Pediatric sepsis clinical evaluation and diagnosis and treatment method based on large model
By breaking down the disease assessment task into indicator analysis and risk prediction stages, and combining large-scale models and machine learning models, the process of generating treatment plans is optimized. This solves the problems of comprehensiveness and personalization in pediatric sepsis assessment, and enables more accurate and real-time risk prediction and personalized treatment plans.
Patent Information
- Application Number
- CN202511009680.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-22
- Publication Date
- 2025-11-14
AI Technical Summary
Existing large-scale models have shortcomings in the clinical assessment of pediatric sepsis, including insufficient comprehensiveness and accuracy, lack of interpretability and real-time capability in risk prediction, and lack of personalization in treatment plans.
The disease assessment task is broken down into two stages: indicator analysis and risk prediction. By combining indicator classification templates and risk adjustment templates with large-scale models and machine learning models, the process of generating treatment plans is optimized. The accuracy and personalization of assessment and treatment plans are improved by designing templates and iteratively modifying queries.
It improves the comprehensiveness and accuracy of clinical assessment for pediatric sepsis, enhances the interpretability and real-time nature of risk prediction, optimizes the personalization of treatment plans, and ensures the quality of diagnosis and treatment protocols.
Smart Images
Figure CN120954746A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of large-scale clinical diagnosis and treatment support, specifically relating to the field of clinical assessment and treatment methods for pediatric sepsis based on large-scale models. Background Technology
[0002] Clinical support is a crucial technology in modern medicine, enhancing healthcare services. Large-scale models offer more flexible support and complex medical recommendations, making them a key factor in hospital modernization and the construction of "smart hospitals." These models are used to enhance clinical support by integrating medical knowledge bases, patient data, and reasoning to analyze and evaluate patient test results, predict risks, and ultimately provide reliable and effective diagnoses and treatment plans, helping doctors improve clinical outcomes. However, this process faces several challenges: the comprehensiveness and accuracy of indicator evaluation cannot be guaranteed; traditional machine learning methods and large-scale model methods suffer from a lack of interpretability, real-time performance, inaccuracy, and instability in risk prediction; and generally applicable treatment recommendations lack personalization. Therefore, constructing a comprehensive clinical assessment and treatment methodology remains challenging.
[0003] Sepsis is a common acute and critical illness in pediatrics, characterized by rapid progression and high mortality, requiring rapid and appropriate diagnosis and treatment. Existing large-scale clinical support models typically provide medical indicators and their ranges directly during assessment, which can lead to numerous omissions and errors at larger scales. Prediction often employs traditional methods, lacking interpretability and real-time dynamic adjustment capabilities, while direct predictions from large models are inconsistent and have low accuracy. In generating treatment recommendations, although retrieval-enhanced generation techniques are commonly used, the illusions inherent in large models and the limitations of existing knowledge retrieval methods cannot guarantee higher quality treatment recommendations. In short, these problems prevent current methods from simultaneously achieving comprehensive assessment and early warning, as well as appropriate personalized treatment. Summary of the Invention
[0004] The purpose of this invention is to provide a clinical assessment and treatment method for pediatric sepsis based on a large model, so as to improve the comprehensiveness and accuracy of clinical assessment of pediatric sepsis and the rationality and personalization of treatment plans.
[0005] To achieve the above objectives, the present invention provides the following solution:
[0006] S1: Obtain the input patient medical records;
[0007] S2: Patient Condition Assessment: The condition assessment task is broken down into two stages: indicator analysis and risk prediction. The indicator assessment stage focuses on the comprehensiveness and accuracy of indicator analysis, while the risk prediction stage summarizes the indicator analysis results and combines them with machine learning models to obtain better predictive results.
[0008] S21: Design indicator classification templates and risk adjustment templates;
[0009] S22: Using the indicator classification template and the input initial medical records, the large model is called to obtain the classified indicators according to the predefined indicator categories;
[0010] S23: Using the categorized indicators, use a large model to analyze the indicators in each category in parallel, and then summarize the overall analysis of each category into a summary analysis report;
[0011] S24: Call the pre-trained random forest model to obtain the patient's initial risk, sort the indicators according to feature importance SHAP, and select the important set of indicators;
[0012] S25: Using the risk adjustment template, the large model adjusts the initial risk to a risk probability based on the real-time situation according to the set of important indicators;
[0013] S3: Quality optimization of treatment plans:
[0014] S31: Using the summary analysis report in S2, select the appropriate database from multiple vector databases. Each vector database contains treatment guidelines for a type of disease or infection.
