Structured interpretation method and device for 24-hour ambulatory blood pressure monitoring report and storage medium

Abstract

The invention provides a 24h ambulatory blood pressure monitoring report structured interpretation method and device and a storage medium. The structured interpretation method for the 24-hour ambulatory blood pressure monitoring report comprises the steps that firstly, a GraphRAG knowledge base is constructed, and self-supervised learning is conducted on an ABPM report interpretation model based on the GraphRAG knowledge base; 2, fine tuning is conducted on the ABPM report interpretation model before use, so that the ABPM report interpretation model has the functions of interpreting a conclusion according to the report and outputting blood pressure management suggestions according to the conclusion; 3, report analysis is conducted through the ABPM report interpretation model subjected to fine tuning before use; and 4, constructing an interpretation report by using a report analysis result, and visually displaying the interpretation report.

Description

Technical Field

[0001] The present invention relates to the field of medical devices, and specifically to a structured interpretation method, terminal device and storage medium for a 24-hour dynamic blood pressure monitoring report based on GraphRAG, which can realize automatic interpretation and visual display of ABPM reports, thereby improving the accuracy and reliability of report interpretation, and the monitored person can perform subsequent blood pressure management according to the report recommendations. Background Art

[0002] 24h ambulatory blood pressure monitoring (ABPM) refers to the automatic and continuous measurement of blood pressure at regular intervals within 24 hours. ABPM is the main means of diagnosing hypertension, evaluating blood pressure levels, and observing the efficacy of antihypertensive treatment.

[0003] ABPM can provide a large amount of blood pressure data outside the medical environment, comprehensively and accurately reflecting the "panorama" of blood pressure: ABPM has 100% specificity and sensitivity for diagnosing hypertension; ABPM can identify "white coat hypertension" and "hidden hypertension"; ABPM reflects the rhythmic fluctuations of blood pressure, understands the overall picture of blood pressure in a day, judges the circadian rhythm, and diagnoses nocturnal hypertension and early morning hypertension; ABPM can effectively predict target organ damage in hypertensive patients; ABPM can effectively evaluate blood pressure fluctuations, predict vascular endothelial damage, and predict atherosclerosis; ABPM provides a T / P ratio to evaluate the long-term effect of antihypertensive drugs, and the T / P ratio is the standard for evaluating the long-term effect of antihypertensive drugs; ABPM can evaluate the parameter SI of blood pressure fluctuations. The higher the SI, the better the antihypertensive effect and target organ protection. Compared with office blood pressure and home self-measurement of blood pressure, ABPM is currently the most accurate way to measure blood pressure.

[0004] ABPM reports can provide a wealth of information that can help doctors and patients determine whether it is hypertension, what type of hypertension it is, and evaluate the effectiveness of non-drug intervention / drug treatment.

[0005] However, there is currently no system that can display 24-hour blood pressure monitoring reports through GraphRAG multimodal visualization. Summary of the invention

[0006] The technical problem to be solved by the present invention is to address the above-mentioned defects in the prior art and provide a structured interpretation method, terminal device and storage medium for 24h dynamic blood pressure monitoring report based on GraphRAG, which can realize automatic interpretation and visual display of ABPM report, thereby improving the accuracy and reliability of report interpretation. At the same time, the monitored person can perform subsequent blood pressure management according to the report recommendations.

[0007] According to the present invention, a structured interpretation method for a 24-hour dynamic blood pressure monitoring report based on GraphRAG is provided, which is characterized by comprising:

[0008] The first step is to build a GraphRAG knowledge base and perform self-supervised learning on the ABPM report interpretation model based on the GraphRAG knowledge base.

[0009] Step 2: Fine-tune the ABPM report interpretation model before use, so that the ABPM report interpretation model has the function of drawing conclusions based on the report interpretation and outputting blood pressure management suggestions based on the conclusions:

[0010] Step 3: Analyze the report using the ABPM report interpretation model that has been fine-tuned before use;

[0011] Step 4: Use the report analysis results to construct an interpretation report and present it visually.

[0012] Preferably, the first step includes: collecting medical knowledge related to ABPM, constructing a GraphRAG knowledge base containing all ABPM report interpretation terms, taking the category of each interpretation term as a node, and the interpretation item content corresponding to each interpretation term as an edge, to form a hyper-relational knowledge graph containing main triples and subsidiary attribute key-value pairs; and performing semantic search based on GraphRAG, and using the collected medical knowledge to perform self-supervised learning on the ABPM report interpretation model.

