0105] "source": "Inspection Report Inquiry",

[0106] "children": []

[0107] }

[0108] S2: Parse the first-level logical node (AND), configure the outer logical operator to AND, generate a LOGIC type node, and attach it to the root node. The JSON format is as follows:

[0109] {

[0110] "type": "ROOT",

[0111] "source": "Inspection Report Inquiry",

[0112] "children": [

[0113] {

[0114] "type": "LOGIC",

[0115] "operator": "AND",

[0116] "medical_domain": "blood routine",

[0117] "children": []

[0118] } ]

[0120] }

[0121] S3: Parse the basic condition node: Convert "absolute neutrophil count (NE#) ≥ 1.5 × 10^9 / L" into a CONDITION node: Then generate the complete AST structure, in JSON format:

[0122] {

[0123] "type": "ROOT",

[0124] "source": "Inspection Query",

[0125] "children": [

[0126] {

[0127] "type": "LOGIC",

[0128] "operator": "AND",

[0129] "medical_domain": "blood routine",

[0130] "children": [

[0131] {

[0132] "type": "CONDITION",

[0133] "test_item": "NE#",

[0134] "test_name": "neutrophils",

[0135] "operator": "≥",

[0136] "value": 1.5,

[0137] "unit": "×10 9 / L"

[0138] } ]

[0140] } ]

[0142] }

[0143] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for cascading filtering and form configuration of medical data based on an AI large-scale model, characterized in that, Includes the following steps: Intent recognition: The Agent planner breaks down the natural language commands input by the user into rule recognition tasks and AI recognition tasks that are related to each other but are configured independently. Form configuration: Generate form configuration parameters that match the task type; Multi-dimensional data processing: Based on the form configuration parameters, medical data is retrieved from the medical database and data preprocessing is performed; Data filtering: A cascading data filtering process is built based on form configuration parameters, and rule-based filtering and AI recognition filtering are performed on the preprocessed data to output the filtering results; The cascaded data filtering process is constructed as a dynamic directed acyclic graph, where each node corresponds to an atomic data processing operation, and the directed edges between nodes represent data or state dependencies. The filtering process is scheduled and executed by performing topological sorting on the dynamic directed acyclic graph. The atomic data processing operations include at least one of the following: OCR (Optical Character Recognition) is used to convert images into text data. Speech-to-text processing is used to convert speech data into text data; Document parsing tools are used to convert unstructured text into structured content; Semantic reasoning based on a large language model is used to determine whether medical texts meet the screening criteria. In the aforementioned form configuration, for rule recognition tasks, a rule form containing logical operators, data tables, fields, and values ​​is generated; for AI recognition tasks, an AI recognition form containing prompt words and contextual scope is generated.

2. The method for cascading filtering and form configuration of medical data based on an AI large model according to claim 1, characterized in that, In the form configuration, the AI-recognized form includes prompts automatically generated by a large language model based on the natural language instructions and corresponding data context ranges. The data context ranges include time ranges, data services, fields, and data pruning parameters.

3. The method for cascading filtering and form configuration of medical data based on an AI large model according to claim 1, characterized in that, In the data filtering process, before AI recognition and filtering, the large language model identifies paragraph anchors in the medical data based on the data clipping parameters in the form configuration, dynamically slices the unstructured text, and removes redundant information that is irrelevant to the current clipping range.

4. The method for cascading filtering and form configuration of medical data based on an AI large model according to claim 2, characterized in that, After the form is configured, form compliance is validated. Consistency, compliance and logical self-consistency checks are performed on the generated form configuration parameters, and the validated target form configuration parameters are output. The form compliance verification includes: Input domain compliance verification: The large language model determines whether the natural language commands input by the user belong to the medical field; Field association validation: Determine whether the selected table and fields are associated using a preset table and field mapping dictionary; Logical self-consistency verification: Based on the medical knowledge graph, the natural language commands input by the user are mapped to entity nodes of the medical knowledge graph, and the logic is verified by calculating the path reachability or relation constraints in the graph.

5. The method for cascading filtering and form configuration of medical data based on an AI large model according to claim 4, characterized in that, In the aforementioned form compliance validation, form configuration parameters are parsed and stored through an abstract syntax tree.

6. The method for cascading filtering and form configuration of medical data based on an AI large model according to claim 1, characterized in that, In the aforementioned AI-based identification and screening process, the screening procedure includes: Based on predefined data services and field mapping relationships, the form configuration parameters are validated for legality; after the legality validation is passed, the corresponding unstructured medical data is obtained. Perform structuring transformation and data trimming on the acquired unstructured medical data; Based on the processed unstructured medical data, a large language model determines whether the medical requirements are met and outputs the results in JSON format.

7. A method for cascading filtering and form configuration of medical data based on an AI large model according to any one of claims 1-6, characterized in that, After data filtering, there is also result output and closed-loop feedback: integrate the data and results of the filtering process, generate a form including filtering rules, data sources and configuration items, output to the user review interface, and receive user feedback for optimization and iteration of the aforementioned process steps; A configuration correction mechanism is also provided, the process of which includes: Configuration correction step one: Based on the user's confirmation, modification or rejection of the automatically generated form configuration, the user interaction behavior is mapped into discrete policy feedback signals; Configuration correction step two: Construct a multi-dimensional reward function that includes rule method accuracy, rule configuration accuracy, data filtering accuracy, and the policy feedback signal. Different reward dimensions are applied to different policy subspaces in the Agent planner for rule recognition method, rule parameter generation, and AI recognition parameter generation. Configuration correction step three: Through multiple rounds of iterative training, the Agent planner will adaptively adjust the rule threshold selection order, field mapping priority, and AI recognition prompt word generation strategy in subsequent form configuration tasks based on historical reward distribution.

8. The method for cascading filtering and form configuration of medical data based on an AI large model according to claim 7, characterized in that, The aforementioned multidimensional reward function is confirmed in the following way: Confirmation Step 1: Select at least two general-purpose large language models, input the question, reference answer, and system answer, and evaluate the consistency of the method selection results based on the reference answer; Confirmation Step 2: If there are differences in the evaluation results of different general large models, a conflict resolution mechanism is introduced to generate stable scoring results, and the scoring results are used as input to the reward function in reinforcement learning. The scoring method is as follows: each general large model judges the system to generate answer rating results based on the reference answer. The ratings are converted to a range of [0,1] as the judges' scores, and the average of the judges' scores is taken.

Images