[0015] S32: Using the summary analysis report in S2 as the initial query statement, call the embedded model to generate query vectors, and perform knowledge retrieval in each selected vector database;
[0016] S33: If the maximum number of searches is not reached, use the decision model to determine the relevance of the search results;
[0017] S34: If the model determines that the search results are relevant, the search results will be used as additional knowledge supplements and the large model will generate a treatment plan.
[0018] S35: If the model determines that the retrieval result is irrelevant, the large model is required to reflect on whether the irrelevant result is due to an incorrect database selection or an unreasonable query statement and make modifications accordingly.
[0019] S36: If the maximum number of searches has been reached, the search results will be used as additional knowledge supplements and a treatment plan will be generated by the large model.
[0020] S37: Use the modified query as the initial query, and iteratively call the knowledge retrieval, relevance judgment, and large model reflection and modification steps.
[0021] Optionally, the indicator classification template in step S21 specifically includes:
[0022] The framework includes role setting, task objectives, thought process, category descriptions, reference examples, and output requirements. Role setting prompts the large model to perform the task as a medical information expert. The task objective is to extract indicators from medical records and classify them according to predefined categories. The thought process provides the large model with the specific analytical steps for indicator extraction and classification. Category descriptions provide predefined category names. Reference examples provide standard input and output examples. Given inputs provide specific input content and required outputs.
[0023] Optionally, the risk adjustment template in step S21 specifically includes:
[0024] The framework includes role setting, task objectives, thought process, reference examples, and output requirements. Role setting prompts the large model to perform the task as a disease risk assessment expert. The task objective is to evaluate the risk probability derived from the traditional model, adjust it according to the actual situation, and provide the basis for the adjustment. The thought process provides a detailed analytical process for probability assessment and adjustment of the large model. The reference examples provide standard input and output examples. The given inputs provide specific input content and required outputs.
[0025] Optionally, the step of determining the relevance of the search results by the determination model in step S33 specifically includes:
[0026] Based on the search results of each database, each entry in the results is compared with the patient's medical report for similarity. Irrelevant knowledge entries are filtered out. After comparison, if the proportion of the remaining knowledge entries to the total number of knowledge entries exceeds a certain value, or the proportion of the databases with non-empty search results to the total number of selected databases exceeds a certain value, the search results are judged to be relevant; otherwise, the search results are judged to be irrelevant.
[0027] Optionally, the step of reflecting on the irrelevant reasons and making modifications in step S35 specifically includes:
[0028] If the problem is due to an unreasonable query statement, rewrite the query statement. If the problem is due to an incorrect database selection, choose a more suitable vector database. Generally, if the relevant knowledge cannot be obtained after rewriting the query, it is considered that a new database should be selected. Attached Figure Description
[0029] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:
[0030] Figure 1 This is a flowchart illustrating a clinical assessment and treatment method for pediatric sepsis based on a large model, according to the present invention.
[0031] Figure 2 This is the design of the indicator classification template of the present invention;
[0032] Figure 3 This is the risk adjustment template design of the present invention;
[0033] Figure 4 This is a flowchart of a method for improving the quality of treatment plans through search optimization in a clinical assessment and treatment method for pediatric sepsis based on a large model, according to the present invention. Detailed Implementation
[0034] To enable those skilled in the art to better understand the present invention and to make the above-mentioned objectives, technical solutions and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings.
[0035] like Figure 1 As shown, the clinical assessment and treatment method for pediatric sepsis based on a large model provided by this invention includes:
[0036] S1: Obtain the input patient medical records;
[0037] S2: Patient Condition Assessment: The condition assessment task is broken down into two stages: indicator analysis and risk prediction. The indicator assessment stage focuses on the comprehensiveness and accuracy of indicator analysis, while the risk prediction stage summarizes the indicator analysis results and combines them with machine learning models to obtain better predictive results.
[0038] S21: Design indicator classification templates and risk adjustment templates;
[0039] S22: Using the indicator classification template and the input initial medical records, the large model is called to obtain the classified indicators according to the predefined indicator categories;
[0040] S23: Using the categorized indicators, use a large model to analyze the indicators in each category in parallel, and then summarize the overall analysis of each category into a summary analysis report;
[0041] S24: Call the pre-trained random forest model to obtain the patient's initial risk, sort the indicators according to feature importance SHAP, and select the important set of indicators;
[0042] S25: Using the risk adjustment template, the large model adjusts the initial risk to a risk probability based on the real-time situation according to the set of important indicators;
[0043] S3: Quality optimization of treatment plans:
[0044] S31: Using the summary analysis report in S2, select the appropriate database from multiple vector databases. Each vector database contains treatment guidelines for a type of disease or infection.