[0013] Preferably, the second step includes: fine-tuning the ABPM report parsing model for health advice, so that the ABPM report interpretation model has the function of drawing conclusions based on the report interpretation and outputting blood pressure management advice based on the conclusions.

[0014] Preferably, the third step includes: automatically extracting blood pressure measurements and ABPM report conclusions based on the ABPM diagnosis results; extracting all diagnostic keywords from the blood pressure measurements and ABPM report conclusions, and obtaining all matched interpretation terms based on preset diagnostic keyword and interpretation term matching rules; obtaining ABPM report data to be interpreted, and inputting the ABPM report data into the ABPM report interpretation model for parsing, and outputting report interpretation conclusions and health recommendations.

[0015] Preferably, the fourth step includes: searching for the numerical values ​​corresponding to each category from the ABPM measurement values ​​of the ABPM diagnostic results, and combining the report interpretation conclusions and health recommendations, constructing an interpretation report and visually displaying it, and setting the interpretation report to include the classification interpretation results and the matching interpretation term analysis results.

[0016] Preferably, the first step comprises:

[0017] Use the report parsing model as an analyzer to segment the ABPM report data and output feature vectors;

[0018] For each piece of sample data, its data feature vector is input into the ABPM report interpretation model one by one. The ABPM report interpretation model predicts the subsequent feature vector based on the input feature vector and calculates the vector error of the two feature vectors of the same piece of sample data.

[0019] Update the model parameters according to the vector error;

[0020] The first two steps were repeated until the ABPM report interpretation model converged.

[0021] Preferably, the second step comprises:

[0022] Design multiple manually written instructions representing different tasks, write the corresponding "input" and / or "output" according to each manually written instruction, and combine these data into "instruction, input, output" and "instruction, output" as a seed pool;

[0023] A plurality of manually written instructions are randomly selected from the seed pool and input into the ABPM report interpretation model. The ABPM report interpretation model outputs the model generation instruction, the instruction type of the model generation instruction, and the "input" corresponding to the model generation instruction, wherein the instruction type includes classified instructions and non-classified instructions;

[0024] Select multiple manually written instructions and multiple model generated instructions, and input the selected instructions into the ABPM report interpretation model, and the ABPM report interpretation model outputs the model generated instructions, instruction types, and "inputs, outputs" corresponding to the instructions;

[0025] Determine the ROUGE-L index of the model generation instruction output in the previous step; if the ROUGE-L index of the model generation instruction is less than the predetermined ratio of the ROUGE-L index of the instruction in the seed pool, add the data "instruction, input, output" and "instruction, output" corresponding to the instruction into the seed pool, and update the seed pool;

[0026] Repeat the first two steps until the number of all instructions in the seed pool reaches the predetermined number;

[0027] According to the report interpretation conclusion output by the ABPM report interpretation model, the "input" in the data "instructions, inputs, outputs" in the seed pool is matched, and the "output" corresponding to the "input" is output as a health recommendation.

[0028] Preferably, the predetermined ratio is 0.7 times.

[0029] According to another aspect of the present invention, a terminal device is also provided for implementing the above-mentioned GraphRAG-based 24-hour dynamic blood pressure monitoring report structured interpretation method.

[0030] According to another aspect of the present invention, a storage medium is also provided, on which are stored computer instructions that, when executed, cause a processor to execute the above-mentioned GraphRAG-based 24-hour dynamic blood pressure monitoring report structured interpretation method.

[0031] The present invention uses GraphRAG to structure and automatically interpret ABPM reports, and also displays the report content of ABPM in vivid and interesting popular science graphic texts, audio, video and other multimodal forms, so that non-hypertension specialists can correctly interpret the 24-hour dynamic blood pressure monitoring report, and ordinary people can also understand their own blood pressure monitoring reports. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] A more complete understanding of the present invention and its attendant advantages and features will be more readily appreciated by reference to the following detailed description taken in conjunction with the accompanying drawings, in which:

[0033] Figure 1 The overall process of the structured interpretation method of dynamic blood pressure monitoring report according to the preferred embodiment of the present invention is schematically shown.

[0034] Figure 2 A specific example of the first step of the structured interpretation method of a 24-hour dynamic blood pressure monitoring report according to a preferred embodiment of the present invention is schematically shown.

[0035] Figure 3 A specific example of the second step of the structured interpretation method of a 24-hour dynamic blood pressure monitoring report according to a preferred embodiment of the present invention is schematically shown.

[0036] Figure 4 A specific application example of the structured interpretation method of a 24-hour dynamic blood pressure monitoring report according to a preferred embodiment of the present invention is schematically shown.

[0037] Figure 5 A specific application example of the structured interpretation method of a 24-hour dynamic blood pressure monitoring report according to a preferred embodiment of the present invention is schematically shown.

[0038] Figure 6 Examples of systolic blood pressure (SBP) and diastolic blood pressure (DBP) data are schematically shown.

[0039] It should be noted that the drawings are used to illustrate the present invention, rather than to limit the present invention. Note that the drawings showing the structures may not be drawn to scale. In addition, in the drawings, the same or similar elements are marked with the same or similar reference numerals. DETAILED DESCRIPTION

[0040] In order to make the contents of the present invention clearer and easier to understand, the contents of the present invention are described in detail below in conjunction with specific embodiments and drawings.

[0041] Figure 1 The overall process of the structured interpretation method of the 24-hour dynamic blood pressure monitoring report according to the preferred embodiment of the present invention is schematically shown.

[0042] like Figure 1 As shown, the structured interpretation method of the 24h dynamic blood pressure monitoring report according to the preferred embodiment of the present invention includes:

[0043] The first step S1: construct the GraphRAG knowledge base and perform self-supervised learning on the ABPM report interpretation model based on the GraphRAG knowledge base;

[0044] Specifically, for example, the first step S1 includes: collecting medical knowledge related to ABPM, constructing a GraphRAG knowledge base containing all ABPM report interpretation terms, taking the category of each interpretation term as a node, and the interpretation item content corresponding to each interpretation term as an edge, forming a hyper-relational knowledge graph containing main triples and subsidiary attribute key-value pairs, and performing a relatively complete medical knowledge representation; and performing semantic search based on GraphRAG, and using the collected medical knowledge to perform self-supervised learning on the ABPM report interpretation model;

[0045] Step 2 S2: Fine-tune the ABPM report interpretation model before use, so that the ABPM report interpretation model has the function of drawing conclusions based on the report interpretation and outputting blood pressure management suggestions based on the conclusions:

[0046] Specifically, for example, the second step S2 includes: fine-tuning the ABPM report parsing model for health advice, so that the ABPM report interpretation model has the function of drawing conclusions based on the report interpretation and outputting blood pressure management advice based on the conclusions;

[0047] Step 3 S3: Analyze the report using the ABPM report interpretation model that has been fine-tuned before use;

[0048] Specifically, for example, the third step S3 includes: automatically extracting blood pressure measurement values ​​and ABPM report conclusions based on ABPM diagnosis results; extracting all diagnostic keywords from blood pressure measurement values ​​and ABPM report conclusions, and obtaining all matching interpretation terms based on preset diagnostic keyword and interpretation term matching rules; obtaining ABPM report data to be interpreted, and inputting the ABPM report data into the ABPM report interpretation model for parsing, and outputting report interpretation conclusions and health recommendations;

[0049] Step 4 S4: Use the report analysis results to construct an interpretation report and visualize it.

[0050] Specifically, for example, the fourth step S4 includes: searching for the numerical values ​​corresponding to each category from the ABPM measurement values ​​of the ABPM diagnosis results, and combining the report interpretation conclusions and health recommendations, constructing an interpretation report and visually displaying it, and setting the interpretation report to include the classification interpretation results and the matching interpretation term analysis results.

[0051] GraphRAG is a RAG system that combines the strengths of knowledge graphs and large language models (LLMs). Retrieval-augmented generation (RAG), as a powerful technique, has emerged as an effective means to address the key limitations of large language models (LLMs). By augmenting LLM prompts with relevant data retrieved from various sources, RAG ensures that LLM responses are factually based, accurate, and free from illusions. In GraphRAG, the knowledge graph serves as a structured repository of factual information, while the LLM acts as an inference engine that interprets user queries, retrieves relevant knowledge from the graph, and generates coherent responses. Emerging research shows that GraphRAG significantly outperforms RAG systems based on vector storage. Research also shows that GraphRAG systems not only provide better answers, but also do so at a lower cost and with greater scalability.

[0052] RAG is a term first coined in a 2020 paper and has now become a common architectural pattern for building LLM-driven applications. RAG systems use a retrieval module to find relevant information from a knowledge source (such as a database or knowledge base) and then use a generation module (driven by the LLM) to generate a response based on the retrieved information.

[0053] Knowledge graph reasoning can be based on the clinical record RDF triple relationship in the patient information model and the clinical recommendation reasoning rules in the chronic kidney disease knowledge graph to infer the patient's missing chronic kidney disease diagnosis, missing disease risk diagnosis, missing disease follow-up requirements and prescription adjustments, and provide corresponding clinical diagnosis recommendations based on the reasoning results.