[0045] S32: Using the summary analysis report in S2 as the initial query statement, call the embedded model to generate query vectors, and perform knowledge retrieval in each selected vector database;
[0046] S33: If the maximum number of searches is not reached, use the decision model to determine the relevance of the search results;
[0047] S34: If the model determines that the search results are relevant, the search results will be used as additional knowledge supplements and the large model will generate a treatment plan.
[0048] S35: If the model determines that the retrieval result is irrelevant, the large model is required to reflect on whether the irrelevant result is due to an incorrect database selection or an unreasonable query statement and make modifications accordingly.
[0049] S36: If the maximum number of searches has been reached, the search results will be used as additional knowledge supplements and a treatment plan will be generated by the large model.
[0050] S37: Use the modified query as the initial query, and iteratively call the knowledge retrieval, relevance judgment, and large model reflection and modification steps.
[0051] Among them, the indicator classification model and the risk adjustment model can be general large-scale language models, such as ChatGPT, DeepSeek, etc.
[0052] To ensure that the large model correctly understands the disease assessment requirements at each stage, an indicator classification template and a risk adjustment template were designed to classify each indicator to complete a comprehensive assessment and obtain better predictive results.
[0053] Among them, constructing indicator classification templates such as Figure 2 As shown, the indicator classification template in step S21 specifically includes:
[0054] The framework includes role setting, task objectives, thought process, category descriptions, reference examples, and output requirements. Role setting prompts the large model to perform the task as a medical information expert. The task objective is to extract indicators from medical records and classify them according to predefined categories. The thought process provides the large model with the specific analytical steps for indicator extraction and classification. Category descriptions provide predefined category names. Reference examples provide standard input and output examples. Given inputs provide specific input content and required outputs.
[0055] The indicator classification template is a question template designed based on the specific indicator classification task and the thinking habits of the large model. It is used to classify various indicators to aid in subsequent indicator analysis. The template first clarifies the model's role, requiring the large model to complete the task as a professional medical information expert. It then describes the task objective: to extract indicators from medical records and classify them according to predefined categories. Next, it outlines the thought process for completing this task, requiring the large model to analyze step-by-step according to these steps: First, extract the names of clinical indicators from the medical records; second, for each indicator, consider which category it should be classified into; finally, find the corresponding values of the indicators from the medical records and fill them into their respective categories. The template then lists category descriptions, including predefined specific category names. Following this are reference examples, showing the large model standard input and output examples. Finally, it provides specific input content, including the initial medical records.
[0056] Among them, constructing a risk adjustment template is as follows Figure 3 As shown, the risk adjustment template in step S21 specifically includes:
[0057] The framework includes role setting, task objectives, thought process, reference examples, and output requirements. Role setting prompts the large model to perform the task as a disease risk assessment expert. The task objective is to evaluate the risk probability derived from the traditional model, adjust it according to the actual situation, and provide the basis for the adjustment. The thought process provides a detailed analytical process for probability assessment and adjustment of the large model. The reference examples provide standard input and output examples. The given inputs provide specific input content and required outputs.
[0058] The risk adjustment template is a question template designed based on specific risk adjustment tasks and the thinking habits of large-scale models. It's used to adjust initial risks into risk probabilities based on real-time conditions. The template first clarifies the model's role, requiring the large model to complete the task as a pediatric disease risk assessment expert. It then describes the task objective: to adjust probabilities based on the child's current condition and traditional machine learning results, providing the rationale. Next, it outlines the thought process for completing this task, requiring the large model to analyze step-by-step: first, based on existing analysis results, consider whether the given probability is too high or too low; second, adjust the probability based on the levels of key feature values; finally, provide the rationale for adjustment, i.e., the key features affecting the probability. Following this is a reference example, showing the large model a standard input-output example. Finally, it provides specific input content, including analysis results, adverse prognostic probabilities, and key features.
[0059] Optionally, the step of determining the relevance of the search results by the determination model in step S33 specifically includes:
[0060] Based on the search results of each database, each entry in the results is compared with the patient's medical report for similarity. Irrelevant knowledge entries are filtered out. After comparison, if the proportion of the remaining knowledge entries to the total number of knowledge entries exceeds a certain value, or the proportion of the databases with non-empty search results to the total number of selected databases exceeds a certain value, the search results are judged to be relevant; otherwise, the search results are judged to be irrelevant.