[0054] Figure 2A specific example of the first step of the structured interpretation method of a 24-hour dynamic blood pressure monitoring report according to a preferred embodiment of the present invention is schematically shown.

[0055] Specifically, the first step S1 includes:

[0056] Step S1-1: Use the report parsing model as an analyzer to perform word segmentation on the ABPM report data and output a feature vector;

[0057] Step S1-2: For each piece of sample data, its data feature vector is input into the ABPM report interpretation model one by one. The ABPM report interpretation model predicts subsequent feature vectors based on the input feature vectors, and calculates the vector error between two feature vectors of the same piece of sample data;

[0058] Step S1-3: Update the parameters of the model according to the vector error;

[0059] Step S1-4, repeating steps S1-2 to S1-3 until the ABPM report interpretation model converges.

[0060] Figure 3 A specific example of the second step of the structured interpretation method of a 24-hour dynamic blood pressure monitoring report according to a preferred embodiment of the present invention is schematically shown.

[0061] Specifically, the second step S2 includes:

[0062] Step S2-1-1, design multiple manually written instructions representing different tasks, write corresponding "input" and / or "output" according to each manually written instruction, and combine these data into "instruction, input, output" and "instruction, output" as a seed pool;

[0063] Step S2-1-2, randomly extract multiple manually written instructions from the seed pool and input them into the ABPM report interpretation model, the ABPM report interpretation model outputs the model generation instruction, the instruction type of the model generation instruction, and the "input" corresponding to the model generation instruction, wherein the instruction type includes classified instructions and non-classified instructions;

[0064] Step S2-1-3, select multiple manually written instructions and multiple model generation instructions, and input the selected instructions into the ABPM report interpretation model, and the ABPM report interpretation model outputs the model generation instruction, the instruction type, and the "input, output" corresponding to the instruction;

[0065] Step S2-1-4, determining the ROUGE-L index of the model generation instruction outputted in step S2-1-3; if the ROUGE-L index of the model generation instruction is less than a predetermined ratio (e.g., 0.7 times) of the ROUGE-L index of the instruction in the seed pool, then adding the data "instruction, input, output" and "instruction, output" corresponding to the instruction into the seed pool, and updating the seed pool;

[0066] Step S2-1-5, repeating steps S2-1-3 and S2-1-4 until the number of all instructions in the seed pool reaches a predetermined number (e.g., 10,000);

[0067] Step S2-1-6, according to the report interpretation conclusion output by the ABPM report interpretation model, the report parsing model matches the "input" in the data "instructions, inputs, outputs" in the seed pool, and outputs the "output" corresponding to the "input" as a health recommendation.

[0068] According to another embodiment of the present invention, the present invention further provides a terminal device for implementing the above-mentioned 24-hour dynamic blood pressure monitoring report structured interpretation method.

[0069] According to another embodiment of the present invention, the present invention further provides a storage medium on which are stored computer instructions which, when executed, cause a processor to execute the above-mentioned 24-hour dynamic blood pressure monitoring report structured interpretation method.

[0070] The present invention uses GraphRAG to structure and automatically interpret ABPM reports, and also displays the report content of ABPM in vivid and interesting popular science graphic texts, audio, video and other multimodal forms, so that non-hypertension specialists can correctly interpret the 24-hour dynamic blood pressure monitoring report, and ordinary people can also understand their own blood pressure monitoring reports.

[0071] Figure 4 and Figure 5 A specific application example of the structured interpretation method of a 24-hour dynamic blood pressure monitoring report according to a preferred embodiment of the present invention is schematically shown. Figure 6 The systolic blood pressure (SBP) and diastolic blood pressure (DBP) data examples are schematically shown. For example, the ABPM diagnostic criteria for hypertension: 24h average value ≥ 130 / 80 mmHg, daytime average value ≥ 135 / 85 mmHg, nighttime average value ≥ 120 / 70 mmHg.

[0072] In addition, it should be noted that, unless otherwise specified, the terms "first", "second", "third", etc. in the specification are only used to distinguish the various components, elements, steps, etc. in the specification, and are not used to indicate the logical relationship or sequential relationship between the various components, elements, steps, etc.

[0073] It is to be understood that, although the present invention has been disclosed as a preferred embodiment, the above embodiment is not intended to limit the present invention. For any person skilled in the art, without departing from the scope of the technical solution of the present invention, the technical content disclosed above can be used to make many possible changes and modifications to the technical solution of the present invention, or modified into equivalent embodiments of equivalent changes. Therefore, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of protection of the technical solution of the present invention.