[0061] Optionally, the step of reflecting on the irrelevant reasons and making modifications in step S35 specifically includes:
[0062] If the problem is due to an unreasonable query statement, rewrite the query statement. If the problem is due to an incorrect database selection, choose a more suitable vector database. Generally, if the relevant knowledge cannot be obtained after rewriting the query, it is considered that a new database should be selected.
[0063] Specific embodiments of the present invention have been described above with reference to the accompanying drawings. However, those skilled in the art will understand that various modifications and substitutions can be made to the specific embodiments of the present invention without departing from the spirit and scope of the invention. All such modifications and substitutions fall within the scope defined by the claims of the present invention.
Claims
1. A clinical assessment and treatment method for pediatric sepsis based on a large model, characterized in that, S1: Obtain the input patient medical records; S2: Patient Condition Assessment: The condition assessment task is broken down into two stages: indicator analysis and risk prediction. The indicator assessment stage focuses on the comprehensiveness and accuracy of indicator analysis, while the risk prediction stage summarizes the indicator analysis results and combines them with machine learning models to obtain better predictive results. S21: Design indicator classification templates and risk adjustment templates; S22: Using the indicator classification template and the input initial medical records, the large model is called to obtain the classified indicators according to the predefined indicator categories; S23: Using the categorized indicators, use a large model to analyze the indicators in each category in parallel, and then summarize the overall analysis of each category into a summary analysis report; S24: Call the pre-trained random forest model to obtain the patient's initial risk, sort the indicators according to feature importance SHAP, and select the important set of indicators; S25: Using the risk adjustment template, the large model adjusts the initial risk to a risk probability based on the real-time situation according to the set of important indicators; S3: Quality optimization of treatment plans: S31: Using the summary analysis report in S2, select the appropriate database from multiple vector databases. Each vector database contains treatment guidelines for a type of disease or infection. S32: Using the summary analysis report in S2 as the initial query statement, call the embedded model to generate query vectors, and perform knowledge retrieval in each selected vector database; S33: If the maximum number of searches is not reached, use the decision model to determine the relevance of the search results; S34: If the model determines that the search results are relevant, the search results will be used as additional knowledge supplements and the large model will generate a treatment plan. S35: If the model determines that the retrieval result is irrelevant, the large model is required to reflect on whether the irrelevant result is due to an incorrect database selection or an unreasonable query statement and make modifications accordingly. S36: If the maximum number of searches has been reached, the search results will be used as additional knowledge supplements and a treatment plan will be generated by the large model. S37: Use the modified query as the initial query, and iteratively call the knowledge retrieval, relevance judgment, and large model reflection and modification steps.
2. The method for clinical assessment and treatment of pediatric sepsis based on a large model according to claim 1, characterized in that, The indicator classification template in step S21 specifically includes: The framework includes role setting, task objectives, thought process, category descriptions, reference examples, and output requirements. Role setting prompts the large model to perform the task as a medical information expert. The task objective is to extract indicators from medical records and classify them according to predefined categories. The thought process provides the large model with the specific analytical steps for indicator extraction and classification. Category descriptions provide predefined category names. Reference examples provide standard input and output examples. Given inputs provide specific input content and required outputs.
3. The method for clinical assessment and treatment of pediatric sepsis based on a large model according to claim 1, characterized in that, The risk adjustment template in step S21 specifically includes: The framework includes role setting, task objectives, thought process, reference examples, and output requirements. Role setting prompts the large model to perform the task as a disease risk assessment expert. The task objective is to evaluate the risk probability derived from the traditional model, adjust it according to the actual situation, and provide the basis for the adjustment. The thought process provides a detailed analytical process for probability assessment and adjustment of the large model. The reference examples provide standard input and output examples. The given inputs provide specific input content and required outputs.
4. The method for clinical assessment and treatment of pediatric sepsis based on a large model according to claim 1, characterized in that, Step S33, which involves determining the relevance of the search results by the model, specifically includes: Based on the search results of each database, each entry in the results is compared with the patient's medical report for similarity. Irrelevant knowledge entries are filtered out. After comparison, if the proportion of the remaining knowledge entries to the total number of knowledge entries exceeds a certain value, or the proportion of the databases with non-empty search results to the total number of selected databases exceeds a certain value, the search results are judged to be relevant; otherwise, the search results are judged to be irrelevant.
5. The method for clinical assessment and treatment of pediatric sepsis based on a large model according to claim 1, characterized in that, Step S35, which involves reflecting on the irrelevant reasons and making corrections, specifically includes: If the problem is due to an unreasonable query statement, rewrite the query statement. If the problem is due to an incorrect database selection, choose a more suitable vector database. Generally, if the relevant knowledge cannot be obtained after rewriting the query, it is considered that a new database should be selected.