Claims

1. A structured interpretation method for 24-hour dynamic blood pressure monitoring report based on GraphRAG, characterized in that include: Step 1: Construct the GraphRAG knowledge base and perform self-supervised learning on the ABPM report interpretation model based on the GraphRAG knowledge base; Step 2: Fine-tune the ABPM report interpretation model before use, so that the ABPM report interpretation model has the function of drawing conclusions based on the report interpretation and outputting blood pressure management suggestions based on the conclusions: Step 3: Analyze the report using the ABPM report interpretation model that has been fine-tuned before use; Step 4: Use the report analysis results to construct an interpretation report and present it visually.

2. The method according to claim 1, characterized in that The first step includes: collecting medical knowledge related to ABPM, building a GraphRAG knowledge base containing all ABPM report interpretation terms, using the category of each interpretation term as a node and the interpretation item content corresponding to each interpretation term as an edge, to form a hyper-relational knowledge graph containing main triples and subsidiary attribute key-value pairs; and performing semantic search based on GraphRAG, using the collected medical knowledge to perform self-supervised learning on the ABPM report interpretation model.

3. The method according to claim 1 or 2, characterized in that: The second step includes: fine-tuning the ABPM report parsing model for health recommendations, so that the ABPM report interpretation model has the function of drawing conclusions based on the report interpretation and outputting blood pressure management recommendations based on the conclusions.

4. The method according to claim 1 or 2, characterized in that: The third step includes: automatically extracting blood pressure measurements and ABPM report conclusions based on the ABPM diagnosis results; extracting all diagnostic keywords from the blood pressure measurements and ABPM report conclusions, and obtaining all matching interpretation terms based on the preset diagnostic keyword and interpretation term matching rules; obtaining the ABPM report data to be interpreted, and inputting the ABPM report data into the ABPM report interpretation model for parsing, and outputting the report interpretation conclusions and health recommendations.

5. The method according to claim 1 or 2, characterized in that: The fourth step includes: finding the corresponding values ​​for each category from the ABPM measurement values ​​of the ABPM diagnostic results, and combining the report interpretation conclusions and health recommendations to construct an interpretation report and display it visually, setting the interpretation report to include the results of the classified interpretation and the analysis results of the matched interpretation terms.

6. The method according to claim 1 or 2, characterized in that: Specifically, the first step includes: Use the report parsing model as an analyzer to segment the ABPM report data and output feature vectors; For each piece of sample data, its data feature vector is input into the ABPM report interpretation model one by one. The ABPM report interpretation model predicts the subsequent feature vector based on the input feature vector and calculates the vector error between the two feature vectors of the same piece of sample data. Update the model parameters according to the vector error; The first two steps were repeated until the ABPM report interpretation model converged.

7. The method according to claim 1 or 2, characterized in that: The second step includes: Design multiple manually written instructions representing different tasks, write the corresponding "input" and / or "output" according to each manually written instruction, and combine these data into "instruction, input, output" and "instruction, output" as a seed pool; A plurality of manually written instructions are randomly extracted from the seed pool and input into the ABPM report interpretation model, and the ABPM report interpretation model outputs the model generation instruction, the instruction type of the model generation instruction, and the "input" corresponding to the model generation instruction, where the instruction type includes classified instructions and non-classified instructions; a plurality of manually written instructions and a plurality of model generation instructions are selected, and the selected instructions are input into the ABPM report interpretation model, and the ABPM report interpretation model outputs the model generation instruction, the instruction type, and the "input, output" corresponding to the instruction; Determine the ROUGE-L index of the model generation instruction output in the previous step; if the ROUGE-L index of the model generation instruction is less than the predetermined ratio of the ROUGE-L index of the instruction in the seed pool, add the data "instruction, input, output" and "instruction, output" corresponding to the instruction into the seed pool, and update the seed pool; Repeat the first two steps until the number of all instructions in the seed pool reaches the predetermined number; According to the report interpretation conclusion output by the ABPM report interpretation model, the "input" in the data "instructions, inputs, outputs" in the seed pool is matched, and the "output" corresponding to the "input" is output as a health recommendation.

8. The method according to claim 7, characterized in that The predetermined ratio is 0.7 times.

9. A terminal device, used to implement the GraphRAG-based 24-hour dynamic blood pressure monitoring report structured interpretation method according to any one of claims 1 to 8.

10. A storage medium storing computer instructions which, when executed, cause a processor to execute the structured interpretation method for a 24-hour dynamic blood pressure monitoring report based on GraphRAG according to any one of claims 1 to 8.

Images